<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Dr. Brian Scott Glassman on Medium]]></title>
        <description><![CDATA[Stories by Dr. Brian Scott Glassman on Medium]]></description>
        <link>https://medium.com/@drglassman?source=rss-4584d7533d83------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*WIzrMWS-ZT7dJkSBbyrvcA.jpeg</url>
            <title>Stories by Dr. Brian Scott Glassman on Medium</title>
            <link>https://medium.com/@drglassman?source=rss-4584d7533d83------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 23:54:48 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@drglassman/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The Risks of Using JEV AI in Enterprise Software]]></title>
            <link>https://levelup.gitconnected.com/the-risks-of-using-jev-ai-in-enterprise-software-5d39ffc72842?source=rss-4584d7533d83------2</link>
            <guid isPermaLink="false">https://medium.com/p/5d39ffc72842</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[enterprise-architecture]]></category>
            <dc:creator><![CDATA[Dr. Brian Scott Glassman]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 15:22:45 GMT</pubDate>
            <atom:updated>2026-10-16T15:22:45.851Z</atom:updated>
            <content:encoded><![CDATA[<h4>Technical Issues with JEV in Practice</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EjNLiA4PQch8_M1wleJ0jA.jpeg" /><figcaption>The Risks of Using JEV AI in Enterprise Software</figcaption></figure><p>JEV AI, provided by <a href="https://typesafe.ai/">TypeSafe</a>, represents a new category of AI with tremendous potential for enterprise and high-volume applications. If you are not familiar with this new category of AI, please read on, as it will likely be popular and copied quickly by other AI providers. Most importantly, if you work in an enterprise, you need to understand the issues with this AI before moving it into development and production applications.</p><p><strong>What is the JEV category of AI models?</strong></p><p>JEV is an AI that is configured to take in a large amount of data and then answer only from a set of predefined choices. For example, you might ask, “Does this credit card transaction look fraudulent? Yes, no, or maybe?” and provide past transaction data for comparison. It then provides a confidence score for its selection (e.g., “Yes, 95% confident”). This is very different from LLMs, which generate free-form text in their answers. JEV is still smart, but it does not generate free-form text; in other words, you cannot chat with it. You simply get an answer to your structured question.</p><p><strong>How can enterprises use the JEV category of AI models?</strong></p><p>JEV AI is best suited to large workloads made up of many small tasks that require intelligence but must be done very quickly and at high volume. Examples include checking for fraud at checkout on e-commerce websites, screening an online post for abuse before it is published, and directing actions, such as deciding the best time to send a promotional code to an online shopper.</p><p><strong>Why is JEV AI beneficial?</strong></p><blockquote>The biggest benefits of JEV AI are cost and speed. TypeSafe claims JEV is roughly 400 times cheaper and 200 times faster<strong> </strong>than frontier LLMs.</blockquote><p>This speed allows AI to be used in real time and at high volume in day-to-day operations. JEV’s cheap, fast, and smart design opens up thousands of new corporate use cases that were previously too expensive or too slow for LLMs like ChatGPT or Claude. However, do not use it until you have read my warnings below.</p><h3>The Major Issues with JEV</h3><p>I did a deep dive into JEV by using and evaluating it in my own business applications. In particular, I applied it to my legal software, <a href="https://contracthulk.com/">ContractHulk</a>, to customize employment contracts for specific job titles, and evaluated its performance. Here is what I found.</p><h3>1. JEV’s Outputs Are Not Easily Auditable (Black Box)</h3><p>JEV outputs its selection as one of the predefined choices, followed by a confidence score. Ask it, “How many legs does the following animal [dog] have?” with the choices two, three, or four, and JEV selects “four” with 100% confidence. You know it is right from your own knowledge. But feed JEV, say, your last 100 credit card transactions and ask whether the most recent one seems fraudulent, and it might answer “Yes” with 65% confidence. JEV will not tell you why, and this is a significant flaw. You could add follow-up prompts that categorize the fraud and identify what flagged it (time of day, category of purchase, vendor, vendor location), but this adds a lot of extra work.</p><blockquote><strong>Here is the core problem: when you set up JEV AI, you are trusting that the developer wrote the prompt and options correctly and provided the right context and evaluation data. Without written feedback from JEV, you are left running trials and comparing results to confirm it is performing correctly. Further, JEV’s self-assessed confidence score does not help confirm that the setup is correct.</strong></blockquote><p>Edge cases are a second concern. Say JEV is set up and operating as intended. If it encounters an edge case, where the answer could go either way, getting the right result requires the developer to add more context to the prompt or background requirements. Without written feedback from JEV, it is hard to tell when an edge case is even occurring, let alone what should be tuned to increase confidence in the answer from JEV.</p><p><strong>Solution: Currently, the only practical option for validating JEV in pre-production is to test it against datasets that have been pre-validated by another AI, such as ChatGPT or Claude, then cross-compare the results and adjust JEV’s configuration accordingly. This process is laborious and time-consuming.</strong></p><h3>2. Drift in Responses Cannot Be Evaluated</h3><p>AI output varies, even from top models like ChatGPT. Some days you log on and the response to the same question is different from usual. Although settings like temperature and seed help reduce variation, there are other settings, controlled by the provider, that can change without notice. Consequently, there is no easy way to measure JEV’s drift, especially as new versions of its models are released.</p><p><strong>Solution: Cross-validating JEV against new AI models will require rerunning the validation datasets and specifically examining the drift across models, so be sure to retain your past validation results.</strong></p><h3>3. Initial Configuration of JEV Should Be Left to Senior AI Developers</h3><p>Because JEV does not output written explanations, setup and configuration must be done very carefully with expert knowledge. I would argue that it requires a senior AI developer who knows where the potential edge cases and setup issues lie for that particular use case. I used Fable 5.1 to set up JEV for <a href="https://contracthulk.com/">ContractHulk</a>, and even then I had to anticipate issues ahead of time, such as JEV not fully evaluating a legal clause, I had to build safety nets for JEV to operate within.</p><blockquote>Unlike other AI, JEV will run and return answers whether it is set up correctly or not and a junior developer may assume it is working correctly.</blockquote><p>Structuring JEV for early system testing is a must, because validation is required, as is careful prompt design. This requires knowing both the application you are building and the potential points of failure in how JEV is used and what data it will see.</p><p><strong>Solution:</strong> Ask whichever AI you are using to configure JEV to plan for edge cases and to create a testing and validation dataset with expected answers and edge cases. Then run those validation tests, iteratively tuning JEV’s inputs. Document everything and save both the testing and validation results in case you need to revisit them.</p><h3>4. TypeSafe Is Missing Key Compliance Certifications</h3><p>JEV is operated by TypeSafe, which, as of October 14, 2026, holds none of the core enterprise security certifications, which is a huge issue. Currently, TypeSafe has no ISO/IEC 42001 AI governance certification, no ISO/IEC 27001 certification, no SOC 2 Type II report, and no external verification against Cyber Essentials or the NIST CSF. On privacy, JEV is hosted only in a single US West Coast region, which may fall short of GDPR data-residency expectations despite its DPA, and it has no ISO/IEC 27701 certification for how personal data is isolated or anonymized. It also lacks industry-specific authorizations, including FedRAMP/StateRAMP for government workloads, HIPAA for health data, and PCI-DSS for payment card data, so regulated enterprises should not send it sensitive data today.</p><p><strong>Solution: Subscribe to the Typesafe.ai newsletter, as they are expected to announce compliance certifications soon. Additionally, look out for similar models released by other AI providers.</strong></p><h3>Updates to JEV</h3><p>I strongly believe JEV needs either a <strong>second AI model</strong> or a <strong>developer mode</strong> that specifically provides written explanations for its answer selections, so that those answers can be properly validated during setup. Once setup is complete, JEV can be switched to its production model, which does not provide these reasoning outputs, and runs fast and cheap. Until such a feature is available, and after Typesafe.ai obtains its compliance certifications, enterprises should limit JEV to low-risk use cases.</p><h3>Conclusion</h3><p>JEV is fast, cheap, and smart, but for now only senior AI developers should set it up, and I would not put it into critical workflows until drift, and certification and uptime guartents can be put in place.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5d39ffc72842" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/the-risks-of-using-jev-ai-in-enterprise-software-5d39ffc72842">The Risks of Using JEV AI in Enterprise Software</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Belongs in a Strong Employment Contract?]]></title>
            <link>https://drglassman.medium.com/what-belongs-in-a-strong-employment-contract-51e8a960e892?source=rss-4584d7533d83------2</link>
            <guid isPermaLink="false">https://medium.com/p/51e8a960e892</guid>
            <category><![CDATA[small-business]]></category>
            <category><![CDATA[human-resources]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[legal]]></category>
            <dc:creator><![CDATA[Dr. Brian Scott Glassman]]></dc:creator>
            <pubDate>Thu, 01 Oct 2026 16:31:49 GMT</pubDate>
            <atom:updated>2026-10-01T16:31:49.946Z</atom:updated>
            <content:encoded><![CDATA[<h4>Many of the greatest risks facing an employer do not become visible until after an employee has been hired!</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dj8tebXo-6M3_yFKsRtNuQ.png" /></figure><h3>Introduction</h3><p>A strong employment contract is important because many of the greatest risks facing an employer do not become visible until after an employee has been hired. An employee may challenge a disciplinary decision, file a lawsuit after being terminated, mishandle confidential information, exceed the responsibilities of the position, or expose the business to liability through interactions with customers, patients, or other employees. If the applicable duties, standards of conduct, and consequences are not clearly documented, the employer may have difficulty defending its decisions or holding the employee accountable for harm caused to the business. Hence, <a href="https://contracthulk.com/?utm_source=medium&amp;utm_medium=social&amp;utm_campaign=2026_10_medium_medical_assistant&amp;utm_content=home"><strong>ContractHulk.com</strong></a> was created to address these risks by generating customized employment contract templates based on the responsibilities of the position and the laws of the applicable U.S. state.</p><p>The same shortcomings appeared repeatedly.</p><ul><li><strong>All poor employment contrats are only two to four pages long.</strong></li><li><strong>Many originated from general template websites and addressed little beyond compensation, working hours, and the employment start date.</strong></li><li><strong>Or the contracts were written by lawyers who did not address the specific of that job’s duties and risk inside the contract.</strong></li></ul><p>It is important to recognize that disputes often arise from unclear job responsibilities, workplace misconduct, or behavior that puts the company at risk. If these matters are not addressed in the employment contract, the company may have difficulty justifying its decision to terminate the employee, particularly if the employee retains an employment attorney. The company may also find it difficult to hold the employee legally accountable for actions that caused harm to the business.</p><p>Thus, a well-constructed employment contract serves a much broader purpose. For many small businesses, <strong>it can fulfill many of the functions of an employee handbook that was never formally created.</strong> More importantly, it presents those expectations in a document that the employee has reviewed and signed.</p><p>This article explains the essential functions of an effective employment agreement, introduces a practical method for evaluating its completeness, and examines one occupation in detail. I selected the <a href="https://contracthulk.com/employment-contract/medical-assistant?utm_source=medium&amp;utm_medium=social&amp;utm_campaign=2026_10_medium_medical_assistant&amp;utm_content=ma_page"><strong>medical assistant employment contract</strong></a><strong> </strong>because the position demonstrates how much a generic agreement can overlook when an employee works with patients, medical information, clinical procedures, and insurance billing.</p><h3>Phase 1: Know the Purpose of the Agreement</h3><p>Before determining whether an employment contract is adequate, it is important to understand the three primary functions it should perform. Agreements that accomplish all three are uncommon. Agreements that address only the first are everywhere.</p><h3>Establish Expectations Before an Issue Arises</h3><p>An employment agreement should identify the risks associated with the position and clearly communicate the employer’s requirements in each area before work begins.</p><p>For a medical assistant, those subjects include patient confidentiality, authorized responsibilities, professional qualifications, medical documentation, workplace safety, and billing practices. When these standards are documented before the employee’s first day, a future dispute can focus on whether an established requirement was satisfied, rather than whether the requirement existed at all.</p><h3>Provide the Employee With a Detailed Understanding of the Position</h3><p>The agreement should also explain the employee’s responsibilities, the behavior expected in the workplace, and the possible consequences of failing to meet established standards.</p><p>Many organizations operate with a limited employee handbook or without one entirely. As a result, the employment contract may be the only formal document containing these requirements. Employees are generally better positioned to perform successfully when expectations are specific and accessible. Managers can also respond more consistently to performance or conduct problems when they can refer to a written responsibility instead of relying on an informal understanding.</p><h3>Explain How Discipline and Separation Will Be Managed</h3><p>The third function is frequently absent from standard templates. A thorough agreement should explain how performance concerns, disciplinary measures, notice requirements, documentation, and separation from employment will be handled when the working relationship begins to deteriorate.</p><p>The objective is to ensure that corrective action or termination is connected to responsibilities the employee previously reviewed and accepted.</p><p>Consider a medical assistant who does not report a workplace hazard. If the agreement contains a clear incident reporting requirement, the manager can enforce an existing contractual obligation rather than creating a rule after the incident has already occurred.</p><h3>Phase 2: Use the Length Test</h3><p>When evaluating an employment agreement, I apply a straightforward guideline. If the document is fewer than ten pages, there is a strong possibility that entire subject areas have been excluded.</p><p>Compensation, scheduling, and the employment start date can easily occupy the first page. The remaining pages are needed to address responsibilities specific to the occupation, workplace policies, compliance obligations, and provisions required or influenced by state law.</p><p>An agreement designed for an actual position and a particular state will commonly range from ten to thirty pages. Length alone does not make a contract effective. Unnecessary language does not improve legal protection or workplace clarity. However, an unusually short agreement is often evidence that a broad template was used instead of a document developed around the requirements of the position.</p><h3>Phase 3: A Practical Example, the Medical Assistant</h3><p>The medical assistant position provides a useful example because a single workday may involve direct patient contact, protected health information, regulated clinical activities, workplace safety requirements, and insurance documentation.</p><p>A well developed agreement for this role should address the following clause categories, organized into three major areas. Each provision establishes an obligation or standard of conduct. It does not guarantee a particular result.</p><h3>Patient Privacy and Medical Documentation</h3><p>A patient confidentiality provision establishes the employee’s responsibility to safeguard protected health information and manage it in compliance with HIPAA requirements.</p><p>A medical records provision establishes standards for completing patient files accurately, thoroughly, and within the required time. Both responsibilities are fundamental to the operation of a healthcare practice, yet both are absent from nearly every general employment template I have examined.</p><h3>Authorized Duties and Professional Qualifications</h3><p>Medical assistants are restricted in the clinical activities they are permitted to perform, and the boundaries of those activities differ by state.</p><p>A scope of work provision distinguishes authorized duties from prohibited activities. This gives both the medical assistant and the supervising healthcare provider a defined framework instead of leaving either party to interpret the boundaries of the position independently.</p><p>A credentials provision requires the employee to maintain applicable certifications, licenses, and CPR training. It should also require prompt notification to the employer if a credential expires, is suspended, or otherwise becomes invalid.</p><h3>Workplace Safety, Incident Reporting, and Regulatory Compliance</h3><p>A workplace safety provision establishes the employee’s responsibility to comply with infection prevention measures, sharps procedures, biohazard requirements, and other clinical safety protocols.</p><p>An incident reporting provision explains which accidents, injuries, errors, and safety concerns must be reported, when the report must be made, and who must receive it.</p><p>A billing provision sets expectations for the accurate recording of services, charges, codes, and supporting information. A drug, alcohol, and substance testing provision addresses impairment reporting and describes any testing the employer is legally authorized to require.</p><p>Each subject represents a potential source of disagreement between a healthcare practice and an employee. That possibility is exactly why the applicable responsibilities should be established in writing before a dispute develops.</p><h3>Phase 4: Incorporate State Specific Requirements</h3><p>The same position may require different contract language depending on where the employee performs the work. Requirements involving notice, final wages, employee leave, workplace testing, and other employment practices differ among states.</p><p>A properly prepared agreement should contain at least two provisions tailored to the laws and employment practices of the applicable state. An agreement that does not address the employee’s work state cannot be considered complete for that jurisdiction.</p><h3>Conclusion</h3><p>A strong employment contract is more than an administrative formality. <strong>It documents the responsibilities attached to the position, the conduct required from the employee, and the procedures that will apply when a workplace problem occurs. </strong>Most importantly, it establishes those terms before circumstances make them urgent.</p><p>Employers who approach employment agreements this way are less likely to spend time reconstructing expectations after a disagreement. Instead, they can rely on standards that were clearly communicated and accepted at the beginning of the employment relationship.</p><p>To review an example of an employment contract developed using this approach, visit <a href="https://contracthulk.com/?utm_source=medium&amp;utm_medium=social&amp;utm_campaign=2026_10_medium_medical_assistant&amp;utm_content=home"><strong>ContractHulk</strong></a><strong>. </strong>Users can enter a job title and a United States work state, examine the complete agreement without charge, and download both an editable Microsoft Word document and a PDF version.</p><p>ContractHulk is not a law firm and does not offer legal advice. This article is provided solely for informational purposes. Before using any employment contract, have it reviewed by an attorney who understands your business, the position, and the laws of the applicable jurisdiction.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=51e8a960e892" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Become an AI‑Native Software Developer]]></title>
            <link>https://pub.towardsai.net/how-to-become-an-ai-native-software-developer-8a868c954d78?source=rss-4584d7533d83------2</link>
            <guid isPermaLink="false">https://medium.com/p/8a868c954d78</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-native]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Dr. Brian Scott Glassman]]></dc:creator>
            <pubDate>Wed, 04 Feb 2026 16:31:01 GMT</pubDate>
            <atom:updated>2026-02-04T16:31:01.346Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2hWSwDUmS_l9j7Muiic45A.jpeg" /></figure><h4><em>The perspectives, abilities, traits, and knowledge required to be an AI‑Native Software Developer</em></h4><h3>How to Become an AI‑Native Software Developer</h3><p>Software development is undergoing a seismic shift driven by AI coding tools, and skilled AI‑native software developers represent the core of this future. This article offers the next generation of developers my best insights into becoming lead developers in an AI‑first, agent‑driven software landscape. As a corporate AI technology strategist, a daily AI coding practitioner, former college professor at NYU, former VP of Engineering managing software development for over 20 years, I aim for these insights to be clear and grounded. My hope is to motivate developers to up skill and meet this next monumental shift.</p><p>The article is structured to flow naturally. It begins by explaining three new perspectives an AI‑native developer should adopt, as these drive high‑level behaviors. It then transitions to three abilities that should be developed through learning and practice with AI‑coding tools. Next, it discusses three natural traits that give some developers an advantage over others. The article then outlines five key areas of knowledge that AI‑native developers should deepen. Finally, it delivers a clear warning to hiring managers and VPs that two areas of knowledge and skill are no longer relevant in an AI‑native software development organization. Let’s start with the core perspectives that an AI‑native developer should adopt.</p><h3>Key Perspectives of AI‑Native Developers</h3><p>To prepare yourself for a shift to an AI‑native coding approach, you must first internalize the following three perspectives. While one could argue that many other perspectives also matter, I have found these to be the most important for reorienting from an Agile‑based philosophical mindset to an AI‑native development perspective.</p><h3>Shift Your Thinking From Writing Code to Leading AI Agent Teams</h3><p>It is vitally important to realize that you are no longer acting as an individual contributor. Instead, you should internalize that you are a virtual department head who directs multiple AI agents through all the aspects of software development for your projects. Hence, your role is to break up the project into clear phases and tasks, assign them to AI agents, and ensure the AI team produces high-quality results. Think of yourself as a manager of a department of AI agents where you establish clear processes, structure, well-defined responsibilities, measurable outcomes, and ongoing monitoring of resource use such as token consumption and quality of their work products.</p><h3>Build Codebases That AI Can Easily Understand and Evolve</h3><p>You must shift your thinking that you are writing code for execution and possibly other people to read to realizing you are writing code so AI systems can read, understand, and change it, with very little effort. Hence, your codebase should be simple for AI models to work with at both the individual file level and the overall system level. Ask yourself: if an AI were reading this file quickly with minimal attention to detail, what would be clear and what might be confusing? Once you accept that AI is now the primary reader and modifier of your code, the priorities of a codebase change. High-level context must be obvious without deep inspection. Project-level markdown documents should explain the codebase’s form and function. File-level headers should clearly state the core functions and relationships. Directory structure, variable naming conventions, and comments should be very obvious and descriptive. The goal is not documentation for its own sake, but reducing cognitive load so an AI can quickly comprehend and confidently make edits.</p><h3>A New Mindset: Designing Processes Instead of Performing Them</h3><p>Second, you must no longer focus on carrying out individual tasks yourself. Instead, think and act like a task designer who directs AI agents through clear processes, guidelines, best practices, reference examples, explicit rules, and unmistakable prompts. The faster you realize that you are not doing the work, the AI team is, and that your role is to guide it accurately, the faster you will gain massive productivity. For example, when designing a user interface, first outline the steps required to achieve a strong design result, then create opportunities to evaluate results and insert your input to creatively guide the outputs and produce a superior UI.</p><h3>Abilities to Practice and Develop for AI-Native Developers</h3><p>With the mindsets adopted, next, you install an AI CLI (Claude Code, Cursor, Codex CLI, Gemini CLI) or install one as an extension in your IDE. Then practice the abilities below until you feel relatively comfortable with each. After that, make a plan to improve each ability as you perform your software development tasks.</p><h3>Learn How to Manage Many AI Tasks Without Losing Control</h3><p>Your first new ability should be high-quality multi-tasking. You are no longer limited to working on one task at a time. AI agents give you the ability to run multiple tasks in parallel, and your new ability is to manage this concurrency effectively. This means keeping track of progress, inputs, and outputs across several AI-driven tasks at once. You may be concurrently evaluating architecture trade-offs in one command-line session, guiding code generation in another, and running QA checks in a third. The goal is not to be busy or maximize your output, but to remain in control without having too much idle time.</p><h3>Build the Ability to Move Fast Without Getting Stuck</h3><p>Your second ability is rapid solution generation and quick implementation. You must avoid over-focusing on a single problem or chasing the most elegant solution at the cost of progress. In an AI-driven workflow, this leads to lost productivity. Thus, your new ability is to quickly generate possible solutions with the help of AI, implement them, test them, and evaluate results. If a solution does not work after a brief debugging effort, the correct response is to pivot to an alternative approach. The priority is to deliver a high-quality outcome efficiently while maintaining momentum, experimenting quickly, and moving away from unproductive paths without hesitation.</p><h3>Learn How to Think Like an AI Agent and Predict Its Outputs</h3><p>Your third ability is learning to think like the AI models you use. Each AI model has its own strengths, limitations, and behavior patterns. As you gain experience with a specific model, you will begin to anticipate how it responds to tasks and instructions of increasing complexity and adjust your tasks and instructions accordingly. It is also critical to recognize that different AI models are better suited to different types of work, such as lightweight models for structured edits and more advanced models for system architecture planning and complex reasoning.</p><h3>Natural Talents and Traits of AI-Native Developers</h3><p>Unfortunately, AI‑native programming is challenging, arguably more challenging than traditional programming, and some individuals are more suited for it than others due to their natural abilities. This section will give individuals the chance to see whether they are up to the task.</p><h3>Long Periods of High‑Intensity Thinking</h3><p>Running multiple AI agents concurrently, quickly reviewing results, making adjustments, identifying patterns, and analyzing outcomes all require a high level of intelligence and the ability to learn quickly. In my experience, many developers are smart and perform well in this area. However, if your preferred work style is to avoid engaging in five to ten hours of sustained, complex thinking each day, an AI‑native developer role may not be a good fit for you.</p><h3>Naturally Highly Organized</h3><p>AI systems today rely heavily on deep project context and well‑defined processes. If you tend to avoid cleaning up code, dislike writing documentation, or resist setting up and maintaining processes in favor of flexibility, or avoiding work that feels unproductive, then AI‑native development is likely not for you. AI‑native development requires highly organized thought processes, careful planning, and the willingness to invest in upfront organizational work to effectively guide AI agents in their tasks. This is a core personality trait; you either have it or you don’t, and those who do will see significant success!</p><h3>Strong Willingness to Experiment</h3><p>AI capabilities continue to advance rapidly, with significant jumps in performance occurring monthly. If you are naturally curious and compelled to experiment and continuously improve your methods, you have a significant advantage. Organizations increasingly need AI‑native developers who are willing to refine and improve their practices on a daily basis. However, if you view individuals who deviate from standard workflows to run one‑hour experiments or spend time after work trying new approaches as inefficient or problematic, you may not be a good fit for an AI‑native development role.</p><h3>Required Knowledge Areas</h3><p>AI‑coding tools and the seismic shift they are creating make some traditional knowledge areas far less valuable, while elevating other areas to critical importance. This shift is also introducing entirely new areas of expertise that developers must understand. Below are the key knowledge areas you must master to perform at the highest level in AI‑native software development.</p><h3>System Architecture and Its Trade‑Offs</h3><p>AI‑native developers must possess deep system architecture knowledge and understand the trade‑offs between different architectures and their configurations. This knowledge allows you to guide architectural planning and decision‑making for a software solution rather than blindly deferring to AI suggestions, which I have found can be highly biased. When creating software solutions, early architectural choices have a major influence on which features can be implemented and how much performance can be extracted from the system. For this reason, architectural decisions and the underlying knowledge should not be deferred to AI.</p><h3>Knowledge Across Multiple Technical Domains</h3><p>AI‑native development favors broad technical knowledge over deep specialization in a single area. Comfort moving between frontend, backend, data stores, DevOps, databases, and APIs allows you to guide AI agents as they naturally cross boundaries. Developers with cross‑domain fluency maintain momentum, avoid bottlenecks caused by narrow expertise, and are significantly more valuable in the AI‑native era.</p><h3>Knowledge Beyond AI Prompt Engineering</h3><p>Modern AI prompt engineering is far more than writing a clever prompt. It now involves designing processes, methods, and guidance for AI systems to follow. This includes defining rules, structure, documentation, and validation steps that AI agents can consistently apply. Clear, coherent instructions across planning, implementation, testing, and maintenance are what keep AI‑driven work reliable, and developing this capability requires significant reading, practice, and ingenuity.</p><h3>Knowledge of How to Optimize AI Token Usage</h3><p>In AI‑native development, tokens are a real operating cost. Effective AI developers design workflows that minimize token waste while preserving output quality. Knowing when to use lightweight models, when to escalate complexity, and how to efficiently scope tasks in terms of token usage is now part of everyday engineering judgment. This is a cutting‑edge capability that will become increasingly important as organizations begin actively monitoring AI token budgets.</p><h3>Skills Becoming Obsolete</h3><p>It may seem weird, but some areas of knowledge or areas of expertise will be phased out as AI‑Native software development becomes the norm. This author believes the following will no longer be important:</p><h3>Hand Coding Skills are Soon to Be Outdated</h3><p>The ability to write machine code is already obsolete, and manually writing low-level memory allocation code is similarly outdated. As technology advances, developers naturally move up levels of abstraction, since working with finer implementation details wastes time and effort. The same shift is now occurring with AI tools and hand-written code. While you may be able to write a software function manually, an AI model can typically do it faster, with fewer errors, and run validation checks. As a result, managers who insist on traditional coding exams during interviews must remove this outdated approach and instead focus on evaluating a candidate’s ability to effectively manage and validate code produced through AI-coding tools.</p><h3>Agile Will Soon Be Obsolete</h3><p>I have written a detailed article explaining why Agile is not compatible with AI‑native software development. In short, applying Agile processes to teams of AI‑native developers, who can operate at 100 to 1,000 times the speed of traditional hand‑coding teams, renders many Agile rituals, tasks, and associated skills obsolete. Without going into excessive detail here, Agile’s relevance is rapidly diminishing. Managers who continue to test for or emphasize Agile‑specific practices when interviewing for AI‑native developer roles should strongly consider updating their evaluation criteria.</p><h3>Summary</h3><p>In closing, this article covered four major areas essential to becoming an AI‑native software developer. First, it outlined the core mindset shifts required to move from writing code to leading AI agent teams. Second, it described the practical abilities developers must practice to operate effectively with AI‑coding tools. Third, it examined the natural traits that determine who is best suited for AI‑native work. Finally, it detailed the key knowledge areas developers must master while also identifying legacy skills that are rapidly becoming obsolete. It is strongly hoped you take these to heart and push yourself into the next major evolution in software development.</p><h3>About the Author</h3><p>Dr. Brian Scott Glassman is an AI strategy and implementation expert who is VP at AInspire Technology Consulting. He writes for Forbes, and other publications and does consulting for Fortune 500 organizations.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8a868c954d78" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/how-to-become-an-ai-native-software-developer-8a868c954d78">How to Become an AI‑Native Software Developer</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SpaceX Valuation Model: Estimating Fair Value for an IPO]]></title>
            <link>https://medium.com/predict/spacex-valuation-model-estimating-fair-value-for-an-ipo-17601d5290f2?source=rss-4584d7533d83------2</link>
            <guid isPermaLink="false">https://medium.com/p/17601d5290f2</guid>
            <category><![CDATA[space]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[ipo]]></category>
            <category><![CDATA[spacex]]></category>
            <dc:creator><![CDATA[Dr. Brian Scott Glassman]]></dc:creator>
            <pubDate>Mon, 26 Jan 2026 01:56:11 GMT</pubDate>
            <atom:updated>2026-03-04T23:59:36.916Z</atom:updated>
            <cc:license>http://creativecommons.org/licenses/by/4.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-EIS4rbyreCuHiIzz38_6A.jpeg" /><figcaption>SpaceX Valuation Model: Financial Forecasts for Revenues and Profits</figcaption></figure><p>A Clear Revenue &amp; Profit Forecast Model For Financial Analysts, Prospective Investors, and Space Enthusiasts Investors</p><h3>Outline of Article</h3><p><strong>Executive Summary</strong></p><p><strong>Section 1: Introduction to the Model </strong><br>1.1 Target Readers<br>1.2 No Outside Interests Were Involved<br>1.3 Accuracy of the Valuation Model <br>1.4 Supporting Excel Revenue, Operating, and Valuation Model<br>1.5 Summary of the SpaceX Valuation Model<br>1.6 Estimate of Fair Value for SpaceX for an IPO<br>1.7 Other SpaceX Valuation Models <br>1.8 Limitations on Available Information<br>1.9 Helping the Community Build a Valuation Model <br>1.10 Expertise of the Author: Dr. Brian Scott Glassman</p><p><strong>Section 2: SpaceX’s Near &amp; Long-Term Revenue Opportunities</strong><br>2.1 SpaceX Revenue Forecast Tables in 2029 and 2036<br>2.2 SpaceX Comparable IPOs<br>2.3 SpaceX Valuation Results</p><p><strong>Section 3: Projected Valuation of SpaceX’s Starlink Internet Services</strong><br>3.1 Market Position and Competitive Landscape<br>3.2 Pricing and Capacity Constraints — Starlink<br>3.3 Subscriber Base and Demand Dynamics — Starlink<br>3.4 Revenue Estimates for Starlink<br>3.5 Unit Economics and Servicing Costs for Starlink<br>3.6 Rapid Expansion of Starlink via Access to Large Countries<br>3.7 Growth of the Subscriber Base<br>3.8 Ancillary Revenues from Aviation and Marine are Small<br>3.9 Starlink Expected Revenues in 2029<br>3.10 Starlink Revenue Projections for 2036<br>3.11 Future Competition for Starlink<br>3.12 Major Risks to Starlink</p><p><strong>Section 4: Telecom Services Provided via StarLink</strong><br>4.1 Pricing Model for Telecom via Starlink<br>4.2 Number of Subscribers for Telecom via Starlink<br>4.3 Valuation of Starlink Telecom Services</p><p><strong>Section 5: SpaceX Launch Services</strong><br>5.1 SpaceX Cost Per Ton to LEO and Margins<br>5.2 The Starship Platform is Changes the Game<br>5.3 Revenues from Launch Services in 2029<br>5.4 Revenue Forecasts for Starship in 2036<br>Section 6: Future Speculative Revenue Streams<br>6.1 Mars and Moon Missions<br>6.2 Space-mining<br>6.3 Space-Based AI Compute Data Centers</p><h3>Executive Summary</h3><p>This article and analysis present a transparent revenue and valuation forecast model for SpaceX, anchored first by enterprise value estimates derived from a 2027 revenue multiple approach and framed under both conservative and optimistic revenue cases for each major revenue line. Under the conservative case, the implied enterprise value is approximately $518B, while the optimistic case supports an enterprise value of approximately $675B, based on a 25× revenue multiple consistent with high growth technology IPO comparables.</p><p>All revenue assumptions, operating cost estimates, and valuation calculations are fully documented and available in the accompanying Excel financial model.</p><p>These enterprise values are supported by projected 2027 revenues of approximately $20.7B in the conservative scenario and $27.0B in the optimistic scenario, built from individual revenue estimates for Starlink internet services, Starlink enabled telecom services, and SpaceX launch activities. The model also extends these conservative and optimistic revenue scenarios forward, with total company revenues forecast to expand meaningfully in 2029 and again in 2036.</p><p>The analysis then examines SpaceX’s core operating segments, beginning with Starlink, which is identified as the primary long term value driver. Starlink internet services are projected to scale from approximately $7B to $8B in annual revenue by 2027 to roughly $20–$29B by 2029, driven by subscriber growth, expanded country level regulatory approvals, and increasing satellite bandwidth capacity. Longer range projections extend to approximately $53–$105B in annual revenue by 2036, reflecting a wide but plausible outcome range. The article also evaluates Starlink enabled telecom services; however, this segment remains relatively small, representing less than 1% of projected future revenues.</p><p>Finally, the article assesses SpaceX’s launch services business, emphasizing the structural shift enabled by the Starship platform and its ability to dramatically lower the cost to orbit for payloads. Launch revenues are forecast to grow from approximately $5B in 2025 to $6.8B–$8.1B in 2027, expanding further to roughly $10–$14B by 2029, with potential to reach approximately $13B–$20B by 2036.</p><p>Beyond these core businesses, the model explicitly addresses longer term opportunities such as lunar missions, Mars exploration, space mining, and space based compute, concluding that while strategically interesting, they remain too speculative to include in a disciplined valuation today.</p><p>Taken together, the analysis positions SpaceX as a company with credible near term revenue visibility, unusually strong long term optionality, and a valuation grounded in clear, auditable financial assumptions rather than narrative driven speculation. An Enterprise valuation in the range of $500B to $700B for an upcoming IPO in 2026 would be reasonable.</p><h3>1. Introduction to the Model</h3><h4>1.1 Target Readers</h4><p>The target readers for this article are financial analysts, investors, and space enthusiasts. First, financial analysts will understand the business and see the economics of their activities stripped down to the most important factors driving profitability and growth. Second, investors will learn about the major revenue lines and how SpaceX can specifically drive growth in the future. Finally, space enthusiasts will understand how the complex engineering boils down to a few key factors that satisfy customers, investors, and drive profits.</p><h4>1.2 No Outside Interests Were Involved</h4><p>The author created this valuation unprompted, unpaid, and has no association with SpaceX or other space-related companies. The author performed this evaluation for their own fair assessment of the potential of the SpaceX IPO with the intent of being fair, unbiased and accurate.</p><h4>1.3 Supporting Excel Revenue, Operating, and Valuation Model on GitHub</h4><p>This article is accompanied by an Excel valuation workbook on GitHub with 8 worksheets that detail the revenue assumptions and projections across SpaceX’s major business segments. The workbook models revenues (and direct operating costs) for Starlink broadband, SpaceX launch operations, and Starlink-enabled telecom services, providing segment-level breakdowns that support the analysis in the article, and can be downloaded for free but not resold.</p><p><a href="https://github.com/briang1621/SpaceX-Valuation-Model-Unbiased">GitHub - briang1621/SpaceX-Valuation-Model-Unbiased: This high quality, unbiased SpaceX valuation model by Dr. Brian Scott Glassman shows the core drivers of revenue, profitability, and growth for finance analysts, investors, and space enthusiasts, and is accompanied by a 40 page article. The author is a former senior mechanical engineer, NYU professor, and PhD in Business Innovation.</a></p><h4>1.4 Accuracy of the Valuation Model</h4><p>This article is meant to be broad; the accuracy of the private information is hard to verify given that SpaceX is currently private, but the author has taken efforts to do background research from leading online sources tracking SpaceX and its activities and to make grounded estimates and set key assumptions. The model and excel workbook highlights the key assumptions so readers can make adjustments as they see fit.</p><h4>1.5 Summary of the SpaceX Valuation Model</h4><p>SpaceX is fundamentally a large bet on global Starlink services and space launch activities, with Starlink accounting for approximately 65% of total revenue in 2025 and roughly 70% of total profits. The analysis presented below shows that it is highly feasible that Starlink could grow from approximately $9.6B in estimated annual revenue in 2025 to $20B–$29B by 2029. Further forecasting shows Starlink could possibly reach $50B–$105B in global revenues by 2036, with operating margins as high as 85% due to economies of scale and advances in satellite bandwidth capacity; hence, this is a very valuable revenue stream.</p><p>Similarly, SpaceX’s launch activities will increasingly be driven by Starship and could generate approximately $10B–$14B in revenue by 2029, scaling to roughly $14B–$22B by 2036. With the help of the new launch platform, Starship (7x launch capacity), margins could move from an estimated profit margin of 55% to potentially 80%. Telecom services provided by Starlink also represent another revenue stream; these are estimated to be in the tens of millions until 2029 and grow to potentially $20B by 2036.</p><p>In aggregate, SpaceX has the potential to become an exceptionally valuable and profitable company if it continues its record of operational excellence and low-cost execution.</p><p>From a valuation standpoint, SpaceX is best analyzed using a revenue multiples model. Under this perspective, projected revenues from 2027 and a multiple of 25x selected from comparable IPOs put today’s value between <strong>$518B and $675B.</strong> Given the potential for monopoly-like positioning in both global satellite internet and launch services, this estimate seems reasonable. However, space operations do have significant risks, and going higher than a 25x multiple seems a bit excessive.</p><h4>1.6 Estimate of Fair Value for SpaceX for an IPO</h4><p>This author would push for a lower valuation of 20× to 25× of 2027 forecasted revenues. This is due to limited financial information, risks of space operations, risks associated with the CEO’s public image and popularity, and the market’s experience valuing companies heavy in space operations.</p><h4>1.7 Limitations on Available Information</h4><p>There is limited verification of financial costs, as SpaceX is a private company; it is expected that the IPO filing will clarify this. Risks due to space operations include proving out the Starship platform and space-based risks associated with large satellite constellations, like solar flares impacting the network.</p><p>Additionally, Elon Musk is a visionary and also controversial figure; love or hate him, one can obviously see that Tesla’s stock price was impacted due to the myriad of political attacks against him by politicians, left leaning media, Delaware courts rulings, and even the FAA delaying launches under Biden, and this must be factored in.</p><p>Finally, the market has limited experience valuing mega-cap satellite and space launch operations. Currently traded companies as of January 2026 are SATS ($35.5B), PL ($9B), VSAT ($6.2B), IRDM ($2B), and SPIR ($410M), with none operating their own space launch services to dramatically lower launch costs.</p><p>Again, This author would advocate for an IPO valuation toward the lower end of this range, 20× to 25×, around $500B and $675B.</p><h4>1.8 Other SpaceX Valuation Models</h4><p>ARK Invest has an open-source valuation model for SpaceX on GitHub; however, in this author’s opinion, it is not realistic. First, it is overly detailed and represents dramatic overkill due to its complexity, and second, it values opportunities such as Moon and Mars colonization at trillions of dollars. Specifically, when a stock is over-modeled, such as through extensive Monte Carlo simulations, there is a high risk of attributing growth and revenue outcomes to faulty assumptions. Second, the Moon and Mars opportunities mentioned below are, in the author’s view, a minimum of 10 years away from generating any revenue. Finally, with ARK Invest being a fund that holds SpaceX, there is a major conflict of interest in ARK Invest’s publicly released valuation.</p><h4>1.9 Helping the Community Build a Valuation Model</h4><p>Ultimately, the investment community must build its own valuation models for SpaceX, and industry experts tend to converge on clear, simple metrics that directly drive revenue and profitability. Note this valuation appears simple, but it is informed by substantial analysis not shown herein. Further, it was built in the hoped that this clear yet realistic revenue estimation methodology will help guide others in further refining their own estimates.</p><h4>1.10 Expertise of the Author: Dr. Brian Scott Glassman</h4><p>I have been following SpaceX operations fervently for 15 years. As a trained senior mechanical engineer with a Ph.D. in Business and Innovation Management, I understand the technical details of the company’s operations and engineering and place them in perspective in terms of their true business value. In short, fancy engines or large rocket ships are nice, but they ultimately must serve the goal, which is to reliably beat the competition, generate profitable revenue, and fuel growth, hence my focus to highlight these aspects.</p><h3>2. SpaceX’s Revenue Forecast Results</h3><p>This analysis focuses on the high revenue opportunities for SpaceX with estimates focused on the valuation year taken in 2027, and then projected 10 years (2036) from the current day.</p><p>To accomplish this, the article will start by reviewing SpaceX’s near-term revenue opportunities from largest to smallest, beginning with Starlink Internet Services, followed by Space Launch Services, and then telecom services provided by Starlink. Current, non-revenue-generating opportunities are also discussed, such as the Moon, Mars, space mining, and space-based AI data centers.</p><h4>2.1 SpaceX Revenue Forecast Tables in 2029 and 2036</h4><p><strong>Below is the author’s estimates of SpaceX’s Revenue Projections in 2025 to 2029.</strong></p><p>The second is the Optimistic Scenario, with 2027 highlighted as the basis for the enterprise valuation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*apUhTEi4vANdbsb7QNTjcw.png" /><figcaption>SpaceX Conservative Revenue and Profit Projections: 2025 to 2036</figcaption></figure><p>The second is the Optimistic Scenario for revenue and profit projections.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rD2VuekYqFA5_IdUcxaX-A.png" /><figcaption>SpaceX Optimistic Revenue and Profit Projections: 2025 to 2036</figcaption></figure><h4>2.2 SpaceX Comparable IPOs</h4><p>The valuation for SpaceX is based on a 2027 revenue multiple approach, using a selected multiple of 25x. This multiple is informed by large IPOs of high-growth, comparable technology companies, including Anduril Industries at 30.5x, Cloudflare at 27.6x, CrowdStrike at 20.6x, and Slack Technologies at 21.0x revenue multiples at IPO, with an average multiple of approximately 24.9x. Based on these comparables, a 25x multiple represents a reasonable and balanced assumption for valuing a category-defining company like SpaceX.</p><p>The author selected 2027 as the base year because NTM+1 is the standard convention for high-growth IPO valuations, balancing near-term visibility with growth trajectory while remaining within the credible forecasting window institutional investors accept.</p><h4>2.3 SpaceX Valuation Results</h4><p>Moving on to the estimated enterprise value, under a conservative scenario, projected 2027 revenues of $20.7B and a 25x revenue multiple result in a base-case enterprise value of approximately $518.3B. Under the optimistic scenario, higher projected 2027 revenues of $27.0B combined with the same 25x multiple imply a valuation of approximately $674.5B.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C5KHFRJZ5sL0kGVsXlSqBA.png" /><figcaption>SpaceX Conservative Valuation Based on 2027 Revenues Multiple Approach</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mzUxnKPufh-V3m_ZzCQzZQ.png" /><figcaption><strong>SpaceX Optimistic Valuation Based on a 2027 Revenue Multiple Approach</strong></figcaption></figure><p>Again, the author does not believe a multiple higher than 25x is warranted given limited financial disclosure, the inherent risks of space operations, risks associated with the CEO’s public image, and the market’s limited experience valuing space-based operations companies.</p><h3>3. Revenues from Starlink Internet Services</h3><h4>3.1 Market Position and Competitive Landscape</h4><p>Starlink offers a high-profit revenue stream with little to no comparable competition. Competitors such as HughesNet and Viasat may advertise download speeds of up to 100 Mbps, but in practice they fall far short due to latency of 0.5 to 1 second. By comparison, Starlink offers communication speeds of around 200 Mbps and as high as 300Mbps with approximately 0.045 seconds of latency, which consumers will always favor (<a href="https://starlink.com/legal/documents/DOC-1470-99699-90">Link to Starlink Speed Specs</a>).</p><p>In addition, these competitors’ satellite internet services are geographically limited, whereas Starlink offers near-global coverage, excluding polar regions; however, availability in certain countries remains pending regulatory approval. Most importantly, Starlink has secured access in regions with high-value customers (see <a href="https://starlink.com/map">their coverage map</a>).</p><p>As of 2027, Starlink operates approximately 9,400 satellites in low Earth orbit, which can be viewed in real time on a live satellite map. These satellites provide internet service via a user installed dish that communicates with low Earth orbit satellites, which then relay data to an Earth based ground station typically located within 200 miles of the customer. Over oceans and remote regions, data is relayed through high bandwidth laser links between multiple Starlink satellites until it reaches a ground station, resulting in slightly higher latency.</p><h4>3.2 Pricing and Capacity Constraints</h4><p>Prices currently range from $50 to $150 per month in the United States, with new $50 plans released in January 2026. Pricing is tiered and based on download speeds. More importantly, the number of customers that can be served is hard limited by total available satellite bandwidth and the number of satellites in orbit. As a result, deploying next generation Starlink satellites with higher capacity will be a primary driver of expanded internet availability, increased speeds, and future subscriber growth.</p><h4>3.3 Subscriber Base</h4><p>Outside of dense urban centers, Starlink’s current deployment has been able to mostly keep pace with rural demand. SpaceX does not regularly publish official subscriber numbers, but in the event of an IPO, this metric will become public and closely scrutinized. SpaceX announced in Nov 2025 that they reached 8 million active subscribers.</p><h4>3.4 Revenue Estimates for Starlink</h4><p>Subscriber statistics are tracked at <a href="https://idemest.com/reports/starlink-country-data-tracker/">IDEMEST</a> but this report was not verified by the author. Revenue figures are estimates and are not officially published. For 2024, estimated revenue is approximately $8.2B. With roughly 4.63 million subscribers, this implies an average annual subscription price of about $1,700 per user, or approximately $150 per month. With 8 million subscribers in November 2025 and the same pricing, expected revenue would be approximately $9.6B. However, the author expects the average subscription price per customer to decline as Starlink introduces lower cost plans and bandwidth capacity increases.</p><h4>3.5 Unit Economics and Servicing Costs</h4><p>The largest cost for servicing customers is the cost of maintaining and growing the Starlink constellation of satellites. The cost to service each subscriber is fairly difficult to calculate and not published by SpaceX. Here SpaceX has some excellent developments to lower costs per customer served like increase bandwidth capacity with next generation Starlink satellites, lower launch costs due to Starship.</p><p>As a rough estimate, a Falcon 9 launch costs SpaceX approximately $45M and is reported to carry about 29 Starlink satellites each costing $200k, implying roughly $1.75M per satellite in launch costs. Second generation (V2) Starlink satellites are estimated to cost approximately $800K each to manufacture and weigh about 800 kg, allowing Falcon 9 to deploy roughly 25 per launch with a deploy cost of $2.6M per V2 satellite. Unpublished estimates suggest a capacity of approximately 1,000 users per first generation satellite and about 4,000 users per second generation satellite.</p><p>Using second-generation Starlink satellites as a reference, this implies a total upfront cost of about $2.6M per satellite, which amortized over five years equates to approximately $130 per user per year, or about $11 per month per user. This represents roughly 7% of the average annual subscription fee. Users also require a satellite dish on their premises, which is included free at signup and retails for $299, though it is likely to cost SpaceX closer to $100.</p><p>The cost to service a customer with internet via Starlink is expected to decline significantly. The author estimates it could fall toward roughly $2 per user per month as satellite user capacity improves and future satellites are deployed using the lower cost Starship launch platform. Accordingly, the author expects a portion of revenue, approximately 10%, to be allocated to launching new Starlink satellites, managing the satellite constellation, and operating land based ground stations.</p><h4>3.6 Rapid Expansion of Starlink via Access to Large Countries</h4><p>The rapid addition of large subscriber bases will be driven by market openings in large countries pending regulatory approval such as India, Pakistan, Vietnam, Thailand, Morocco, Egypt, and Turkey. Note that China does not look like a likely prospect due to preferred government internet controls. These agreements are likely to add between big segments of customers 2 million and 10 million additional users over the next three years. At a discounted average subscription price of $100 per month, this would contribute approximately $2.4B to $12B in additional annual revenue after 3 years which is quite sizable.</p><h4>3.7 Growth of the Subscriber Base</h4><p>Some online sources cite annual subscriber growth of approximately 40% in developed markets, but this appears aggressive. A more conservative assumption is declining growth rates over time. Starting from a base of 8 million subscribers in November 2025 and assuming conservative 20% annual growth from 2025 to 2029 results in approximately 16.6 million subscribers, while an optimistic 25% growth rate results in approximately 19.5 million subscribers.</p><p>Taken together, over a three year horizon, average subscription prices may decline modestly to approximately $100 to $125 per month in order to attract more customers. With a subscriber base of 16.6 million to 19.5 million users, this implies annual Starlink revenues in 2029 in the range of approximately $19.9B to $29.3B.</p><p>For projections from 2030 to 2036, annual growth rates of 15% and 20% were assumed for conservative and optimistic cases, respectively, yielding revenues in 2036 between approximately $53B with 40 million subscribers and $105B with 70 million subscribers. These estimates may be overstated, as growth rates that far into the future are inherently difficult to forecast, although reaching 70 million global subscribers is feasible given strong worldwide demand for internet services and continued development across many nations.</p><h4>3.8 Ancillary Revenues from Aviation and Marine are Small</h4><p>Additional revenue from aviation and maritime customers is meaningful but relatively small compared with core land based consumer internet subscriptions. As of 2025 there are roughly 370 cruise ships worldwide, with equipped ships typically using two to four Starlink terminals. Pricing ranges from $4,000 to a negotiated average of approximately $8,000 per month per ship. It is estimated that 100% of cruise ships will be equipped with Starlink by 2027, generating approximately $35 million in annual revenue.</p><p>Similarly, long haul commercial aircraft represent another opportunity. With approximately 4,000 aircraft globally equipped at $4,000 per month, this segment would generate approximately $192 million in annual revenue in 2029, assuming 100% saturation that year. From 2029 to 2036, growth rates are estimated at 2% to 3%, broadly in line with fleet growth.</p><p>Short haul aircraft are less attractive candidates for Starlink upgrades due to higher installation costs and shorter flight times, which reduce passenger willingness to pay for in flight internet access outside of premium airline carriers. However, with 10% to 15% saturation expected by 2029 and a lower monthly subscription rate of $2,000, revenues are estimated at $48 million to $72 million in 2029. Overall, both maritime and aviation Starlink revenues are expected to remain approximately 1% of land based revenues and may not be material enough to warrant detailed modeling by other analysts.</p><h4>3.9 Starlink Expected Revenues in 2027 &amp; 2029</h4><p>Starlink appears to be a strong cash generator, and given the limited public information available on costs, it is likely highly profitable, with estimated operating margins of approximately 70% to 85%. Under these assumptions, 2027 revenues are projected to range from $13.9 billion to $18.8 billion under conservative and optimistic scenarios, while 2029 revenues are estimated at $20.1 billion to $29.6 billion under similar assumptions.</p><h4>3.10 Starlink Revenue Projections for 2036</h4><p>Assuming slower annual subscriber growth from 2029 to 2039, it is reasonable to project 2036 annual revenues in the range of $53 billion to $105 billion for the conservative and optimistic scenarios respectively, with operating margins of approximately 70% to 84%. The wide revenue range reflects uncertainty in annual subscriber growth, which was modeled linearly at 15% to 20% for convenience, though in reality it is more likely to taper lower year over year.</p><p>Consequently, this would make Starlink a highly valuable revenue stream, but again these estimate are harder to project.</p><h4>3.11 Future Competition for Starlink</h4><p>There is current competition for Starlink, as noted above, from Viasat and HughesNet, but the largest future entrant that could pose a challenge is Amazon LEO. Amazon is currently launching satellites at a cadence of 27 per month and has deployed 180 satellites, or approximately 6% of the total. The FCC requires 1,618 satellites to be deployed for operations; therefore, based on the timing of this article, it would take roughly 4.4 years for the system to become operational, unless the deployment rate increases significantly via Blue Origin’s New Glenn rocket. Unfortunately, New Glenn has only completed one test-to-orbit flight to date.</p><p>Another potential future competitor is China, which could deploy a large constellation. The author believes these systems will remain primarily domestic or regionally limited, as most foreign governments in developed nations have major privacy and security concerns about routing their internet data through a Chinese data carrier. Developing nations will be less concerned, and a low price point would put China in a strong position with customers in these countries.</p><p>Again, barriers to entry in satellite internet are extremely high. Replicating a high-bandwidth, low-latency, low-Earth-orbit internet constellation would require tens of billions of dollars in satellite manufacturing and launch costs and would be most constrained by limited launch vehicle availability and cadence.</p><p>Blue Origin also plans to develop a competing satellite internet system for governments and data centers, the TeraWave space-based network, with an initial deployment schedule beginning in Q4 2027. The timeline to achieve meaningful scale remains extended. Assuming they use their top-line New Glenn launch vehicle’s approximately 45 metric ton capacity to low Earth orbit and satellite specifications broadly comparable to Starlink V2 at roughly 800 kilograms per satellite, approximately 50 TeraWave satellites could be deployed per launch.</p><p>To reach the announced constellation size of 5,280 satellites in low Earth orbit would therefore require on the order of 106 launches. Given that New Glenn has completed only a single test flight to date, it is reasonable to estimate that it will take a further four to seven years from the end of 2027 to complete constellation deployment, implying a 2032 to 2034 completion timeframe.</p><p>Despite the mentioned new entrants, Starlink is expected to have achieved significant global penetration, creating a decent competitive moat. Notably, Blue Origin is considering contracting SpaceX to launch a portion of its satellites, which could accelerate that timeframe. As a result, the author believes SpaceX is unlikely to face meaningful near-term competition from rival satellite internet constellations.</p><h4>3.12 Major Risks to Starlink</h4><p>However, competitive dynamics are not the only risk. Space is an inherently hostile operating environment, where low probability but high impact events such as extreme solar activity or debris cascade scenarios, often referred to as Kessler Syndrome, could disrupt satellite operations, forcing replacement satellites to be deployed and potentially reducing total network bandwidth.</p><h3>Section 4. Telecom Services Provided via StarLink</h3><p>Telecom ground coverage for cell phones is generally good to excellent in large cities across the world due to the significant profits generated from offering 5G and 4G cellular services (see the <a href="https://verizon.cellmaps.com/">cell phone coverage global map</a>). However, outside of major population centers, especially in developing nations, cell phone coverage becomes far more challenging. To erect a cell phone tower, there must be sufficient local traffic to justify the expense; as a result, many rural areas in less developed nations are not currently covered and are unlikely to be covered in the future.</p><p>Consequently, SpaceX launched a telecom service offering in 2025, enabled by newer generations of Starlink satellites that can support modern cell phones via the 4G/LTE protocol at 5 to 10Mbps. This allows near global coverage for phone calls and messaging, albeit with lower bandwidth data, and opens vast, previously underserved rural populations to cellular coverage. Furthermore, the number of customers that can be served concurrently per satellite is estimated to be significantly higher than for land based internet services in the 3,000 to 10,000 range, and thousands of SMS per second.</p><h4>4.1 Pricing Model for Telecom via Starlink</h4><p>The pricing and business model will most likely take the form of an add-on service offered through existing cellular providers in each country and may be priced by the carriers in the range of $2–$10 per month, depending on the country. SpaceX is likely to get 30% to 50% of the local cell carriers price. It is estimate that SpaceX will get $2/month or $24 per year per subscriber is taken as the conservative case with $3 or $36 per year as the optimistic case.</p><h4>4.2 Number of Subscribers for Telecom via Starlink</h4><p>The range of potential subscribers for Starlink’s Telecom services is difficult to estimate today, as it depends heavily on the quality of service, whether country-based telecom providers and national governments choose to promote or block the offering, the affordability of the service for rural populations, and if SpaceX puts a significant marketing push behind this service line. For example, a country such as Brazil alone could ultimately support millions of rural subscribers.</p><h4>4.3 Valuation of Starlink Telecom Services</h4><p>Shifting to the valuation, for scale, 8 million subscribers at $24 per year would generate approximately $192M in annual revenue, while 100 million subscribers would generate roughly $2.4B. Even at very large subscriber counts, this revenue stream would be small in comparison with Starlink’s core internet service revenues.</p><p>For 2027 the revenue will be less than $5 million. However, for 2029, the conservative case estimates roughly 8 million subscribers generating $194 million in revenue, while the optimistic case assumes 15 million subscribers generating approximately $540 million. As with other non core segments, this revenue remains small relative to Starlink’s primary services and may therefore receive limited marketing emphasis. In 2036, revenue projections range between $500M to 2.5B.</p><h3>Section 5. SpaceX Launch Services</h3><p>Valuing SpaceX’s launch services is complex because of an upcoming shift to a new launch vehicle named Starship. In 2025, SpaceX performed 170 launches, most with Falcon 9; of those, 75% did not generate direct revenue because they deployed Starlink satellites, but 42 of them were paid for by clients. With an average launch cost of $70M, rough math shows that it could have generated launch-related revenues of $2.94B, but published revenue for 2024 shows <a href="https://payloadspace.com/estimating-spacexs-2024-revenue/">$4.2B in launch services</a>. So in reality, 2025 launch revenues are probably closer to $5B. The additional revenues would be do to government contracts, manned missions, and priority missions priced at 2–4X normal due to <a href="https://spaceexplored.com/spacex-launches-2025/">expensive launch integration services. </a>However, with margins around 40% to 50%, expect $2.5B gross profit from launch operations from Falcon 9 in 2025.</p><p>For financial analysts, the core economic metric to monitor is the cost per metric ton deployed to low Earth orbit (LEO), as this provides a consistent baseline for comparing launch providers across countries. Deployments to medium Earth orbit, high Earth orbit, or geosynchronous orbit are effectively derived from this baseline, reflecting reduced payload capacity relative to LEO for the same launch vehicle.</p><p>The second number to watch is the launch cadence, or launches per year, as that dictates how much tonnage can be put into space. SpaceX performed <a href="https://spacestatsonline.com/launches/year/2025">170 launches in 2025</a>, with the next closest being China’s CASC with an impressive 93 launches. Note that growth in launch cadence is slow, as manufacturing and readying rockets for launch is a lengthy procedure.</p><h4>5.1 SpaceX Cost Per Ton to LEO and Margins</h4><p>After a detailed analysis not shown here, the SpaceX Falcon 9 launch platform is effectively priced for customers at around $3M to $5M per metric ton to low Earth orbit (cost variation is due to payload configuration variations) with one rocket putting up 22.8 metric tons. The author’s estimates put margins at 40%, so one metric ton costs SpaceX between $1.2M to $2M to launch to an equivalent LEO.</p><h4>5.2 The Starship Launch Platform is Changes the Game</h4><p>Starship, due to its large size and reusability, fundamentally disrupts the launch industry by putting 150 metric tons to LEO or 6.5x Falcon 9 in one launch. Based on this author’s internal calculations, Starship could dramatically lower the cost to Low Earth Orbit to approximately $150k per ton while deploying 150 tons per launch if fully reusable; about $300k per ton while deploying 200 tons if only the booster is reused; and roughly $450k per ton while deploying 250 tons if the entire vehicle is expended.</p><p>Starship is entering version 5, and has proven in the last flight to complete its mission objectives. The $5B to $10 in R&amp;D expenses for its development are now shifting toward a rapid manufacturing line for producing hundreds of Starships. Per unit cost for a single Starship are not public but can be estimated at approximately $100M ($50M for the upper stage and $50M for the lower stage), with an additional fixed $8M per launch for fuel and support costs. These assumptions underpin the payload cost estimates in the first paragraph.</p><p>With per-ton launch costs measured in the hundreds of thousands rather than millions, Starship fundamentally changes the economics of launch. This gives SpaceX the ability to aggressively undercut competitors while retaining pricing control. For example, reducing customer pricing by half from $3M per ton to $1.5M per ton would still represent an approximate 90% gross margin for SpaceX, assuming an underlying cost of $150k per ton for a fully reusable Starship.</p><h4>5.3 Revenues from Launch Services in 2029</h4><p>Currently, SpaceX launches a Starship approximately once every three months, but with two launch towers at its Texas facility, a monthly cadence is achievable. It is expected that Starship could reach a launch rate of twice per month by 2027. At that pace, and assuming a hypothetical price of $1M to $1.3M per metric ton to low Earth orbit, annual Starship launch revenue would be approximately $3.6B. With additional customer integration and mission-specific service charges, total revenue could reasonably increase to roughly $6B. Given the significantly lower marginal launch costs, operating margins could expand to approximately 85%, implying retained profits of around $5B.</p><p>For 2027, the projections assume approximately 10 paid Starship launches and between 35 and 45 paid Falcon 9 launches, generating combined launch services revenue of roughly $6.8B to $8.1B. By 2029, total launch volume could increase by approximately twofold as lower launch prices stimulate incremental demand, allowing launch services revenue to expand to an estimated $10B to $14B annually.</p><h4>5.4 Revenue Forecasts for Starship in 2036</h4><p>Estimating future launch cadence or the incremental demand generated by lower launch pricing is inherently difficult. If the U.S. government were to pursue its own high speed Starlink-like satellite constellation, dubbed “<a href="https://www.spacex.com/starshield">Starshield</a>” equipped with high resolution radars and optics for use by the US air-force, this alone could represent a $10B opportunity on top of existing launch demand. Conservatively, launch revenues could reach approximately $18B by 2036, roughly doubling from 2029 levels, while maintaining profit margins near 80% due to Starship’s exceptionally low operating costs.</p><h3>Section 6: Future Speculative Revenue Streams</h3><h4>6.1 Mars and Moon Missions</h4><p>If the U.S. government were to fully fund a lunar base, it could represent a profitable endeavor for SpaceX. However, extensive data collected from rovers over more than a decade strongly indicate that both the Moon and Mars are largely barren environments. Claims to the contrary are optimistic beyond the evidence currently available.</p><p>Hence, It is this author’s view that SpaceX’s public emphasis on Mars missions serves primarily as a long-term motivational narrative for engineers, investors, and the public rather than a near-term revenue opportunity.</p><h4>6.2 Space-mining</h4><p>This opportunity remains highly speculative. While asteroid mining could theoretically yield hundreds of billions of dollars in rare-earth minerals, it is likely at least a decade away from assessing technical and economic viability through robotic missions to asteroids. As such, space mining cannot be responsibly incorporated into any current valuation estimates.</p><h4>6.3 Space-Based AI Compute Data Centers</h4><p>The author is skeptical about deploying space-based compute and AI data centers as large satellite constellations or as couple massive single deployments. In this concept, Elon proposes using highly effective space-based solar arrays to eliminate ground-based electricity bills, office and land leasing costs, and government taxes, but this simply trades operating costs for expensive, heavy satellites that are difficult to maintain and expensive to launch.</p><p>Put plainly, it would require deploying the most power-hungry, heat-generating computing satellite ever built, with power need roughly factor of 100x the next closes satellite, into the hostile environment of space.</p><p>Any space-based AI cluster must also compete, on an ROI basis, with land-based data centers. Lets take the large AI satellite approach, a space-based 20 MW solar array might weigh around 300 tons to power a decent-sized space-based AI compute cluster, while the compute hardware itself could weigh another 700 tons. Even under an highly optimistic assumption of $100k per ton to orbit using Starship, launch costs alone would be about $100M, merely to avoid roughly $15M-$40M per year in electricity costs, office and land leasing costs, and government taxes, in the United States. This estimate does not even include the cost of the specialized hardware being launched.</p><p>If we consider the smaller AI satellite approach, at roughly 4x the size of the Starlink V2 satellites (then approximately 3,200 kg each), the economics may work out better. A Starship could launch 45 satellites at once, and governments and defense systems would likely pay a premium for 4x to 8x land-based AI compute costs in systems that offer privacy and are resistant to tampering or sabotage. Here, each satellite would have its AI compute reserved, guaranteeing an immediate return on investment. However, cost-sensitive consumer and business applications will be less interested at those higher price point. Consequently, calculating the revenue stream AI space based AI compute of Starlink will be difficult without specifications for a Starlink AI satellite for comparison, as there are too many design variables to consider. Hopefully, those specifications will be released soon.</p><p>No matter how you look at it, the compute costs of AI in space will be higher, but the other benefits will be very attractive to some customers. Innovation is difficult to predict, but over several years the cost of space-based AI compute would come down. However, the author does not expect it to fall below terrestrial AI compute costs, which is likely to see rapid innovation, including improvements in AI compute efficiency and reductions in electricity and other associated costs via simple negotiations.</p><h3>About the Author Dr. Brian Scott Glassman</h3><p>The author is a fan of space operations and closely follows developments at SpaceX and other space companies. He holds a Ph.D. in Product Innovation from Purdue and three engineering degrees. He currently specializes in AI strategy and software. He can be contacted on <a href="https://www.linkedin.com/in/drbrianglassman/">LinkedIn</a>, and his written work can be viewed on his personal website, <a href="https://drbrianglassman.com/">https://drbrianglassman.com</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=17601d5290f2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predict/spacex-valuation-model-estimating-fair-value-for-an-ipo-17601d5290f2">SpaceX Valuation Model: Estimating Fair Value for an IPO</a> was originally published in <a href="https://medium.com/predict">Predict</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Eight AI Thought Leadership Articles from AInspire & Dr. Brian Scott Glassman]]></title>
            <link>https://drglassman.medium.com/eight-ai-thought-leadership-articles-from-ainspire-dr-brian-scott-glassman-eeda0a81da22?source=rss-4584d7533d83------2</link>
            <guid isPermaLink="false">https://medium.com/p/eeda0a81da22</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[corporate]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Dr. Brian Scott Glassman]]></dc:creator>
            <pubDate>Tue, 02 Dec 2025 23:01:56 GMT</pubDate>
            <atom:updated>2025-12-02T23:01:56.170Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7eQGowaaqRXj89FC_db4nQ.jpeg" /><figcaption>AInspire.ai Technology Consulting Articles for 2025</figcaption></figure><p>AI Thought Leadership for CTOs</p><p>At <a href="https://ainspire.ai/">AInspire.ai</a>, we publish research and frameworks to help organizations successfully adopt and govern artificial intelligence. This article provides summaries and key takeaways from our latest whitepapers, each available as a free PDF download from our website.</p><p>Articles Discussed Herein:</p><ol><li>A Framework for Managing a Company Wide AI Program</li><li>20 Tips for Implementing Effective AI Workflow Automations</li><li>Top 12 Abilities for AI-Native Software Developers: What to Look For When Hiring</li><li>What A CTO Must Budget For AI Coding Tools</li><li>How AI Is Revolutionizing Product Management</li><li>Why AI-Driven Software Development Makes Agile Obsolete</li><li>Governing AI in Healthcare: Six Critical Imperatives</li><li>The CTO’s Essential Guide to Securing AI MCP Servers: Four Audits That Can’t Wait</li></ol><h3>1. A Framework for Managing a Company Wide AI Program</h3><p>PDF Link: (<a href="https://ainspire.ai/whitepapers/ai-strategy-framework">https://ainspire.ai/whitepapers/ai-strategy-framework</a>)</p><p>Author: Dr. Brian Scott Glassman</p><p>Short Summary: This whitepaper provides executives and AI practitioners with a comprehensive three-layer framework covering stakeholders, AI leadership, and portfolio management to systematically guide corporations from initial AI use case exploration through pilot projects and test deployments to successful enterprise-wide implementation.</p><h3>Key Takeaways</h3><ul><li>This article provides a structured three-layer framework that integrates stakeholders, AI leadership, and AI portfolio management into a cohesive corporate AI program.</li><li>Readers can expect a comprehensive understanding of the four-phase AI portfolio approach that progresses from use case exploration through pilot projects, test deployments, and company-wide rollouts.</li><li>The framework offers clear guidance on establishing AI governance structures including ethics committees, risk management protocols, and regulatory compliance processes.</li><li>Organizations will learn how to systematically mitigate risk while maximizing learning opportunities at each phase of their AI journey.</li></ul><h3>2. 20 Tips for Implementing Effective AI Workflow Automations</h3><p>PDF Link: (<a href="https://ainspire.ai/whitepapers/ai-workflow-automation-tips">https://ainspire.ai/whitepapers/ai-strategy-framework</a>) Author: Dr. Brian Scott Glassman</p><p>Short Summary: This practical guide provides CTOs and operations leaders with 20 battle-tested tips across four phases — strategic selection, system design, deployment transition, and continuous improvement — to help organizations deploy AI-powered workflow automations that balance innovation with risk management and deliver measurable ROI.</p><h3>Key Takeaways</h3><ul><li>This article provides a four-phase framework covering strategic selection, system design, deployment, and continuous improvement for AI workflow automation projects.</li><li>Readers will learn how to evaluate workflows using a risk-tiered approach from Level 1 (minimal harm) to Level 4 (catastrophic consequences) to prioritize automation candidates.</li><li>The guide offers a practical ROI calculation methodology using conservative multipliers: 10x throughput, 3x labor reduction, and 5x quality improvement.</li><li>Organizations will understand how to build AI redundancy and runaway prevention mechanisms to control costs and ensure reliability in production environments.</li><li>Teams will learn deployment strategies including side-by-side replication and gradual transition approaches that minimize risk during the automation rollout.</li></ul><h3>3. Top 12 Abilities for AI-Native Software Developers: What to Look For When Hiring</h3><p>PDF Link: <a href="https://ainspire.ai/whitepapers/ai-native-developer-abilities">https://ainspire.ai/whitepapers/ai-native-developer-abilities</a> <br>Author: Dr. Brian Scott Glassman</p><p>Short Summary: This comprehensive hiring guide identifies the 12 key abilities, perspectives, and knowledge areas that make software developers highly effective at using AI coding tools like Claude Code and Cursor — essential reading for CTOs and recruiters building teams in the AI-first era where traditional coding skills are no longer the primary hiring criteria.</p><h3>Key Takeaways</h3><ul><li>This article demonstrates that AI coding tools deliver 20x to 100x productivity gains, fundamentally changing what CTOs should prioritize when evaluating developer candidates.</li><li>Readers will learn the three key perspectives AI-native developers must possess: acting as a virtual department head, optimizing code for AI consumption, and using processes to de-risk AI outputs.</li><li>The guide identifies critical abilities including multitasking at scale across multiple AI agents, rapid problem solving, and building accurate mental models of AI behavior.</li><li>Organizations will understand that traditional skills like expert hand-coding and Agile methodology knowledge are no longer necessary for AI-native development roles.</li><li>Recruiters will learn to evaluate candidates for essential knowledge spanning solution architectures, cross-area expertise, prompt engineering, and token budget management.</li></ul><h3>4. What A CTO Must Budget For AI Coding Tools</h3><p>PDF Link: <a href="https://ainspire.ai/whitepapers/ai-coding-budget">https://ainspire.ai/whitepapers/ai-coding-budget</a></p><p>Author: Dr. Brian Scott Glassman</p><p>Short Summary: This detailed financial analysis provides CTOs with practical per-developer budget projections ranging from $1,500 to $3,000 monthly, comprehensive cost comparisons between Claude AI and GPT-5 for coding tasks, and a methodology to forecast AI coding expenses for teams of any size from 1 to 100 developers.</p><h3>Key Takeaways</h3><ul><li>This article provides specific monthly budget recommendations of $1,500-$3,000 per developer based on realistic usage patterns and team size considerations.</li><li>Readers will gain a comprehensive cost comparison between Claude AI (Opus/Sonnet) and GPT-5 for various coding tasks and use cases.</li><li>The analysis demonstrates how both conservative and aggressive AI adoption scenarios result in daily coding costs stabilizing around $70 per hour as efficiency gains balance price deflation.</li><li>CTOs will receive detailed monthly and yearly budget tables for teams ranging from 1 to 100 developers with clear cost projections.</li><li>Organizations will understand multi-agent coding cost analysis with projections for running 1, 3, and 5 concurrent AI agents simultaneously.</li></ul><h3>5. How AI Is Revolutionizing Product Management</h3><p>PDF Link: <a href="https://ainspire.ai/whitepapers/ai-revolutionizing-product-management">https://ainspire.ai/whitepapers/ai-revolutionizing-product-management</a></p><p>Author: Dr. Brian Scott Glassman</p><p>Short Summary: This industry insight reveals how the ACID framework (AI Context Informed Decisions + Product Management) enables product teams to compress what traditionally required 2–3 months of research, strategy development, and documentation into just 14 days without sacrificing quality, while generating prototypes 1,000 times faster than traditional designers.</p><h3>Key Takeaways</h3><ul><li>This article introduces the ACID + Product Management framework that compresses traditional 2–3 month product planning cycles into just 14 days.</li><li>Readers will learn how AI-powered insight generation synthesizes social media comments, support tickets, and market data into comprehensive 50–100 page contextual documents.</li><li>The guide demonstrates how to create 20–40 page AI-ingested strategy documents that enable real-time market analysis and strategic vision development.</li><li>Product managers will discover that AI generates graphical prototypes 1,000 times faster than traditional designers and creates superior agile feature tickets with user stories and unit tests.</li><li>Teams will understand the four-process framework covering insight generation, strategic vision development, product formation, and execution iteration.</li></ul><h3>6. Why AI-Driven Software Development Makes Agile Obsolete</h3><p>PDF Link: <a href="https://ainspire.ai/whitepapers/ai-driven-development-agile-obsolete">https://ainspire.ai/whitepapers/ai-driven-development-agile-obsolete</a> Author: Dr. Brian Scott Glassman</p><p>Short Summary: This provocative future vision article argues that after 25 years of dominance, Agile methodology’s core principles — designed for human-paced development — are being systematically broken by AI models that generate hundreds of lines of code per minute, making sprint-based planning, daily standups, and traditional team structures obsolete.</p><h3>Key Takeaways</h3><ul><li>This article demonstrates how AI compresses development cycles from months to weeks and weeks to days, rendering sprint-based planning fundamentally obsolete.</li><li>Readers will understand why AI mesh agents coordinating simultaneously across codebases move faster than human daily standups can possibly manage.</li><li>The analysis shows how development velocity becomes completely uncoupled from human effort when AI generates hundreds of lines of code per minute.</li><li>Organizations will learn that the marginal cost of generating extra code becomes minimal, eliminating Agile’s simplicity principle and enabling better results through multiple iterations.</li><li>CTOs will discover that AI-powered customer simulation and intent analysis resolves the new bottleneck: understanding genuine customer needs rather than coding capacity.</li></ul><h3>7. Governing AI in Healthcare: Six Critical Imperatives</h3><p>PDF Link: <a href="https://ainspire.ai/whitepapers/ai-governance-healthcare">https://ainspire.ai/whitepapers/ai-governance-healthcare</a></p><p>Author: David Rivkin Ph.D.</p><p>Short Summary: This industry insight provides healthcare organizations with a comprehensive governance framework addressing six critical imperatives — patient safety, regulatory compliance, data privacy, algorithmic fairness, organizational accountability, and stakeholder trust — essential for responsible AI deployment in clinical settings where systems directly impact patient care.</p><h3>Key Takeaways</h3><ul><li>This article provides guidance on patient safety and clinical risk mitigation through rigorous validation, continuous monitoring, and human oversight mechanisms for AI systems.</li><li>Readers will understand how to navigate legal and regulatory compliance including FDA approval requirements, HIPAA privacy protections, and emerging AI-specific regulations.</li><li>The framework addresses data privacy and security protection with comprehensive protocols, privacy safeguards, and breach response mechanisms for sensitive patient health information.</li><li>Healthcare organizations will learn to ensure algorithmic fairness and health equity through demographic bias analysis and continuous monitoring across patient populations.</li><li>Leaders will understand how to establish organizational accountability structures, clear roles and responsibilities, and incident response protocols for AI governance.</li></ul><h3>8. The CTO’s Essential Guide to Securing AI MCP Servers: Four Audits That Can’t Wait</h3><p>PDK Link: <a href="https://ainspire.ai/whitepapers/securing-mcp-servers">https://ainspire.ai/whitepapers/securing-mcp-servers</a></p><p>Authors: David Rivkin Ph.D., Brian Scott Glassman Ph.D.</p><p>Short Summary: This security whitepaper provides CTOs with essential audit methodologies for Model Context Protocol (MCP) server deployments, examining four critical vulnerabilities — widespread misconfiguration, retrofitted security features, type safety failures, and vast attack surfaces — across the 15,000+ MCP servers deployed worldwide where rapid adoption has outpaced security maturity.</p><h3>Key Takeaways</h3><ul><li>This article reveals that research analyzing over 7,000 publicly accessible MCP servers found hundreds vulnerable to the NeighborJack exploit, with approximately 70 servers exhibiting severe compound flaws enabling complete host machine compromise.</li><li>Readers will learn that OAuth 2.1 support was only added to MCP in March 2025 after production adoption, leaving enterprises to retrofit critical security protections.</li><li>The guide demonstrates how schemaless JSON without type safety allows data type mismatches to propagate silently, creating catastrophic consequences in financial, healthcare, and industrial systems.</li><li>Security teams will understand the 20+ distinct attack vectors spanning supply chain compromises, authentication bypass, privilege escalation, cross-agent contamination, and prompt injection attacks.</li><li>CTOs will receive actionable audit methodologies including network scanning, OAuth verification, penetration testing with OWASP ZAP and Burp Suite, SBOM scanning, and ML-based anomaly detection.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eeda0a81da22" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Top 12 Abilities for AI-Native Software Developers: What to Look For When Hiring]]></title>
            <link>https://medium.com/predict/top-12-abilities-for-ai-native-software-developers-what-to-look-for-when-hiring-cf4c3e4cdb56?source=rss-4584d7533d83------2</link>
            <guid isPermaLink="false">https://medium.com/p/cf4c3e4cdb56</guid>
            <category><![CDATA[hiring]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Dr. Brian Scott Glassman]]></dc:creator>
            <pubDate>Thu, 27 Nov 2025 03:08:44 GMT</pubDate>
            <atom:updated>2025-11-27T03:08:44.660Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bcNZP5mR_zYuuHgcMdTwAQ.jpeg" /><figcaption>Top 12 Abilities for AI-Native Software Developers: What to Look For When Hiring</figcaption></figure><p>The Key Abilities and Knowledge that Software Developers should have to Excel at Using AI Coding Tools</p><h3>Top 12 Abilities for AI-Native Software Developers: What to Look For When Hiring</h3><p>The Key Abilities and Knowledge that Software Developers should have to Excel at Using AI Coding Tools</p><h3>Overview</h3><p>AI-Native Software Developers represent the future of Software Development and technology departments. This article focuses on the key attributes that make an excellent AI-Native Software Developer and engineer.</p><p><strong>Author:</strong> Dr. Brian Scott Glassman, VP of Product Management at <a href="https://ainspire.ai/">Ainspire.ai Technology Consulting</a></p><h3>Sections in this Article</h3><ol><li>Why CTOs Need to Look for AI-Native Developers</li><li>Key Perspectives Required by AI-Native Software Developers</li><li>Key Abilities Needed by AI-Native Software Developers</li><li>Natural Talents of AI-Assisted Software Developers</li><li>Required Knowledge for AI-Savvy Software Developers</li><li>Skills No Longer Needed by AI-Driven Software Developers</li><li>An Interview Process for Finding an AI-Native Software Developer</li></ol><h3>AI-Assisted Software Development Tools Are The Future</h3><p>AI-assisted software development tools, such as Anthropic’s Claude Code, Anysphere’s Cursor, and Google’s Gemini, are rapidly becoming foundational to the future of software development, DevOps, and IT infrastructure optimization. The productivity gains reported by experts at Boston Consulting Group are remarkable, with them citing evidence that development teams are achieving 20x to 100x increases in feature output and productivity compared to traditional manual coding workflows. Honestly, these productivity gains are obvious to anyone who uses AI coding tools in 2025.</p><h3>CTOs Need to Find the Best Engineers at Using AI</h3><p>As a result, CTOs and HR recruiters must reorient themselves around AI-first methodologies and focus on hiring and retaining engineers whose perspectives, traits, skills, and knowledge make them highly effective at using AI coding tools.</p><p>The group of PhDs and researchers at <a href="https://ainspire.ai/">Ainspire.ai</a> have been exploring how to scale and de-risk the use of AI software development tools for larger organizations. In this pursuit they studied many high-performing developers who effectively leverage AI coding tools. What was discovered is that these individuals possess a distinct combination of abilities, traits, knowledge, and perspectives that make them dramatically more effective than their peers at using AI coding tools.</p><p>To present these findings and to provide clarity, this article is structured into these sections: key perspectives, key abilities, natural talents, required knowledge.</p><h3>Key Perspectives Required by AI-Native Software Developers</h3><h4><strong>1. Act as a Virtual Department Head</strong></h4><p>An AI-native developer should see themselves as the director of a virtual department of AI software agents. This role involves organizing these AI agents as effectively as possible to produce the highest quality end product. An AI-native software developer is no longer an individual contributor tasked with solving a small job or a couple of problems per day. Instead, their role requires guiding AI agents by setting processes, establishing best practices, creating plans, selecting tools, and creating rules to keep these agents in constant alignment and creating quality outputs.</p><h4><strong>2. Optimize the Code and Processes for AI, Not for Humans</strong></h4><p>AI-native developers need to tailor their codebase and coding process to suit AI systems. The codebase should be streamlined for AI to ingest and edit, with high-level READMEs built into all files. Long files should be split up for AI to ingest more easily. Database column titles should be self-explanatory with detailed comments, and there should be a comprehensive markdown file explaining the full code operations and goals which can be loaded into the start of every chat conversation. Given the complexity of optimizing code for AI, a more detailed explanation of this topic may be warranted if sufficient interest develops. Contact me to express your interest in such an article.</p><h4><strong>3. Use the Process to De-Risk the Results</strong></h4><p>AI is imperfect and will fail in many creative and unexpected ways. Developers need to keep a watchful eye on AI outputs and set up processes and safety catches for these failures. Assuming the resulting code is good simply because it works is a recipe for disaster. Creating tailored processes for the AI to follow, with multiple checks, best practices, and code examples, will help de-risk the resulting AI-generated code and enforce consistent methods of action. Having a standardized set of checks is key for each specific phase of software development.</p><h3>Key Abilities Needed by AI-Native Software Developers</h3><h4><strong>4. Multitasking at Scale</strong></h4><p>Multitasking while maintaining excellent quality of work is essential. Not everyone multitasks well, and some prefer not to multitask. The best AI-first developers will run multiple command-line interface chats controlling multiple AI teams concurrently via integrated development environment. They also will wear multiple hats, jumping between Product management for planning features, system architecture for tech selection, QA for code review, quality testing, and technical product management for documentation multiple times every half hour. Keeping track of everything will be challenging and stressful, yet at the same time feel hyper productive if done correctly.</p><h4><strong>5. Rapid Problem Solving and Quick Pivoting</strong></h4><p>AI will create problems, and things surely will come out wrong. Developers will need to quickly problem-solve and pivot by creating new solutions to address the same challenges. Avoiding the iterative bug-fix loop is key to preventing going down unproductive, token-burning rabbit holes. Great AI-native developers will possess the ability to recognize when a path is unproductive and quickly pivot by shifting AI to execute a different solution.</p><h4><strong>6. Think Like the AI Model Your Using</strong></h4><p>This is fundamentally about building a mental model around the AI’s specific limitations, common behaviors, known abilities, and unmentioned preferences. Using that deep understanding of how your specific AI model behaves, a developer can then write guiding prompts, coding plans, or rule documents and design processes to minimize the errors that can occur. Not everyone can develop an intuitive feeling for how other people think, let alone a synthetic AI. It is a unique ability, and frankly, some people simply do not care to do this.</p><h3>Natural Talents and Innate Traits of AI-Native Software Developers</h3><h4><strong>7. Have a Very High IQ</strong></h4><p>Being highly intelligent helps with most tasks, but when commanding groups of powerful AI agents, having the intellectual capacity to see where things are going while they are being executed and adjusting in real time is vitally important. From creating processes, rules, documents, and prompts to predicting AI agent outcomes, having a tremendous ability to quickly learn and identify hidden causes and effects is vitally important in AI-native software development. This will not be a slow, methodical game where the diligent worker wins. It will be an intense race to deploy and guide multiple AI agents and quickly course-correct while managing many sensitive tasks concurrently.</p><h4><strong>8. Have a Natural Tendency to be Highly Organized</strong></h4><p>Our team clearly found that enforcing clear organization on AI by documenting practices, processes, and rules for AI agents to follow is priority number one. Thus, individuals with an innate need to be organized are vital to being AI-Native Software Developers. Developers cannot delegate organization to their AI; they must do it themselves and actively maintain the structures. A developer without self-motivation to organize will find that their AI systems quickly accumulate bugs, errors, and unexpected behavior, and engineers have repeatedly proved that lacking organizational structure for AI results in low-quality results.</p><h4><strong>9. Be Naturally Curiosity and Willing to Experiment</strong></h4><p>AI best practices and processes for coding are nowhere near perfected and probably never will be. Being curious about an AI model’s behaviors and running tests to see if new methods can increase productivity is essential for the continuous improvement needed to work with AI. AI-native developers will constantly be thinking about how to refine their instruction documents to better guide their AI agents to produce the best outcomes for the least tokens.</p><h3>Required Knowledge for AI-Native Software Developers</h3><h4><strong>10. Know Solution Architectures and Trade-Offs Evaluation</strong></h4><p>AI-native developers need a broad understanding of systems, system architectures, and high-level platforms, along with their benefits and downsides. This understanding allows them to choose the best approach for the job instead of letting the AI decide. AI allows developers to move up a level of abstraction and focus on high-value architectural decisions. Knowing which AWS services exist or which open-source packages can be integrated to save time is extremely valuable.</p><h4><strong>11. Have Cross Area Expertise</strong></h4><p>AI-native developers should be comfortable with front-end, back-end, databases, DevOps, APIs, and more. An AI-native developer who is comfortable across all these areas is far more valuable than an engineer who focuses on a single area such as databases. The new software developer role demands constant context switching and multitasking, and cross-system knowledge allows for more work to be completed highly efficiently. In fact, lacking expertise across multiple areas could prove detrimental, as AI agents naturally traverse technology boundaries and require guidance across diverse technical domains.</p><h4><strong>12. Excel at Prompt Engineering</strong></h4><p>Prompt engineering is far more complex than it was even six months ago. Modern prompt work includes designing documentation, rules, processes, and checks that a network of AI agents will read, follow, and use. Conflicting instructions across any part of the documents can cause serious issues. Prompting knowledge must be holistic and cover the entire plan from start to finish: feature planning, architecture, documentation, quality assurance, testing, bug fixing, and upgrading. This level of prompt engineering is a skill unto itself.</p><h4><strong>13. Understand Token Usage Schemas</strong></h4><p>Yes, it says 12, but 13 is unlucky, so here is a little more to consider. This author performed a detailed analysis of the annual cost of AI coding tools and published the results in this <a href="https://medium.com/towards-artificial-intelligence/what-a-cto-must-budget-for-ai-coding-tools-a-detailed-forecast-of-ai-coding-costs-4f2cfd0f70bf">Medium Article</a>. The resulting costs ranged from $15,000 to $30,000 per engineer per year. AI-native engineers will have to work within a token budget. Burning millions of tokens to fix a simple bug is very costly. They will need to design their tasks and processes to minimize token usage while still producing high-quality code, bug fixes, and updates. In an AI-first environment, tokens are becoming the daily currency for programmers and a measure of their AI proficiency.</p><h3>Skills That Are No Longer Needed by AI-Native Software Developers</h3><h4><strong>14. Hand Coding Experts NOT Needed</strong></h4><p>This may be controversial, but expert coders are much less critical now that AI allows engineers to work at a higher level of abstraction. Being overly obsessive about hand-written code quality would be a major hindrance. If the AI-generated code meets the requirements for output, reliability, error handling, and security, the exact form the code takes is much less important compared to the productivity requirements for each engineer. As a result, traditional coding exams, on-the-spot coding tests, and timed whiteboard challenges are what the other considers unnecessary. Organizations should abandon these remnants from traditional hiring practices. Running 10 AI agents writing 40,000 lines of code, developers will have neither the time, nor bandwidth to run code-level reviews or fix code by hand. Again, this level of detail is best left to the AI.</p><h4><strong>15. Agile Practices Are No Longer Needed</strong></h4><p>With the introduction of AI-driven software development tools, this author argues in detail in this <a href="https://medium.com/towards-artificial-intelligence/why-the-agile-philosophy-fails-in-the-face-of-ai-powered-software-development-1e563fa1e835">Medium Article</a> that agile practices have become obsolete. Requiring an AI-native developer to follow Agile methodologies is asking them to ride a bicycle in order to drive a race car. The knowledge is totally irrelevant in the context of AI-first development setting running multi-AI Agent teams.</p><h3>Interview Process for Finding the Best AI-Native Software Developers</h3><p>The interview process for identifying and recruiting the best AI-native software developers is vastly different from most current interview practices. This deserves comprehensive exploration in a follow-up article, if I get enough interest. It is worth noting that stories in the news have documented developers using AI to assist (CHEAT) with coding interviews and still landing major roles later to be fired. Message me or comment if you think this is a worthy article.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cf4c3e4cdb56" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predict/top-12-abilities-for-ai-native-software-developers-what-to-look-for-when-hiring-cf4c3e4cdb56">Top 12 Abilities for AI-Native Software Developers: What to Look For When Hiring</a> was originally published in <a href="https://medium.com/predict">Predict</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[20 Tips for Implementing Effective AI Workflow Automations]]></title>
            <link>https://medium.com/predict/20-tips-for-implementing-effective-ai-workflow-automations-2fe365bcaa48?source=rss-4584d7533d83------2</link>
            <guid isPermaLink="false">https://medium.com/p/2fe365bcaa48</guid>
            <category><![CDATA[workflow-automation]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Dr. Brian Scott Glassman]]></dc:creator>
            <pubDate>Thu, 20 Nov 2025 03:05:01 GMT</pubDate>
            <atom:updated>2025-11-20T03:05:01.668Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N8tuzZGWdQJR_4gs9dky1g.jpeg" /></figure><p>Help Your Company Avoid Major Mistakes in Its AI Deployments</p><h3>Building Effective AI Workflow Automation</h3><h4>Introduction</h4><p>AI workflow automation has become one of the most transformative developments in business technology. McKinsey’s recent survey found that 22 percent of organizations have reached the expanding stage, where AI is deployed across departments and actively transforming workflows. This positions AI alongside historic innovations such as the steam engine and the internet. According to Gartner’s Fall 2024 report, by 2026, 30 percent of enterprises will automate more than half of their network activities, a sharp rise from fewer than 10 percent in mid 2023. Despite this rapid adoption, true success requires more than technical capability. It demands a strategic and systematic approach that balances innovation with risk management.</p><p>Through extensive testing with my team at AInspire.ai and through validation across multiple client implementations, we identified key considerations that must be addressed. Now that the system has been designed and validated, the next step is to move AI from controlled environments into real operational use. This phase focuses on deploying automations safely, observing their behavior closely, and ensuring that teams gain confidence as AI begins taking on production workload.</p><h3>Phase 1: Strategic Selection and Planning for AI Automations</h3><p>When planning for AI automation of workflows inside your company, consider the following recommendations. They will help direct efforts toward the most beneficial workflows and streamline the planning process.</p><h4>Evaluate and Prioritize Workflows by Risk Levels</h4><p>Begin by categorizing potential automation candidates into risk tiers. Level one workflows cause minimal harm if automated imperfectly. Level two affects efficiency or costs noticeably. Level three risks brand damage. Level four carries catastrophic consequences and should remain under human oversight until AI maturity is firmly established.</p><p>This evaluation should be a rapid collaborative process completed through focused meetings with department experts who understand both the workflows and their potential failure impacts. The goal is speed and practical judgment, not exhaustive analysis. The output of this assessment is a departmental inventory of workflows organized by business unit and rated from level one to four.</p><h4>Calculate Business Benefits and ROI</h4><p>For each workflow, estimate throughput improvements, reductions in human labor, and quality enhancements. Use conservative multipliers across the board for all workflows such as ten times throughput, three times human reduction, and five times quality improvement to calculate potential value. Then calculate the ratio of risk-to-return, rank order the workflows in their respective risk category. This will reveal a target list of high value lower risk AI automation that can be deployed.</p><h4>Bring Business Process Experts into the Redesign Cycle Early</h4><p>Technical expertise alone is not enough. The most effective implementations merge deep AI knowledge with a clear understanding of business processes and their underlying logic. When reforming a workflow with AI you must partner with a department expert to understand each departmental workflow processes and the underlying logic before even thinking about how to proceed with the redesign. Together the AI expert and the business process expert can architect a solution that handles the inputs and outputs in the most logical and low risk way.</p><h4>Select AI Champions in Each Business Process Group</h4><p>Additionally, early in the process one must identify a few key supporters within each department who will advocate for change, guide adoption, and work with teams to resolve problems. These AI implementation champions are vital because without a formal evangelist and an on the ground presence there are countless ways an AI implementation can go wrong, and having someone available to help solve issues is essential. Also, when successful the AI evangelist can spread knowledge to the rest of the organization on how to move AI forward.</p><h4>Remove Naysayers from the Process</h4><p>Because AI implementation is a hot topic there will inevitably be individuals who actively or covertly resist AI adoption, and they can seriously disrupt implementation efforts, especially if they feel their jobs are at risk. While concerns should be heard initially, those who continually refuse to engage constructively should be removed from planning and execution teams.</p><h4>Partner with an Experienced AI Integrator</h4><p>Organizations new to AI automation benefit significantly from working with integrators who have solved similar challenges. Their experience helps avoid mistakes and accelerates time to value. For example, top consulting firms understand the limitations of AI systems and where they typically break, saving time in AI integrations by reducing unnecessary experimentation.</p><h3>Phase 2: Design and Architecture of AI Systems</h3><p>Now that your company has identified some workflows to automate, it is time to shift toward designing AI automation for those workflows. In doing so, please consider the following items as they will reduce your risk, improve your ability to ship workflows, and help you track progress.</p><h4>Optimize for the Majority, Not Edge Cases</h4><p>Initially, design AI systems to handle fifty to eighty percent of the most common workflow cases. Edge cases outside that range should be intentionally routed to human teams, where they can be documented and used as future training data. This approach enables faster deployment and clearer success metrics.</p><h4>Build in AI Redundancy</h4><p>Design AI automations with careful attention to situations where the AI fails to produce the desired results or stops functioning entirely. For example, if the AI begins generating nonsensical output, a final validation step should check the results. Even more serious issues can occur when an AI inference provider goes offline. Consequently, implementing automatic switching between AI models so that another engine takes over when one fails is a smart and reliable design choice.</p><h4>Implement Comprehensive AI Monitoring and Alerts</h4><p>AI systems require real time monitoring that tracks performance, detects anomalies, and triggers alerts when thresholds are breached. Visibility into AI behavior is essential for safety and cost control, and without it you run a significant risk of AI runaways.</p><h4>Establish AI Auditing Capabilities</h4><p>Another valuable function is the ability to review AI performance through a robust audit and logging capability. Here the AI workflows should leave audit trails that log every decision, action, and outcome. These are not merely compliance tools. They will be extremely valuable when it comes time to upgrade the AI and improve the process, as that data will be used directly in pre upgrade analysis.</p><h4>Deploy AI Runaway Prevention Mechanisms</h4><p>Unlike software code, which executes reliably, AI, especially agentic AI systems, have the ability to enter loops where the cloud hosting bill will increase while no meaningful work is being produced. To avoid this, designers must implement automated cutoff and timeout mechanisms to prevent runaway processes. Even for normal tasks it is best to set clear execution timeouts to prevent unintended overuse.</p><h4>Create AI Workflow Dashboards</h4><p>It may seem unnecessary to create a dashboard to track automated tasks, but remember the system will be upgraded over time, and having the ability to see where the AI automations fall short will be extremely valuable for planning future upgrades. Teams should therefore plan from the beginning to have an AI dashboard that shows in near real time which jobs are queued, in process, completed, and which were routed to humans.</p><h4>Define Process-Specific AI Metrics</h4><p>It is fairly important when setting up an AI automation workflow to establish metrics around expected throughput, process times, error rates, cost per job, and similar factors and to track them over time. This will be a key mechanism for improving the processes as new generations of AI are released and implemented or as AI prompts are updated.</p><h3>Phase 3: Deployment and Transition of AI Workflows</h3><p>Now that the AI workflows are designed and ready to implement, the following considerations are vital to a successful rollout and the future upgrades of the processes.</p><h4>Start with Side-by-Side Replication</h4><p>To see how the workflows will perform without incurring risk, one option is to run AI workflows in parallel with existing human processes. By doing this your teams can compare results, identify discrepancies, and refine AI logic before transitioning entirely. This builds confidence and catches issues early.</p><h4>Conduct Simulated Stress Testing</h4><p>Before going live, test AI workflows under realistic load conditions. Simulate peak volumes, unusual edge cases, and system failures to identify weaknesses before they affect operations. It is often easiest to feed in old job backlogs to see how the system performs and where any hangups occur. Limit this to two iterations to avoid getting stuck in a loop of endless stress testing, as these do not fully reflect real world conditions.</p><h4>Choose Your Transition Method</h4><p>There are two primary approaches to integrating an AI workflow. One is a full quick transition, where the change occurs overnight. The other is a gradual transition, routing easy jobs or randomly assigning jobs to the AI workflow. The right choice depends on corporate risk tolerance and operational constraints.</p><h4>Implement Live Monitoring During Launch</h4><p>It may seem redundant, but having people on call for the first couple of days of an AI workflow deployment, especially to monitor the deployment’s activities, is important since each AI workflow will have its own unique tasks. This provides a monitoring and catch system for immediate issues.</p><h3>Phase 4: Evolution and Continuous Improvement of AI Capabilities</h3><p>After AI workflows are deployed, and as both the technology and organizational expertise mature, there is a clear path for enhancing them to cover a greater share of incoming work and improve performance metrics. Consider the following points.</p><h4>Plan for Incremental Coverage Expansion</h4><p>After the initial deployment covers fifty to eighty percent of the workload, work on expanding AI capability in a systematic and controlled manner to address additional scenarios or harder cases related to that workflow. Move up the complexity curve gradually, carefully recording what works well and where adjustments are needed and reviewing audit logs.</p><h4>Accept That the Last Five or Ten Percent May Always Need Human Intervention</h4><p>The final edge cases in your AI workflows, especially jobs that are unusual, high stakes, or ambiguous, may never be cost effective to fully automate with AI. Incorporate this reality into your architecture and cost models rather than pursuing unattainable perfection in your AI efforts. Remember that deploying AI at scale requires you to be selective about where you invest your time integrating AI.</p><h4>Build for Upgradeability and Long Term Evolution</h4><p>AI technology evolves quickly. Designing automation systems that can swap AI models, update logic, and incorporate new capabilities without requiring a rebuild is smart, though it is also more complex and costly. Accept that a degree of upgradeability must be built into the AI automations, especially around the AI models and safety and security systems. Also understand that unlike traditional software, AI upgrades may occasionally break processes that previously worked.</p><h3>Conclusion</h3><p>Effective AI workflow automation is not about technology alone. It requires integrating technical excellence with a deep understanding of business processes. Organizations that succeed are those that balance technology with business reality and build for evolution rather than perfection.</p><p>By following these points from strategic selection through continuous improvement, organizations can avoid many of the risks and errors associated with AI and enjoy the immense benefits of automation while managing the challenges that have stalled less thoughtful implementations.</p><p>Dr. Brian Scott Glassman is VP of Product Management at <a href="https://ainspire.ai/">AInspire.ai</a> and Head of AI Strategy.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2fe365bcaa48" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predict/20-tips-for-implementing-effective-ai-workflow-automations-2fe365bcaa48">20 Tips for Implementing Effective AI Workflow Automations</a> was originally published in <a href="https://medium.com/predict">Predict</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What a CTO Must Budget for AI Coding Tools: A Detailed Forecast of AI Coding Costs]]></title>
            <link>https://pub.towardsai.net/what-a-cto-must-budget-for-ai-coding-tools-a-detailed-forecast-of-ai-coding-costs-4f2cfd0f70bf?source=rss-4584d7533d83------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f2cfd0f70bf</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[budget]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[management]]></category>
            <dc:creator><![CDATA[Dr. Brian Scott Glassman]]></dc:creator>
            <pubDate>Thu, 21 Aug 2025 18:01:31 GMT</pubDate>
            <atom:updated>2025-08-21T18:01:31.080Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xGCEV28SGf1uaHVD2VQwqg.jpeg" /><figcaption>Image By Author Dr. Brian Scott Glassman</figcaption></figure><p>A CTO’s guide to AI coding tool costs, including per-developer projections, model pricing trends, and sustainable budgeting strategies.</p><h4>Executive Summary</h4><p>CTOs, CEOs, and software development leaders need a clear view of the rising costs of AI coding tools such as Claude Code and OpenAI’s Codex, and what to budget for the next quarter and next year. This analysis reviews costs for Claude and ChatGPT-5 and highlights a methodology that can be applied to your organization to project future AI coding tool expenses, because not everyone has a budget of a Fortune 500 company.</p><p>The analysis arrived at the $1,500 or $3,000 monthly per-developer budget range by calculating the hourly costs of running AI coding tools (GPT-5 vs Claude for single agents), then modeling realistic usage scenarios where developers would run AI coding assistants for total time of 3–5 hours daily with 1, 3, or 5 concurrent AI agents. The $1,500 budget tier assumes developers use AI coding tools with up to 3 concurrent agents for approximately 3 hours per day (reflecting current usage patterns), while the $3,000 tier accounts for more aggressive usages with up to 5 concurrent agents running for up to 5 hours daily. These projections factor in expected price decreases of 40% and 70% respectively over the next six months due to model optimization and competition, while also accounting for increased usage and productivity gains, ultimately predicting a stable cost equilibrium around $70 per hour of AI used.</p><p>The analysis further concludes that establishing a structured monthly AI coding budget for developers can promote more strategic and efficient use of compute resources before executing AI-driven coding prompts. The following table presents my recommended monthly AI coding budgets for teams of various sizes, based on two tiers: $1,500 per developer per month with access to Claude AI (Opus (5% usages) with Claude Sonnet 4.1 (95%), once released, utilizing up to three concurrent agents for a total coding time of three hours per day) and premium tier of $3,000 per developer per month with premium Claude AI (Opus with Claude Sonnet 4.1 utilizing up to five concurrent agents for a total of five hours per day coding). These are significant additional costs, but are to a large degree a must due to the massive efficiency gains from using a frontier AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UHKU-DbVvfK5Uu_1IMny_Q.png" /><figcaption><strong>Table: Monthly Budgets for AI powered Coding Tools</strong></figcaption></figure><h4>The key takeaway from this analysis is as follows:</h4><ol><li>Frontier LLMs are expected to continue producing more tokens per minute while requiring fewer computational resources. As a result, the cost per input and output token will decline until it reaches what I see as a baseline (approximately $1 per million input tokens and around $10 per million output tokens), after which pricing will primarily reflect perceived value and reputation, and base on some level of competition, rather than be based on operational expenses or R&amp;D recovery.</li><li>Frontier models will likely be offered in two tiers: one optimized for maximum accuracy and quality, and another designed for bulk coding tasks with solid but highly cost-efficient performance. The premium tier will be priced higher 2X the lower model, while the bulk coding model will be more affordable. This approach balances computational demands and cost, allowing the two models to complement each other effectively, expected to average the costs with 5% being on the premium model.</li><li>The quality of coding outputs from these models is projected to improve steadily, with anticipated gains of 20% to 40% per quarter in improvements in first time accurate completions. Over the next several years, it will be challenging for open-source alternatives to replace leading frontier models such as those from OpenAI and Claude. The mentioned closed source frontier models are expected to generate significant revenues from customers, which will further fund accelerated training, enhancements, and release cycles. The business model of open-source alternatives may struggle to remain competitive in this environment (it was nice but I don’t see it lasting).</li><li>Token generation rates (tokens per minute) are expected to increase gradually (20 to 40%), unless advancements in hardware, such as LLM-optimized chips from companies like Groq (not xAI’s Grok), or Inference only Cards from NVIDIA are applied to OpenAi and Anthropic then we can see 100% to 200% improvements in token rate generation.</li><li>Using GPT-5 for five consecutive hours per day costs about $14, compared to $75 for Claude. With three agents over the same period, costs rise to roughly $42 for GPT-5 and $225 for Claude. Claude’s pricing, however, is expected to decrease significantly with the release of Sonnet 4.1.</li><li>Both conservative and aggressive adoption of Claude result in daily AI coding costs stabilizing around $70 per hour, with efficiency gains balanced by price deflation, ultimately delivering far greater productivity without significantly increasing expenses (see that section for clarification).</li><li>Projected workforce costs scale rapidly: at $1,500 per developer per month, teams of 100 can reach $1.8M annually, while at $3,000 per developer per month, costs double to $3.6M per year, underscoring the need for strict budget controls and strategic deployment of AI coding tools so budgets are not blown out. Last thing we want is coders funding their own unauthorized side projects with AI coding hours stolen from their parent company.</li><li>Note: Use this full analysis as a template within your own AI system to calculate your own detailed corporate AI coding cost projections. If you are dissatisfied with any of the assumptions, you can adjust them within your own model. Ask the AI to recreate the model by stating the core assumptions, then generate three scenarios and make the necessary adjustments accordingly specific to your corporation.</li><li>Note: This analysis does not include a detailed review of productivity gains. While it may seem logical to perform a cost‑benefit analysis, in practice the productivity impact varies greatly from task to task. Establishing a reliable baseline is nearly impossible without incorporating an organization’s specific AI coding use cases. Therefore, each company must conduct its own testing to evaluate productivity improvements accurately.</li></ol><h4>Introduction</h4><p>CTOs and Engineering VPs must begin factoring the substantial cost of AI coding programs into their departmental planning today. While adoption may seem early, it is critical to allocate budgets now for upcoming quarters and the next year to sustain momentum in AI-based software development and avoid falling behind.</p><p>AI coding providers, along with their cloud inference partners, are set to capture significant revenues as software developer productivity and effectiveness increase dramatically through the use of AI (tens of billions in compute bills). However, this leap in performance will come at a steep price. Technology leadership teams should prepare their CEOs in advance for these additional costs, as the bill for AI-driven productivity will be very high and likely remain that way.</p><p>Consequently, the central question for every CTO is: What will the cost per developer be? This article aims to provide that answer.</p><p>Let’s jump into it.</p><h4>A Quick Look Back at the Cost of OpenAI and Anthropic Coding Models</h4><p>When ChatGPT and Claude first became effective at coding at scale, their costs were significantly higher per million tokens. This was because their large foundational AI models contained vast amounts of information, lacked optimization for output speed, and were missing many of the structural improvements that make today’s LLMs faster. As a result, they were expensive to run on inference compute, leading to higher prices. The table below shows that pricing with ChatGPT-5 dropped dramatically, primarily because the model is less CPU intensive while still delivering high-quality results, a clear example of innovation driving efficiency.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qirwIYR-BtSG7IbJf3cOlw.png" /><figcaption><strong>Table 1: Cost of OpenAI model use for coding as they were released</strong></figcaption></figure><p>Next, we examine the historic pricing of Anthropic’s Claude models, which has generally trended upward. However, the upcoming September release of Claude 4.1 Sonnet is expected to include a price reduction to better align with OpenAI’s pricing. The most capable model for coding, Claude Opus 4.1, remains quite expensive. Consequently, Anthropic recommends using Opus for planning and complex coding tasks while leveraging Sonnet for large-scale coding due to its higher token output rates. <strong>I also expect to see a multi-model approach, where the more tedious, complex, or important the task, the more likely organizations will rely on the most expensive AI models, because in practice switching models can be done seamlessly.</strong></p><h4>Anthropic Claude Models</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q4CnQ4OeA7pc8DuB32Yg9g.png" /><figcaption><strong>Table 2: Cost of Anthropic’s model use for coding as they were released</strong></figcaption></figure><h4>Pricing Trends of AI Coding Models</h4><p>Below is the pricing for each frontier model at the time of its release.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8TYgTU1H__avne1lhtyahQ.png" /><figcaption><strong>Table 3: Pricing trends of major AI coding models at release, showing shifts across OpenAI and Anthropic</strong></figcaption></figure><p>We are beginning to see competition between OpenAI and Anthropic drive down per-million-token pricing, alongside the evolution of smaller, more efficient models. OpenAI took a different approach with GPT-5 by introducing reasoning effort = {low, medium, high}, which directs the model to spend more compute cycles before producing an answer, generally improving quality. Nevertheless, this is a less optimal solution compared to deploying two distinct models, one of which is smaller and requires a lighter compute footprint.</p><p><strong>Opinion:</strong> I expect the input token cost to stabilize at around $1, while the output cost will likely remain in the $10 to $20 range for frontier coding models over the next year. The detailed reasoning behind this expectation is beyond the scope of this article, but I would be glad to discuss it further.</p><h4>Coding Performance and Benchmark Comparisons</h4><p>Next, we need to examine coding performance. From the table, it is clear that each model shows excellent improvement on its SWE-bench coding tasks, advancing at a steady pace. Soon SWE-Benchmark test will need a new version. Improved output quality naturally supports higher pricing, particularly when sensitive tasks demand high-quality coding and DevOps AI agents, plus human developers get used to a certain quality of code and do not like to go backwards, as it feels highly counterproductive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*61VsUNrb1RPV7poBLi8DNw.png" /><figcaption><strong>Table 4: Coding performance benchmark results across models</strong></figcaption></figure><p>However, output quality is only part of the equation. We must also consider the speed at which these outputs are generated.</p><h4>Coding Speed / Token Generation Rate</h4><p>As developers, we value rapid code generation, but more importantly, we expect the process to continue reliably until all planned tasks are completed. The ability to step away briefly, such as taking a break while tasks are executed seamlessly, fosters a sense of productivity and efficiency. Smaller models like Claude Sonnet often achieve faster token generation, while GPT-5 also delivers impressive speeds through its architecture. However, at higher reasoning levels, GPT-5 performance slows noticeably. In my assessment, attempting to make a large language model function as a universal multi-tool, as seen with GPT-5, is not ideal because the components inevitably influence one another, creating complex internal dependencies.</p><p>The key lesson is that AI agent models (like Claude Sonnet) should be optimized for speed since they perform the majority of the work, while the orchestrator LLM (Opus so on) can afford to operate more slowly, as its accuracy is crucial to guiding the overall process. Expect the majority of the token generation speed to be around 30 to around 50 token/sec in practice for the next 6 months. Unless a significant breakthrough emerges to challenge the frontier models, such as Mamba or Diffusion coding models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nYS7oZlbfggxZS2B54KRJA.png" /><figcaption><strong>Table 5: Token Generation Speed Comparison</strong></figcaption></figure><p>The table below provides a more accurate representation of coding speed, as token outputs per minute of coding tend to be more stable and are relatively close to one another. These figures are based on user-reported data rather than questionable company metrics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*STktOmrAzq_i4FAdsuPmzg.png" /><figcaption><strong>Table 6: User Reported Token Generation Speed Comparison (Per Minute and Per Hour for GPT-5, Claude Opus, and Sonnet)</strong></figcaption></figure><h4>Cost Per Hour of Coding at Full Speed</h4><p>Currently, no AI system codes continuously for a full hour, though this may change in the near future (give it 4 months). Today, software developers typically run a coding plan against a large product specification, then evaluate the results, suggest changes and updates, or execute the code to analyze outcomes. In practice, my team operates at full coding speed for roughly 30 to 40 percent of each hour. Over an eight-hour workday, this equates to approximately three hours of sustained AI-driven coding. For a baseline calculation, however, we need to determine the per-hour costs. A referenced source cited an 8-to-1 input-to-output ratio based on context window size and related factors. Using this, I multiplied the output tokens by eight to estimate the <strong>per-hour cost</strong> at full read-write capacity for the three models. <br>Although GPT-5 offers the lowest cost per hour, it is relatively new, and developers still require time to evaluate its effectiveness before committing to such a transition. While a CTO may be tempted to mandate adoption purely for cost savings, the strategic wisdom of such a decision warrants careful consideration like a unbiased blind code analysis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dEH7BM8noHsA5SFDGATepQ.png" /><figcaption><strong>Table 6: Cost Per Hour of Coding at Full Speed</strong></figcaption></figure><h4>Agent Cost Comparison</h4><p>Claude Code with agents demonstrated that the future of software development <strong>WILL</strong> rely heavily on multi-agent coding. In this model, agents work independently on their respective pieces of code and then close once their portion of the plan is complete. For the purposes of this analysis, I have assumed scenarios with 1, 3, and 5 agents operating concurrently. These are the respective costs. While GPT-5 is generally more cost-efficient, the trade-off between cost and quality must always be considered. It is also important to note that operating with 5 agents drives costs into a range comparable to the hourly wage of many software professionals.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8g_tSlhzLc7ykXh0My186w.png" /><figcaption><strong>Table 7: Cost Per Hour of Programming Assuming You Add More Agents</strong></figcaption></figure><h4>Senior Software Developer Daily AI Costs</h4><p>Again you can not expect and coder to run the AI all day, so I divided it up into running for a total of 3 hours, 5 hours, and 8 hours. to show expenses for both GPT-5 and Claude Sonnet. I expect the Claude Sonnet 4.1 to be release to drop the pricing by 40% so the 5 and 8 hour pricing for 5 agent will be around ~$150 to ~$350 per day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QMRBX567F8GAzpWeEoGYbQ.png" /><figcaption><strong>Table 8a: Daily Cost of GPT-5 for 3, 5, and 8 Hours Across 1, 3, and 5 Agents</strong></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YdQ7vWnmD1mhXXiKqXawcQ.png" /><figcaption><strong>Table 8b: Daily Cost of Claude Sonnet 4 for 3, 5, and 8 Hours Across 1, 3, and 5 Agents</strong></figcaption></figure><h4>Scenario Analysis: Conservative vs Aggressive AI Adoption</h4><p>I have considered several scenarios that may unfold over the next six months. As a baseline, we assume AI is used for 3 hours per day with 1 agent, which is close to what many programs are experiencing today in their token spend.<br>In a conservative six‑month projection, AI usage remains at 3 hours per day but expands to 3 agents working concurrently. In this case, token speed is expected to increase by 25%, while prices decrease by 50%, resulting in daily costs of around $70 per hour, but with significantly greater output (over 300% increase in code writing compared to 1 agent for 3 hours).</p><p>For the aggressive forecast scenario, AI coding is projected to run for 5 hours per day with 5 agents (~730% increase compared to 3 agents for 3 hours).</p><p>Although token generation speeds would also increase, I anticipate costs to decline more rapidly, reducing to about 70% of their current levels. Despite higher performance and productivity, the net result is that the cost per hour of AI coding remains roughly $70, but with significantly greater output (over 730% increase in code writing over the conservative case).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OUN72lA2kqe8fN6XOJZK0A.png" /><figcaption><strong>Table 9: Scenario Analysis of Daily AI Usage Costs</strong></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ADcZIlbx5xVEherTBISj0Q.png" /><figcaption><strong>Table 10: Formula Impact Analysis of Usage and Pricing</strong></figcaption></figure><p><strong>Key Insight:</strong> Massive usage increases (+300% to +733%) are almost entirely offset by projected AI price deflation (-50% to -70%), resulting in similar cost outcomes.</p><h4>Monthly Costs Analysis (20 Working Days)</h4><p>The most important part of this analysis is expected cost per month of deploying these agent will be high, look at current sonnet 4 vs. expected 4.1 (I took half the cost) still at $1,400 per month per programmer only running 3 agents is significant.</p><p>Across a development team of 50 programmers, the monthly bill could reach approximately $70,000 (or $840,000 annually) when using Claude Sonnet, while GPT-5 would still amount to about $25,000 per month (or $300,000 annually). Although these costs are substantial, the corresponding productivity gains are expected to be significant. Organizations should prepare their budgets accordingly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GnwxkH6KyAN2GvbKsJPJEw.png" /><figcaption><strong>Table 11: Monthly cost for AI running 3 Hours/Day with 1 to 3 agents</strong></figcaption></figure><p>Now say we bump it up to 5 hours per day with 1 to 5 agents, Across a development team of 50 programmers, the monthly bill could reach approximately $190k (or $2.34M annually) when using Claude Sonnet 4.1, while GPT-5 would still amount to about $70k per month (or $840k annually). Yep this is still very expensive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rudLCR3SioCSaSOw-jjCEg.png" /><figcaption><strong>Table 12: Monthly cost for AI running 5 Hours/Day with 1 to 5 agents</strong></figcaption></figure><h4>Projected Cost for the Workforce</h4><p>You can manage AI deployment by limiting the number of AI coding hours allocated per developer. While this approach may reduce maximum productivity, it encourages more strategic use of compute resources and better task planning per developer. A budget can be set on a per-developer basis, with the following examples:<br>Running AI coding for 3 hours a day with 3 agents would likely result in costs ranging between $1,000 and $1,700 per employee per month.<br>For highly productive developers running AI for 5 hours a day with 3 agents, the monthly budget would be approximately $3,000.</p><p>What organizations should avoid, among other issues, is allowing developers to press ‘go’ on a poorly conceived AI coding plan that consumes 20 to 30 minutes of expensive compute only to generate code that must be rolled back. Basically call it AI coding waste, and doing this repeatedly over teams of developers is a great way to waste money and time. One simple way I can see to stop this is to give them a token budget and let them get smarter about using them to maximize their work.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j9892X9pKWmfp7pN-NEH1A.png" /><figcaption><strong>Table 13a: Projected Team AI Budget ($1500 Monthly Budget)</strong></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iLdWQODYvUq27zPdJvVcpA.png" /><figcaption><strong>Table 13b: Projected Team AI Budget ($3000 Monthly Budget)</strong></figcaption></figure><h4>Conclusion</h4><p>The financial impact of deploying AI coding tools is substantial and will require careful planning by CTOs and technology leaders. While prices may gradually stabilize, they are still likely to remain expensive on a per-user basis. A smart strategy is to enforce budget caps per developer, ensuring that organizations extend their AI budgets effectively while still capturing the productivity benefits of AI-driven software development.</p><p>Note: I have deliberately chosen not to include Google Gemini 2.5 or xAI Grok in this analysis. Google’s track record as a steward of sensitive data raises concerns about entrusting them with proprietary corporate code, and for that reason I do not consider them a viable option. In addition, Grok remains a very new entrant to the market, and its relative immaturity brings the risk of significant security vulnerabilities. As a result, both platforms are excluded from the projections in this report.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f2cfd0f70bf" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/what-a-cto-must-budget-for-ai-coding-tools-a-detailed-forecast-of-ai-coding-costs-4f2cfd0f70bf">What a CTO Must Budget for AI Coding Tools: A Detailed Forecast of AI Coding Costs</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why the Agile Philosophy Fails in the Face of AI-Powered Software Development:]]></title>
            <link>https://pub.towardsai.net/why-the-agile-philosophy-fails-in-the-face-of-ai-powered-software-development-1e563fa1e835?source=rss-4584d7533d83------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e563fa1e835</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[agile]]></category>
            <dc:creator><![CDATA[Dr. Brian Scott Glassman]]></dc:creator>
            <pubDate>Tue, 12 Aug 2025 12:02:00 GMT</pubDate>
            <atom:updated>2025-08-12T12:02:00.230Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZSYPbTmhUhGXEzPW1rF5ww.png" /><figcaption>Image by Dr. Brian Scott Glassman</figcaption></figure><h3>Why the Agile Philosophy Fails in the Face of AI Development</h3><h4>Agile Is Dead, Long Live AI (Powered Software Development)</h4><h4>It is time to confront the World’s Predilection to Agile. It had an impressive run, but with the arrival of AI Powered Software Development, the rules are about to change whether anyone likes it or not.</h4><p>Sorry, but Agile Software Development had a good twenty five year run as the premier framework for organizing and managing software development. However, the new AI powered software coding models, with their ability to generate hundreds of lines of code per minute, are breaking Agile’s core principles and rendering its foundations irrelevant (Agile is soon to be a “floppy disk analogy” of the way we did things back in the day ).</p><p>When I was earning my Ph.D. and studying frameworks and philosophies, Agile stood out as one of the most transformative. However, over the past year, through hands-on work with my own software development teams and ongoing conversations with CTOs at large enterprises, I am seeing the very foundations of Agile being undermined by AI powered software development practices.</p><p>That shift, in my view, calls for a clear and timely (and hard conversation) about what should supersede Agile. But first, let us touch briefly on its history.</p><h4>Short History of Agile Software Development</h4><p>Agile was introduced in 2001 by a group of software developers (clearly disgruntled developers) who wrote the Agile Manifesto to address the limitations of rigid, plan-heavy development methods like Waterfall. It was designed to promote flexibility, faster delivery, and better alignment with customer needs in rapidly changing environments, with an emphasis on human collaboration.</p><h4>The Structure of My Argument Against Agile</h4><p>To clearly make the argument for why Agile will be obsolete in the face of AI-powered software development, I want to dissect the core principles of Agile in light of the new capabilities of AI, because the processes and methods of Agile are built entirely around those principles.</p><p>So, to keep it brief, let’s jump into the first principle.</p><h4>1. AI’s Impact on Agile’s First Principle</h4><blockquote>Agile’s 1st Principle: “Our highest priority is to satisfy the customer through early and continuous delivery of valuable software”</blockquote><p>Agile first emerged as a reaction to the slow, opaque, and rigid model of Waterfall-based software development. By making customer feedback and quick delivery cycles central to its philosophy, Agile enabled software to be developed and refined in full view of the customer. In doing so, the customer could help guide the development with a clear understanding and confidence in the resulting product (hopefully not the kind of customer who takes a photo of the computer screen with their phone instead of using a screenshot to dictate what they want changed to the software).</p><p>However, due to AI, development now moves at breakneck speed, and entire projects can be completed before the next scheduled customer feedback session, making the wait for live customer feedback the new bottleneck (particularly when the customer is busy attending a week-long “urgent” leadership retreat in Maui).</p><p>Fortunately, there is a powerful alternative. An AI, equipped with a detailed understanding of your customers built from comprehensive documentation such as multiple customer personas, that can simulate hundreds of rich customer interactions and generate deep, actionable insights. This accelerates feedback cycles to unprecedented speeds. In this hyper-rapid iteration environment, the original Agile principle feels less like an indispensable guide and more like a relic from a slower era of development. I therefore propose replacing this principle with the following AI-driven alternative:</p><blockquote>The First Reformed AI Agile Principle: “”Continuously refine the solutions using AI-created customer feedback, derived from high-quality models of the customers.”</blockquote><h4>2. AI’s Impact on Agile’s Second Principle</h4><blockquote><strong>Agile Principle 2: “</strong>Welcome changing requirements, even late in development. Agile processes harness change for<br>the customer’s competitive advantage.<strong>”</strong></blockquote><p>Agile’s second principle was a remedy for the rigid, change-resistant Waterfall processes that made changing requirements difficult. Agile empowered teams to adapt to the latest customer needs and, in doing so, course correct the software solution to make it more valuable. However, in the new age of AI, adapting the software solution is orders of magnitude easier than if it were done entirely by human developers alone, making the pain of change essentially a moot point.</p><p>The deeper issue, as anyone who interacts frequently with customers in an Agile environment knows, is that customers often provide limited or convoluted feedback and usually need guidance to identify the core intent underlying their request. Consequently, I trained my product manager never to react to the first request but, instead, like Sherlock Holmes, to dig deeper and uncover the real problem motivating the requested change (and occasionally the customer’s “feedback” is just a way to remind everyone they are still in charge, but that is for another conversation).</p><p>Here comes the kicker. Our AI-powered change request process can collaborate with customers, product managers, and stakeholders to uncover the true issues driving a requested change, validate whether those changes are necessary, and then brainstorm holistic change requests that deliver significantly greater value. At the same time, the AI can calculate and balance the broader impact of those changes across the entire solution and code base. I therefore propose replacing this principle with the following :</p><blockquote><strong>The Second Reformed AI Agile Principle:</strong><br>“Use customer change requests as an opportunity to leverage AI to uncover deeper insights and ideate, refine, and implement changes that deliver greater value.”</blockquote><h4>3. The Impact of AI on the Third Agile Principle</h4><blockquote>Agile Principle 3: “Deliver working software frequently, from a couple of weeks to a couple of months, with a preference to the shorter timescale.”</blockquote><p>The third Agile principle was created to push teams to deliver value faster than the long development cycles of the past. However, AI has now shattered those timelines, compressing development from months to weeks and from weeks to days. From this point forward, AI-driven development speeds will only continue to accelerate. For example, teams of AI mesh coding agents will soon be able to update a single codebase simultaneously, coordinating, checking, and validating multiple sections of code at once. The speed of software development will no longer be the bottleneck in bringing products to market.</p><p>Consequently, I feel the third principle of Agile is no longer important enough to be mentioned in the age of AI.</p><h4>4. The Impact of AI on the Fourth Agile Principle</h4><blockquote><strong>Agile Principle 4: “Business people and developers must work together daily throughout the project.”</strong></blockquote><p>At the founding of Agile, daily collaboration and meetings between business people and delivery partners were designed to solve the major handoff problem in Waterfall, provide tactical-level guidance on elements such as features and technical performance requirements, and make decisions on how to prioritize work, since not everything can be developed and trade-offs must be made to optimize the use of developers’ time. Daily meetings with business, or stand-ups as we know them (and many of us gravely dread), represented a titanic shift in the way of working back in 2001. However, this is changing with AI.</p><p>If you provide a highly intelligent AI with comprehensive project documentation along with the software’s strategy, vision, product requirements, customer profiles, and business operation preferences, it can make informed feature decisions, guide work tasks, prioritize delivery, and even fill in gaps. This is what I call AI’s tactical-level value-added activities.</p><p>It is important to remember that every hour a business person spends on tactical activities is an hour not spent on high-value strategic work, such as researching and planning new product line. To maximize impact, daily developer and business meetings should shift away from tactical discussions and focus instead on high-level strategy and vision-setting for product updates or new proudcts, leaving execution details to AI and its human developer overlords (HDOs).</p><blockquote><strong>The Fourth Reformed AI Agile Principle:</strong> “<em>Let AI manage the continuous alignment between business goals and tactical level execution, freeing business people and developers to focus on high-value strategic activities.”</em></blockquote><h4>5. The Impact of AI on the Fifth Agile Principle</h4><blockquote><strong>Agile Principle 5: “Build projects around motivated individuals; give them support and trust.”</strong></blockquote><p>Agile’s fifth principle was born in an era when building a solution around motivated individuals was not just important, it was essential. Success depended on a team’s sustained energy, drive, and discipline to push projects forward despite complexity, ambiguity, and multi-month or even multi-year development timelines. It was a grind. In that context, progress often relied on the sheer willpower of human developers to carry the load from concept to completion without perpetually creating new issues along the way.</p><p>However with the collapsed times of AI development, that dynamic has fundamentally changed. Software development no longer depends on heroic team efforts. Instead, developers need to manage the orchestration of AI to do heavy lifting, writing code, scaffolding systems, generating tests, and maintaining documentation.</p><p>What we now have, with AI, is the capability to transform software development into a highly repeatable efficient process. The nature of the work is shifting from labor-intensive craftsmanship to an intelligent AI production line. Human contributions are becoming less about relentless motivation to execute and more about guiding, curating, and refining outcomes created by AI. <strong>Motivation still matters to some degree</strong>, but the role of the modivated developer team has evolved into that of an orchestrator of AI, which is why I emphasize the “human developer overlords (HDO).” For these reasons, I see the fifth principle being killed off by the new AI development systems.</p><h4>6. The Impact of AI on the Sixth Agile Principle</h4><blockquote><strong>Agile Principle 6: “Face to face conversation is the most efficient and effective way to share information.”</strong></blockquote><p>In traditional Agile, face-to-face conversation solved many problems far more quickly than the low-context back and forth of emails, chats, or comments on a document. Instant feedback, combined with the emotional weight that comes from live conversation, was a powerful problem-solving method and a way of focusing on what truly mattered (unless the meeting was derailed by the occasional mental breakdown episode from the junior developer).</p><p>However, meetings during the workday come with costs. They take time away from other tasks, they often drift into topics relevant to only a few people in the virtual meeting room, and most importantly they are a slow and often imprecise way to communicate information. This is because conversations are fleeting and not easily recalled, even if recorded, which is why clear written documentation has always been critical.</p><p>Now AI changes the equation. A highly capable AI can ingest thousands of pages of documentation in seconds. Product strategy decks, customer personas, architecture diagrams, workflow charts, annotated codebase summaries from GitHub or local repositories, feed it all to the AI and it can follow the instructions within to develop a solution efficiently with some guidance from its HDO. It is hard to ignore that face to face meetings are now among the least effective ways to share information in the new world of AI-driven development process.</p><p>Every month, AI systems are improving their ability to navigate high-context documents. From product strategies and customer personas to architecture diagrams, workflow charts, and annotated codebase summaries, leading AI tools are now able to easily ingest this information, form coding plans, and generate development outlines for new software projects. Personally, I have taught my teams to create high-context documentation that enables AI development agents to fully understand our products and how we want things done. I therefore propose replacing this principle with the following AI-driven alternative:</p><blockquote><strong>The Sixth Reformed AI Agile Principle:</strong> <em>“High context document sets are the most efficient way to enable AI powered development.”</em></blockquote><h4>7. The Impact of AI on the Seventh Agile Principle</h4><blockquote><strong>Agile Principle 7: “Working software is the primary measure of progress”</strong></blockquote><p>At the founding of Agile, working software was considered the ultimate measure of progress because if the customer could click through it, test it, and ship it, then the project was visibly on track (unless, of course, you had one of those annoying customers or bosses who never attended the milestone demos). Of course, this made sense in a time when building software was the longest lead item on the product development Gantt chart. However, this is changing with AI.</p><p>If you provide a highly capable AI with a set of well-structured contextual documentation that includes the business’s needs, customer personas, technical constraints, architecture preferences, technical features, and user flows, it can generate working software. Amazingly, even amateurs (including annoying TikTok influencers) are using vibe coding tools to produce functional software programs. Whether the code meets best practices is another matter entirely, but it works, and by Agile’s seventh principle, that is all that counts. The problem is that working software will no longer be the true measure of progress. In the era of AI-driven software development, we will have to dig deeper to define what progress actually means, and it will not be as simple as “the software runs” because that is a very low bar to set. Bye bye principle 7.</p><h4>8. The Impact of AI on the Eighth Agile Principle</h4><blockquote>Agile Principle 8: “<strong>Agile processes promote sustainable development. The sponsors, developers, and users should be able to maintain a constant pace indefinitely.”</strong></blockquote><p>In traditional Agile, sustainability was about maintaining the workforce (because that last two week coding blitz probably took years off your programmers’ lives). Teams were expected to move at a steady, manageable pace so they could deliver working software consistently over the long haul. Agile borrowed from the assembly lines of the past and measured coding efficiency in story points and burn down charts (because everyone loves their boss standing behind them with a stopwatch).</p><p>But with AI, velocity becomes uncoupled from human effort. You can now build more software in two weeks with AI than the same team of humans could in six months. The constraint is no longer pace. It may instead be something else such as quality, customer fit or satisfaction, application speed, AWS app operational costs, or even the AI inference bill from OpenAI or Anthropic. In other words, the limiting factor will be something other than pace. Once again, we see another Agile principle losing its relevance in the age of AI.</p><h4>9. The Impact of AI on the Ninth Agile Principle</h4><blockquote>Agile Principle 9: “<strong>Continuous attention to technical excellence and good design enhances agility.”</strong></blockquote><p>In traditional Agile, technical excellence was about craftsmanship. Clean code, smart architecture, and thoughtful designs gave teams a logical and adaptable code base, one that could be modified more easily in response to customer requests. Then came AI.</p><p>We are now seeing many coding best practices applied by AI, from the low level design of call functions to the high level organization of entire code bases. However, AI does not truly understand which practices need to be enforced within the broader context of the code base and the company’s needs (at least till the new CTO is hired and changes the best practices). This is where the human developer overlords (HDOs) come in. They can review the code generated by AI identify which best practices should be enforced, and which can reformed to better suit AI. (Wait, what, suit AI?)</p><p>Yes, every LLM coding model has its quirks, and they all have limitations. A new and growing set of best practices for using AI code development models is emerging and discussed on forms like Medium. For example, my development teams tend to keep files smaller and provide a high level written description of the file at the top of each file. These practices are just as important, if not more so, for maintaining code with AI . We will continue to see these AI coding best practices grow and solidify into a new framework, supported by training programs, college courses, (and some Wealthy YouTube Influencers) over the next few years. Consequently, I would argue that the ninth principle of Agile is still vitally important, but it needs a vital tweak.</p><blockquote>The Ninth Reformed AI Agile Principle: “Continuous attention to good design and AI driven coding best practices enhances agility.”</blockquote><h4>10. The Impact of AI on the Tenth Agile Principle</h4><blockquote>Agile Principle 10: “<strong>Simplicity — maximizing the amount of work not done — is essential.”</strong></blockquote><p>In traditional Agile, simplicity was a safeguard. Writing less code meant fewer bugs, easier maintenance, and faster delivery. Teams were trained to minimize effort and complexity because every extra line of code meant more time, more risk of bugs, and more human labor. There was also a certain pride in the elegance of it all (think of the scruffy programmer who looked like he had not bathed bragging about replacing 2000 lines of code with 10). But with AI writing the code, the work shifts.</p><p>With AI, writing extra functions, building additional modules, or creating features only to disable them later is easy. I have my dev teams mock up multiple user interfaces and cross compare them with insights generated from AI simulated customer reviews, simply to gather more perspective before committing to the final UI design. This approach openly violates the tenth Agile principle (no court summons from the Agile Police yet), but it is valuable because it refines and improves the end user experience.</p><p>But I do apply this principle in the planning phase before the AI starts on its next batch of coding efforts. It may not seem like much, but influencing the design during the AI code planning phase saves time compared to letting the AI free flow its design decisions. Knowing what is not needed helps. <strong>But is it essential? Dear fellow, I think not.</strong></p><p>The reason is simple. In an AI driven environment, the real cost is no longer tied to the number of features or lines of code. The marginal cost of generating an extra module or redesigning an interface is minimal in terms of human labor. <strong>What is essential is optimizing the code base to work effectively with AI.</strong> Yes, AI systems are getting smarter, faster, and better, but applying the best practices for AI driven development to the code base will help prevent errors, or worst critical blunders. Let me explain..</p><p>One day, while sitting with my laptop on the back of my private yacht in the Mediterranean (I wish — in reality, I was in my crummy windowless office), I watched as Claude Code deleted several key functions in a controller while rewriting the file. Running up against the limitations of the context window and working with a large file size went directly against my corporate best practices, and in doing so I invited chaos and watched it unfold in real time. From that moment on, I always run post coding checks to see if any code has been unintentionally deleted.</p><p>Funny stories have emerged of coding LLMs going rogue (not like HAL 9000, just a little rogue), deleting production databases, and then admitting to their mistakes (like a naive young schoolboy). Errors happen, and AI works best when you operate within its performance limits and apply AI best practices. Consequently, I strongly feel what is essential is this:</p><blockquote>The Tenth Reformed AI Agile Principle: “Applying AI coding best practices — is essential, to avoiding AI induced errors.”</blockquote><h4>11. The Impact of AI on the Eleventh Agile Principle</h4><blockquote>Agile Principle 11: “<strong>Self‑organizing teams produce the best outcomes, architecture, and design”</strong></blockquote><p>In classic Agile, self-organizing teams thrived on autonomy. The belief was that teams closest to the work, empowered to make their own decisions, would design better systems and produce better outcomes (ignore the nonexistent effects of Google-style offices with cool bean bag chairs and free snacks).</p><p>But with AI entering the picture, the size and structure of development teams is shrinking. A few experienced developers working with high performance AI coding models can accomplish what once required entire departments. In my view, the best outcomes now come from highly experienced developers working with cutting edge AI coding models, organization of the teams is, to me, very much becoming irrelevant. Adios, eleventh principle.</p><h4>12. The Impact of AI on the Eleventh Agile Principle</h4><blockquote><strong>Agile Principle 12: “Regularly reflect on how to become more effective, then tune and adjust behavior accordingly.”</strong></blockquote><p>In traditional Agile, effectiveness was a human trait. Teams looked inward, reflecting on their habits, communication, and collaboration. Retrospectives were about behavior, tweaking how people worked together to improve outcomes (yes, Frank should seriously reflect on how he berated Jim during his last code review — bad Frank). But AI does not respond to behavior. It runs processes, executes instructions, and follows patterns in data. To become more effective in an AI driven development environment, the focus must shift from people to AI technology systems and how they are performing.</p><p>One thing we now have with AI is the ability to inspect its performance directly. You can look under the hood at the code it writes, track where it drifts off course, and understand how well it interprets your instructional document sets. Effectiveness now depends on two things: first, how well you craft contextual documentation and instructions, and second, how you run your AI. For example, do you ask it to pre-plan major coding tasks, breaking out the process into identifying functions, coding changes, and then validating or fixing those changes? Or do you tell the AI to go for it, run a test after the code is complete, and then use the AI to correct bugs (and for the love of all that is good, do not admit to choosing the second option). To put it bluntly, principle 12 should be rewritten to be:</p><blockquote>The Twelfth Reformed AI Agile Principle: “Regularly check on your AI’s effectiveness, then tune and adjust your best practices to optimized your use of AI.”</blockquote><p>I hope my review of the twelve principles, and how they need to be altered or will fail to work in light of coding-based AI systems, underlines the weaknesses and eventual fall from grace of Agile as the predominant coding philosophy. But who knows, some people do like their antiques.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e563fa1e835" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/why-the-agile-philosophy-fails-in-the-face-of-ai-powered-software-development-1e563fa1e835">Why the Agile Philosophy Fails in the Face of AI-Powered Software Development:</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Business Leaders Need to Know About China’s AI DeepSeek R1 & Janus]]></title>
            <link>https://medium.com/predict/what-investors-and-business-leaders-need-to-know-about-chinas-deepseek-1-ai-model-c82ddc14f6a4?source=rss-4584d7533d83------2</link>
            <guid isPermaLink="false">https://medium.com/p/c82ddc14f6a4</guid>
            <category><![CDATA[google]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[microsoft]]></category>
            <dc:creator><![CDATA[Dr. Brian Scott Glassman]]></dc:creator>
            <pubDate>Tue, 28 Jan 2025 19:30:16 GMT</pubDate>
            <atom:updated>2025-02-15T00:15:47.997Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>What Business Leaders &amp; Investors Need to Know About China’s AI DeepSeek R1 &amp; Janus</strong></h3><p><strong>Dr. Brian Scott Glassman</strong></p><p>Original Publish Date Jan 28th 2025, Revision on Jan 29th 2025</p><p><a href="https://www.linkedin.com/in/drbrianglassman/">LinkedIn</a> | <a href="https://drbrianglassman.com/">Personal Website</a> | <a href="https://medium.com/me/stories/public">Gen AI Articles from this Author</a><br>Generative AI Expert and Founder of a Generative AI Consulting Firm in New York City, Executive Product Leader, Ph.D. in Product Innovation, Purdue University, Former Professor of Innovation, New York University. <br><a href="#9561">Jump to the Author’s Biography Below</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oJoLs9uvrT5yDPzNAc-zBA.png" /><figcaption>DeepSeek’s Impact on AI Technology Companies, Technology Leaders, and Investors</figcaption></figure><h3>Update Note</h3><p>One day after the publication of this article, on January 29, 2025, DeepSeek released a new multimodal generative AI model named <strong>DeepSeek Janus </strong>to celebrate Chinese New Year. This model is said to surpass the image generation capabilities of ChatGPT-4o, which utilizes DALL·E 3. With this open-source release, in this author’s opinion, DeepSeek AI has positioned itself among the top three leaders in generative AI.</p><h3>Overview</h3><p>This article is designed to address the most critical questions investors and business leaders are asking, <strong>delivering memorable answers in three minutes or less if you only read the short answer sections,</strong> while providing deeper analysis for those seeking a more detailed perspective.</p><p>As a hands-on leader in generative AI development, with a Ph.D. in product innovation, I’ll break down the key takeaways from China’s groundbreaking DeepSeek-1 AI model and explore its far-reaching implications on global markets, technology competition, and business strategy.</p><p>If you find this content valuable, be sure to <em>clap</em> and f<em>ollow on Medium</em>, and feel free to connect with me on <a href="https://www.linkedin.com/in/drbrianglassman/">LinkedIn</a>. The <a href="#9561">Author’s Bio is Below</a>.</p><h4>Outline of Topics</h4><ol><li><a href="#40b3">What is the new DeepSeek R1 Model? Why is it important?</a></li><li><a href="#c015">Does the DeepSeek R1 Model pose a potential cybersecurity threat?</a></li><li><a href="#3f63">Why is DeepSeek Causing Stock Market Turmoil?</a></li><li><a href="#f75f">Does DeepSeek challenge the United States’ dominance in AI?</a></li><li><a href="#1e53">How is DeepSeek affecting AI technology stocks?</a></li><li><a href="#1c30">Is DeepSeek a positive or negative for corporations?</a></li><li><a href="#6336">How is DeepSeek Impacting the Open-Source vs. Closed-Source AI Race?</a></li><li><a href="#6d69">How is Nvidia going to be impacted by DeepSeek?</a></li><li><a href="#499b">How is Meta / Facebook going to be impacted by DeepSeek?</a></li><li><a href="#b992">How is Microsoft going to be impacted by DeepSeek?</a></li><li><a href="#37fa">How is Google going to be impacted by DeepSeek?</a></li><li><a href="#eb29">How is Amazon going to be impacted by DeepSeek?</a></li></ol><h4>1. What is the new DeepSeek R1 Model? Why is it important?</h4><p><strong>Short Answer:</strong> It is a new generative AI model from China that surpasses OpenAI’s best-in-class flagship model GPT-4. Its unexpected release and free availability for download and use took the AI community by surprise. Note: On January 29, DeepSeek released <strong>Janus</strong>, a text-to-image generation model. This addition strengthens their offerings, solidifying DeepSeek as a true leader in the generative AI space.</p><p>If you’re short on time, feel free to skip the <strong>Deeper Insights Sections</strong>.</p><p><strong>Deeper Insight:</strong> China’s DeepSeek R1 model is a groundbreaking, cost-effective, ultra-high-performance AI system, widely regarded by experts in AI as one of the most advanced globally. It is celebrated for its exceptional logic and reasoning capabilities, alongside the standard functionalities expected of cutting-edge AI systems. DeepSeek R1 directly competes with OpenAI’s flagship model, o1, which is 49% owned by Microsoft.</p><p>What sets DeepSeek R1 apart is its unique approach: it is entirely free to download and operate. This innovation allows businesses to bypass the high costs associated with OpenAI’s premium-tier model, making DeepSeek R1 up to 1,000 to 10,000 times more cost-effective to operate. By enabling companies to run a best-in-class AI model on their own infrastructure, DeepSeek empowers organizations to drastically reduce operational expenses while deploying the best-in-the-world AI.</p><h4><strong>2. Does the DeepSeek R1 model pose a potential cybersecurity threat to companies who use it?</strong></h4><p><strong>Short Answer:</strong> No, and if there is, the world will learn about it quickly.</p><p><strong>Deeper Insight:</strong> China’s DeepSeek R1 model is trained on widely published and privately gathered datasets, there may be some copyright questions, but it will be hard to prove it in court and even harder to enforce it. Also, if there is some malicious training the AI experts will uncover it in the next couple weeks, well before it gets integrated into any applications.</p><h4><strong>3. Why is DeepSeek Causing Stock Market Turmoil?</strong></h4><p><strong>Short Answer:</strong> First, DeepSeek shows that OpenAI and Anthropic (supported by Microsoft, Amazon, and Google) can not control the AI leadership; other smaller upstart companies will be leaders as well. Second, the GPU hardware requirements to build future AI are much lower than expected, and plans to invest hundreds of billions in AI data centers are in question.</p><p><strong>Deeper Insight: </strong>First, this new AI model highlights China’s emergence as a formidable player in the development of best-in-class AI systems, defying expectations shaped by U.S. trade restrictions and policies designed to curb China’s progress in advanced technology. These efforts have proven ineffective in slowing China’s advancements in AI. Second, China’s DeepSeek AI, with its highly efficient design requiring significantly less computational resources to develop and operate, challenges the prevailing belief that the future of AI depends on investing hundreds of billions of dollars into specialized AI data centers. Finally, this development positions Chinese technology as a potential disruptor to the profit margins of major U.S. technology firms deeply invested in American-based AI solutions. However, this concern may be overstated, as innovation in software, particularly within the open-source community tends to thrive on collective progress rather than fostering a zero-sum competition.</p><h4><strong>4. Is China’s DeepSeek challenging the United States’ dominance in AI?</strong></h4><p><strong>Short Answer:</strong> Yes, and we should expect China to be a leader, along with other emerging powers like India that have both the talent pool and ability to attract capital. This shift suggests a future where AI leadership is more globally distributed and correlated with talent pools with sufficient capital backing.</p><p><strong>Deeper Insight: </strong>China’s leadership in AI was inevitable, regardless of U.S. compute import restrictions. Due to technological advancements, a new wave of countries is poised to become significant players in AI, with India emerging as a strong contender given its expansive talent pool, adaptive education system, and ability to attract capital. As AI systems become more powerful while requiring fewer resources to develop and operate, the competitive landscape will continue to expand.</p><h4><strong>5. How is DeepSeek affecting AI technology stocks?</strong></h4><p><strong>Short Answer:</strong> The AI technology sector is in shock as a small upstart company has produced a best-in-class AI model. This development is shattering the widespread belief that AI excellence is restricted to mega-cap technology companies, causing stockholders and institutional investors to reevaluate these companies’ growth narratives.</p><p><strong>Deeper Insight: </strong>Investors are beginning to realize that mega-cap companies such as Microsoft, Google, Meta, Amazon, and NVIDIA may no longer retain near-exclusive control over the development of best-in-class AI software. Additionally, their dominance in AI hardware, particularly the reliance on NVIDIA GPUs as critical components for advanced AI systems, is being increasingly challenged. This uncertainty has unsettled markets, creating significant ripple effects across the technology sector as institutional investors begin to question their underlying growth model assumptions for these mega cap stocks composing the majority of the Nasdaq. These disruptions are likely to persist, with further market shifts anticipated as China capitalizes on the strong reception of its DeepSeek AI model and intensifies state-backed efforts to advance its AI capabilities. Similar changes are expected as India draws on its substantial intellectual human capital and growing focus on AI innovation to release their own best-in-class models.</p><h4><strong>6. Is DeepSeek a positive or negative for all corporations?</strong></h4><p><strong>Short Answer:</strong> The development is overwhelmingly positive, as companies seeking best-in-class AI now have a low-cost alternative from China and soon from Meta. This competition puts tremendous downward pressure on OpenAI and Anthropic’s pricing.</p><p><strong>Deeper Insight: </strong>For most companies worldwide, across various industries looking to integrate AI into their workflows and processes, DeepSeek R1 offers a significant advantage and a major step forward for AI. It provides a substantial boost in performance at a fraction of the operating costs compared to OpenAI’s ChatGPT o1 model, which previously dominated the market. Also in comparison to existing open source AI models it is far superior in logic and reasoning. Furthermore, since the inner workings of DeepSeek have been published academically, it is likely that a free, open-source competitor to DeepSeek may emerge from Meta or other entities in Q2 of 2025, further accelerating accessibility and innovation in the AI space.</p><h4><strong>7. How is DeepSeek Impacting the Open-Source vs. Closed-Source AI Race?</strong></h4><p><strong>Short Answer:</strong> This gives a tremendous boost the the open-source players, and will likely spawn new open source players from other countries like India.</p><p><strong>Deeper Insight: </strong>The competition between open-source and closed-source AI models is delivering significant benefits for both corporations and individual users. On one side, proprietary models from companies like OpenAI and Anthropic impose high usage fees for their corporate users, while on the other, open-source models, like DeepSeek, offer the ability to deploy a best in class AI model directly on private servers without licensing cost and offers organization the ability to tune the model to their needs. This release is accelerating the pace of AI by placing significant pressure on the market leaders and showing that deep pockets are ultimate .</p><p>Moreover, DeepSeek is exerting downward pressure on proprietary providers to make their best-in-class models more affordable while simultaneously driving them to innovate at a faster rate to maintain their competitive edge.</p><p>Many corporations have ambitious plans to integrate AI across their departmental workflows, and the success of these implementations hinges on the logic and reasoning capabilities of the AI models they adopt. Smarter AI systems that are more energy-efficient, such as those requiring less electricity, deliver better results by enhancing performance and reducing costs. This ongoing competition is ultimately propelling AI technology forward, creating better outcomes for businesses and consumers alike.</p><h4><strong>8. How is NVIDIA going to be impacted by DeepSeek?</strong></h4><p><strong>Short Answer:</strong> It will push the stock price of NVIDIA down, because demand projections for their AI GPUs for training are in question.</p><p><strong>Deeper Insight:</strong> NVIDIA’s AI GPU profit projections — and those of institutional investors — are closely tied to AI demand forecasts from leading consulting firms, major cloud providers like AWS and Azure, and market leaders such as OpenAI and Anthropic. These forecasts consistently project a massive demand for GPUs to power future AI advancements.</p><p>However, DeepSeek has claimed that its groundbreaking R1 model was developed using an exceptionally small number of NVIDIA GPUs. While some speculate that U.S. export controls may obscure the true number of GPUs used, the claim has sparked widespread speculation that DeepSeek has achieved a major technological breakthrough. Such advancements could significantly disrupt institutional forecasts for NVIDIA’s GPU-driven AI growth.</p><p>This author strongly believes that AI technology leaders will drastically reduce their GPU requirements in the coming years, both for developing and operating AI models. Consequently, NVIDIA’s current AI GPU sales projections for generative text AI are likely overstated. However, the anticipated generative AI revolution in video creation, which demands far greater GPU capacity, could offset some of this decline and remain a key driver of future growth for NVIDIA.</p><h4><strong>9. How is Meta / Facebook going to be impacted by DeepSeek?</strong></h4><p><strong>Short Answer: </strong>The stock reaction remains uncertain, but within the next three months, Meta is expected to respond with a new open-source competitor to DeepSeek, which could match or slightly surpass its capabilities. This development has the potential to positively impact Meta’s standing in the AI space.</p><p><strong>Deeper Insight: </strong>Meta, renowned for its Llama 3 series of open-source AI models, now faces the challenge of DeepSeek potentially surpassing it as the preferred open-source solution for fine-tuning and deployment across corporate and startup AI applications. However, it is likely that Meta is already working on a competitive response to DeepSeek, and its release is anticipated to generate significant industry attention. Despite this, Meta continues to grapple with a broader challenge: translating their open-source AI expertise into a sustainable business model beyond deploying AI within their existing platforms like Facebook, Instagram, and WhatsApp.</p><h4><strong>10. How is Microsoft going to be impacted by DeepSeek?</strong></h4><p><strong>Short Answer:</strong> The impact on Microsoft is mixed. While the rise of DeepSeek introduces additional competitive pressure for OpenAI, in which Microsoft holds a 49% stake, it could also drive greater demand for AI hosting services on Microsoft’s Azure Cloud Services, presenting a positive opportunity for the company.</p><p><strong>Deeper Insight:</strong> DeepSeek represents both a challenge and an opportunity for Microsoft. On one hand, competition from DeepSeek could devalue Microsoft’s significant investment in OpenAI. On the other hand, Microsoft’s diversified business model positions it to benefit from the growing and profitable demand for AI compute and hosting services via Azure (market share 24% in 2025). As the second-largest cloud provider, Azure is well-suited to support solutions like DeepSeek and other open-source AI models through offerings such as Azure AI Studio. The rise of more affordable and capable AI models like DeepSeek could accelerate the mass adoption of AI in enterprise applications, ultimately benefiting Microsoft’s broader cloud hosting and AI strategy. Conversely, this model will support the internal development of Microsoft CoPilot giving it added logic and reasoning bettering their consumer and office AI product offerings.</p><h4><strong>11. How is Google going to be impacted by DeepSeek?</strong></h4><p><strong>Short Answer:</strong> The impact on Google is mixed. While DeepSeek directly challenges Google’s flagship Gemini model, the company could still benefit from increased demand for AI hosting services provided through its cloud platform, which holds a 10% global market share, similar to competitors like Microsoft and Amazon.</p><p><strong>Deeper Insight:</strong> DeepSeek’s emergence as a best-in-class AI model is likely to affect the adoption and pricing of Google’s Gemini, pushing down enterprise usage fees as competition intensifies. On the cloud hosting side, however, Google Cloud stands to gain from the growing demand for hosting powerful AI solutions like DeepSeek, presenting a profitable opportunity to expand its presence in the AI compute market.</p><p>On the search front, DeepSeek’s advancements could exacerbate existing challenges. Both Microsoft’s Bing and Google Search offer AI-powered summarization capabilities, but these features currently lack integration with advertisements, negatively impacting ad revenues, a core profit driver for Google. This trend poses a significant challenge to Google’s search advertising model. For a deeper exploration of AI’s impact on search advertising, refer to my detailed article on this topic, <a href="https://medium.com/predict/navigating-the-ai-stocks-landscape-by-dr-brian-glassman-4a6ff7c01688">Navigating the AI Stock Market</a>.</p><h4><strong>12. How is Amazon going to be impacted by DeepSeek R1?</strong></h4><p><strong>Short Answer:</strong> Amazon Web Services (AWS), with a market share of 45% to 50% and a specialized AI hosting service, stands to benefit significantly from the adoption of AI models like DeepSeek R1, as the increasing compute requirements for hosting these models will drive higher hosting revenues.</p><p><strong>Deeper Insight:</strong> AWS, through its Bedrock AI hosting platform, is in direct competition with Azure AI Studio in the race to support the hosting of generative AI models. Amazon holds a dominant position due to its substantial market share and seamless integration of existing AWS services with internal AI hosting tools. With a growing catalog of models and competitive AI hosting pricing, AWS Bedrock is well-positioned to serve as the preferred platform for hosting DeepSeek R1 and other cutting-edge AI models.</p><p>Unlike competitors heavily invested in developing proprietary AI models, AWS has focused on providing infrastructure and support for third-party models. This strategy positions AWS to capitalize on the rise of more powerful and widely adopted AI solutions like DeepSeek R1, further fueling the broader adoption of AI across industries while strengthening its leadership in cloud-based AI hosting.</p><p>Thank you and Clap and Follow for more, and please connect on <a href="https://www.linkedin.com/in/drbrianglassman/">LinkedIn</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1MxwJ88LfGoM6XwFQCSPIg.png" /></figure><h4><strong>About the Author Dr. Brian Glassman</strong></h4><p><a href="https://www.linkedin.com/in/drbrianglassman/"><em>LinkedIn</em></a><em> | </em><a href="https://drbrianglassman.com/"><em>Personal Website</em></a><em> | </em><a href="https://medium.com/me/stories/public"><em>Additional Gen AI Articles</em></a></p><p>— <br>Brian Glassman, Ph.D., is a highly accomplished Technology and Innovation leader, celebrated for his combination of hands-on technical expertise, team leadership, and strategic vision. With over 25 years of experience in technology leadership across software development, product management, and engineering, Dr. Glassman has consistently delivered transformative results.</p><p>In the past five years, he has specialized in driving the development of advanced AI products across a range of industries, including finance, insurance, e-commerce, customer service, and automotive. An innovation strategist and technology thought leader, Dr. Glassman combines a deep technical understanding with the ability to lead diverse, international teams effectively.</p><p>A former professor at NYU, a patented inventor, and a dedicated coach and mentor, he is passionate about fostering growth and innovation. Open to new ideas and opportunities, Dr. Glassman can be reached via LinkedIn or through his website.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c82ddc14f6a4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predict/what-investors-and-business-leaders-need-to-know-about-chinas-deepseek-1-ai-model-c82ddc14f6a4">What Business Leaders Need to Know About China’s AI DeepSeek R1 &amp; Janus</a> was originally published in <a href="https://medium.com/predict">Predict</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>