The ink was barely dry on Anthropic's Mythos announcement when OpenAI fired back. Not publicly. Not in a blog post. In a private memo to its own investors, obtained by Bloomberg and CNBC, that reads like a corporate hit piece disguised as a shareholder update.

The core claim: OpenAI has 30 gigawatts of computing infrastructure either deployed or committed. Anthropic has 7 to 8. That is a 4x gap, and OpenAI wants its investors to know it.

The memo, sent within hours of Anthropic unveiling Claude Mythos Preview as "the most capable AI model ever built," describes Anthropic as "meaningfully smaller" and claims OpenAI has been "rapidly and consistently" scaling infrastructure while its rival struggled to keep pace. It specifically calls out Anthropic's history of server capacity problems. You know, the ones that had developers screaming into Slack channels for months.

Here is what is actually happening: both companies are heading toward IPOs this year, and this memo is not really about compute. It is about narrative control.

Anthropic just dropped a model that scores 93.9% on SWE-bench, 97.6% on math olympiad problems, and found thousands of zero-day vulnerabilities that no human security team had ever caught. Forbes called it the most capable AI model ever built. The model is so dangerous that Anthropic will not even release it to the public, restricting access to a handful of cybersecurity companies through its new Glasswing program.

OpenAI's response was not to release a better model. It was to send a letter to investors saying "sure, their model is good, but we have more power outlets."

That tells you everything about where this race actually stands.

The numbers are real. OpenAI's Stargate partnership with SoftBank, Oracle, and the UAE's MGX gives it an enormous infrastructure pipeline. And having 30 gigawatts of committed compute is genuinely significant. Anthropic's infrastructure deal with Google, Broadcom, and Microsoft for 3.5 gigawatts of next-gen TPUs does look smaller on paper.

But compute is not everything. OpenAI is burning $85 billion a year. Its own CEO and CFO reportedly cannot agree on how to handle the cash bleed. Bloomberg reports that in the secondary market, "not a single institution was willing to buy OpenAI" shares when one firm contacted hundreds. Meanwhile, Anthropic's revenue tripled to $30 billion last year and investors are lining up for its IPO.

The memo also comes with context the investors did not get: OpenAI is simultaneously racing to build its own cybersecurity model to counter Mythos, according to an Axios scoop. Both companies are now so terrified of their own creations that restricted access is becoming the norm for frontier models.

This is what pre-IPO panic looks like in the AI industry. You do not send private memos trashing your competitor to investors when you are winning comfortably. You do it when you are worried the narrative is slipping.

And right now, the narrative is very much slipping.

First reported by Bloomberg and CNBC.