OpenAI Paid $100 for a 4.2% Cerebras Stake Weeks Before Ultrafast Launch

OpenAI paid about $100 in cash in July to exercise vested warrants for a 4.2% stake in Cerebras, weeks before it previewed Ultrafast for GPT-5.6 Sol on Aug. 13. At Cerebras's Class A trading price of about $229 on Aug. 13, the holding had an implied value of roughly $2.3 billion, though the shares carry no votes. OpenAI is buying computing capacity from Cerebras while holding equity in the chipmaker, which ran or characterized every competitive speed comparison in the launch.

The equity filing

A quarterly filing released after the market closed on Aug. 12 disclosed that OpenAI acquired 10,033,508 Class N shares at a contractual price of $0.00001 each. OpenAI exercised every vested warrant share, giving it a 4.22% stake based on Cerebras's 237,564,041 shares outstanding as of Aug. 5. The Class N shares are not liquid and generally convert to Class A only when transferred.

The full warrant covered 33,445,026 Class N shares when granted under a December 2025 master relationship agreement. Cerebras valued it at $82.02 per share at grant and recorded $822.9 million in customer-warrant assets during the first half of 2026. OpenAI committed to buy 750 megawatts of inference capacity in tranches through 2028, with an option on another 1.25 gigawatts by the end of 2030. A secured working-capital loan of approximately $1 billion in January 2026 triggered the first warrant tranche. The filing showed OpenAI retained warrants for another 23,411,518 shares tied to capacity, payment or market-value milestones.

Ultrafast

The inference capacity secured under that agreement now powers the Ultrafast preview. Ultrafast is a service tier, not a new model. It runs the same GPT-5.6 Sol on Cerebras wafer-scale systems, with OpenAI advertising a maximum of 750 output tokens a second at the preview, up to 14 times its Standard processing tier. Cerebras says each wafer-sized chip holds 44 GB of SRAM, keeping model weights on-chip instead of shuttling them to off-chip storage between tokens.

On the preview date, OpenAI's standard GPT-5.6 Sol API list price was $5 per million input tokens and $30 per million output tokens. Its existing Fast mode cost roughly $10 and $60 per million, respectively, for up to 2.5 times Standard speed.

Speed evidence

The quality and competitive comparisons came from Cerebras. In tests run on different dates and with different software harnesses, Cerebras said Ultrafast completed the 2,500-question Humanity's Last Exam on July 10 in 11 hours and 11 minutes, while Claude Fable 5 took 78 hours and 27 minutes from July 13 through July 15. The company called the result nearly seven times faster with comparable accuracy, but it did not publish the accuracy scores. A separate Cerebras test on July 31 reported a 5.6-fold end-to-end speedup against Standard processing with no quality loss.

The advertised ceiling of up to 14 times Standard speed disclosed no baseline workload, prompt set or reasoning-effort setting. Because access was limited to select customers, no third-party quality evaluation was available at the preview, and OpenAI had published no Ultrafast price, separate model ID or general-availability date.

Within OpenAI, Ultrafast is being tested in incident response and research loops that previously ran overnight. Jane Street, Podium, Basis and Rogo are among the early customers. Access will expand as Cerebras capacity grows.

"Whereas formerly I might have to wait a couple minutes for a task to finish, it now finishes for me before I even have the opportunity to context-switch. It makes me way more productive," OpenAI researcher Jeffrey Wang said.

Anthropic Weighs Microsoft Maia Chip Deal as Claude Demand Grows

Anthropic is in early talks to rent servers powered by Microsoft’s Maia AI chips, The Information reported Thursday, citing two people familiar with the discussions. The talks would give the Claude ma

The Implicator

Nebius Buys Eigen AI for $643 Million to Strengthen Token Factory

Nebius agreed Friday to acquire Eigen AI, a California inference-optimization startup, for about $643 million in cash and stock. The deal would fold Eigen's serving, system and kernel work into Nebius

The Implicator

Nvidia still wins per chip. Google just changed what counts.

Google Cloud on Wednesday unveiled two new TPUs at Cloud Next 2026, splitting its eighth-generation design into a training chip and an inference chip for the first time in the program's decade-long hi

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai