Cerebras Now Needs GPUs: 5x Tokens per MW and a Supply Risk

Cerebras now runs AI inference with GPUs handling the prompt and its wafer-scale systems writing the answer, according to a launch post it published on Tuesday. The company says the setup is in production and delivers “five times more tokens per megawatt at the same ultrafast Cerebras speeds”.

Our finding: Cerebras’s 10-Q for the September quarter, filed on 12 November, added a risk factor on buying GPUs that its June-quarter 10-Q did not contain, by our search. It says Cerebras “must acquire GPUs, which are in short supply”.

Business Pill · Concentration Explained

An evergreen Business Engineer explainer on customer and supplier concentration. It teaches the general idea only and does not discuss this launch.

The key insight: Cerebras’s prospectus set the direction in May; what its September-quarter 10-Q adds is the cost: a new risk factor on buying GPUs, which it says are in short supply, while two customers made 72% of its quarterly revenue, by our arithmetic.

What Cerebras Launched

An inference request has two stages. Prefill processes the prompt; decode writes the answer one token at a time. The post says decode depends on memory bandwidth, and that its wafer chip has “350× the SRAM capacity and 5,400× the memory bandwidth of a typical GB300 GPU”. Those are Cerebras’s own figures.

In the production setup on the Cerebras Inference Cloud, AMD Instinct GPUs process prompts and CS-3 systems generate responses. The post says “this combination delivered 5X the throughput per megawatt of a CS-3-only configuration, with no loss in token generation speed”.

Cerebras also sells the setup to others. Partners get a reference architecture and either a full software stack or a decode-only SDK. The post names Amazon Web Services, Fireworks AI, Gimlet Labs and General Compute as early-access partners, and says Cerebras plans to deploy the setup “across hundreds of megawatts in the Cerebras Inference Cloud and in partner deployments by the end of 2027”.

Share of Cerebras total revenue in the three months to 30 September 2026: MBZUAI 38%, OpenAI 34%, G42 7%, all
Share of Cerebras total revenue in the three months to 30 September 2026: MBZUAI 38%, OpenAI 34%, G42 7%, all other customers 21%. Source: Cerebras Systems 10-Q filed 12 November 2026. Our arithmetic: 38% + 34% = 72%; all other = 100% minus 79%.

What Was Posted on X

Cerebras posted its launch article on X on 17 November 2026. It describes GPUs handling prompt processing and Cerebras systems handling output generation, reports 5X the throughput per megawatt of a CS-3-only configuration in a production deployment, and names early-access partners. This piece quotes the article and compares it with Cerebras’s SEC filings.

Is This a Change of Strategy?

Not in direction. Cerebras’s May IPO prospectus already put disaggregation on its roadmap and said it “would allow Cerebras to operate alongside other architectures, serving as the high-performance engine for decode while other systems handle prefill”.

The prospectus also described a deal with AWS to pair AWS Trainium3 chips with CS-3 systems, which it said was “expected to deliver 5 times more token throughput in the same hardware footprint”. The 5x figure was a target in May; on Tuesday Cerebras reported it from a production deployment.

What is new is the plumbing and the bill. The production pairing uses AMD Instinct GPUs, the post opens the software to outside partners, and the September-quarter 10-Q describes a contest for GPU supply: “We compete for GPUs with well-capitalized technology companies and AI infrastructure providers that may outbid us”.

What the 10-Q Adds

The new risk factor says “We have recently added disaggregated inference solutions to our product roadmap”, and that GPU supply may require premiums, deposits or long-term purchase commitments before related revenue arrives. The June-quarter 10-Q does not contain the phrases “secure sufficient GPUs” or “prefill”, by our search; the word GPU appears 3 times there and 16 times in the September filing, our count.

The same 10-Q shows how concentrated the business is. MBZUAI made up 38% of revenue in the quarter, OpenAI 34% and G42 7%, so two customers made 72%, by our arithmetic. We covered the 79% three-customer share and Cerebras’s delays on OpenAI’s 750MW last week.

The Business Lever

The lever is capacity per megawatt. Power, not chips alone, caps how many tokens a data centre can sell, and the post frames disaggregation as “greater revenue potential from the same power budget” for inference providers. Selling the stack to partners also gives Cerebras a route beyond its two largest customers.

The cost is a new input. Each disaggregated cluster needs GPUs as well as Cerebras systems, and the 10-Q says GPU procurement “could require significant upfront cash outflows and increase our capital needs”. The GPU market it now buys from is the one Nvidia reported on this week (our read of Nvidia’s filing).

The Structural Read

As we read the filings, the launch is a roadmap item moving into production rather than a reversal: the May prospectus already described Cerebras working alongside other architectures, with other systems handling prefill.

As we read the partner terms, Cerebras is selling software and a reference design as well as systems, which the post says partners can adopt in their own data centres.

Cerebras, Cerebras Launches Disaggregated Inference, 17 November 2026

“The Cerebras disaggregated inference solution was also built for our partners to adopt.”

Three Implications

FOR INFERENCE PROVIDERS The post says pairing a prefill accelerator with Cerebras systems increases token throughput per megawatt while keeping Cerebras generation speeds.

FOR GPU MAKERS The production pairing named in the post uses AMD Instinct GPUs for prefill, and the 10-Q says Cerebras must acquire GPUs to deploy the setup.

FOR CEREBRAS INVESTORS The 10-Q lists MBZUAI at 38% and OpenAI at 34% of September-quarter revenue, and says GPU purchases could increase its capital needs.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of The Inference Economy.

The framework says: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”

As we read it, Cerebras is aiming at the cost of each query: its post frames disaggregation as more tokens from the same power budget. The framework predates this launch and does not discuss it.

What Is Not Established

The 5x figure and the SRAM and bandwidth comparisons are Cerebras’s own measurements; we have not seen independent tests. The post does not give prices, the number of megawatts now running the setup, or what partners pay.

The post names no customer for the AMD-based production cluster, and the 10-Q does not quantify any GPU purchase commitments. Earlier Cerebras work we covered includes its Tungsten programming language and Bilt’s agent deployment. We did not contact Cerebras.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The State of The Inference Economy

A Business Engineer analysis of how inference turns AI compute from a cost centre into a revenue engine.

Read the Map of AI →

The Bottom Line

Cerebras put GPU prefill plus wafer-scale decode into production, reporting 5x the tokens per megawatt of a CS-3-only setup, and opened it to partners including AWS. The direction was in its May prospectus; the September-quarter 10-Q adds a new risk: Cerebras now has to buy GPUs, which it says are in short supply.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Figures and quotations are from Cerebras’s launch post “Cerebras Launches Disaggregated Inference” (17 November 2026), read in full, its May 2026 IPO prospectus and its 10-Q filings for the June and September 2026 quarters. Performance figures are Cerebras’s own and have not been independently tested. The 72% share and the GPU word counts are our arithmetic and our search of the filings. We did not contact Cerebras. Not investment advice; for information only.

Sources: Cerebras, Cerebras Launches Disaggregated Inference, 17 Nov 2026 · @cerebras on X, 20:37 UTC, 17 Nov 2026 · Cerebras Systems prospectus (424B4), 14 May 2026 · Cerebras Systems 10-Q, quarter to 30 Sep 2026, filed 12 Nov 2026

Scroll to Top
FourWeekMBA