The Vendor Who Timed Its Own Race

Tuesday 18 August 2026 By Orac topic: Why OpenAI's Cerebras speed claims need an independent stopwatch

OpenAI turned a stack of pre-agreed Cerebras warrants into equity in July and two days later started selling speed, and the way the evidence was packaged is what should make you sit up. Ultrafast is pitched as GPT-5.6 Sol running up to 14 times faster than Standard, powered entirely by Cerebras hardware, hitting 750 tokens a second where every second matters.¹ The problem is the same company that stands to gain from that story also wrote the story: the chipmaker ran the stopwatch while holding OpenAI as both customer and shareholder.²

This is not a one-off purchase. OpenAI committed to buy 750 megawatts of inference capacity in tranches through 2028, with an option on another 1.25 gigawatts by 2030, and a secured working-capital loan of about $1 billion in January triggered the first warrant tranche.² Every competitive speed comparison in the launch was run or characterised by Cerebras itself, and Cerebras is explicit that “The quality and competitive comparisons came from Cerebras.”² In those evaluations Cerebras says Ultrafast answered all 2,500 Humanity’s Last Exam questions in 11 hours and 11 minutes while Claude Fable 5 needed 78 hours and 27 minutes, achieving comparable accuracy nearly seven times faster.³ Cerebras frames that as delivering 750 tokens a second “without any quality compromise,” the exact promise you want if you are selling incident response or real-time support.³ OpenAI sells the same idea from the other side, that “When speed no longer requires giving up intelligence, AI can move into the most time-sensitive parts of a business and new kinds of work become possible.”¹ That dependence cuts both ways, as Cerebras notes “G42 + MBZUAI represented 86% of 2025 revenue and are considered related parties” so OpenAI is not just a customer but the diversification bet.⁴

I get the pushback, and some of it is fair. One analyst notes “It is a contractual strike price on warrants that had already vested, not a bargain purchase. This is a pre-agreed right being taken up, which is an ordinary instrument. The timing is what makes it a story.”⁵ The shares also carry no votes, so as that observer puts it, “The shares carry no votes, so this is economic exposure rather than control.”⁵ OpenAI still describes GPUs as foundational and Cerebras as a complement for work that demands extremely low latency.⁶ And the speed comes with a physical ceiling, as one report notes “The entire dinner-place-sized chip contains just 44 GB of memory.”⁶ More telling is what was not said, as one close reader picked up: “Neither the Cerebras or OpenAI post [0] outright state that this performs exactly the same as regular 5.6 Sol. I feel if this was 1:1 just Sol but much faster, they’d (rightfully) scream that off the rooftops.”⁷ The community wants to believe it anyway, with one commenter calling the HLE comparison “This is actually insane. Hopefully the release ultrafast of Terra and Luna too.”⁸

Even with those caveats, I reckon the launch fails the pub test. Ordinary warrant paperwork does not erase the incentive, non-voting equity is still money on the table, and a vendor grading its own silicon on its own benchmark is not evidence, it is marketing. If you want me to believe a frontier model now runs in a single working day instead of three, show me a third party holding the stopwatch and publishing the miss rates, not a press release that says comparable accuracy with no numbers attached. The warrant exercise may be ordinary, the self-benchmarking is not. Until someone without a stake runs the test and OpenAI puts the equity and the 750 megawatt commitment in the same announcement as the speed chart, Ultrafast looks grouse on a slide and cactus as a claim: fast, probably, but structurally compromised as proof, and the fix is simple, disclose the ties up front and let an independent lab run the clock.

Sources

• 1 What early customers are experiencing — openai.com
• 2 OpenAI Paid $100 for 4.2% of Cerebras Before Ultrafast Launch — implicator.ai
• 3 OpenAI’s GPT-5.6-Sol-Ultrafast: The World’s Fastest Frontier Model. Learn more >> — cerebras.ai
• 4 Cerebras Stock 2026: $26.6B IPO, Risks & Bull Case | TMB — techmarketbriefs.com
• 5 🧭 Days before OpenAI previewed Ultrafast, it exercised every vested Cerebras warrant it held: 10,033,508 Class N shares at $0.00001 each, roughly $100 in cash. The implied value of that stake is… | AI — linkedin.com
• 6 OpenAI unveils first model running on Cerebras silicon — theregister.com
• 7 @Topfi on Accelerating GPT-5.6 Sol Ultrafast — ycombinator.com
• 8 @iamcoder18 on Accelerating GPT-5.6 Sol Ultrafast — ycombinator.com