Get real with Chi on AI
#AIFomo #AlSlop #AIBts #AIOpenWts #AICompute
Whats going on right now across the board in AI?? What are some strategic things to watch out for among the noise — which is a lot to be honest. I am covering above topics that I think are things to be aware of.
#AIFomo
Fig 1 : [AIG] Gemini generated
There is a lot of FOMO about AI in every organization. It’s Fear Of Missing Out for non GenZ. There is also lot of this “free advice” floating around that — BETTER to GET ON IT QUICKLY. Although it’s a good advice from an early adoption point of view it can also cause false influences. AI is still in its infancy when it comes to governance, TCO etc. and WILL go through lot of iterations before it stabilizes to a happy median point. AI is not going anywhere faster believe me I read concept of neural networks in 1999 and it took 2 decades to get to a point of transformers. A technology moving fast doesn’t mean it’s all Right. Here are few challenges that will take their course.
Cost
It’s multifold. It’s more trickier than it sets out to be. First you have to pay for an enterprise base license from the providers to really get good benefits, then the token cost for practical use like API etc , then the cost of Human In The Loop reviews which can drastically vary depending on the use case. Typical enterprise user licenses come with unlimited token usage for regular user chat interactions, for now, which can be used for day to day efficiencies for employees. But true practical applications like RAG APIs, customer service etc are token based. Teams are reporting spending $4k-$10k in tokens per month per person — somone of Sr AI/Software Engineer role. Thats like a .5FTE impact , so unless there is ROI out of the bat the AI software is costly investment unlike regular software like MS office or Airtable.
Governance
AI can not be used anywhere where there is a requirement for strict reproducibility. Even where reproducibility is not a requirement still it needs additional controls that were previously non existent like guardrails( external threat, leaking of PII etc), AI transparency(need to know what the underlying components are) , AI explainability ( how did it come to a conclusion) etc. These are critical for all PUBLIC companies as a part of SOX guidelines. Basically you need an audit trail of how a particular financial output is derived, so if you cant explain in a series of steps that are reproducible then you violated SOX. There is also clause named “Indemnification” clause responsible for vendor AI companies to be responsible for all third party lawsuits that originated from enterprise using their product. But all AI labs keep a very narrow tolerance for this clause for example included only if you are taking their “Enterprise” plan and that too strictly operating within the mentioned “guardrails”. Think of a new drug in market, where the ad focuses 5 secs about the actual product and 25 secs about the side effects.
Use Case dependent
AI is very “Emo” when it comes to problems its solving. Sometimes its super happy and efficient for some its like Kramer working at an office(Seinfeld reference). Let me explain.
AI is #VeryGood at below right now
#CustomerSupport bots both chat as well as voice. They can handle 50–70% of standard inquiries operating 24/7 with zero scaling costs. And probably never getting tired. Only redirecting 30% of tickets that needs higher level of troubleshooting to human agents.
#DocumentAnalysis Whether its a knowledge management center within a company or a legal firm where information sources are documents, blogs, wikis this does a great job of plug and play. Also in industries like legal where documents are kind of center of everything, whether that is attorney needing to review 1000 documents to find certain context from deposition or contractual risks in a M&A deal , AI is great and in fact provides 10–20x or more benefits.
AI is #NotGood at below
#MultiStepReasoning Due to its lack of grounding to real world it can not do complex multistep reasoning well yet. It might drift off course. Also the accuracy degrades exponentially with each step. So areas like medical diagnosis or legal argument are not good to use AI for.
#CodeHelp For some reason this is talk of the town. CoPilot, Cursor, Claude Code all seems the favorite buddies for our developers-developers-developers-developers (Steve Balmer’s thing) but its still questionable on ROI. It can be efficient in some cases, when converting legacy mainframe code to Java or upgrading code from one version to later version etc. However when you are coding on brand new idea its great for prototyping but not for production code creation. Reasons being review tax, prone for security risk, cognitive debt as developers dont really understand depth of the code as AI wrote it so when time comes for maintaining and fixing the effort keeps adding up.
#AllOthers if a use case is not covered above by default assumg AI is not great for it.
So going back to #AIFomo don’t be FOMO. Simply sit back, relax and watch as AI stabilizes. You are better off waiting an year or even two to see what tools stay what companies stay and then make the transition. In the meanwhile, explore purely from a research perspective. What every enterprise can do now is get your organization AI ready, invest in modernizing data stack that can support AI etc.
Now lets talk about sloppy AI.
#AISlop
Thats what the Gen Z is calling the content that is AI generated and of low quality. It also does not simply mean content is of low quality, but also means that there is not much human thinking and effort went into it. Think of it this way. If a movie is made with scenes where they are using AI to replace a green screen with generated AI and actually mention that they used AI to do this, consumers are generally more acceptable of that content JUST because of all other efforts around it. There was a cast of people, equipment etc that preceded that scene and followed that scene so people are accepting it. So it is a real thing. If you are non Gen Z and for all right reasons get excited to see AI generated content dont think the newer generation is actually feeling the same way. For some reason they are moving away from it. May be I understand. People that are non Gen Z like me, we have seen tech from mainframes, databases, data warehouses to excel with 65000 limitation the new AI tech is fascinating to see in our lifetimes but for newer generation as they grew up with smartphones their brain is equipped to simply brush off AI generated content as slop. So be careful especially if you are in entertainment business and targetting for next generation content the kids are liking human generated content more intrinsically. So stay away from using AI for #ContenGeneration especially #Video #Audio , however if its something supportive function like creating subtitles in different languages etc feel free to use.
Fig 2: [AIG] Gemini generated — I really love that it says “Trump’s ginger tea” on the board I didn’t even ask for it to make a shop board
#AIBTS
So whats the behind the scenes happening across AI players right now. Lot of competition, lawsuits, rat race literally.
Get Chiranjeevi Mallavarapu’s stories in your inbox
Join Medium for free to get updates from this writer.
OpenAI , Anthropic are in deep stress right now. Both are preparing for their upcoming IPOs. Both face massive cash burn, rising enterprise demands for efficiency, and intense pressure from recent security incidents where development models briefly escaped their sandboxes. Anthropic and OpenAI both evaluated at close to $1trillion with Anthropic slightly leading the way.
Fig 3 : [AIG] Gemini generated
Both companies confidentially filed S-1 initial public offering documents in June 2026 to capitalize on current valuations before market cooling.
Corporate clients are shifting focus from high-volume token consumption (“tokenmaxxing”) to cost efficiency, squeezing profit margins for both. So it is a tricky time for them to go IPO but also might be right time because if they wait longer their valuations might actually drop. Again a big gamble. Both companies are burning a lot of cash right now without IPO moves they cant afford to do more investment rounds with being already at14–18 rounds. Any further private funding means VC market would not be able to invest in upcoming startups anymore if these blackhole size companies keep sucking private capital. So IPO is the only way to go.
Now comes the kicker . “AI Bubble”. The investors are going to be looking to cashout at these IPOs. So it will be a good gameshow to watch to see if they are cashing out or still continuing to keep the bet on. If they cashout it will mostly indicate a potential bubble. This means your 401k, pension funds will get hit if not losses but at least flatline the growth. So yes these companies will impact you ultimately IF THEY DONT BREAK EVEN. Because both are $1Trillion evaluations, these massive evaluations will have considerable impact on stock market both S&P 500 and Nasdaq. Any negative quarterly earnings report could trigger a massive sell-off, dragging down the entire public stock market. public crash would destroy the valuations of healthy, revenue-positive software companies that happen to use AI, halting industry innovation for years. Historically Amazon, Uber went into IPO with a net loss but with massive IPO capital they turned into net positive. So it is a big bet on AI labs IPOs. It can be wither GOOD or BAD in both instances one thing will be common it will be BIG impact.
#AIOpenWts
You heard the latest news. Notable leaders in AI have signed a pact that they officially support LLMs moving towards Open Weights models. This means a lot of upcoming transitions in the industry. Hence my point abot Hakuna Matata on the #AIFomo. Things are going to change.
Open weights means any company can download the LLM on their private private clouds, avoiding sending sensitive data to outside APIs and eliminating vendor lock-in. So they are only paying for compute. As alternative open models drive down the market rate per token, labs face lower margins on basic text generation and inference tasks.
Fig 4: [AIG] Gemini generated and no there are no clones of them in real world
This pushes AI labs and tech giants to provide enterprise guarantees, service level agreements (SLAs), dedicated security compliance, and custom vertical agents to make applied AI not just AI.
What this means is a larger share of total market profit risks shifting away from the model layer itself and toward cloud infrastructure providers and hardware suppliers. We are talking Azure, GCP and NVIDIA.
Again lot of flux happening in few weeks. My take on this is as China broke the ice on open weights models with Deepseek, Kimi K the infra giants NVIDIA, AMD are forced to support the transition but held back by few tech giants to a point where they felt comfortable to be on their own models and released this into ether.
Again there are lot of moving pieces right now. But what I am happy about is the leaders we have are ETHICAL leaders that are at heart are technologists came from humble beginnings like Satya, Jensen so they will take a bite but wont get greedy to a point the tech is doing harm. The open weights pact such a display of human
Open Weights Signatories [50 and counting]
AI21 • AMD • American Innovators Network • AMP • Andreessen Horowitz • Arcee AI • Arena • Baseten • Black Forest Labs • Block • Box • Cisco • Cloudflare • Cohere • CrowdStrike • Dell Technologies • DoorDash • Emergence Capital • Fireworks AI • Genspark • GitHub • Google • Hugging Face • IBM • Inferact • Interconnects AI • The Linux Foundation • Mariana Minerals • Meta • Microsoft • Mistral • Morph • Mozilla • Nebius • Nous Research • NVIDIA • Ollama • OpenAI • OpenClaw • Palantir • Palo Alto Networks • Periodic Labs • Perplexity • Prime Intellect • Reflection • Replit • ServiceNow • Telnyx • Trajectory • Y Combinator
#AICompute
Now lets talk about AICompute in general.
OpenAI recently raised its projected computing spend to a staggering $750 billion through 2030. To keep pace, Anthropic signed a monumental $45 billion, three-year compute deal with Elon Musk’s SpaceX to access its infrastructure, alongside a joint $30 billion cloud and hardware commitment split between Microsoft and Nvidia.
Fig 5: [AIG] Gemini generated
While securing Nvidia GPUs remains highly competitive, the primary constraints holding back AI compute have shifted down the physical supply chain.
- The “Transformer Crunch”: Over half of the planned 2026 data center expansions in the U.S. face severe structural delays. The bottleneck is no longer just the AI chips, but a critical shortage of high-voltage electrical transformers, switchgear, and grid connectivity.
- Severe Memory Shortages: AI clusters consume immense amounts of High-Bandwidth Memory (HBM). This high demand has crowded out standard semiconductor manufacturing, causing conventional server DRAM prices to spike.
