Everyone's calling it the AI bubble. There's been endless talk about the circular economy in AI, the money moving in loops between the labs, the clouds, and the chipmakers. I think we're looking at something more specific and more dangerous. A valuation trap that's forcing companies like OpenAI and Anthropic to cannibalize their own customers.
The numbers don't add up
Consider what's happening right now. The SpaceX IPO just dropped with eye-watering numbers. $75 billion raised at a $1.77 trillion valuation, the largest IPO in history. Anthropic confidentially filed on June 1. OpenAI followed a week later. Then news broke that OpenAI might push its listing to 2027. The reporting says advisers warned Sam Altman that a 2026 debut would price below $1 trillion, and OpenAI's own filing admits "there are things we want to do that are likely easier as a private company." The books themselves are already public. A $38.5 billion net loss in 2025 on $13.1 billion of revenue.
That revenue number should sound familiar. Last October on the BG2 podcast, with Satya Nadella sitting right there, Brad Gerstner asked Sam Altman a simple question. How do you justify over a trillion dollars in compute commitments when your annual revenue is $13 billion? Altman's response: "if you want to sell your shares, I'll find you a buyer." He seemed peeved and it didn't seem to be the response of someone comfortable with their unit economics. And by February 2026, OpenAI quietly cut its compute target to around $600 billion by 2030, down from the $1.4 trillion Altman had been touting. His response during the podcast hasn't aged well so far.
We might remember the Loopt story. Sam Altman's first startup claimed 4 million registered users at its peak. When it sold to Green Dot in 2012, it reportedly had around 500 daily active users left. I read him as an eager salesperson at heart. He has a tendency to exaggerate and inflate.
Cornered by their own valuations
Both Sam Altman and Dario Amodei seem seem to be caught in a vicious cycle. Smack dab balancing responsibilities to their customers, investors, and employees. They're at the receiving end of extremely high valuations. Anthropic's last round priced it at $965 billion. There's no doubt they have incredibly valuable offerings. But those offerings are now being commoditized and delivered at scale. The challenge? They need to find avenues to justify their valuations and demonstrate returns that match their revenue and burn. The compute requirements, the talent they're hiring, various other factors. There's massive demand at their end for how to solve it.
They have to feed the monster they've created.
Eating their customers' lunch, dinner, and breakfast
Here's where it gets uncomfortable. These companies can't operate purely as utility providers. They're starting to solve for specific verticals. Claude for Financial Services launched in July 2025. Claude for Legal followed in May 2026, walking straight into the market occupied by Harvey and Legora, two legal AI startups built on Claude itself. They're essentially eating the lunch, dinner, and breakfast of their customers by creating solutions that take away business from the very companies paying them.
The pattern has form. In June 2025, days after OpenAI moved to acquire the coding startup Windsurf, Anthropic cut off most of Windsurf's access to Claude models. All while selling Claude Code to the same market. And look at what happened with Figma. On April 17, 2026, Anthropic launched Claude Design and Figma's stock fell 7% that same day. Three days earlier, Anthropic's chief product officer Mike Krieger had resigned from Figma's board. To be fair, Figma's longer slide has other causes, lock-up expiries and insider selling among them. But that same-day 7% drop was the market pricing exactly this dynamic.
The factory workers wearing cameras
One of the most striking stories I've seen recently: garment workers outside Delhi wearing head-mounted cameras through their shifts, paid around 250 rupees an hour, recording first-person footage that trains AI models to essentially do their work in the future.
Many knowledge workers are inadvertently doing the same thing. They're just not wearing the cameras on their heads. The camera they're offering is their system access, their business data access. They're training these AI companies to outmode their workflows.
This won't affect the majority of subject matter experts and mid-to-top management. Their work and value add goes far beyond what an AI might replicate. A financial expert or legal advisor with experience, relationships, network, and credentials is far more valuable than just an AI tool that can churn out responses. Very often people are buying in for the holistic human access.
But there will definitely be a correction for routine, repeatable roles. A classic example: Figure AI was challenged in May to run its humanoid robot through a single 8-hour package sorting shift. Its robots ran for 200 hours straight, sorted 249,560 packages, and had zero hardware failures. Amazon's own leaked planning documents point to avoiding over 600,000 hires by 2033 through automation, though Amazon disputes that framing. I've written before about how Amazon is fumbling the ball. Robotics is one area where it isn't. Meanwhile Chinese companies like Unitree already control roughly 90% of humanoid robot sales, Japan is committing $6.3 billion to robotics against its labor shortfall, and Tesla's Optimus, for all the noise Elon Musk makes, missed its 2025 production target. The robots are coming either way. The only question is whose.
AI sovereignty and the coming jolt
This is where AI sovereignty enters the analysis. Palantir's Alex Karp went on CNBC recently and said "something has gone completely wrong" with how OpenAI and Anthropic sell AI, calling their token-based business model into question.
Strip away the theatrics and he's pointing at the same thing those garment workers learned firsthand. Renting intelligence from a frontier lab while handing it your data and workflows is training your own replacement. As more enterprises internalize that, the correction won't necessarily act out like a bubble bursting. But it will definitely send a jolt down the markets. A single report of OpenAI's IPO delay in late June was enough to knock the entire AI trade for a day. The valuations, the investors, the companies caught up in this circular ecosystem will have to re-correct for the market paradigm shift.
The inevitability of it all
The starting premise holds. Valuations are creating an inevitability for these AI companies to generate revenue by all means necessary. Dario Amodei and Sam Altman might not necessarily be in a position to choose. They might not wish to attack their own customers. Rather, they might find themselves cornered and compelled to pursue this approach.
That's the real story. Not a bubble. A trap we must all find our way out of.
If you're a founder preparing a raise in this market, here's how I help founders get raise-ready.
Follow me on X for more thoughts on growth and marketing. Subscribe to my newsletter for new posts.