OpenAI and Cerebras: Altman Reaffirms Chip Deal

Public reassurance lifts the chipmaker’s stock after rumors suggested a pivot toward Nvidia GPUs for real-time serving.

OpenAI chief executive Sam Altman stepped forward to quell market speculation of an operational split with chipmaker Cerebras Systems. In a post on X published on October 2 and digested by financial markets on October 5, Altman stated that OpenAI and Cerebras remain close partners engaged in pushing the limits of inference speed. The clarification brought relief to investors: Cerebras shares (CBRS) climbed between 7% and 9.5% during trading on Monday, October 5, regaining ground surrendered during the previous week.

First: understanding real-time inference

Inference describes the execution phase where an already trained AI model processes a user prompt and generates an answer. For asynchronous tasks, a delay of several seconds causes little trouble. In interactive programming assistants and autonomous agents, however, each token (the word fragment generated by the model) must appear on screen with near-zero latency to preserve a usable conversational flow.

Cerebras addresses this by manufacturing wafer-scale processors cut from an entire silicon disk rather than individual chips. Keeping memory and computing cores on a single giant slab minimizes interconnect delays. The architecture aims to deliver high-speed token generation, contrasting with heavy model training that runs across distributed GPU clusters for months.

What triggered the drop in Cerebras stock

Altman intervened after Wall Street research notes circulated in early October. Analyst commentary, including notes associated with Barclays, raised concerns about whether Cerebras could manufacture enough hardware on time to satisfy OpenAI’s deployment schedule.

Those reports indicated OpenAI had begun routing workload from its newly launched Ultrafast tier toward conventional Nvidia graphics cards. That speculation sent Cerebras shares tumbling almost 20% into the end of last week, given that the startup relies on the OpenAI contract as its primary commercial showcase for wafer-scale hardware.

How OpenAI and Cerebras divide workloads

In practice, OpenAI manages its computing infrastructure through what it terms load-matching systems rather than replacing one vendor with another. Sachin Katti, head of compute infrastructure at OpenAI, previously explained that the company brought in Cerebras hardware specifically for real-time token decoding, where low-latency delivery is essential.

Meanwhile, Nvidia GPU clusters handle pre-training foundational models, executing batch tasks, and ingesting expansive context windows. Rather than picking a single winner, the setup distributes specialized silicon across the parts of the pipeline where each design performs best.

The 750 MW contract and vendor claims

The two companies signed a multi-year agreement on January 14, 2026. The contract targets the staged delivery of 750 megawatts of computing capacity dedicated to model inference through 2028. Financial reporting from The Wall Street Journal and Reuters estimates the deal exceeds $10 billion over its lifetime, although neither company disclosed exact transaction values in official SEC filings.

In benchmarks published directly by Cerebras, the vendor claims that GPT-5.6 Sol reaches up to 750 tokens per second in Ultrafast mode, while GPT-5.3-Codex-Spark tops 1,000 tokens per second. Because these measurements originate solely from the hardware supplier, they warrant the typical skepticism applied to promotional vendor benchmarks.

What remains unconfirmed

While Altman’s post halted the selloff on Wall Street, operational uncertainties persist. OpenAI has not revealed what proportion of its Ultrafast query traffic actually executes on Cerebras silicon versus traditional Nvidia hardware.

Neither party has addressed potential manufacturing bottlenecks that could complicate delivering the 750-megawatt plan by 2028. For observers tracking enterprise AI infrastructure, the episode demonstrates the ongoing effort by major labs to build multi-vendor supply chains, trading the simplicity of a single ecosystem for lower unit costs and faster token delivery.

Sources

• Official: Sam Altman’s post on X
• Official: OpenAI partnership announcement
• Vendor: Cerebras high-speed inference announcement and Cerebras benchmarks
• Press: Market coverage on CNBC
• Press: Stock analysis on Yahoo Finance
• Press: Market update on Investopedia
• Press: Investment reporting on Silicon Republic
• Market analysis: Barclays Research and market reaction on TradingView