Shares of Cerebras Systems Inc. (CBRS) slid between roughly 7% and 9% this week after a report raised questions about whether the chipmaker still has a role in powering OpenAI's fastest AI model — a service tier it had been publicly named to support. The stock dropped as much as 8.87% during Wednesday's session and closed at around $180.22, remaining down close to 8% as of Wednesday afternoon.

The trigger came from SemiAnalysis, a semiconductor research firm that said in a social media post that OpenAI's new GPT-6.1 Sol Ultrafast model is running on standard Nvidia GPUs at low batch size, not on Cerebras' flagship Wafer-Scale Engine.

The claim carries weight because of what Ultrafast was supposed to represent. OpenAI launched the tier with a promise of speeds up to eight times faster than normal — roughly 300 tokens per second, the standard measure of how quickly a model generates text — and publicly announced in August that Cerebras hardware would power the mode.

Neither Cerebras nor OpenAI has confirmed or denied the SemiAnalysis claim, and neither company has issued any public statement on the matter. The silence has left analysts guessing rather than working from anything official.

What the SemiAnalysis Report Said

The pivotal detail is a phrase buried in the report: “low batch size.” In AI inference, batching refers to grouping requests together so that a chip processes many of them in a single pass and runs more efficiently. Nvidia GPUs typically need large batches to reach peak performance, while a low batch size means fewer requests handled at once, favoring speed over raw efficiency.

That is precisely the niche Cerebras has marketed itself around. Serving individual users quickly, rather than working through large queues of requests, is exactly where the company has claimed an advantage. Its Wafer-Scale Engine is a massive chip that fits billions of cores and huge pools of memory onto a single piece of silicon, a design that avoids the slowdowns caused by shuttling data between hundreds of separate GPUs.

If OpenAI has found a way to run a top-tier model on Nvidia hardware at low batch size without sacrificing cost efficiency, it erodes one of Cerebras' central selling points — which helps explain why traders reacted so quickly.

A Steep Slide From the IPO

Cerebras went public in May at an IPO price of $350, and the stock has struggled since. At Wednesday's levels, shares were changing hands at roughly half their debut price. The listing came with considerable hype: investors were drawn to a chip company that claimed it could outperform Nvidia on raw inference speed for certain workloads. That enthusiasm has since cooled, and earnings reports after the IPO failed to match the excitement investors showed on day one.

Why OpenAI's Involvement Matters So Much

OpenAI is not just another client for Cerebras. It is the company's largest customer by revenue backlog, according to Barron's, and Cerebras has leaned heavily on the relationship to justify its valuation. Losing ground with OpenAI, even temporarily, therefore raises real questions for investors.

There is also a chance the situation is not permanent. Capacity constraints or technical optimization issues could explain why Nvidia chips are handling the workload at the moment, and Cerebras could still end up powering Ultrafast mode down the road once any technical kinks are worked out. Nothing from either company rules that out.

For now, the stock reaction reflects uncertainty more than confirmed bad news. Investors are pricing in the risk that Nvidia's dominant CUDA software ecosystem and continued hardware improvements make it difficult for any challenger to lock in exclusive deals with major AI labs. Cerebras has not issued any statement confirming or denying its current role in OpenAI's infrastructure plans. Until one of the two companies breaks its silence, any official comment — or a disclosure clarifying which hardware serves Ultrafast — would be the most direct way to establish whether the report reflects a lasting change or a temporary arrangement.

This article first appeared on CoinCentral.