Don’t Build an AI Strategy Around One Company

Something is off in the AI market, and it is not the products. The models keep getting better. The revenue keeps climbing. But the people running these companies keep walking out the door, and the money keeps flowing into things nobody has proven will pay off. Let me catch you up.

The deep cut

• Watch the fundamentals, not the headlines. OpenAI's models improve while a dozen executives leave and its IPO slips to 2027.
• Cheaper models win the daily work. Spending on Anthropic's Fable 5 stalled at 11 percent as buyers reached for cheaper tools.
• Hype is not a hedge. Situational Awareness went all-in on AI stocks, lost two-thirds of its value, and drew an SEC probe.

The revolving door at the top

OpenAI shipped GPT-5.6, one of the strongest models out there, and added about 15 million coding subscribers in two months. Good news. Then read the exit list. Since January, more than a dozen executives have left, including the COO, the chief revenue officer, the chief marketing officer, and the de facto second-in-command, Fidji Simo.

The latest was Chris Malone, head of data centers, who joined in early 2024 and left last week. Compute is OpenAI's main edge over rivals, which makes that seat an odd one to lose. The pattern is real, not noise.

One man ends up holding everything

As people left, Greg Brockman scooped up their jobs. He now runs consumer, enterprise, product, scaling, and infrastructure. As Verge reporter Hayden Field put it, there are four arms of the company and he controls all of them. One API lead told TechCrunch, "everyone reports to Greg at the end of the day."

Read the departures next to the timeline. The IPO, once expected this year, is now pushed to 2027. OpenAI's losses are growing with its revenue, while Anthropic is reportedly profitable. The reorg looks less like strategy and more like a company cutting weight before it opens its books. When you pick a vendor to build on, you are also betting on that vendor's roadmap surviving its own turnover.

The cheap model ate the frontier

Here is the surprise for anyone budgeting on top-tier AI. Spending data from 70,000 US companies shows outlay on Fable 5, Anthropic's priciest model, plateaued at 11 percent of what buyers spend on its tools. Companies save the flagship for hard problems and let cheaper models handle the rest.

"Most people don't need to operate at the frontier," said Accel's Miles Clements, whose firm put $1 billion into Anthropic. That is a problem for a business built on selling the biggest, most expensive model. OpenAI's cheaper GPT-5.6 helped push its annualized revenue to $40 billion, while Anthropic hit $65 billion, short of the $80 billion bulls wanted. For your team, the lesson is plainer: match the model to the task, and stop paying frontier prices for work a small model does fine.

The bill for all this compute

Even as buyers trade down, the labs keep spending like demand only goes up. Anthropic just signed a $45 billion deal with Nscale, on top of $10 billion with Volta, $5 billion with AMD, and more with SpaceX, Amazon, and Google. Amazon tripled its Nvidia order to 2 million more GPUs, a deal worth tens of billions.

Nvidia's Jensen Huang says the spend is justified because "AI is generating profitable tokens." Maybe. But that math assumes more compute keeps turning into more profit, and the Fable 5 numbers suggest buyers are getting cheaper, not hungrier. That gap is the whole story.

When the bet was just the story

For a warning about betting on the narrative, look at Situational Awareness. Founded by a 24-year-old former OpenAI researcher with no investment experience, it went all-in on AI stocks and once managed $45 billion in assets. When AI stocks dipped in July, it lost over two-thirds of its value, saved only by a fire sale to Citadel.

The fund ran on borrowed money and a skeleton staff of eight. Now the SEC is subpoenaing its banks over its trades and borrowing. It sold investors the tidy story that AI only goes up. The products are real. The certainty was not.

Three questions for your team

• For each AI feature we run, are we paying frontier prices when a cheaper model would do the job? Where can we trade down this quarter?
• If our main vendor's leadership or roadmap changed in six months, how exposed would our product be, and what is our fallback?
• Are our AI bets tied to real usage and revenue we can point to, or to a story about where the market is going?