Nvidia has published its own account of the OpenAI relationship that sent Cerebras shares to record lows, confirming on Friday that GPT-6 Astra Ultrafast, the speed optimized variant of OpenAI’s new flagship, runs on Nvidia Blackwell GPUs and is now available in the OpenAI API to eligible ChatGPT Work and Codex users.
The confirmation is direct rebuttal to a fast moving dispute over whose hardware powers OpenAI’s newest release. A day earlier, a widely shared report said OpenAI runs GPT-6.1 Sol Ultrafast on Nvidia chips rather than on Cerebras hardware, and Cerebras stock fell 20 percent to its lowest level since its May IPO, erasing much of the premium the market had priced in around the company’s OpenAI order flow.
Nvidia’s post goes beyond a vendor mention. It describes Blackwell GPUs running the full OpenAI speed tier in production, with availability in the OpenAI API and to eligible enterprise and developer users already live. Nvidia published the item on its newsroom alongside a broader October slate, including a 64GB DGX Spark desktop for local AI work at $4,999 and telecom partnerships building inference grids on distributed networks.
For Cerebras, whose wafer scale inference systems had been ports of call for AI labs chasing token throughput no GPU cluster could match, the report lands at a delicate time. The company’s IPO in May was sold to investors in large part on the OpenAI relationship, and any sign that OpenAI routes its fastest, highest volume serving tier back to Nvidia chips erodes the differentiated demand story. The stock closed at post IPO lows on the news.
Why serving tier, not training, decides the verdict
OpenAI buys from both camps, but the two purchases carry different weights. Training runs consume a fixed compute bill for a bounded stretch of months, and analysts argue about efficiency gains. Serving consumes variable compute at market scale, every day, forever, which is why each new serving tier assignment moves valuations more than any training deal.
GPT-6 Astra Ultrafast is OpenAI’s low latency serving variant, and routing it onto Blackwell tells customers the fastest responses on the market now come off Nvidia racks. That is a direct, revenue level signal about which hardware wins when latency carries a price.
The 64GB DGX Spark at $4,999, shipping October 23, rounds out the picture on the other end of the market. Nvidia is compressing local AI hardware pricing while DRAM costs climb, betting that developer mindshare at the low end defends share at the high end, where the Blackwell racks serving GPT-6 variants live. The existing 128GB Spark jumped to $6,950 in the same announcement, a sign that memory supply, not chips, is the material bottleneck for the next twelve months.
What this means for the compute market
Three forces are likely to shape the rest of the year, none of them friendly to Cerebras.
| Force | Direction | Why it hurts now |
|---|---|---|
| OpenAI serving mix | Shifted toward Nvidia for the ultrafast tier | Removes the very contract the IPO story rested on |
| Nvidia pricing and supply | Blackwell and DGX Spark pricing expanding across tiers | Cerebras has to win on economics, not on the fact of availability |
| Customer concentration | A small set of labs accounts for most of Cerebras revenue | One order change moves the whole model |
None of this means Nvidia ownership clears the field. Cerebras still sells real throughput advantages, and its wafer scale systems win on token per dollar at specific batch sizes. But a serving tier that runs on Blackwell removes the urgency that once pushed labs toward exotic inference hardware, and the stock’s reaction shows how much of the company’s valuation had been built on urgency rather than on arithmetic.
The other labor question dwarfs them all. If OpenAI’s fastest serving tiers cycle back onto Nvidia inventory, the Cerebras thesis needs either a second large lab customer or a self hosted enterprise pipeline that consumes its hardware directly. Neither has materialized at scale so far, and the stock is now pricing that absence.
The deal currently under strain
A second problem compounds the first. Cerebras’ IPO filing and subsequent disclosures show how heavily its order book leans on OpenAI and, more recently, on the compute brokers that intermediate between the labs and the hyperscalers. An Nvidia disclosure this week showed OpenAI has committed to pay up to $4 billion through April 2029 under an additional agreement with the Nvidia backed cloud provider Coreweave, which some of the initial investors once treated as an indirect Cerebras channel.
When OpenAI announces speed tiers on Nvidia hardware and deepens its Coreweave agreement in the same two week window, the inference hardware market is repricing OpenAI’s actual supplier map, not just a rumor about one contract. That is the headwind Cerebras faces as it tries to broaden its base.
The timing is also awkward because Cerebras has spent the quarter positioning its systems as throughput accelerators for exactly this class of model variant. Whatever the long run economics, the first half of the GPT-6 rollout belongs to the incumbent, and proving out a share gain requires a customer willing to publish results, something no large lab has volunteered in this cycle.
For Nvidia, meanwhile, the story is almost boring: the market leader keeps taking the workloads that matter and charges accordingly. Blackwell racks, DGX boxes for the garage, telecom inference grids, and now a named OpenAI serving tier running on its silicon, all inside one month. The question the market will ask next is not whether Nvidia wins the serving cycle but how long Cerebras can survive a demand curve that no longer bends toward it.