OpenAI now decides what Cerebras's wafer is worth
Cerebras proved this month that its wafer serves trillion-parameter models fast, and the model it showcased was open-weights, not OpenAI's. With about $20 billion of $25 billion in contracted business tied to OpenAI and the last lock-ups expiring before the November 12 report, one buyer now sets the wafer's price.
Richard Tang · 3 min read
On October 22 Cerebras published the engineering behind its sales pitch: the open-weights Kimi K3 runs across 59 of its dinner-plate wafers, whole layers of the model resident in on-chip memory, only intermediate results moving between wafers 1. The post reached Hacker News over the weekend 2. The telling detail is the model it chose to showcase: open-weights, not OpenAI's$500B — OpenAI, private, latest valuation $500B. The wafer's speed is settled; its price is not, and one customer sets it.
That customer is OpenAI, behind about $20 billion of the $25 billion Cerebras+1.52% — Cerebras, up 1.52 percent today has contracted for future delivery, against roughly $210 million of core revenue last quarter 34. The pairing with AMD+1.59% — AMD, up 1.59 percent today, announced in July when AMD chief executive Lisa Su unveiled her Helios server system 5, deepens that dependence rather than easing it.
A second rival's silicon now sits in front of the wafer
In the joint system, AMD's Helios rack, holding up to 72 MI400-series accelerators, handles prefill, the first pass over a prompt; the Cerebras wafer handles decode, writing the answer token by token, with up to five times better energy efficiency claimed 6. Amazon Web Services−0.72% — Amazon Web Services, down 0.72 percent today already runs the same split with Cerebras, its Trainium chip on prefill and the wafer on decode 3.
That gives Cerebras a second way to sell decode, and OpenAI a second way to buy it. CBRS rose about 4% on Thursday to a one-month high, its fourth straight up session, and still trades roughly 72% below the $386 reached after its May listing 63.
Speed is settled, so the fight has moved to price
The two halves of inference have different physics. Prefill digests a long prompt in parallel, which racks of GPUs already do well; decode emits one token after another and is bound by memory bandwidth, which the wafer's on-chip memory supplies 63. Outside measurement backs the company's method: Artificial Analysis clocked the wafer at 981 output tokens a second on Kimi K2.6, an earlier Kimi release, 6.7 times the next-fastest GPU cloud 7. Coding agents live on fast decode. On speed, the argument is over.
One buyer can now buy decode two ways
What is unsettled is price. AMD says OpenAI plans to deploy Helios at scale 6. Reports keep asking whether the "Ultrafast" tier of OpenAI's GPT-6.1 Sol ran on Nvidia GPUs rather than Cerebras hardware, and Microsoft's+0.64% — Microsoft, up 0.64 percent today Maia 300 inference chip is arriving 8. If OpenAI can put its prompts on an AMD rack and still buy decode behind it, it no longer needs Cerebras's whole stack. A buyer holding two ways to buy pays less than one holding a single supplier.
Cerebras's contracted promises tower over what it bills each quarter
Data
| Value | |
|---|---|
| Total obligations | $25.4B |
| Obligations tied to OpenAI | $20B |
| Quarterly core revenue | $0.21B |
The calendar prices the stock before the earnings do
Nineteen million shares become tradable on October 28 and 14 million more on November 10, after which every share can trade 9. Insiders, among them the chief technology officer and chief financial officer selling under pre-arranged plans, took out nearly $388 million across three months 4. A lock-up expiry makes shares sellable; it does not show they were sold, and several sellers are funds illiquid for a decade taking profit rather than verdicts on the hardware 89.
The test is the third-quarter report on November 12 7. The number to watch is revenue tied to OpenAI against last quarter's $210 million base 3. If it is climbing, the backlog is turning real and AMD is a second sales channel. If it is flat, the wafer has a price ceiling that no speed benchmark removes.
Get Chips by email, free
How AI demand is reshaping chipmakers, memory and the fabs behind them. In our daily email, every weekday.
