<?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 Parth Chhaparwal on Medium]]></title>
        <description><![CDATA[Stories by Parth Chhaparwal on Medium]]></description>
        <link>https://medium.com/@pdchhaparwal?source=rss-c93454f16f39------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*DrRmNQU30gxzc69MIznVSA.jpeg</url>
            <title>Stories by Parth Chhaparwal on Medium</title>
            <link>https://medium.com/@pdchhaparwal?source=rss-c93454f16f39------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 08:04:27 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@pdchhaparwal/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 Rise of the AI-Native Product Builder]]></title>
            <link>https://pdchhaparwal.medium.com/the-rise-of-the-ai-native-product-builder-b34bc05606ae?source=rss-c93454f16f39------2</link>
            <guid isPermaLink="false">https://medium.com/p/b34bc05606ae</guid>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[design]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Parth Chhaparwal]]></dc:creator>
            <pubDate>Sat, 23 May 2026 12:07:17 GMT</pubDate>
            <atom:updated>2026-05-23T12:07:17.310Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>AI is not just changing how software gets written. It is changing who gets to build.</em></h4><p>A few months ago, Jensen Huang made a simple but important point: software developers now need to become problem solvers. That line stuck with me because it captures what is changing in product creation. The most valuable people in technology are not just people who can write code. They are people who can take a messy problem, break it down, decide what matters, and build something useful.<br>For years, software companies solved this by dividing the work across functions. Product decided what to build. Design decided how it should feel. Engineering decided how to build it.<br>That model made sense when software creation was slow, expensive, and highly specialized. But AI tools are changing the structure of work itself.<br>Today, a strong builder using tools like Cursor, Claude, Codex, Lovable, v0, Bolt, and Replit can often move from a vague idea to a working prototype much faster than before. What used to require multiple people and several handoffs can now sometimes be done by one person with enough judgment, technical fluency, and product sense.</p><h3>A new talent category is emerging: the AI-native product builder.</h3><p>By that, I mean someone who can take an ambiguous problem, break it into smaller tasks, use AI tools effectively, and ship a working product.<br>I think this role will matter a lot over the next few years. I also think the market is not ready for it yet.</p><blockquote><strong><em>The old model of building products</em></strong></blockquote><p>In the pre-AI era, building software usually required a sequence of specialists. A PM would write the requirements. A designer would create the flows and mockups. A frontend engineer would build the interface. A backend engineer would build the logic. QA, DevOps, and iteration would follow.<br>This was not inefficient for its time. It was the right structure for a world where each part of the product-building process required specialized labor and tooling.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aED3H2PHCYf9CZsJL71KdA.png" /><figcaption>The old workflow of building products</figcaption></figure><p>The company structure reflected this workflow. Product, design, and engineering were separate functions, each with its own leadership, processes, and handoffs. At the team level, the PM, designer, and engineering manager would coordinate to move work forward.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1BRpvqk1WkCbIRUp5PIOGw.png" /><figcaption>The traditional problem-solving structure</figcaption></figure><p>This model created clarity. Product owned what to build. Design owned how it should feel. Engineering owned how to build it.<br>For a long time, this worked well.</p><blockquote><strong><em>AI is compressing the product-building loop</em></strong></blockquote><p>What has changed is not only coding speed.<br>The bigger change is that AI is reducing the cost of execution across the product stack. A builder can use AI tools to help with ideation, wireframing, frontend generation, backend scaffolding, debugging, documentation, and deployment.<br>Earlier, the bottleneck was often: Can we execute this at all?<br>Now, the bottleneck is increasingly: do we know what to build, how to break it down, and how to use AI tools well enough to ship something useful?<br>That is a very different bottleneck.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FnqwZA_ZAeaTl8QYDElcUA.png" /><figcaption>The compressed AI-era workflow</figcaption></figure><p>This is why I do not think the new edge is simply knowing how to prompt.<br>The edge is judgment.<br>The best builders in the AI era will not be the people who blindly outsource thinking to AI. They will be the people who know how to use AI as leverage while still owning the problem, the product, and the outcome.</p><blockquote><strong><em>A new role is emerging: the AI-native product builder</em></strong></blockquote><p>The AI-native product builder is not just a software engineer. They are also not just a PM, designer, or prompt engineer.<br>They sit at the intersection of product, design, engineering, and AI-tooling fluency. They can move across the full loop from problem to product.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G96Zw1348EaA1N5tkHtG0Q.png" /><figcaption>A new talent category is emerging</figcaption></figure><p>The role combines several capabilities that have historically been spread across different functions.</p><pre>Capability            What it looks like<br>---------------------------------------------------------------<br>Problem solving       Can take a vague idea and structure it clearly<br><br>Product judgment      Knows what to build first and what to ignore<br><br>Design sense          Can create usable and intuitive flows<br><br>Technical fluency     Understands frontend, backend, data, and<br>                      deployment well enough to build intelligently<br><br>AI-tool usage         Uses AI to move faster without losing ownership<br>                      of the work<br><br>Execution             Ships working products, not just ideas or mockups</pre><p>This is especially valuable in early-stage startups, AI-native companies, founder offices, venture studios, internal innovation teams, and lean product teams.<br>These teams do not just need more people who can complete tickets. They need people who can own ambiguous problems end-to-end.</p><blockquote><strong><em>The market is already showing a talent gap</em></strong></blockquote><p>The data points all point in the same direction.<br><a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-how-organizations-are-rewiring-to-capture-value">McKinsey’s 2025 State of AI survey</a> found that 78 percent of respondents said their organizations used AI in at least one business function. <a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai">In a later 2025 survey</a>, that number rose to 88 percent. The same 2025 State of AI work found that no more than 10 percent of respondents reported scaling AI agents in any individual function.<br>A separate <a href="https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/superagency-in-the-workplace-empowering-people-to-unlock-ais-full-potential-at-work">McKinsey workplace report</a> found that almost all companies invest in AI, but only 1 percent believe they are mature. <a href="https://www.bain.com/about/media-center/press-releases/20252/widening-talent-gap-threatens-executives-ai-ambitions--bain--company/">Bain also found that 44 percent of executives cite lack of in-house AI expertise</a> as a key barrier to implementing generative AI.<br>The message is clear: companies have access to the models and the tools, but they do not have enough people who can turn them into reliable, adopted products.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L-Eqo2elal0Smg9sQH_Ikw.png" /><figcaption>The AI adoption bottleneck</figcaption></figure><p>This is not just a tooling problem. It is a talent problem.<br>Many organizations now have access to the same AI tools. The question is whether they have people who can apply those tools with taste, discipline, technical judgment, and product sense.</p><blockquote><strong><em>The hiring system is not designed for this role</em></strong></blockquote><p>If this talent category is real, the next question is obvious: how do companies identify these people?<br>That is where the market feels weak.<br>Companies still rely on resumes, standard interviews, GitHub profiles, LeetCode-style coding rounds, referrals, and sometimes portfolio work. These signals are useful, but they are incomplete.<br>A resume does not show whether someone can take an ambiguous product problem and ship something useful in seven days. LeetCode does not show product judgment. GitHub does not always show design taste, user empathy, prioritization, or ability to work under ambiguity. Interviews often reward communication polish more than actual building ability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w1SFC-gLjp6LorR5XvvDiQ.png" /><figcaption>Existing signals vs. the signal companies actually need</figcaption></figure><p>The market needs a better signal for AI-native product-building ability.<br>Not just a better resume. Not another generic coding test. Not another AI course certificate.<br>A real signal should show what someone can build, how they think, what tradeoffs they made, how they used AI tools, and whether the final product is thoughtful, usable, and real.</p><blockquote><strong><em>The whitespace: hiring, training, and benchmarking</em></strong></blockquote><p>The opportunity can be summarized in three gaps.<br>First, there is a hiring gap. Companies do not know how to identify strong AI-native product builders.<br>Second, there is a training gap. Candidates do not know how to move from being only an engineer, PM, designer, MBA student, or operator into becoming a full-stack product builder.<br>Third, there is a benchmarking gap. There is no trusted standard for what good AI-native product building looks like.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XDJyiB4YsjKinZbcDVanUg.png" /><figcaption>The three gaps in the market</figcaption></figure><p>We have LeetCode for coding. We have Kaggle for machine learning. We have case interviews for consulting.<br>But we do not yet have a trusted benchmark for people who can use AI tools to turn ambiguous problems into working products.<br>That is the gap I keep coming back to.</p><blockquote><strong><em>What the future may look like</em></strong></blockquote><p>I do not think specialization disappears.<br>Large organizations will still need deep specialists. Many products will still require strong PMs, designers, engineers, researchers, data scientists, and infrastructure teams.<br>But I do think the most leveraged early-stage talent will increasingly be people who can combine product, design, engineering, and AI-tooling fluency in one person.<br>In that world, the best talent signal will not only be where someone worked, what title they had, or how well they interviewed.<br>It will be much simpler: what can they actually build?</p><blockquote><strong><em>Introducing Builders Fellowship</em></strong></blockquote><p>This is why I have started working on Builders Fellowship.<br>Builders Fellowship is a 7-day product-building challenge for AI-native builders.<br>The idea is simple: candidates receive a real product case, build a working prototype using AI tools, submit their demo, product explanation, and architecture, and get evaluated across product, design, engineering, AI usage, and execution.<br>Top performers may be invited to Builder Demo Day, where they get the opportunity to present their work to selected startups and tech companies.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L0fA_qgDW2ICZZO9phJWIg.png" /><figcaption>How Builders Fellowship works</figcaption></figure><p>The goal is not to run another generic hackathon.<br>The goal is to create a trusted hiring and benchmarking signal for AI-native product builders.<br>If you are a student, engineer, PM, designer, MBA student, indie hacker, or startup operator who wants to prove you can build in the AI era, this might be relevant.<br>If you are a startup or tech company looking for high-agency builders who can turn ambiguous problems into working products, this might be relevant too.<br>You can learn more here: <a href="https://www.builderfellowship.org/">https://www.builderfellowship.org/</a> or reach out to me at <a href="/service/mailto:parth.chhaparwal@yale.edu">parth.chhaparwal@yale.edu</a>.</p><blockquote><strong><em>Final thought</em></strong></blockquote><p>I think one of the biggest talent shifts of the next few years will not just be more people using AI.<br>It will be the rise of people who can use AI to build real products from messy problems.<br>That is a different kind of talent.<br>The companies that learn how to identify it early will move faster. The candidates who learn how to demonstrate it clearly will stand out.<br>That is the thesis behind Builders Fellowship.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b34bc05606ae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Where AI Applications Will Actually Win: An Investment Thesis]]></title>
            <link>https://pdchhaparwal.medium.com/where-ai-applications-will-actually-win-an-investment-thesis-480ce81e6b95?source=rss-c93454f16f39------2</link>
            <guid isPermaLink="false">https://medium.com/p/480ce81e6b95</guid>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[investment-opportunities]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Parth Chhaparwal]]></dc:creator>
            <pubDate>Sun, 22 Mar 2026 00:40:05 GMT</pubDate>
            <atom:updated>2026-03-22T00:40:05.614Z</atom:updated>
            <content:encoded><![CDATA[<p>Eighty percent of the global workforce does not sit at a desk. Yet historically, less than one percent of enterprise software investment has been built for them. This article explains what happens when that changes.</p><h3>Software was built for one kind of work. The economy runs on two.</h3><p>Almost every software application ever built, and most AI being funded right now, was designed for the same person. Someone sitting at a desk, on a screen, with a keyboard in front of them and a moment of stillness to type into it. Let’s call this <strong>screen-based work</strong>. Software developers. Accountants. Lawyers. Financial analysts. The assumption was so deeply baked in that nobody noticed it was an assumption at all.</p><p>The money followed the same logic. In the last two quarters alone, the largest funding rounds went to companies targeting this exact slice of the economy. Pace AI bringing agentic process outsourcing to insurance back-offices, Sandstone AI building AI-native legal departments, Accordance AI running agents for tax and audit workflows, Day AI acting as a cursor for GTM teams. Smart bets, well-executed, going after real pain. But they are all, without exception, going after the same person: the knowledge worker at a screen.</p><p>Here is what that misses. The majority of economic work, by volume and by number of people, doesn’t happen at a desk. It happens in the field, in unstructured environments, with incomplete information and no moment of stillness. Let’s call this <strong>field-based work</strong>. The home health aide documenting a visit in a patient’s living room, working off a phone with poor cell connectivity. The restaurant manager at 7pm on a Thursday trying to understand why food cost jumped three points this week. The retail associate who doesn’t know off the top of their head whether the router on the shelf is compatible with the customer’s mesh setup at home. The field technician diagnosing an HVAC unit by sound and smell before pulling up a single manual.</p><p>These are not edge cases. They are the majority of work by volume. Teachers, home health aides, retail workers, manufacturing supervisors, field service technicians, restaurant operators, administrative staff. This is where the economy actually runs. And almost none of it has been meaningfully touched by software, let alone AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0usqwWaVdw6nMlN2sDfR3g.png" /><figcaption>Opportunity Size for Field-Based Work vs Screen-Based Work</figcaption></figure><p>The reason isn’t that the problems were too hard. The workflows are often repetitive, domain-specific, and highly documentable, exactly the kind of work that technology should have reached decades ago. The reason is simpler: the interface was wrong. Every software wave assumed a keyboard, a screen, and a user with the time to type a structured query. Workers in the field don’t have that. They have their hands, their voice, their camera, and the situation directly in front of them. Software was never built for them, so they were never counted as a market.</p><p>Until now, that constraint held. What changed isn’t that the problems got easier. What changed is that the tools for solving them finally exist. And they arrived quickly enough that most of the investment world hasn’t caught up to what they make possible.</p><h3>AI Removed the Constraints That Defined the Last Software Era…</h3><p>There have been many trends shaping how AI impacts work today. Model context protocols. Cursor and vibecoding. Multimodal models. The list grows every quarter and it can be hard to know which shifts are meaningful and which are noise.</p><p>Five of them are definitely not noise. Five of them are closing off a retreat to the old way of doing things entirely. Not because they are the most talked about, but because each one removes a specific constraint that previously made field-based work unreachable by software. Together, they don’t just improve what software can do. They change who software can reach.</p><ol><li><strong>Voice and camera became first-class interfaces.</strong> For three decades, software assumed you would type. Field workers don’t type, not because they are less capable, but because their hands are occupied, their environments are noisy, and the work doesn’t pause to let them sit down at a keyboard. Multimodal infrastructure changed this. A field technician can now narrate what they see into a phone. A home health aide can speak a two-minute summary of a patient visit and have it structured, documented, and filed. In the near future, a video of a broken piece of equipment becomes the input, not a description of it. The interface constraint that excluded the majority of the workforce for three decades dissolved in roughly eighteen months.</li><li><strong>The cost of a complex thought is approaching zero. </strong>A reasoning model can now interpret a 400-page regulatory corpus, apply it to a specific fact pattern, and produce a legally coherent output for $0.10. Work that previously required a specialist billing $400 an hour is now a fraction of a cent per inference call. This is not an efficiency gain. It is a different production function entirely. The economics of expertise have changed in a way that cannot be undone.</li><li><strong>Organizational memory has become holdable. </strong>AI can now hold a million context tokens in a single call. That means an entire organization’s institutional knowledge, every contract, every patient history, every equipment manual, every regulatory filing, can sit inside a single reasoning call simultaneously. The AI stops being something you constantly query and starts being something that remembers everything the organization knows. That shift changes what kind of switching cost a software product can build.</li><li><strong>Agents can now own a goal, not just answer a question. </strong>The difference between a good AI tool and an agent is scope. A tool responds when asked. An agent receives a goal on Tuesday, takes actions on Wednesday, follows up on Friday, and escalates on Monday if nothing happened. Multi-step execution across time is now reliable enough to ship in production workflows. This is what makes it possible to hand an entire job function to software rather than just a task within one.</li><li><strong>The marginal cost of creation has hit zero for entire output categories.</strong> The fiftieth ad variant costs the same as the first. The thousandth service report costs the same as the first. The thousandth grant proposal costs the same as the first. This doesn’t just make production cheaper. It restructures who can compete in markets where production volume was previously the moat. Companies that used to win by producing more can now be matched by a two-person team with the right system.</li></ol><p><a href="https://pdchhaparwal.medium.com/infrastructure-trends-shaping-the-next-generation-of-ai-applications-7488187a2aeb">If you want to read more about these trends in detail, check out my article on infra trends impacting AI applications here.</a></p><p>None of these trends are reversible. No version of the future puts the cost of reasoning back to $400 an hour, or shrinks context windows back to thousands of tokens, or requires field workers to sit at a keyboard again. The direction is set.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H8jSLbmk_PvsaaBmNc2_ig.png" /><figcaption>Defining Trends that Shaped the Future of AI Applications</figcaption></figure><p>What isn’t set is which economic structures these trends disrupt first, or in what order, or how fast. That depends on where the constraints were thickest, where the current tools are weakest, and where the gap between what software can now do and what software has actually built is widest.</p><h3>…But Product Success in AI Often Depends on Deeper Business Mechanics</h3><p>The five shifts mentioned don’t automatically translate into durable businesses. Plenty of companies will be built on top of them that fail, not because the technology didn’t work, but because of decisions made early about how to build, what to charge, and what to own. The way these trends interact with business behavior determines whether value compounds inside a company or leaks out of it.</p><p>There are seven mechanics that govern this. They are not equal, and they are not independent. They have a sequence. Getting some of them wrong early makes the others impossible to get right later.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WrB2ia1M0xs3S0g5M-07ag.png" /><figcaption>The 7 Mechanics Defining the Success of AI Apps</figcaption></figure><p>The first three are mechanics of scale. They determine how many customers you can reach, how deeply the product embeds, and whether customers can leave once they are in.</p><ol><li><strong>Voice and camera are the new distribution. </strong>Whether you build for voice and camera or for text determines which workforce you can reach. The smartphone, cheap internet, and the app store already put a capable device in the pocket of every field worker on earth. What was missing was software that worked the way they work. Audio recordings, photos, and video are the natural modalities of field-based work. Text and code are not. Founders who build the right infrastructure for multimodal input create a genuine right to win, because the infrastructure itself is hard to replicate. This means building data ingestion workflows that preserve fidelity and reduce translation loss, letting models interact with information in its native form. The closer the model stays to the original input, the closer its output stays to the truth. You can see this mechanic playing out already in how people use ChatGPT, where the proportion of voice-based inputs is rising fast.</li><li><strong>Own the function, not the feature.</strong> Whether you own a function or a feature determines whether you get bundled away. Agentic workflows have made it possible to automate entire functions within and across organizations, not just single tasks within them. Building for functions resists bundling. Building for features doesn’t. A feature gets commoditized or absorbed by a platform product. A function, done well, becomes the thing the platform depends on.</li><li><strong>Institutional memory is the new lock-in.</strong> Whether your product accumulates memory determines your switching cost. Large context windows at low prices make it possible for an application to hold an organization’s entire operational history, every contract, every patient file, every equipment manual, in a single reasoning call. The longer that memory accumulates, the harder it is to leave. This is the mechanic that creates durable switching costs in products like executive assistants, sales agents, and home health documentation tools.</li></ol><p>These three work together. Voice and camera maximize the addressable user. Owning the function unlocks high-value complexity that was previously out of reach. Institutional memory ensures customers cannot churn once the product has learned their context. Together they drive adoption and retention.</p><p>The next three are mechanics of profitability. They determine whether the business that grows is actually worth owning.</p><ol><li><strong>Charge per thought, not per seat.</strong> Flat subscription pricing is a margin trap for AI businesses. Reasoning models have variable inference costs. Charging a fixed monthly fee for variable compute means your margins deteriorate as your best customers use the product more. The right pricing unit is the output, a claim processed, a report generated, a proposal drafted. Usage-based pricing aligns what you charge with what it costs to deliver, which is the only way gross margins survive at scale.</li><li><strong>Orchestration quality is margin. </strong>Poor data pipelines and AI workflows bleed margin to inference costs. Good pipelines capture it. The difference between a company with 70% gross margins and one heading toward a unit economics crisis at Series B is often not the model they chose. It is how well they built the retrieval, context management, and workflow execution underneath it. This is where founders can build business model moats that are genuinely hard to replicate.</li><li><strong>Zero marginal cost restructures categories</strong>. When the fiftieth ad variant costs the same as the first, or the thousandth service report costs the same as the first, the economics of entire markets change. This is only an advantage if the company owns the pricing. Founders who don’t think carefully about this early will find that the value created by zero marginal cost flows to their customers or to the model providers, not to them.</li></ol><p>The mechanic that cuts across all six of these is the simplest one to understand and the most underrated in practice.<strong> </strong><a href="https://pdchhaparwal.medium.com/rethinking-product-strategy-in-the-age-of-ai-1b33e8fc81d8"><strong>If the output is checkable, adoption is fast. </strong>When a customer can verify AI output immediately against their own knowledge, something such as a tax return that either gets accepted or doesn’t, or a lesson plan the teacher can read and judge in sixty seconds, or a service report the customer can confirm on the spot, trust builds quickly. Verifiability is user-facing. It drives retention before any deeper switching cost has had time to form. It is the mechanic that connects scale and profitability into something that compounds</a>.</p><p>Together these seven mechanics define the difference between an AI company that grows and one that grows durably. The founders who understand them early make product and pricing decisions that are very hard to undo later. The ones who don’t will face the consequences at Series B in the form of problems that look like growth issues but are actually structural ones.</p><p>Which brings up the obvious question. Where exactly do these mechanics apply with the most force, against the largest untouched markets, with the least competition in the way?</p><h3>The Best Opportunities Are Where the Shifts Land and the Moat Builds Itself…</h3><p>These mechanics describe the conditions for building a durable AI business. The thread running through my <em>investment thesis on AI Applications</em> is finding exactly those conditions. Back exceptional founders building AI applications in large, undercrowded markets where the right to win compounds naturally as the product does its job. Those conditions hold most clearly across four groups of markets right now.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5gleV5DgbXS7ldUpzXKZbA.png" /><figcaption>Investment Thesis: The 11 Lucrative Startups Ideas</figcaption></figure><p>The first group shares one characteristic: the workers in these markets cannot be reached without voice and camera as the primary interface, which means every product built for them is a genuinely new category with no incumbent to displace.</p><ol><li><strong>Home Health AI Agent: </strong>An agent that listens to a home health aide speak a two-minute summary after a patient visit, structures it into compliant documentation, and coordinates the next visit automatically.</li><li><strong>Field Service Intelligence Agent: </strong>An agent that a technician narrates to at the job site, which pulls the relevant equipment manual, diagnoses the fault, orders the right part, and generates the service report before the technician leaves the driveway.</li><li><strong>Product Intelligence AI / Frontline Retail Agent: </strong>A voice agent on the retail floor that tells an associate in real time whether a product is compatible with a customer’s specific setup.</li><li><strong>Restaurant Operations Agent:</strong> A restaurant intelligence layer that connects point-of-sale, labor, and food cost data and tells the manager on Thursday evening exactly why margin dropped this week.</li></ol><p>The second group shares a different characteristic: zero marginal cost doesn’t just make production cheaper in these markets, it changes who can compete in them entirely, which means the founder who owns the pricing rewrites the economics of the category rather than just building a better product. These are not productivity tools. They are new competitive realities for entire industries.</p><ol><li><strong>Video Ad Creative Agent: </strong>A performance marketer briefs an agent on audience, product and goal, and gets fifty ad variants tested and iterated without a human editor touching them.</li><li><strong>Music Production AI for Brands: </strong>A brand gets a licensable music track generated to brief for fifty cents instead of five thousand dollars.</li><li><strong>Game and 3D Generation Agent: </strong>A solo game developer ships a title that previously required a ten-person art team.</li></ol><p>The third group shares the characteristic that the value only exists if the entire job function is owned, because anything less than the function is a feature that a platform product will eventually bundle away for free. The difference between these products and what exists today is the difference between a tool that waits to be used and one that holds a goal and sees it through.</p><ol><li><strong>AI Executive Assistant and Meeting Execution: </strong>The executive assistant agent holds an executive’s goals, attends every meeting, executes action items across days without being prompted, updates the CRM, drafts the follow-up, and escalates what hasn’t been resolved by Friday.</li><li><strong>SMB Manufacturing OS: </strong>AI runs shift reporting, tracks OEE, manages maintenance schedules, and flags anomalies across a facility that previously ran on spreadsheets and whiteboards.</li></ol><p>The fourth group shares the characteristic that the product must hold an enormous and constantly evolving body of knowledge in context at all times, which means the switching cost is the memory itself and grows automatically the longer the product runs. In these, leaving the product means starting over. That is a moat that no competitor can shortcut.</p><ol><li><strong>AI Compliance Agents across all Non-IT Regulatory Domains:</strong> An agent that holds an entire regulatory corpus across OSHA, EPA, SEC and employment law, monitors for changes, prepares filings, and maintains audit trails simultaneously.</li><li><strong>AI Education Assistant across the Full Teaching and Learning Ecosystem:</strong> An AI agent that knows every student’s learning history, understands the district’s curriculum standards, generates differentiated lesson plans, and prepares the documentation that teachers currently spend hours on every week.</li></ol><h3>…And Here Is What Each of Those Markets Actually Looks Like…</h3><p>The diagram below shows the labor market size and the existing players in each of the eleven spaces.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CLyog-zkaGnHk_QNZmeIoA.png" /><figcaption>Total Labor Spend Across the 11 Ideas and Core Competitors</figcaption></figure><p>What it doesn’t show is the product that needs to be built and why the gap is structural rather than accidental.</p><ul><li>In education, nobody has built the productivity layer for the teacher. Carnegie Learning built for the student.</li><li>In product intelligence, nobody has built for the person standing on the shop floor. Salesforce and Intercom built for the manager behind the screen.</li><li>In executive assistance, the gap is not scheduling (Reclaim and Motion handle that), the gap is autonomous execution of everything that follows a meeting.</li><li>In home health, Ambience and Abridge built for the hospital. The aide working off a phone in a patient’s living room has nothing.</li><li>In SMB manufacturing, Tulip and Tractian built for the enterprise. The two hundred thousand mid-market facilities running on Excel have nothing.</li><li>In field service, the product that works the way a technician actually works has never been built.</li><li>In video ad creative, Runway built the infrastructure but nobody built the B2B product on top of it.</li><li>In compliance, four hundred million dollars has gone into IT security compliance. Every other regulatory domain is untouched.</li><li>In game and 3D, the fragments exist but the full development environment for a solo developer does not.</li><li>In restaurant operations, four platforms hold the data and none of them talk to each other. In music production for brands, the consumer products exist and the B2B brief-to-track product does not.</li></ul><p>In every case, the adjacent company validated the problem. The direct product remains unbuilt.</p><h3>…But Not All of Them Build Moats the Same Way</h3><p>The following maps the eleven opportunities by how the moat builds and how quickly. The three clusters that emerge have a different risk profile, a different capital requirement, and a different thing that has to go right for the business to be durable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KZN_CkU51xgXLwufy9sZlg.png" /><figcaption>Building a Portfolio Across Moats</figcaption></figure><p><strong>Compounders</strong> take time. The data flywheel in field service, home health, education, compliance, and manufacturing builds quietly over years, and the switching cost only becomes truly punishing once the accumulated context is deep enough to be irreplaceable. These are the businesses worth the most patience.</p><p><strong>Category Creators </strong>lock in faster because the moat is integration depth rather than data accumulation. The product intelligence agent, the restaurant intelligence layer, and the executive assistant embed across tools the customer already depends on. Removing them is painful from the first year of serious use.</p><p><strong>Restructurers</strong> are a different bet entirely. Video ad creative, music production for brands, and game and 3D generation don’t win by locking customers in. They win by changing the economics of the category so completely that the old way of doing things stops making sense. The founding team’s most important early decision is pricing, not product.</p><p>A portfolio with only Compounders is diversified by sector but uniform in how the moat builds and what can go wrong. Adding a Category Creator and a Restructurer creates genuine diversification across moat mechanics, which is the risk management that actually matters in AI investing right now.</p><p>The markets are large, the products are unbuilt, and the window is open. If you are a founder building in any of these spaces, or an investor thinking through the same questions, reach out at <a href="/service/mailto: pdchhaparwal@gmail.com">pdchhaparwal@gmail.com</a>.</p><p>Check out other relevant articles on AI Applications here:</p><ol><li><a href="https://pdchhaparwal.medium.com/rethinking-product-strategy-in-the-age-of-ai-1b33e8fc81d8">Rethinking Product Strategy in the Age of AI</a></li><li><a href="https://pdchhaparwal.medium.com/infrastructure-trends-shaping-the-next-generation-of-ai-applications-7488187a2aeb">Infrastructure Trends Shaping the Next Generation of AI Applications</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=480ce81e6b95" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Infrastructure Trends Shaping the Next Generation of AI Applications]]></title>
            <link>https://pdchhaparwal.medium.com/infrastructure-trends-shaping-the-next-generation-of-ai-applications-7488187a2aeb?source=rss-c93454f16f39------2</link>
            <guid isPermaLink="false">https://medium.com/p/7488187a2aeb</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[venture-capital]]></category>
            <dc:creator><![CDATA[Parth Chhaparwal]]></dc:creator>
            <pubDate>Mon, 16 Mar 2026 03:43:06 GMT</pubDate>
            <atom:updated>2026-03-16T03:43:06.884Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How founders and operators should think about reasoning, multimodality, context, and where value will accrue in AI applications</em></p><p>Over the last four years, the way people consume AI has changed dramatically. As a venture capitalist, I have seen this shift up close. In 2022, a wave of startups looked genuinely exciting. By 2024, many of them had already become redundant because reasoning models and the underlying AI stack had improved so quickly. I also saw a very different conversation emerge in FinOps circles, where the focus started shifting to something far more practical: how to bring down the cost of token usage at scale. The promise of AI is real, and its ability to unlock scale is holding up. But adoption is still far from frictionless. Both startups and enterprises are running into infrastructure constraints that shape what can actually be built, deployed, and sustained. This article looks at the infrastructure trends that are quietly, but decisively, shaping the next generation of AI applications.</p><h3>Trend 1: Reasoning models have changed the promise of AI from a response to a result</h3><p>Until 2024, standard large language models such as GPT-4, Claude 3, and Gemini 1.5 defined how most humans interacted with AI. In hindsight, the core mechanism was fairly simple. These models were trained to predict the next token as accurately as possible, with the final output optimized to sound coherent, logical, and syntactically correct.</p><p>That was enough to make AI feel impressive. It could answer, summarize, draft, and explain. But it was still largely operating in the business of producing plausible responses.</p><p>The arrival of reasoning models such as OpenAI o1, o3, and DeepSeek R1 changed that equation. These models were not just designed to sound right. They were designed to work toward getting the right answer. That shift matters more than it may seem.</p><p>Instead of jumping straight to an output, reasoning models break a problem down internally before responding. They generate a long scratchpad of intermediate thinking, often referred to as thinking tokens, where the model works through the problem step by step, checks its own logic, revisits earlier assumptions, corrects mistakes, and sometimes even tries alternate paths before landing on a final answer. On the surface, this can feel brute-force. And it does come with tradeoffs, which are important. But it also unlocks a very different class of outcomes, especially in areas where earlier models often struggled. Multi-step math, complex code generation, scientific reasoning, and other high-precision tasks suddenly became far more viable.</p><p>In other words, the promise of AI shifted. It was no longer just about getting a polished response quickly. It became about reaching a reliable result, even if that required more work under the hood.</p><p>But reasoning models also introduced a very new set of constraints.</p><ol><li>Latency went up sharply. Because the model now spends time thinking before answering, response times often stretch to 10 to 90 seconds, compared to the 1 to 5 second experience users had come to expect from standard models. That creates a very different product experience. Users are no longer having a quick back-and-forth with a chatbot. In many cases, they are waiting for a system to complete a task.</li><li>Costs rose materially. More thinking means more computation, and more computation means more tokens consumed. What may once have cost roughly $0.005 to $0.015 per 1K tokens started moving closer to $0.06 to $0.15 per 1K tokens, making reasoning-heavy workloads meaningfully more expensive. In practice, that can mean a 10x to 30x jump in cost depending on the task.</li><li>These models can overthink. Ironically, the very mechanism that makes them powerful on hard problems can make them inefficient on simple ones. For lightweight conversational tasks, fast responses, or routine actions, reasoning models can be unnecessary and sometimes even worse.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WEcz6mnCy6yaY083rE-txA.png" /><figcaption>Tasks and Ideal Model Type</figcaption></figure><h4>Implications for AI application builders</h4><p>This changes how builders need to think about product architecture.</p><ol><li><strong>Routing:</strong> Not every task deserves a reasoning model. Builders now need to make deliberate decisions about which model should handle which job. The best founders will not just pick a model. They will build intelligent model routers that decide when a standard model is enough, when a reasoning model is necessary, and how to control cost without compromising quality. If every task defaults to reasoning mode, economics can spiral very quickly.</li><li><strong>Pricing:</strong> Flat monthly subscriptions start to break when heavy users run dozens or even hundreds of expensive reasoning queries. In a world where compute consumption varies dramatically across users, usage-based pricing becomes much harder to avoid. Cursor is a useful example here, with its distinction between fast and slow requests. That is not just a pricing choice. It is a reflection of the underlying economics of reasoning systems.</li><li><strong>Trust: </strong>Reasoning models are increasingly being used for tasks that are higher stakes, more complex, and harder for the end user to verify instantly. In that world, it is not enough for an application to simply present an answer. Users need visibility into why the model reached a conclusion, what evidence it relied on, and how the reasoning unfolded. Inspectability, citability, and auditability are quickly becoming table stakes, especially in serious workflows.</li></ol><p>The broader recommendation to founders is that if you are building with reasoning models, you need a dedicated infrastructure stack built for them. These systems are not just better versions of old chat products. They behave differently, cost differently, and need to be orchestrated differently. The ideal stack looks something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*peOvq9CsRqShBctx-2TtFw.png" /><figcaption>Ideal Tech Stack in the Age of Reasoning Models</figcaption></figure><h3>Trend 2: Multimodality is substitutive, not additive, in how humans interact with AI</h3><p>Most application builders think about multimodality as added optionality. The assumption is simple: if users can interact with AI through text, audio, vision, and video, the product becomes more flexible. That framing misses the deeper shift.</p><p>Multimodality is not just about giving users more ways to provide input. It is about replacing entire upstream pipelines that were built only to convert the real world into text. In many software systems, text was never the natural source of information. It was just the format machines were best able to process. So companies built layers of manual effort, structured workflows, and brittle tooling to translate voice, images, documents, and video into text before any intelligence could be applied.</p><p>That is now changing.</p><p>When a model can work directly with the native modality, the old translation stack starts to disappear. And that matters because every conversion step introduces friction. Going from video to image to text is not a neutral transformation. It creates loss. Context gets stripped away, nuance gets flattened, and errors compound across each stage. By the time the model finally sees the input, it is often working with an incomplete version of reality.</p><p>That is why the real promise of multimodality is substitutive, not additive. It does not just expand the interface. It collapses the infrastructure sitting behind the interface.</p><p>For builders, this changes the goal. Winning in multimodality is not about sprinkling in more input types and calling the product multimodal. It is about building robust data ingestion workflows that preserve fidelity, reduce translation loss, and let models interact with information in its native form whenever possible. The closer the model stays to the original modality, the closer its output stays to the truth of the input.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kpfXbWfl0TTFekRt2AmOEQ.png" /><figcaption>Multimodality: Maturity and Use Cases</figcaption></figure><h3>Trend 3: Expanding context windows are eliminating an entire layer of brittle AI infrastructure</h3><p>For the first few years after ChatGPT launched, Retrieval Augmented Generation, or RAG, became the default architecture for a large share of AI applications. The logic was simple. Models could not hold much context, so if you wanted a useful answer, you first had to retrieve the most relevant pieces of information and feed only those into the model. That constraint created an entire ecosystem. Vector databases, embedding models, chunking strategies, and similarity search all emerged to solve the same basic problem: context was scarce, so relevance had to be engineered before reasoning could begin.</p><p>That world is now starting to change.</p><p>As context windows expanded, the underlying assumption behind much of this infrastructure began to weaken. Models moved from handling a few thousand tokens to hundreds of thousands, and then to a million tokens or more. At that scale, a model is no longer just reading a page or two. It can work through something closer to 750,000 words, which is enough to ingest massive document sets, large codebases, or even a meaningful slice of an organization’s institutional memory in one shot.</p><p>The speed of this shift has been easy to underestimate. But its implications are enormous.</p><p>Once context becomes abundant, retrieval stops being the only viable path to useful output. The model no longer always needs a carefully pre-selected set of fragments. In many cases, it can absorb the broader corpus directly and reason across it as a whole. That opens up a very different class of applications. Legal and financial analysis can move from narrow document lookup to full corpus understanding. Consumer products can move beyond short-term memory and start behaving more like persistent, lifetime AI companions that actually remember enough context to be useful over time.</p><p>This does not mean retrieval disappears entirely. But it does mean retrieval is no longer the unquestioned center of the stack.</p><p>The companies most exposed to this shift are the ones still built too rigidly around old assumptions. If a product remains overly dependent on chunking, embedding, and retrieval logic without rethinking what larger context windows make possible, it risks becoming infrastructure for a world that is fading. The winners will be the teams that see the shift early, adapt their architecture quickly, and rebuild around the fact that context is becoming dramatically less scarce.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yJtazXU45t2T24NjUN5QxA.png" /><figcaption>History of Expanded Context Windows</figcaption></figure><h3>Trend 4: Model commoditization is forcing startups back to the old playbook of data and distribution moats</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LHEU5RCNohN8W-Co2aHoFA.png" /><figcaption>Model Commoditization &amp; Rapid Reduction in Token Costs</figcaption></figure><p>As competition intensifies and innovation keeps compounding, AI models that once felt scarce, differentiated, and expensive are becoming cheaper, more abundant, and increasingly interchangeable. That has set off a race among LLM providers to push token prices down and capture market share. In just the last two years, token prices have fallen dramatically, from roughly $30 per million tokens to $0.15 per million tokens. For AI application companies, this is great news on one level. Lower model costs make business models more viable and improve margins. But there is a catch. As models get cheaper and more accessible, capability itself stops being a durable moat.</p><p>That is the real shift.</p><p>When the underlying intelligence layer becomes widely available, startups can no longer rely on access to the model as their edge. Everyone gets access to the same or similar capabilities. And once that happens, the question stops being who has the model and becomes who owns the customer, the workflow, and the feedback loop around the model.</p><p>Jasper AI is a fascinating case study here. In 2022, it was one of the hottest companies in the market, positioned as an AI copywriter at a time when that felt novel and highly differentiated. But underneath the product was essentially GPT-3 wrapped in a clean interface, plus a useful template library. That worked well in the beginning. Customers liked the experience and were willing to pay for it. But as the base models improved and ChatGPT itself became easier and better to use, Jasper’s defensibility started to weaken. The company later moved toward enterprise workflows and tried to build more around the product, but by then, the window to build a real moat had narrowed. Eventually, it pivoted again into an AI-first content creation tool through a browser extension form. With that new direction, Jasper raised a $125 million Series A round in 2026.</p><p>The lesson here is subtle, but important. Model commoditization can actually help incumbents more than startups.</p><p>If everyone has access to the same models at a lower cost, then value shifts upward. Distribution becomes more important. Workflow integration becomes more important. The companies that already sit inside daily user behavior gain the advantage. Microsoft’s decision to embed Copilot into Office 365 is a perfect example of this. It is not just a product decision. It is a distribution decision. Microsoft took a rapidly commoditizing layer of intelligence and turned it into a feature inside a product suite that hundreds of millions of people already use and pay for.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VpvReu0SmBoAZaGec0FiaA.png" /><figcaption>Where Value Lives (Increasing Bottom to Top)</figcaption></figure><p>So where does that leave startups? Where can they still build real value?</p><p>The first answer many founders reach for is fine-tuning and customization. That can matter, but even that is getting easier with time. A more interesting near-term moat lies in orchestration: routing queries well, chaining calls intelligently, managing memory and state, and building agentic systems that actually work in production. Teams that do this well can build a meaningful head start, perhaps six to twelve months ahead of the market.</p><p>But even that advantage may not last forever.</p><p>And that is what makes this trend so important. If orchestration also becomes more standardized over time, then the enduring moats start to look very familiar. Proprietary data. Strong feedback loops. Distribution. Deep workflow integration. In that sense, AI is not rewriting the laws of defensibility as much as it is bringing founders back to them. The tools are new, but the winning logic is old. The companies that endure will not just have impressive model outputs. They will own something around the model that gets stronger as the product gets used.</p><p>The next generation of AI winners will not be built by chasing raw model capability alone. Models will keep getting better, cheaper, and more interchangeable. What will matter is whether you can see the infrastructure shift early and build on top of it faster than everyone else.</p><p>That means knowing when to use reasoning and when not to. It means designing products around native modalities instead of forcing everything through text. It means rethinking retrieval in a world of expanding context windows. And above all, it means building where moats actually compound: proprietary data, strong feedback loops, distribution, and deep workflow integration.</p><p>Every major infrastructure shift resets the application layer. Some companies get unlocked by it. Others get abstracted away by it. The founders who win will be the ones who understand that AI infrastructure is not just the plumbing beneath the product. It is increasingly the thing that decides what kind of product can exist at all.</p><p>If you are an entrepreneur or investor building agents or copilots for enterprise or consumer use and want to brainstorm, reach out to me at <a href="/service/mailto:pdchhaparwal@gmail.com">pdchhaparwal@gmail.com</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7488187a2aeb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rethinking Product Strategy in the Age of AI]]></title>
            <link>https://pdchhaparwal.medium.com/rethinking-product-strategy-in-the-age-of-ai-1b33e8fc81d8?source=rss-c93454f16f39------2</link>
            <guid isPermaLink="false">https://medium.com/p/1b33e8fc81d8</guid>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cursor]]></category>
            <dc:creator><![CDATA[Parth Chhaparwal]]></dc:creator>
            <pubDate>Sat, 14 Mar 2026 01:20:58 GMT</pubDate>
            <atom:updated>2026-03-14T01:20:58.220Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How founders and executives should think about product form, workflow design, and the defensibility of AI products</em></p><p>One of the most important forces shaping AI applications has been product form. In 2023, when I was investing as a venture capitalist, one of my biggest reasons for passing on many AI application startups was that they felt like ‘wrappers’ around ChatGPT. I had picked up that term on Twitter, though in this piece I want to define it in a more thoughtful way.</p><p>That hesitation was not theoretical. It came from experience. I had already burned my fingers investing in startups building domain-specific AI image products for businesses on top of foundational image models like Stable Diffusion and DALL·E. Startups like these usually run into two core problems.</p><p>The first is that customers care deeply about output quality. In the example I saw, the requirement was hyper-realistic furniture images. In practice, customers wanted near-perfect realism, often closer to 99 to 100 percent quality than something merely good enough. That was the real problem. The product could only be as strong as the underlying image generation models powering it, and at that stage those models were simply not reliable enough. A founder building on top of any underlying foundational model has only limited control over improving the result. In many cases, the company’s success ends up being tied to capabilities it does not actually own.</p><p>The second problem is the exact opposite. Sometimes the underlying model improves so fast, and adds so many new features and product experiences, that the startup built on top of it starts to look unnecessary. That is what is happening across a large part of the AI startup landscape today. Quite a few VC-backed companies are potentially headed toward irrelevance. That, in my view, is where the real bubble sits.</p><p>But the smartest AI entrepreneurs saw this risk early. They realized that succeeding in AI would require going beyond traditional product management playbooks and rethinking how humans and AI interact. If applications built on top of foundation models could not fully control accuracy and efficiency, and those had already become table stakes anyway, then they needed to build defensibility somewhere else. That somewhere else was experience.</p><p>Reimagining experience in the age of AI came down to three critical choices.</p><ul><li>The first was about initiation. Who actually kicks off the AI? Is it the human? Is it a software system that triggers AI when context changes? Or is it the AI itself, acting on its own?</li><li>The second was about scope. What is the AI allowed to touch? Is it limited to text? Does it only have read-only access to outside sources? Or can it take editable action across the many tools an enterprise already uses?</li><li>The third was about how errors compound. In a one-step workflow, a mistake is usually visible, diagnosable, and fixable. But in a twelve-step process, an error at step two can quietly carry forward all the way to step twelve, creating fresh problems along the way.</li></ul><p>This third parameter is especially important. Solving for data, context, and execution is what creates a near-zero-error experience, and that is what actually delights customers. And those near-zero-error experiences do not happen by accident. They depend on how well the AI system is architected underneath, which is where real product defensibility begins to emerge.</p><p>Out of these choices came five core product archetypes in AI apps:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YJVXrAeYPjsKFYMF89H46g.jpeg" /><figcaption>5 Core Product Archetypes in AI Applications</figcaption></figure><p>The progression shown in the image above has shaped product strategy across the entire AI ecosystem, not just for startups, but also for incumbents and frontier labs like OpenAI and Anthropic. The first wave was simple and powerful. Companies won users through chat-based products built on top of their models. Once ChatGPT and Claude saw broad adoption as consumer-facing assistants, both companies opened up APIs. That unlocked a flood of startups building narrow applications on top of these foundation models. These were the products many people casually started calling ‘wrappers’.</p><p>But the bigger shift came when the model stopped being just something you talk to and started becoming something that could plug into the rest of the software world. Anthropic’s introduction of the Model Context Protocol mattered because it made it much easier for models to securely connect with external tools, applications, and data sources. That changed the role of the model itself. OpenAI and Anthropic were no longer just offering chat interfaces and APIs for text generation. They were moving toward assistants that could understand context and begin interacting with the tools people already use every day.</p><p>As model quality improved, this shift accelerated. Models became better at handling larger context windows, better at working across tools, and better at carrying memory across sessions and even outside the session itself. That mattered because it helped break one of the biggest limits of the copilot era. Copilots were useful, but they were usually trapped inside a specific product, workflow, or ecosystem. Agents changed that. With agents, the human sets the goal, but the system can decide how to get there, break the task into steps, move across tools, and execute with far more independence. That is where the current product innovation wave is really happening in 2025 and 2026.</p><p>What makes this evolution so important is that every shift in product form expands both capability and risk. Moving from API to chatbot, from assistant to copilot, and from copilot to agent increases what the product can do, but it also increases how many things can go wrong. The surface area for error gets wider at every step. And that is exactly why the winners in AI will not just be the ones that make these systems more powerful. They will be the ones that design them to be reliable from day one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DPzBWBpHZ5NhzkzutF-CDg.png" /><figcaption>Spectrum of AI products from API to Agents</figcaption></figure><p>The final phenomenon that meaningfully changed how organizations think about AI product strategy was Cursor.</p><p>Before Cursor, incumbents had already started responding to the rise of ChatGPT and Claude by layering AI into products they already owned. GitHub Copilot was one of the strongest early examples. Inside the VS Code ecosystem, it showed just how valuable AI could be for developers. But it also revealed something deeper. Adding AI into an existing product is not the same as rebuilding the product around AI. That gap created the opening Cursor went after.</p><p>There were three reasons this opening mattered.</p><ol><li>GitHub Copilot lived inside VS Code as an extension, which meant it had to operate within the boundaries of what the extension layer allowed. It could read files, suggest completions, and open a chat sidebar. That was powerful as a starting point, but it still left AI sitting at the edge of the developer experience rather than at the center of it. It did not fundamentally redesign how the editor behaved. It could not natively hold the entire repo in context, rethink core interactions like the tab key or command palette, or move fluidly across terminal, file tree, and multi-file rewriting as a deeply integrated system. That was the ceiling of building on top of someone else’s product surface.</li><li>Microsoft’s scale was an advantage in distribution, but a constraint in product speed. GitHub Copilot’s roadmap was only one input among many shaping the broader direction of Microsoft and the VS Code ecosystem. In a large company, product decisions move through more layers, more tradeoffs, and more internal alignment. And when the company is public, those decisions are judged even more carefully because they tie back to revenue, customer expectations, and platform stability.</li><li>Third, and perhaps most importantly, product choices at Microsoft had to remain aligned with Microsoft’s broader business model. GitHub Copilot could not evolve in ways that too quickly disrupted Azure, GitHub’s enterprise packaging, or other parts of the stack. Large companies almost always innovate with one eye on what must be protected. That creates space for startups. Startups do not have to defend the old system. They just have to build a better one.</li></ol><p>Cursor saw that opening clearly. It recognized that developers already loved VS Code because it was familiar, flexible, and deeply embedded in daily workflows. It also recognized something equally important. Coding is unusually well-suited to AI. It is structured, testable, and full of feedback loops. Outputs can be checked. Errors can be caught. Improvements can compound quickly.</p><p>What Cursor did was not just add AI to the editor. It rebuilt the editor around AI. By forking VS Code at the application level, Cursor gave itself the freedom to redesign the product much more deeply. It could rethink how key interactions worked across the editor, the command palette, the file tree, and the terminal. That architectural decision is what allowed Cursor to solve the earlier constraints in a way extensions could not.</p><ul><li>It gave AI awareness of the broader codebase without having to wait on Microsoft. That created room for richer agentic workflows that would have been much harder to deliver if the product had remained just another layer on top of the existing ecosystem.</li><li>It also benefited from organizational focus. Cursor was not one priority among a hundred. The company was the product. A lean, aligned team could move fast and ship features like multi-file editing, codebase indexing, and agent mode with far greater urgency.</li><li>And its incentives were simple. Cursor did not need to protect a larger platform strategy. It just needed to build something developers loved enough to pay for. That is a very different motivation from the one driving Microsoft and GitHub.</li></ul><p>Cursor’s success was not just limited to building a great product for developers. It showed that in the agent era, the winning move is often rebuilding the entire workflow around the agent. In that sense, Cursor introduced a new product form for agentic software. The agent was no longer just a helper sitting beside the product. The agent became deeply embedded in the product surface, the interface, and the way work actually got done. It gave the market a blueprint for a whole new category of products, which is why so many founders and investors started describing new companies as ‘Cursor for X.’</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Fve6Yk8f18Nlrmbxo7TfXg.png" /><figcaption>Opportunities across Cursor/Copilot for X</figcaption></figure><p>With this evolution, both incumbents and startups have become much more deliberate about product form in the age of AI. The real strategic question is no longer just whether AI should be added to a product. It is how that AI should show up. In practice, the choice between building a ‘Copilot for X’ and a ‘Cursor for X‘ usually comes down to two things: verifiability and switching costs.</p><p>That distinction matters. A ‘Cursor for X’ approach means rethinking the primary workspace of a professional from the ground up. A ‘Copilot for X’ approach means leaving the existing tool in place and layering AI into the workflow people already live inside.</p><p>A ‘Cursor for X’ model works best when output can be checked quickly and confidently. Code can be tested. Video renders can be reviewed. SQL queries either return the right result or they do not. In these environments, users can tolerate a new interface because the value shows up immediately and the feedback loop is tight. This becomes even more attractive when the incumbent product is familiar but not deeply locked in, and when its architecture or incentives make it poorly suited for AI-native work. In those cases, the old interaction model and the new one often cannot coexist for long. The better experience usually comes from rebuilding the product around AI rather than attaching AI to the side of it.</p><p>The ‘Copilot for X’ model becomes the default when one of those conditions breaks. Maybe the output is harder to verify because it is subjective. Maybe the tool is deeply embedded not because users love it, but because regulation, compliance, procurement, or company policy makes it very hard to replace. In those markets, people are not going to switch their core environment easily, no matter how elegant the new product is. So AI has to enter as a layer, not as a replacement. That can still create real value, but it is a different kind of business. It carries permanent platform risk, depends on someone else’s distribution, and usually does not produce the kind of explosive product-led adoption that Cursor-style products can.</p><p>For founders, the cleanest way to make this decision is to ask five simple questions.</p><ul><li>Is the output highly verifiable? If yes, that is the strongest argument for a Cursor-style approach.</li><li>Is the incumbent tool just a habit, or is it part of a deep workflow? If it is a weak habit, users can be moved. If it is a deep workflow, a copilot is often the more realistic path.</li><li>Will the platform owner eventually ship its own AI layer? If the answer is yes, that should immediately raise caution for anyone trying to build a lightweight extension business. In that case, a deeper product rethink is often the better bet. If the answer is no, then building a copilot can remain viable.</li><li>Does the user spend most of the day inside one tool? If yes, rebuilding that environment can create enormous leverage. If not, then a copilot layer that travels across workflows may be the better fit.</li><li>Can the model’s context hold most of the relevant project data? If yes, Cursor-style interaction becomes much more powerful. If not, the wiser move is to start with narrower, high-value use cases before attempting a full workflow replacement.</li></ul><p>Using that lens, some categories look especially well-suited for the ‘Cursor for X’ model. Legal work, music production, game development, video editing, software development, frontend development, education, and enterprise search all have strong reasons to move in that direction.</p><p>Other categories are more naturally suited for the copilot model. Content workflows, email, marketing operations, recruiting and HR, CRM and Salesforce optimization, spreadsheets, healthcare systems and EHRs, finance and trading, and CAD often sit inside environments where distribution is entrenched, and switching is harder.</p><p>And then there is a middle ground. Areas like DevOps and infrastructure, Figma-based design, scientific research, and customer support can move either way depending on the exact workflow, the strength of the incumbent, and how quickly users can trust and verify the output.</p><p>That, ultimately, is the deeper point. In AI, product strategy is no longer just about what the model can do. It is about where trust can be built, where behavior can change, and where a new interaction model is strong enough to replace the old one.</p><p>If you are an entrepreneur or investor building agents or copilots for enterprise or consumer use and want to brainstorm, reach out to me at <a href="/service/mailto:pdchhaparwal@gmail.com">pdchhaparwal@gmail.com</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1b33e8fc81d8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Centers Part 3: Ideas of the Future]]></title>
            <link>https://pdchhaparwal.medium.com/data-centers-part-3-ideas-of-the-future-bf16ca1c2a38?source=rss-c93454f16f39------2</link>
            <guid isPermaLink="false">https://medium.com/p/bf16ca1c2a38</guid>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[data-center]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[venture-capital]]></category>
            <dc:creator><![CDATA[Parth Chhaparwal]]></dc:creator>
            <pubDate>Sat, 27 Dec 2025 06:20:12 GMT</pubDate>
            <atom:updated>2025-12-27T08:01:38.505Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>TLDR:</em></strong><em> Building data centers themselves is scale capital; the venture opportunity is the ‘second economy’ around them: products that reduce uncertainty, compress timelines, improve uptime, and convert megawatts into revenue more efficiently. Data Center Ops is the emerging software + workflow + control-plane layer that becomes the operating system for the AI buildout.</em></p><p>AI is forcing the world to rebuild the physical internet. But the most misunderstood thing about data centers is that the core asset itself isn’t a venture business. A modern AI data center is a mega-project: land, power, substations, steel, cooling plants, commissioning, and long-lead equipment, financed with infrastructure capital and supported by long-term contracts. Even if a founder wanted to start from scratch, the project size and timelines make it hard to match the venture equation of speed, iteration, and capital efficiency.</p><p>That’s the good news: because when an industry moves from important to civilization-critical, a second economy forms around it: the tooling and operating system that makes the whole machine build faster, run better, and waste less. That’s where early-stage founders can build billion-dollar companies. We call this universe Data Center Ops.</p><p>Think of Data Center Ops as the set of products that turn a data center from a risky construction project into an industrial process: products that reduce uncertainty in power procurement, compress permitting and construction timelines, increase usable density, improve uptime, and, most importantly in the AI era, convert megawatts into revenue more efficiently.</p><p>A useful way to understand this market map is to follow the actual journey of a data center.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/958/1*xmAKgj-shpcjSUAJ8NRRFw.png" /></figure><blockquote><strong>The journey starts with one question: <em>Can I get power here?</em></strong></blockquote><p>If you cannot reliably energize hundreds of megawatts on a known timeline, everything else is academic. That’s why the first startup category that matters is a Power Feasibility Underwriter: software that doesn’t merely show grid maps, but produces a decision-grade answer, where power is likely to be delivered, by when, and at what upgrade cost and risk. In the AI era, this is not a nice-to-have. Developers can lose a year chasing the wrong site. A product that prevents even one failed bet can justify an enormous price tag, which is why this category is quietly one of the most venture-fundable in the entire stack.</p><p>The challenge is also why the opportunity exists: utility data is fragmented, timelines are uncertain, and the truth is often locked in relationships and local knowledge. The startups that win won’t be dashboards. They’ll be underwriting systems that learn from real projects and become the default language for power risk.</p><blockquote><strong>Once power is plausible, the question becomes: <em>Which exact site is buildable, and what’s the hidden risk?</em></strong></blockquote><p>That’s where Data Center Site Scoring Engines come in. These products take parcel-level inputs (power deliverability, fiber adjacency, zoning probability, flood/fire risk, water constraints, community pushback risk, incentives) and output a ranked shortlist that can actually survive diligence. In older eras, site selection was a real estate game. In the AI era, it’s a constraints game: the best site is not the cheapest land; it’s the land that can be energized and permitted at scale without surprises.</p><p>This category becomes compelling when it moves beyond research into repeatable workflow: a system that a developer uses every quarter to allocate billions in capex. If it stays ‘consulting with a UI,’ it will stall. If it becomes underwriting-grade, it becomes sticky.</p><blockquote><strong>Then comes the slowest, most political part of the process: <em>permitting.</em></strong></blockquote><p>The data center market is running into a new reality: communities and regulators increasingly treat data centers as infrastructure that competes for power and water. This is where a very specific startup archetype wins: a Permitting &amp; Entitlements Compiler built for data centers. The product is not project management. It’s a workflow that prevents rework: it knows jurisdiction requirements, assembles submission packages, runs pre-checks, routes tasks to the right consultants, and standardizes an otherwise bespoke process.</p><p>This is an underrated VC wedge because the ROI is binary: permitting delays don’t cost 3%, they can cost the entire year of capacity delivery. And in a market where demand is exploding, time is the most valuable resource. The whitespace here is large because most permitting software has been built for housing and generic construction, not the unique mix of grid coordination, environmental review, and community scrutiny that data centers face.</p><blockquote><strong>Even after approval, projects fail for a simpler reason: <em>execution.</em></strong></blockquote><p>AI data centers are mechanically complex: power trains, cooling trains, redundancy design, and now the added burden of density and liquid cooling readiness. This creates a large opportunity for a Commissioning &amp; Build Verification System, a product that becomes the system-of-record for whether a facility is actually ready, not just built.</p><p>Reality capture tools exist in construction, but data centers need more than images and progress reports. They need repeatable verification against design specs, automated punch lists, commissioning readiness tracking, and ongoing reliability workflows. The best products here don’t just document the jobsite. They reduce change orders, compress commissioning, and lower the probability of a catastrophic ‘it looked done but it wasn’t’ failure.</p><p>This market is competitive in general construction tech. The way a startup wins is by being unambiguously data-center-specific: commissioning truth, MEP verification, prefabrication coordination, and density readiness, problems generic tools don’t solve deeply.</p><blockquote><strong>After the building is up, the game shifts from <em>building to running.</em></strong></blockquote><p>This is where the AI era becomes fundamentally different. In legacy data centers, performance was mostly about uptime and efficiency. In AI data centers, performance is also about protecting compute output: thermals can throttle GPUs, cooling instability can cause failures, and small inefficiencies become massive costs because power consumption is so high.</p><p>That’s why one of the most attractive categories in Data Center Ops is the Autopilot for Facility Operations, closed-loop optimization software that doesn’t just report PUE or alarms, but actively tunes cooling, liquid loops, airflow, and power distribution based on telemetry and conditions. A dashboard tells you what happened. An autopilot prevents it from happening again.</p><p>This category is especially venture-fundable because the value proposition is crisp and compounding: lower energy cost, higher usable capacity, fewer incidents, better density. The hard part is trust: operators don’t like software touching controls. Winners will earn adoption through safe rollouts, proofs, and increasingly workload-aware control, where facility behavior adapts to what GPU clusters are doing in real time.</p><blockquote><strong>In parallel, the industry is realizing something uncomfortable:<em> the grid will not keep up easily.</em></strong></blockquote><p>This is where Frontier Firm Power enters the story, not as startups building nuclear, but as products that make new power partnerships and energy campuses feasible. The venture-shaped opportunity is often in enablement: software and workflows that help structure co-located power deals, model risk, accelerate siting, and integrate long-dated power sources into phased campus planning. These are not overnight businesses, but the need is real: more and more compute plans are being written around what power can be secured, not what customers want.</p><blockquote><strong>Now we reach the most visible part of the AI boom:<em> selling compute.</em></strong></blockquote><p>Many people assume the future of data centers equals new GPU clouds. But even here, the venture opportunities split into distinct product categories.</p><p>First, there are <strong>GPU Supply Exchanges</strong>, marketplace models that aggregate distributed GPU supply. The logic is obvious: there is fragmented capacity around the world and demand that can be flexible. But the failure mode is also obvious: enterprises don’t buy GPUs the way consumers book Airbnbs. They need SLAs, compliance, security, and predictability. So the marketplaces that can become durable businesses are the ones that build trust primitives: benchmarking, attestation, standardized contracts, scheduling reliability, and insurance-like SLA layers. Without that, marketplaces remain shallow and price-competitive.</p><p>Second, there are <strong>GPU Application Platforms (serverless GPU interfaces)</strong>. These are the ‘Vercel moment’ for compute: developers don’t want to manage clusters. They want to ship. The platform handles provisioning, autoscaling, routing, retries, observability, billing, and cost guardrails. This category can win quickly because developer experience is a distribution wedge. But durability comes from owning the control plane (portability, governance, and optimization) rather than being a thin wrapper around commodity GPUs.</p><p>Finally, there’s a category that looks orthogonal but is increasingly relevant: Train-Where-Data-Lives platforms <strong>(federated ML and distributed governance)</strong>. This is the reality that much of AI value will be created inside regulated data boundaries (healthcare, finance, defense) where data cannot be centralized freely. These products matter because they determine where compute is consumed and how models are governed. This tends to be a more enterprise-style business, but it is likely to become more important as AI spreads beyond tech.</p><h3>Implications for founders:</h3><p>The common pattern across the strongest Data Center Ops opportunities is simple: sell certainty and speed in a world where megawatts and time are scarce. The categories that tend to be most venture-fundable are the ones that:</p><p>· attach to a large, urgent budget (developers, operators, hyperscalers),</p><p>· have measurable ROI (time saved, capacity increased, energy reduced, outages avoided),</p><p>· can scale as software (not as bespoke projects)</p><p>Consequentially, the five ideas entrepreneurs would find tremendous opportunities in are:</p><p><em>1) </em><strong><em>Power Lifecycle OS:</em></strong><em> </em>A system that starts as power feasibility underwriting before you buy land, and then becomes the system-of-record for power after you energize: contracts, hedges, curtailment, PPAs, carbon compliance, load growth planning, and grid services. (<em>Salesforce for power</em>)</p><p><em>2) </em><strong><em>Data Center Permitting &amp; Entitlements OS:</em></strong> A workflow engine that turns permitting from a bespoke, consultant-driven process into a repeatable pipeline: jurisdiction requirement library, automated submission packets, constraint checks (power/water/noise/traffic), task routing across engineers/lawyers, community hearing prep, and real-time “time-to-approval” risk tracking. <em>(Asana + TurboTax for permits)</em></p><p><em>3) </em><strong><em>Commissioning &amp; Reliability System-of-Record:</em></strong> A product that follows the facility from build to go-live to steady-state, capturing design intent: test plans, punch lists, commissioning results, incident logs, so operators always know what’s actually ready, what changed, and what must be verified — especially for high-density power trains and liquid cooling. <em>(Jira + QA for data center readiness and uptime)</em></p><p><em>4) </em><strong><em>Facility Autopilot for Cooling &amp; Power Controls:</em></strong><em> </em>Closed-loop optimization software that doesn’t just monitor PUE, but actively tunes setpoints, airflow, liquid loops, and power distribution in real time based on telemetry, weather, and workload heat maps, raising usable kW/rack, lowering energy cost, and reducing thermal-driven failures. <em>(Autopilot for the plant room)</em></p><p><em>5) </em><strong><em>GPU Workload Control Plane (Serverless + Governance):</em></strong> A developer-facing platform where teams deploy AI workloads without managing clusters — handling provisioning, autoscaling, routing, retries, observability, and cost guardrails — while providing enterprise governance (security, compliance, quotas) and portability across multiple GPU supply sources. <em>(Vercel for GPUs, with “FinOps + Security” built in)</em></p><p>The next decade won’t just be about who builds the most data centers. It’ll be about who builds the operating system that makes data centers faster to deliver, easier to run, and more profitable per megawatt. That is where early-stage founders can still create outsized outcomes, even in an industry dominated by scale capital.</p><p>If you’re a founder building the next generation of data center software and would like to brainstorm / raise capital, reach out to me at <a href="/service/mailto:pdchhaparwal@gmail.com">pdchhaparwal@gmail.com</a>.</p><p><strong>Previous Reads:</strong></p><ul><li><a href="https://pdchhaparwal.medium.com/data-centers-part-1-primer-and-macros-29b6d4625368"><em>Data Centers Part 1: Primer and Macros</em></a></li><li><a href="https://pdchhaparwal.medium.com/data-centers-part-2-public-markets-two-models-22c7b9be5e1a"><em>Data Centers Part 2: Public Market’s Two Models</em></a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bf16ca1c2a38" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Centers Part 2: Public Market’s Two Models]]></title>
            <link>https://pdchhaparwal.medium.com/data-centers-part-2-public-markets-two-models-22c7b9be5e1a?source=rss-c93454f16f39------2</link>
            <guid isPermaLink="false">https://medium.com/p/22c7b9be5e1a</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-center]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[stock-market]]></category>
            <dc:creator><![CDATA[Parth Chhaparwal]]></dc:creator>
            <pubDate>Fri, 26 Dec 2025 05:25:10 GMT</pubDate>
            <atom:updated>2025-12-27T07:56:57.925Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>TLDR:</em></strong><em> Public markets separate core data centers (cash-flow compounding platforms) from new-age AI factories (speed + execution + contract-backed scale stories). The lesson for private capital is: returns often come from owning bottlenecks and project economics, not just buying the operator, and the best structures frequently look like JVs, sale-leasebacks, or structured capital rather than classic high-leverage LBOs.</em></p><p>Public markets have started treating data centers as two different industries that happen to share the same physical reality (land, power, cooling). One category is built to compound cash steadily; the other is built to scale AI compute aggressively. The confusing part is that both can report growth, but the market rewards them for very different reasons, and punishes them for different mistakes. The two types of data center stocks are core and new-age.</p><h3><strong>Core / legacy data center companies (the infrastructure REIT model)</strong></h3><p>Companies like Equinix, Digital Realty, Iron Mountain, American Tower (via CoreSite), and COPT Defense Properties make money in a way that resembles owning high-quality infrastructure. A simple way to explain it: they build or own facilities that offer (a) power, (b) uptime, © security, (d) connectivity, and rent that capacity to customers on multi-year contracts.</p><p>Many of these businesses are structured like REITs. A REIT (Real Estate Investment Trust) is basically a company that owns income-producing real estate and is designed to pay out a meaningful portion of its cash flow to shareholders (often via dividends). That structure is why investors in core data center stocks care a lot about cash generation and per-share cash growth, not just accounting profits.</p><p>There are a plethora of reasons why some ‘steady growers’ get rewarded more than others:</p><blockquote><strong>The core business is not “growth.” It’s per-share cash compounding.</strong></blockquote><p>Take Equinix vs Digital Realty. Both can look like healthy growers. But the market differentiates them based on one simple question: <em>If the business is growing 10%, how much of that growth actually shows up for each shareholder?</em></p><p>Equinix’s cash generation rises in a way that shows up meaningfully per share (Adjusted Funds From Operations / AFFO per share growth is high single digits). Digital Realty’s total cash can grow, but per-share growth is much lower. The intuitive implication is:</p><p>· If a company grows cash flow but issues a lot of new shares to fund expansion, the pie gets bigger, but each slice doesn’t.</p><p>· That doesn’t mean the business is bad. It means the stock can underperform, because shareholders don’t fully capture the growth they’re paying for.</p><p>This one comparison explains a lot of the valuation gap investors are willing to tolerate. Equinix can sustain a premium because it has a clearer record of turning industry tailwinds into shareholder compounding. Digital Realty can look like it’s doing the same thing, but the value capture can be diluted away.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_yiZEkjHCCiZBVByjCSOww.png" /><figcaption>Comparing legacy data center companies</figcaption></figure><blockquote><strong>Margins matter, but they’re not the punchline.</strong></blockquote><p>Look at American Tower vs Iron Mountain. American Tower has extremely high EBITDA margins (it’s a very efficient cash machine), while Iron Mountain’s margin profile is lower. Yet Iron Mountain can still be compelling when the market believes its incremental data center capacity is expanding its growth runway without breaking the balance sheet.</p><p>The story here isn’t higher margin = better stock. It’s:</p><p>· High margins tell you the engine is strong.</p><p>· But the stock outcome depends on whether the company can reinvest at attractive returns and still grow per share.</p><p>That’s why the market cares deeply about how cash flows are allocated: reinvestment vs dividends vs dilution.</p><blockquote><strong>Core winners have moats that improve pricing power.</strong></blockquote><p>A big reason Equinix is treated differently is not just cash flow. It’s the nature of demand. Equinix is often closer to a <em>marketplace of connections</em> (read interconnections) than a generic landlord: once companies connect to the ecosystem, switching can be painful, and expansions become natural.</p><p>Compare that with models that are more space-and-power commodity. Those can still do well, but pricing power and customer stickiness are usually weaker, and the market discounts them accordingly.</p><blockquote><strong>The failure modes are boring, and that’s why they’re dangerous.</strong></blockquote><p>Core data center businesses rarely “blow up” overnight. They underperform when one of these slow leaks shows up:</p><p>· Cost of capital pressure: if rates rise and refinancing gets expensive, development returns compress.</p><p>· Power bottlenecks: the company may have demand, but can’t bring capacity online fast enough.</p><p>· Tenant mix bargaining power: relying heavily on a few giant customers can quietly compress economics over time.</p><p>This is why the market’s core obsession becomes rational: it wants companies that can keep delivering growth without needing aggressive equity issuance and without letting financing costs swallow the upside.</p><p><strong>In plain English: </strong>core data center stocks don’t win by building the most. They win by building profitably, financing prudently, and letting shareholders actually keep the upside.</p><h3>New-age data center companies (the NeoCloud model)</h3><p>CoreWeave, Nebius, IREN, Bitdeer sit in the new-age bucket because they are not just leasing powered space. They are turning infrastructure into a compute product (GPU-as-a-Service / AI cloud). If core data centers are landlords, new-age players are factories. Their factory output is GPU-hours. And like any factory, the market cares about three things:</p><p>· How big is the factory floor?</p><p>· How fast can it ramp to production?</p><p>· Are orders locked in?</p><p>The impact of these questions are understood when we discuss the following:</p><blockquote><strong>Power is the raw material, but only energized power is usable.</strong></blockquote><p>Compare Bitdeer vs Nebius on power: both can talk about huge secured power footprints, but the market doesn’t pay for “secured” in the same way it pays for ‘energized.’ Bitdeer has far more energized MW than Nebius. Yet it trades at dramatically lower value per energized MW. Why would that happen? Because the market isn’t just buying “power online.” It’s buying power that can be converted into premium AI revenue. Which leads to the next comparison.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BSM4h23lsIPxKORDuPWOeQ.png" /><figcaption>Comparing new-age data center companies</figcaption></figure><blockquote><strong>Revenue efficiency tells you whether MW is being monetized as ‘AI’ or as something cheaper.</strong></blockquote><p>Compare CoreWeave vs IREN. CoreWeave generates far more revenue per MW than IREN. That single fact implies a lot:</p><p>· CoreWeave’s capacity is being sold as higher-value AI compute (stronger monetization).</p><p>· IREN’s capacity may be earlier in the ramp, sold at a different mix, or priced with different customer/contract structure.</p><p>This is exactly why two companies can both claim ‘AI data center exposure,’ but one is treated like a premium AI platform and the other like an undervalued capacity owner. The market is basically asking: Is this MW acting like an AI factory, or like generic powered infrastructure?</p><blockquote><strong>Contracted demand is what turns expansion into an investable story.</strong></blockquote><p>Here’s the most intuitive comparison: CoreWeave vs Bitdeer. CoreWeave has enormous contracted demand visibility (backlog/RPO), while Bitdeer has far less demand visibility in the same ‘contracted order book’ sense. That changes how investors interpret growth:</p><p>· With CoreWeave, capacity expansion looks like ‘building to fulfill orders.’</p><p>· With Bitdeer, expansion can look more like ‘building and then finding demand,’ or demand that is structurally more volatile.</p><p>This is why the market often rewards companies that announce large, credible, long-duration contracts. It’s not about PR. It’s about de-risking the factory.</p><blockquote><strong>Concentration can be both strength and fragility.</strong></blockquote><p>CoreWeave’s story also comes with a sharp edge: heavy customer concentration. The market can love it because it signals blue-chip demand and fast scaling, but it can also fear it because one relationship holds enormous pricing power. That’s why these stocks can rip on contract wins and still sell off violently on any hint of renegotiation, pricing pressure, or ramp delays.</p><blockquote><strong>Profitability is interpreted differently in this category, but it still matters.</strong></blockquote><p>Nebius can be growing incredibly fast and still lose money. The market tolerates that only if it believes:</p><p>· the ramp is real,</p><p>· unit economics will improve with scale,</p><p>· and funding won’t become a value-transfer event (dilution/expensive debt).</p><p>Contrast that with CoreWeave, which already shows very high adjusted profitability. That tends to make the ‘AI factory’ story easier to believe, and is a big reason the market can value it aggressively even with leverage concerns.</p><p><strong>In plain English:</strong> new-age data center stocks win when MW becomes revenue quickly, revenue becomes profit credibly, and contracts make the whole machine feel inevitable.</p><h3>CY25 — year in review for both categories</h3><p>This year created a clean divergence:</p><p>· Core names behaved like infrastructure compounders. They moved, but they didn’t behave like rockets. Investors treated them as a mix of growth + yield + rate sensitivity.</p><p>· New-age names behaved like AI beta with execution risk. They were repriced sharply upward on demand validation and repriced downward on any hint of delays, funding stress, or concentration risk.</p><p>That’s why the biggest winners and losers weren’t decided by who grew revenue most. They were decided by who made the market believe the model is scalable, fundable, and contract-backed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*doOXfBVpS4LtEbr1T4swLw.png" /><figcaption>Normalized returns on data center stocks</figcaption></figure><blockquote><strong>The core learning is that across both categories, major moves consistently come from a few triggers:</strong></blockquote><p>· Demand validation events: When a new-age player lands a major hyperscaler relationship or contract, it changes perception instantly: the factory has real orders, not just ambition. That’s why bullish research (like Buy calls) often clusters around major customer validation. It’s a credibility upgrade.</p><p>· Execution reality checks: A single commissioning delay matters more in new-age than in core, because it directly shifts time-to-revenue. These names are priced on ramps. Break the ramp, break the valuation.</p><p>· Financing signals (the silent killer): For core companies, financing shows up as ‘does per-share cash keep compounding?’ For new-age companies, financing shows up as ‘can the ramp be funded without transferring value away from equity holders?’</p><p>This is why the same headline growth rate can produce opposite stock reactions depending on whether the market thinks growth is self-funded or dilution-funded.</p><p>Public markets now treat data centers as two different games:<strong> </strong>cash compounding (core) versus AI factory scaling (new-age). In both, the winners are the ones that turn capacity into durable economics, but in the AI factory model, markets will keep repricing belief every time contracts, ramp speed, or funding reality changes.</p><h3>Implications for PE / Growth Equity investors</h3><p>Public markets may be debating core vs new-age, but for PE the question is more pragmatic: what kind of underwriting and deal structure can actually work in a business where capex is massive and power is the bottleneck? The answer is different across the two categories, and it’s why some strategies fit naturally while others are structurally hard.</p><blockquote><strong>1) Traditional LBOs are structurally challenging, unless the asset looks like infrastructure cash flow, not buildout optionality</strong></blockquote><p>A classic LBO works when three things are true: predictable cash flows, limited ongoing capex needs, and an ability to de-lever steadily. <strong>Data centers often violate at least two:</strong></p><p>· Capex is not discretionary: even maintenance is meaningful (power systems, cooling, refresh cycles), and growth capex can be lumpy and huge.</p><p>· Cash flows can be less predictable than they look: churn can be low, but contract renewals, pricing resets, tenant mix shifts (hyperscaler vs retail), and timing of energization can create volatility, especially during rapid build phases.</p><p>That said, PE can underwrite buyouts when the target resembles a stabilized infrastructure platform: high occupancy/utilization, strong renewal track record, long weighted average lease duration, and visible contracted cash flows.</p><p>The playbook that tends to work: treat it less like a tech LBO and more like an infrastructure-style buyout: moderate leverage, strict covenants around capex, and upside coming from optimization and selective expansion rather than aggressive debt paydown assumptions.</p><blockquote><strong>2) Take-privates: doable, but only when the valuation dislocation is clear and the capex plan is controllable</strong></blockquote><p>Take-privates in this sector are hard for the same reason: once private, the sponsor still has to fund expansion, and public markets often already value quality core cash flow at a premium. Still, <strong>they can make sense when:</strong></p><p>· Rates and macro temporarily compress multiples, even though long-term demand is intact (classic cyclic dislocation).</p><p>· The company’s per-share value capture is broken (e.g., persistent dilution / mispriced capital allocation) and a sponsor believes private ownership can reset the funding strategy.</p><p>· The asset has a portfolio of under-optimized campuses where operational fixes, repricing, and capex discipline can unlock meaningful value.</p><p>In practice, many data center take-private-like outcomes happen through carve-outs, JVs, or platform privatizations rather than a clean 100% buyout, because large checks + ongoing capex needs make full take-privates tougher to finance.</p><blockquote><strong>3) The most natural PE strategy is not always buying the operator; it’s owning the project economics</strong></blockquote><p><strong>For both categories, the most durable returns often come from owning the hard bottlenecks and getting paid for solving them:</strong></p><p>· Powered land / substation rights / interconnection queue position</p><p>· Energization-ready capacity</p><p>· Long-term power procurement / PPAs and the ability to derisk energy cost</p><p>This is why infrastructure and PE investors often prefer JVs at the campus level, build-to-suit / forward purchase agreements, sale-leasebacks or preferred equity / structured capital against a specific pipeline of builds. These structures let investors get exposure to AI-driven demand without underwriting the entire corporate story.</p><blockquote><strong>4) Growth equity fits new-age AI factories better than buyout, but only with contract-backed downside protection</strong></blockquote><p>New-age neoclouds can grow explosively, but their risk is not product-market fit. It’s execution + financing + concentration.<strong> For growth equity, the best opportunities tend to be where there is:</strong></p><p>· Credible contracted demand (backlog / RPO that has real teeth: take-or-pay, prepayments, strong termination economics)</p><p>· A clear plan to turn secured MW into energized MW on schedule</p><p>· A path to unit economics improving with scale (revenue per MW and gross profit per MW trending up)</p><blockquote><strong>5) What PE should learn from the public market divide: value capture matters more than growth</strong></blockquote><p><strong>Public markets are already teaching the underwriting lesson PE should internalize:</strong></p><p>· In core data centers, the good version of growth is per-share cash compounding (value capture without dilution).</p><p>· In new-age data centers, the good version of growth is contract-backed ramp speed (energized MW to GPUs to revenue with visible demand).</p><p><strong>So for PE, diligence should center on:</strong></p><p>· Who captures the economics (customer pricing power, renewals, concentration, power pass-through),</p><p>· What constrains scaling (grid timelines, supply chain, cooling design readiness),</p><p>· and how capital gets recycled (whether growth requires perpetual external funding or becomes increasingly self-funding).</p><p>For core platforms, the attractive private-market angle is ‘infrastructure compounding + disciplined expansion’, but classic high-leverage LBO math is often constrained by capex and refinancing cycles. For new-age AI factories, the attractive angle is ‘contract-backed capacity ramp’, but it is usually better suited to growth equity and structured capital than pure buyouts, because execution speed, customer concentration, and power economics dominate outcomes.</p><p>We’ll next discuss the opportunities for startups and new entrepeneurs around data centers. Public markets clarify what investors reward, but the most interesting implication is what they don’t reward: the operational chaos underneath the buildout. That gap is the opening for founders.</p><p><a href="https://pdchhaparwal.medium.com/data-centers-part-3-ideas-of-the-future-bf16ca1c2a38"><strong><em>Part 3 zooms into Data Center Ops: the tools and systems that turn a data center from a bespoke mega-project into something closer to an industrial process, faster to deliver, easier to run, and more profitable per megawatt.</em></strong></a></p><p><strong>Other Reads:</strong> <a href="https://pdchhaparwal.medium.com/data-centers-part-1-primer-and-macros-29b6d4625368"><strong><em>Data Centers Part 1: Primer and Macros</em></strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=22c7b9be5e1a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Centers Part 1: Primer and Macros]]></title>
            <link>https://pdchhaparwal.medium.com/data-centers-part-1-primer-and-macros-29b6d4625368?source=rss-c93454f16f39------2</link>
            <guid isPermaLink="false">https://medium.com/p/29b6d4625368</guid>
            <category><![CDATA[data-center]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Parth Chhaparwal]]></dc:creator>
            <pubDate>Fri, 26 Dec 2025 05:12:46 GMT</pubDate>
            <atom:updated>2025-12-27T07:54:01.103Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>TLDR:</em></strong><em> AI is converting data centers from background infrastructure into a front-page constraint: power deliverability, cooling readiness, and build timelines now shape what gets deployed and at what economics. The key takeaway is the stack is becoming an integrated system, and the bottlenecks are increasingly upstream (power) and thermal.</em></p><p>AI’s biggest breakthroughs don’t just happen in model papers. They show up in whether the real world can actually scale up capacity fast enough, reliably enough, and at the right economics. That’s why data centers have moved from background infrastructure to a front-page constraint that shapes how AI gets built, priced, and deployed.</p><p>In this edition, we go deep into three things:</p><blockquote><a href="https://pdchhaparwal.medium.com/data-centers-part-1-primer-and-macros-29b6d4625368">(1) <strong>Macro + primer:</strong> a shared foundation and vocabulary for how to think about the data center ecosystem end-to-end.</a></blockquote><blockquote><a href="https://pdchhaparwal.medium.com/data-centers-part-2-public-markets-two-models-22c7b9be5e1a">(2) <strong>Public markets:</strong> how listed players map to the stack and what their positioning reveals about where value is accruing.</a></blockquote><blockquote><a href="https://pdchhaparwal.medium.com/data-centers-part-3-ideas-of-the-future-bf16ca1c2a38">(3) <strong>Venture markets:</strong> where most of the buildout is happening and what new categories are forming around the next wave of demand.</a></blockquote><p>Data centers may be a 40-year-old concept, but they’ve been central to every major wave of tech adoption, through four key eras: (1) the rise of the internet, (2) the post–dot-com reset, (3) the shift to cloud computing, and now (4) the rise of AI.</p><p>As the internet took off in the early 1990s, spending on IT infrastructure surged. Companies started building large facilities to house servers and store data as they realized how transformative the internet could be. But when the dot-com bubble burst, the data center industry was hit hard, with 17 of 27 European data center players reportedly went bankrupt.</p><p>Out of that reset, one defining trend reshaped the industry: virtualization. Virtualization adds a software layer (<strong>virtual machines</strong>) on top of physical servers, storage, and networking, allowing a single physical setup to be split into many smaller, consumable virtual environments (<strong>software-defined data centers</strong>) that can be managed and accessed remotely. The impact was dramatic because multiple virtual workloads could run on the same physical server, and companies could sharply reduce their requirements for power, space, and cooling.</p><p>The next major shift came with cloud computing, which moved data centers from being primarily <strong>on-premise</strong> to increasingly colocated. <strong>Colocation</strong> means renting space and power for your IT infrastructure inside a third-party data center, often alongside access to rich connectivity and interconnection ecosystems.</p><p>With AI adoption accelerating, businesses are generating massive volumes of data, pushing the industry into another reset. Data centers are evolving into <strong>high-performance facilities</strong> built for ultra-dense deployments and very high power delivery, designed to run the most demanding compute workloads.</p><h3><strong>Advances in data centers have led to the creation of 4 ownership models for data center capacity:</strong></h3><blockquote>(i) <strong>Enterprise (self-owned/on-premise)</strong>, where data centers are owned/operated by non-hyperscale companies for their own internal IT, often located on their campuses or in dedicated facilities.</blockquote><blockquote>(ii) <strong>Colocation for non-hyperscalers</strong>, where corporates, SaaS companies, mid-sized cloud providers, and networks use capacity provided by third-party data center providers.</blockquote><blockquote>(iii) <strong>Colocation for hyperscalers</strong>, where hyperscaler cloud-providers like AWS, Google, Azure, Meta, Oracle use capacity provided by third-party data center providers like Equinix, Digital Realty, CyrusOne, QTS, Vantage and STT GDC.</blockquote><blockquote>(iv) <strong>Hyperscaler self-build</strong>, where data centers are owned and operated by hyperscalers themselves.</blockquote><h3>There are four core challenges with on-prem data centers that are pushing businesses experimenting with AI to adopt high-performance data centers:</h3><blockquote>(i) <strong>Performance and scalability</strong>, where massive data generation across the world is breaching physical capacity and increasing demand for global presence of data centers.</blockquote><blockquote>(ii) <strong>Data strategy</strong>, where rapidly increasing penetration of hybrid, multi-cloud environments demands data being stored near the customers, closer to the edge, and compliant with GDPR and PIPE regulations.</blockquote><blockquote>(iii) <strong>Cost</strong>, where constant hardware + software updates require employing tech-savvy network personnel. Vendor-lock in and rising energy costs lead to additional costs.</blockquote><blockquote>(iv) <strong>Sustainability</strong>, where older, inefficient data of on-prem data centers lead to high carbon impact.</blockquote><h3><strong>High-performance data centers (HPDCs)</strong> solve these challenges by providing space, power, cooling, sustainable investment and expertise all in one place, ensuring future scalability.</h3><p>HPDCs typically have strong partnerships with AI providers like Nvidia and HPE, and they are globally distributed, sitting closer to customers and reducing latency (time to output). A lot of HPDCs also experiment with renewable energy sources like fuel cells, hydrotreated vegetable oils (HVOs), and nuclear energy.</p><p>In the age of AI, there are several steps a data center must take to ensure high-performance. Firstly, the data center must provide appropriate ‘sellable power’ which is the standard metric to measure capacity that it can handle. The industry accepted metrics for sellable power include <strong>critical power (MW),</strong> which is the amount of IT power that can be delivered to customer’s equipment at full deployment. Other critical parameters include <strong>commissioned IT power</strong> (MW), which is the total amount that can be provided <em>today</em>. <strong>Leased MW</strong> is the portion of commissioned IT power that is in use and available MW is the proportion currently available and not in use. Second order performance indicators include <strong>power utilization %</strong> (= leased / commissioned) and kW per rack or kW per square feet.</p><p>Secondly, the data center should be ‘reliable’ with minimal downtime. The performance indicator measured is <strong>availability uptime</strong> (= (total time — downtime)/total time) and incident metrics including <strong>mean time between failure (MTBR</strong>) which is total time period in days/months/years between two failures, and <strong>mean time to repair (MTTR)</strong> which is the total time taken to repair per each incident.</p><p>Finally, the data center should be resource efficient, measured using metrics like <strong>power usage effectiveness</strong> (PUE = total facility energy / IT equipment energy), <strong>water usage effectiveness </strong>(WUE = annual water used in liters / IT equipment energy) and <strong>carbon usage effectiveness</strong> (CUE = total CO2 emissions in kgCO2e / IT equipment energy)</p><p>As new-age data centers get themselves AI-ready, the global data center market is expected to reach 163 GW in 2030 from 67 GW in 2024.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gXmx26U1iySAgPqISBV3mQ.png" /><figcaption>Global Data Center — Split by delivery model</figcaption></figure><p>The growth is mostly driven by hyperscalers, which are consumers of data centers having massive capital expenditure, and an ability to absorb compute and scale up AI processes. By 2030, 45%+ of data center capacity would be in use for AI training and inference, which occupied &lt;2% in 2020.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*df8_7p_YMkS4xyQq7G_PBg.png" /><figcaption>Global Data Center Capacity — Split by use case</figcaption></figure><p>While data center owners often capture the most value, especially from AI, there are many value creators in the data center space, including developers of DCs, thermal &amp; cooling providers, and computer/server providers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1014/1*K1kFTmrt9mIvlxsFMtjcEw.png" /></figure><p>In the AI era, data centers are no longer just buildings. They’re integrated power + thermal + network + compute systems. The biggest bottlenecks (and upside) sit upstream in power/cooling and downstream in high-density compute, while the biggest white space is a unified ‘data center OS’ that co-optimizes workloads with power, thermals, and energy markets.</p><h3>In the next section, we will discuss the data center market from the perspective of public market and private equity investors.</h3><p>Part 1 built the shared vocabulary: data centers aren’t real estate with servers, they’re power + cooling + network + compute systems where timelines and constraints matter as much as square footage. The natural next question is: who captures the value when the world is forced to build more of this infrastructure?</p><p><a href="https://pdchhaparwal.medium.com/data-centers-part-2-public-markets-two-models-22c7b9be5e1a">Part 2 moves from the physical stack to the financial scoreboard: how public markets price different data center business models, and what the winners (and disappointments) reveal about durable advantage in the AI buildout.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=29b6d4625368" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Manufacturing in the Age of Intelligence]]></title>
            <link>https://pdchhaparwal.medium.com/manufacturing-in-the-age-of-intelligence-17ba70da0497?source=rss-c93454f16f39------2</link>
            <guid isPermaLink="false">https://medium.com/p/17ba70da0497</guid>
            <category><![CDATA[manufacturing]]></category>
            <category><![CDATA[private-equity]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[venture-capital]]></category>
            <dc:creator><![CDATA[Parth Chhaparwal]]></dc:creator>
            <pubDate>Mon, 22 Sep 2025 16:35:46 GMT</pubDate>
            <atom:updated>2025-09-22T16:47:01.183Z</atom:updated>
            <content:encoded><![CDATA[<p>The factory floor is no longer defined by steel and smoke. It’s becoming defined by data and intelligence. For decades, manufacturing has been understood as a game of physical scale (the size of plants, the efficiency of supply chains, and the weight of assets on balance sheets). Financial performance reflected this logic: steady but slow growth, margins bound by input costs, and valuations tied to capacity and cycles. It was a model that prized predictability over reinvention.</p><p>Today, that model is being tested. AI is entering the factory floor, the supply chain, and even the products themselves. What was once a world defined by steel and machinery is now being reshaped by data, algorithms, and intelligence. This shift is beyond operational: it’s financial. In the sections that follow, we trace how AI is beginning to change the way manufacturing companies perform, how markets value them, and why the very definition of “what a manufacturer is” may no longer look the same.</p><h3>AI Makes Manufacturers Look More Like Tech Companies</h3><p>The financials of traditional manufacturers tell a story of limits. Revenues move slowly, tied to construction cycles, consumer demand, or commodity prices. Margins stay thin because materials and labor dominate costs. Balance sheets are weighed down by plants and inventories, and cash is treated as something to be handed back, not reinvested. Growth is steady, but it rarely escapes the gravity of the broader economy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TOOcjP7o0vSSvgVQw0Lx3A.png" /></figure><p>AI-enabled manufacturers look like they’re playing an entirely different game. Their revenue curves rise faster because embedding intelligence into products creates new markets, not just new models. Tesla’s 50% revenue growth and Nvidia’s 60% surge in 2022 weren’t driven by economic tailwinds; they were pulled by demand for autonomy and AI infrastructure. The revenue lines tilt sharply upward for AI-driven firms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jFmoBWyl5A21-Rbp8sJxSg.png" /></figure><p>Margins are where the contrast sharpens. Traditional players grind in the 25–40% gross margin range: International Paper at ~28%, Whirlpool around 30%, Caterpillar in the low 30s. Those numbers reflect businesses competing on cost. AI-enabled firms break out of that box. Tesla crossed 25% gross margins in 2022, Deere climbed into the mid-30s by selling AI-equipped tractors, and Nvidia operates at more than 70%. When products compete on intelligence, not just weight and steel, customers reward it with higher prices and stickier demand.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RKNmFCe-0U6tmflprodxHw.png" /></figure><p>The same pattern shows up in spending. Old-world manufacturers treat R&amp;D as upkeep, rarely more than 2–4% of sales. On the AI side, R&amp;D becomes the engine of growth (10–20% of revenue at places like Siemens’ digital arm or Nvidia). That investment explains why operating and net margins expand rather than collapse under pressure. Rockwell Automation’s AI-driven software segment, for instance, delivers close to 30% margins, a level unreachable for traditional hardware alone.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cdcpnSdFvm670qVbbTCZbA.png" /></figure><p>Balance sheets and cash flows round out the divergence. Traditional firms load up on tangible assets and tie up working capital in warehouses, while AI-enabled firms hold more intangibles (patents, data, algorithms) that scale without weight. Cash in the old model flows back as dividends and buybacks; in the new model, it’s recycled into gigafactories, chip design, or software platforms. Even capex looks smarter: traditional firms spend 5–10% of sales just to maintain plants, while AI-enabled ones stretch every dollar with predictive maintenance, or in Nvidia’s case, bypass plants altogether by staying fabless.</p><p>Put together, the charts tell a single story: AI isn’t just tweaking line items, it’s rewiring the financial model of manufacturing. Revenues grow like tech, margins expand beyond historic ceilings, balance sheets tilt toward intellectual capital, and cash compounds instead of plateauing. The examples (Tesla’s revenue surge, Deere’s premium margins, Nvidia’s 70% gross margins) are proof points, but the bigger story is structural. With AI, manufacturers stop looking like factories bound to cycles and start looking like technology companies that happen to make physical things.</p><h3>AI Turns Manufacturers Into Tech Companies, and Markets Value Them That Way</h3><p>If financial statements are being rewritten, it’s no surprise markets are rewriting valuations too. When fundamentals shift this dramatically, markets notice. Traditional manufacturers, with their slow revenue growth, capped margins, and asset-heavy balance sheets, are valued accordingly. They trade at modest multiples, often 6–12x EBITDA or 10–15x earnings, because investors see earnings as cyclical and capped. But once AI reshapes the financials (steep revenue curves, gross margins pushing 40–70%, reinvestment-driven cash flows), the market stops pricing these firms like factories and starts treating them like tech companies. That’s why Tesla trades at 40–60x forward earnings while legacy automakers sit at 7–10x, and why Nvidia, without owning fabs, commands EV/EBITDA multiples north of 30x while traditional chipmakers linger closer to 12–15x. The premium isn’t just hype; it’s recognition that the financial DNA, and therefore the risk-reward profile, has changed.</p><p>This re-rating of multiples has a direct effect on enterprise value. In the traditional model, EV is anchored by tangible capacity — plants, machines, and the ability to churn out volume. In the AI model, EV is driven by intangibles: patents, proprietary datasets, algorithms, and software ecosystems that scale without weight. That’s how Nvidia crossed the trillion-dollar mark while outsourcing manufacturing, and how Tesla’s EV at times exceeded the combined value of multiple legacy automakers. The logic is simple: if revenues grow faster, margins expand, and cash compounds rather than plateaus, investors will pay far more for every dollar of earnings.</p><p>The same logic is spilling into private markets. Traditional industrial businesses still get benchmarked against public comps at single-digit EBITDA multiples. But manufacturing startups with credible AI layers are increasingly valued on tech multiples, sometimes 15–30x revenue at growth stage. The bifurcation is widening: firms without a tech angle are being treated like commodity players, while those with AI-driven differentiation are being rewarded as software-like businesses, even if they make physical goods.</p><p>For founders, this shift carries a clear recommendation. To escape industrial multiples and earn tech multiples, the AI story must be more than a buzzword. It has to show up in the numbers: higher gross margins through automation, new revenue streams through AI services, resilient operating margins through predictive efficiency. Investors reward proof of AI leverage, not just promises. That means founders should double down on building intangible moats (IP, data, proprietary algorithms) and show how cash reinvests into compounding growth rather than distributions. The valuation upside is enormous, but only if the fundamentals and narrative align.</p><p>The transformation of manufacturing through AI is more than an operational upgrade. It is a financial re-rating that reshapes how companies grow, how investors value them, and how founders must position themselves. If you are a founder building in the Manufacturing AI space and would like any kind of help, feel free to reach out to me at <a href="/service/mailto:pdchhaparwal@gmail.com">pdchhaparwal@gmail.com</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=17ba70da0497" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Agents Are the New Execution Layer for Enterprise IT]]></title>
            <link>https://pdchhaparwal.medium.com/ai-agents-are-the-new-execution-layer-for-enterprise-it-33e87b785127?source=rss-c93454f16f39------2</link>
            <guid isPermaLink="false">https://medium.com/p/33e87b785127</guid>
            <category><![CDATA[information-technology]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[disruption]]></category>
            <category><![CDATA[saas]]></category>
            <dc:creator><![CDATA[Parth Chhaparwal]]></dc:creator>
            <pubDate>Wed, 20 Aug 2025 02:49:35 GMT</pubDate>
            <atom:updated>2025-08-20T02:49:35.018Z</atom:updated>
            <content:encoded><![CDATA[<p>For decades, enterprise IT has been a patchwork of three things: legacy systems, outsourced operations, and SaaS tools. These setups have kept organizations running, but they’ve also been inefficient, costly, and heavily dependent on human labor.</p><p>Now, with the rise of AI agents (autonomous, context-aware software entities that can observe, reason, and act), we’re at an inflection point. AI agents are poised to rethink how enterprises execute IT workflows.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*W5tyrE1KuGVDVRAU" /></figure><h3>AI Agents are redefining how enterprises approach IT</h3><p>AI agents create a new execution layer in enterprise software. Traditionally, a business either:</p><ol><li>Built or bought custom software, or</li><li>Relied on human services, either in-house or outsourced.</li></ol><p>Agents now provide a third path. They can autonomously perform repetitive, rule-driven tasks, making IT cheaper and more standardized.</p><p>This opens up two types of opportunities:</p><ul><li><strong>AI-Enhanced ITeS:</strong> Service firms that weave agents into delivery, offering faster, cheaper, more standardized execution.</li><li><strong>AI-Agentic Products:</strong> Standalone tools where the agent itself <em>is</em> the product, replacing SaaS/PaaS functionality.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0AtF-ExNU5YBUhQl" /></figure><h3>However, not every IT workflow is ready for agentic disruption</h3><p>Startups and enterprises need to ask three critical questions to understand if they can adopt agents into their IT workflows:</p><ol><li><strong>Workflow Shape: </strong>Is it repetitive, standardized, and rules-driven? Agents thrive in structured environments.</li><li><strong>Enterprise Intent: </strong>Do companies want to control the workflow or just the outcome? This decides whether a product-led or services-led approach is better.</li><li><strong>Cost &amp; Talent Pressure:</strong> Is the process human-heavy and expensive to scale? If yes, agents can unlock massive savings.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cRBfXcvBZYou98mN" /></figure><h3>Every IT process needs to be evaluated independently for agentic adoption</h3><p>By applying the above framework, three opportunity clusters emerge:</p><ol><li><strong>High-Conviction Zones (Most Attractive): </strong>Processes that are outcome-driven, rule-based, and human-heavy.</li></ol><ul><li>BPO &amp; ERP Services — Claims, form processing, ERP configurations.</li><li>Cloud &amp; Infra Optimization — Agents that manage costs, provisioning, and scaling.</li><li>IT Service Desks — L1/L2 ticket resolution with faster, cheaper automation.</li><li>GenAI Ops — Orchestration agents for enterprise LLMs.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MMqWrvXV4Wr0pQ_6" /></figure><p><strong>2. Emerging Plays (Attractive but Competitive/Complex): </strong>Use cases are real, but require sharper wedges to win.</p><ul><li>QA/Test Automation, Dev Productivity, Data Agents.</li><li>Support/HR Workflows, ESG Compliance.</li></ul><p><strong>3. Cautious Zones (Longer-Term): </strong>High-risk, too horizontal, or not agent-ready today.</p><ul><li>Cybersecurity, IoT/Edge, Strategy Copilots, Generic CRM/CX agents.</li></ul><h3>Agents can take a product or services-heavy approach based on the category of IT processes</h3><p>The approach would differ as follows:</p><ol><li><strong>Legacy Processes (e.g., ADM, IMS, ERP, QA, BPO):</strong> Historically outsourced and human-heavy. <em>AI agents can reduce human dependence, scaling work without scaling headcount. AI-enabled services become the right approach here.</em></li><li><strong>Strategic Processes (e.g., Digital Transformation, AI, Data/Analytics, Product Engineering): </strong>These remain in-house but rely on tools and talent. <em>AI agents built on proprietary data can evolve into full product platforms here.</em></li><li><strong>Operational Processes (e.g., Cloud Services, AMS, Sustainability Tech):</strong> Often customized and vendor-heavy.<em> Agents can cut vendor dependence and unlock efficiency. AI-enabled services become the right approach here.</em></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nf8mOt34oB1Mo3Pi" /></figure><h3>Legacy incumbents now have the pressure to redefine how they approach enterprise IT</h3><p>ITeS incumbents recognize the threat from agents and are reshaping their client delivery approaches in various ways, making it hard for startups to establish long-term moats:</p><ol><li><strong>Full-Stack Delivery is the New Standard: </strong>Clients expect end-to-end AI delivery. Mid-sized players without consulting depth are being squeezed out. Startups must own a category end-to-end.</li><li><strong>AI Talent is a Bottleneck: </strong>Incumbents are upskilling, acquiring GenAI studios, and running internal bootcamps. Startups need retention pipelines or acquisition strategies for AI-native talent.</li><li><strong>Domain-Led Accelerators are Emerging as the New IP Moat: </strong>Pre-built, domain-specific agents embedded into ERP/CRM stacks command 30–40% pricing premiums.</li><li><strong>AI Centres of Excellence (CoEs) are the New Differentiator: </strong>PE-backed firms with verticalized CoEs are valued at 11–20x EBITDA, higher than the ~10x median.</li></ol><p>On the other hand, the impact of AI adoption in IT is uneven for SaaS and PaaS products. Application Development (ADM) copilots are the clear winners so far:</p><ul><li><strong>Why ADM works:</strong> Structured tasks, modular workflows, abundant open-source training data, fast feedback loops, and developer openness.</li><li><strong>Why others lag:</strong> QA needs deep context, ERP is highly customized, cybersecurity is too risky, BPOs monetize headcount, and consulting is too judgment-heavy.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/866/0*AvU0PCNHZI4GaCg7" /></figure><h3>There are a plethora of opportunities for startups to capture market share from ITeS incumbents</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Skj_VONo3M0PiEqT" /></figure><p>AI agents are redrawing the boundary between what enterprises build, buy, and outsource. The real opportunity lies in navigating this shifting line with domain depth, trust, and end-to-end ownership. Startups that get this balance right won’t just automate workflows; they’ll redefine categories.</p><p>If you are a founder building in this space and want a sounding board to bounce off ideas, refine strategies, or connect with the right venture capitalists, reach out to me at pdchhaparwal@gmail.com. Coffee’s on me.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=33e87b785127" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[User Interfaces in the Age of AI]]></title>
            <link>https://pdchhaparwal.medium.com/user-interfaces-in-the-age-of-ai-c624cb3ceba0?source=rss-c93454f16f39------2</link>
            <guid isPermaLink="false">https://medium.com/p/c624cb3ceba0</guid>
            <category><![CDATA[ui]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[figma]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ux]]></category>
            <dc:creator><![CDATA[Parth Chhaparwal]]></dc:creator>
            <pubDate>Mon, 11 Aug 2025 02:31:51 GMT</pubDate>
            <atom:updated>2025-08-11T02:31:51.582Z</atom:updated>
            <content:encoded><![CDATA[<h3>Figma’s trajectory is opening new opportunities to rethink UI/UX for the age of AI</h3><p>Figma just went public, and the market reaction was extraordinary. In four years, Figma’s revenue jumped from $40M to $749M, a growth that convinced investors to value it at nearly 300x earnings. That kind of multiple doesn’t happen unless the market believes a company is indispensable to how modern software gets built. And in Figma’s case, it’s easy to see why. It’s where design, product, and engineering already meet to turn ideas into shipped products.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SkUKsqMMyhO-GpOa" /></figure><p>Now that the IPO is done, the bigger question is: what’s next? What does this moment mean for the UI/UX industry? And where does interface design go from here?</p><p>Figma’s own moves give a glimpse into how it’s approaching the future. At its recent Config 2025 conference, the company showed it’s expanding horizontally, embedding itself deeper into the software development lifecycle. New launches like Figma Make (for turning designs into interactive prototypes) and Figma Sites (for designing and publishing full websites) point to a clear goal: make Figma a single workspace where design, prototyping, and even production build happen together. It’s a shift from being a specialist tool to becoming an end-to-end platform in the product creation process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MucdD6dYGz5tet3r" /></figure><p>And this push isn’t just visible in Figma’s own product launches. It shows up in its investment activity too. Through Figma Ventures, the company has backed tools like Graphite, LottieFiles, Jam, and Sprig, all of which sit at different points in the software development process. It’s another sign of Figma’s intent to be embedded across the entire lifecycle of how digital products are built.</p><p>UI/UX was once the undisputed core of Figma’s identity. Now, its focus has widened, from designing interfaces to shaping the entire product creation process. But that shift leaves a bigger question hanging over the industry: In the age of AI, how will user interfaces themselves evolve? Is this the moment for entrepreneurs to go back to the drawing board and build the next generation of interface design tools, ones built specifically for the realities of UI/UX in an AI-driven world?</p><h3>With AI making coding easier, UI/UX’s real challenge is building interfaces that adapt to each user</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mhmBf8SEJDSqlp8f" /></figure><p>Even with all the hype around emerging modalities, most of our digital interactions are still mediated by a few dominant interface patterns. That persistence isn’t accidental. GUIs, menus, forms, and search have been refined over decades, and entire user mental models have been built around them. Replacing them means overcoming not just technical hurdles but also deeply ingrained habits. The result is a kind of <em>interface conservatism</em> where new modalities like voice, natural language, wearables, or AR/VR are growing, but they still serve as adjuncts to the old guard, not full replacements.</p><p>But here’s what’s changing. The hard part of building software is no longer the coding itself. AI-assisted development tools like Cursor have collapsed the time, cost, and skill barriers to turning an idea into a working product. What used to take weeks of engineering can now be assembled in days or even hours. That shift has quietly moved the bottleneck upstream. When the back-end and feature set can be spun up quickly, the limiting factor becomes how well the product’s interface anticipates user needs, adapts to different contexts, and delivers a level of polish that meets today’s sky-high expectations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8gLzCGFXXZbz_C82" /></figure><p>This creates a paradox for UI/UX. On one hand, the dominance of established patterns means there’s a well-understood playbook for “good enough” design. But “good enough” is now table stakes. Users are benchmarked against the most frictionless apps they use daily, and any gap feels jarring. On the other hand, the rise of AI is enabling entirely new interface paradigms, including interfaces that can shift form based on user preference, collapse complex workflows into a single query, or even operate invisibly in the background. The opportunity (and challenge) is no longer just to make something usable, but to make it adaptive, context-aware, and capable of disappearing when it’s not needed. That’s where products will win or lose in the coming decade.</p><p>In the age of AI, interfaces could evolve from fixed structures into adaptive layers that respond to each user’s preferences, context, and workflow. A single product might appear as a traditional GUI for one user, a conversational chat interface for another, or a voice-driven assistant for someone else, all powered by the same core system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KzFM-aC1QdeL5Sjk" /></figure><p>This is more than cosmetic personalization. It’s structural flexibility, where the interface adapts in real time to how a user wants to engage. The challenge for designers will be to create systems that support multiple modalities without losing coherence, usability, or brand identity - a very different design problem from the static, one-size-fits-all screens that have dominated the past few decades.</p><h3>When Interfaces Adapt, So Must the Organization</h3><p>The rise of adaptive, AI-driven interfaces is quietly rewriting what it takes to build a great product. As the interface shifts from a single, static format to a flexible layer that can take many forms, the job of delivering a cohesive user experience has become more complex and more critical than ever. In this environment, the question isn’t just what a product does, but how it meets the user in the right way, at the right time, through the right modality.</p><p>For most organizations, this is a cultural reset. For decades, UX has been treated as the final step before launch, the “skin” on a finished product. That was fine when engineering was the bottleneck. But when the back end can be generated in hours, an afterthought approach to UX creates a massive gap: the product works, but the experience lags behind user expectations. And in competitive markets, that gap is where churn happens.</p><p>To close it, three big shifts are needed:</p><ol><li><strong>Treat UX as a strategic lever, not a decorative layer: </strong>Companies already track acquisition cost, conversion rate, and retention like core business metrics. In the age of adaptive interfaces, UX belongs on that same dashboard. The right interface strategy can cut onboarding time in half, lower support tickets, and even open new revenue streams (think of voice ordering for hands-busy contexts). This means roadmaps must start from desired user and business outcomes, then design the interface strategy to deliver them. “Feature shipped” can’t be the milestone anymore — “friction removed” should be.</li><li><strong>Build ‘technical’ UX talent that understands systems, not just screens: </strong>Post-2022, traditional UX hiring slowed, but that’s misleading. The real demand is for designers who can think beyond static layouts: understanding how a GUI, chat flow, and voice skill should all serve the same intent, when to use each, and how to pass state between them seamlessly. They’ll need to speak API, reason about latency, and design for LLM quirks as naturally as they pick colors or typography. In other words, the next great UX designer is part interaction architect, part systems engineer.</li><li><strong>Expand the product manager’s role to become the interface orchestrator: </strong>When a product can manifest in multiple modalities, someone has to decide which one shows up, when, and why. That’s a product decision, not just a design one. PMs will need to own “modality strategy” alongside the roadmap: setting rules for when to offer a chat handoff versus an inline GUI, measuring the friction of switching modes, and ensuring all variants feel like one product. This adds complexity, but it also creates a new competitive edge: the product that always “feels right” in the moment will win.</li></ol><p>The through-line here is simple: in the AI era, building the feature is the easy part. Orchestrating how that feature reaches the user and ensuring it works across contexts and modalities — is where the real value (and risk) lies. Organizations that get this right will stop thinking in terms of “the interface” and start thinking in terms of interface portfolios: multiple, adaptive entry points to the same product core.</p><h3>AI Is Redefining Interfaces, and Early Movers Will Set the New Rules</h3><p>Every major leap in computing has come with a new interface paradigm, and with it, a new generation of category-defining companies. We’re on the cusp of another one. AI isn’t just adding features to existing products; it’s changing the physics of product creation. When the cost and time to build collapse, the center of competition shifts to how products are experienced, and how quickly they can adapt to a user’s needs in the moment.</p><p>This is the kind of shift that reorders the value chain. Entire ecosystems, from design tools to analytics platforms to customer engagement systems, will need to evolve to support adaptive, multi-modal products. That’s a wide-open lane for founders: building the infrastructure, frameworks, and vertical-specific solutions that make adaptive UX reliable, measurable, and easy to ship.</p><p>It’s also an opportunity to question defaults that have gone unchallenged for decades (read Adobe, Figma). If interfaces no longer need to be one-size-fits-all, why do workflows look the same for every user? Why does onboarding follow the same linear path? Why are most products still designed around the constraints of the screen instead of the possibilities of the context? Founders who explore these questions won’t just improve usability, they’ll invent entirely new ways for people to interact with technology.</p><p>The window is open now, but it won’t stay open forever. Incumbents are smart, have deep resources, and will move quickly once the patterns become clear. As adaptive design patterns mature, the gap between products that feel truly fluid and those that don’t will be obvious and unforgiving. The founders who act early will define the patterns everyone else copies.</p><p>If you’re a founder building in this space and want to connect to brainstorm, have a sounding board, or get connected with the right venture capitalists (including myself), reach out to me at <a href="/service/mailto:pdchhaparwal@gmail.com">pdchhaparwal@gmail.com</a>. Coffee’s on me!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c624cb3ceba0" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>