Cerebras doubles its chip's clock speed as OpenAI's fast tier runs at 300 tokens a second, down from 750

The fast tier's ceiling fell from 750 tokens to 300

OpenAIPVT$1.18T — OpenAI, private, latest valuation $1.18T introduced its Ultrafast speed lane on September 29↗: up to 300 tokens a second, with the version for its newest model, GPT-6.1 Sol, due within days ¹. The previous model's fast lane, previewed in August, had run on Cerebras chips at up to 750 tokens a second ¹. The next day, research firm SemiAnalysis posted that the new tier runs on Nvidia▼−0.38% — Nvidia, down 0.38 percent today GPUs. Cerebras▼−7.72% — Cerebras, down 7.72 percent today fell 7 percent, and neither OpenAI nor Cerebras has confirmed the hardware ²³.

Altman said partner; Microsoft said Maia

OpenAI chief executive Sam Altman called Cerebras "a close partner" on October 2, naming no chip, and the stock rose about 9 percent ³. On October 12, Microsoft▼−1.36% — Microsoft, down 1.36 percent today said its own Maia chips become "the primary inference system for Microsoft products and services" from 2027↗, on silicon that already runs OpenAI models; Cerebras fell 10.69 percent ⁴⁵. The relationship under question is a multiyear agreement worth more than $20 billion for 750 megawatts of capacity ⁶.

Nvidia is serving the lane Cerebras was built for

Inference, running a trained model to produce answers, is Cerebras's whole business. Its Wafer-Scale Engine, more than 50 times the size of a conventional processor, is built to answer first, and the company claims output 10 to 15 times faster than GPU platforms ⁷. Batch size counts how many requests a chip handles at once. GPUs are most efficient with thousands; Cerebras wins with one. SemiAnalysis reported the new tier runs at low batch size↗, meaning Nvidia is answering the fast lane a single request at a time, the game Cerebras's hardware was designed to win ².

The reply: double the clock, keep the wafer

Cerebras's answer, detailed on October 12↗, is not a new chip. CS-4 Nexus ships this quarter with the Wafer-Scale Engine 3's clock doubled to roughly 2.8 GHz from 1.4, on the same 900,000 cores, in a rack built to host three generations of machines through 2029 ⁸. The bet is that agentic software multiplies the demand for fast single answers: one question now spawns "a family of daughter agents" coordinating together, says chief strategy officer Andy Hock, and that is "where speed really makes the difference" ⁷.

Insiders sold the unlock; Ark bought the dip

About 19.4 million shares became eligible to trade on October 14 ⁶, into a stock that opened at $142.17 on October 16, just above its 52-week low of $133.88 and far below the $185 price of its May listing ⁵⁹. Insiders sold 1,864,452 shares for $387.7 million over three months; chief operating officer Dhiraj Mallick halved his holding at $204.04 on September 29 under a pre-arranged trading plan ⁹. Cathie Wood's Ark Invest took the other side, buying about $16.3 million across October 13 and 14 ⁶.

A $25.4 billion backlog, and only a fifth converts by 2028

The $25.4 billion backlog stands against a full-year core revenue guide of $880 to $890 million↗, after second-quarter core revenue more than doubled to about $210 million, and the stock carries a 217-times forward earnings multiple ¹⁰. Cerebras expects only about 22 percent of that backlog, near $5.6 billion, to become revenue by June 2028↗, on timing a customer can shift ¹¹. Three readings settle it: whether OpenAI or Cerebras names the silicon behind Ultrafast, whether CS-4 ships this quarter, and the third-quarter report against its $214 to $216 million revenue guide ⁸¹¹.