OpenAI makes a ‘Code Red’ turn in strategy - Financial Times

OpenAI is running out of runway to be a research lab.

The company that spent 2025 launching Sora, the Atlas web browser, an adult ChatGPT mode, and a chip still in the design and tape-out stage — while burning through an estimated tens of billions of dollars — is now doubling its workforce to 8,000, redirecting compute away from consumer experiments, and talking openly about enterprise productivity as the only path that matters. The Financial Times reported this week that OpenAI plans to grow from 4,500 to 8,000 employees by the end of 2026, part of a broader pivot that Chief Executive Sam Altman signaled with an internal code red in early December 2025, pausing non-core projects in response to Google Gemini 3.

The headline number — $840 billion valuation on a $110 billion funding round backed by SoftBank and Big Tech — masks the pressure underneath. OpenAI reported $20 billion in annualized revenue for 2025, a 233 percent increase from the prior year; CNBC reported $13.1 billion in actual collected revenue over the same period — the gap reflects the difference between run-rate and cash received. Either figure sounds like growth. It is growth. It is also nowhere close to covering the infrastructure costs of running frontier AI at scale, costs the company is estimated at tens of billions of dollars annually. The math that made OpenAI worth $840 billion is the same math that makes it dependent on institutional buyers who can pay enterprise rates, not consumers who cancel after a weekend.

The clearest evidence for where OpenAI thinks the money is: Codex. By the end of January 2026, OpenAI's coding agent was bringing in just over $1 billion in annualized revenue, according to a person with direct knowledge of the matter, per Wired. In September 2025, Codex was running at 5 percent of the usage of Anthropic's Claude Code. By January 2026, it had climbed to 40 percent. That trajectory is the enterprise story in miniature — OpenAI was behind, closed the gap fast, and the gap matters because coding agents are now the clearest proof that AI generates revenue at the enterprise layer.

Claude Code itself is pulling in more than $2.5 billion in annualized revenue, accounting for nearly a fifth of Anthropic's business, the company said in February. Anthropic is valued at $380 billion with $14 billion in revenue and is planning an IPO this year. The comparison is not incidental. Anthropic CEO Dario Amodei rejected a $200 million Defense Department contract in late February, citing concerns about autonomous weapons and mass surveillance. "We cannot in good conscience accede to their request," Amodei said. The Trump administration responded by effectively blacklisting Anthropic from government work. OpenAI took the same category of contract within hours. The boycott tracking site quitgpt.org reported that more than 1.5 million people took action — cancelling subscriptions, sharing boycott messages, or signing up via the site — in the days following the announcement. The causal link to the Pentagon deal specifically is editorial inference drawn from the timing of the surge; Euronews notes the QuitGPT spike followed the broader Anthropic-Pentagon standoff rather than establishing direct causation.

Fidji Simo, OpenAI's CEO of applications, told staff in January that the company could not let the enterprise moment slip away because of "distractions from fringe activities." A strategy of trying to do everything at once, senior officials told the Wall Street Journal, was putting the company on the defensive. The Sora shutdown illustrates this with unusual clarity: the video generation app peaked at 3.3 million downloads in November 2025 and generated roughly $1.4 million in lifetime net revenue, against estimated compute costs of about $15 million per day, per Sensor Tower data cited by CNN Business and Forbes. ChatGPT made about $1.9 billion in the same period. Sora was discontinued in late March so its compute could be redirected. It was not a product decision. It was an accounting decision.

The Pentagon boycott is the detail that changes the shape of this story. The narrative that OpenAI simply overreached into too many consumer products — a neat cautionary tale about scope — elides that the moment users actually organized in numbers was over a defense contract, not a chatbot feature. Anthropic drew a line on the same deal. Its valuation and revenue have only grown since. OpenAI crossed the line and is now in code red. The divergence is not about product strategy. It is about what the company is willing to be.

OpenAI is not walking the line back. Simo said the company is "very much acting as if it's a code red," though she added that declaring codes for everything did not necessarily make sense. The company is hiring for enterprise, building for enterprise, and hoping the consumer base that remains is large enough and loyal enough to subsidize the transition. ChatGPT reached 900 million weekly active users as of February. That number is real. It is also the subscriber base that staged a visible, organized protest over a single contract decision. The 1.5 million who took action were not churning because something better existed. They left — or signaled they would leave — because of a choice. The code red is the admission that the research lab days are over.