Cerebras Systems

Overview

Conventional semiconductor manufacturing cuts a wafer into hundreds of individual dies. Cerebras does not. The Wafer Scale Engine is a single processor occupying almost an entire 300mm wafer, and everything else about the company follows from that one decision — the cooling, the power delivery, the packaging, the software model and the sales motion are all consequences of building a chip roughly fifty times the area of a large GPU.

The company was founded in 2015 by Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie and Jean-Philippe Fricker, several of whom had previously built SeaMicro, a server company acquired by AMD. It spent its first several years positioned mainly around training and scientific computing. The pivot to inference — marketed as "fast inference" and sold both as systems and as cloud capacity — is what turned it into a company with the revenue trajectory it now reports.

On 18 August 2026 the company launched the CS-4, its first rack-scale system, built around a new WSE-3 Turbo processor. Notably, both are derived from existing silicon rather than a new wafer-scale generation: the WSE-3 Turbo has the same 4 trillion transistors, 900,000 cores and 44GB of on-wafer SRAM as the WSE-3, on the same TSMC 5nm process, but runs at double the clock. Cerebras is extracting a further generation of performance from a design it has already paid to develop, which is a favourable position for gross margin — though the reported roughly doubled power draw means the efficiency gain is smaller than the performance gain.

Cerebras listed on Nasdaq on 14 May 2026 at $185 per share, raising $5.55 billion. Shares rose 68% on debut, briefly lifting market capitalisation close to $100 billion. It is the first of the current generation of AI inference specialists to complete an IPO, which makes its disclosures unusually valuable for anyone trying to understand the economics of the sector.

Chip and product families

Role in inference

The wafer-scale argument for inference is straightforward. The decode stage of language model inference is bound by how fast weights and cached attention state can be moved to compute units. On a GPU cluster that movement crosses memory buses, package boundaries and network links. On a wafer-scale engine, much of it does not leave the wafer.

In practice this produces very high single-stream token generation rates — the property Cerebras markets as fast inference, and the reason its results are most compelling for latency-sensitive workloads: interactive assistants, agentic systems that chain many sequential model calls, voice, and reasoning models that generate long outputs before returning an answer. Andrew Feldman has framed the company's competitive position specifically around tasks requiring low latency rather than around aggregate throughput.

The corresponding weakness is that wafer-scale is an unusual thing to put in a data centre. Power density, cooling and physical integration are non-standard, the unit of purchase is large, and there is no path to buying a small quantity to test. This is a systems sale with a long cycle, which is very different from adding GPU instances to an existing cloud account.

Financials and the 2026 results

Cerebras reports both GAAP revenue and a "core revenue" measure that excludes pass-through revenue and the amortisation of customer warrant assets. The distinction matters: in Q2 2026 GAAP revenue was $180 million against roughly $194 million expected, while core revenue was $209.9 million, up 103% year over year. Warrant amortisation associated with the OpenAI agreement was the principal reason for the gap.

Management stated that the $25.4 billion remaining performance obligation balance does not include any backlog from AWS or other hyperscalers, and projected that core revenue would more than triple in 2027. Despite the raised guidance, the stock fell around 14% following the Q2 report — a reminder that for a company priced on growth, beating guidance and satisfying the market are not the same thing.

Core revenue is a non-GAAP measure defined by the company. It is reported here because Cerebras uses it for guidance and because market data providers have adopted it, but readers comparing Cerebras to other companies should be aware that the GAAP and non-GAAP figures differ materially and that the difference is driven by customer warrant accounting.

Customers and partners

The defining commercial relationship is with OpenAI. Cerebras has disclosed a multi-year agreement valued at more than $20 billion, under which it has committed to deliver 750 megawatts of AI inference compute capacity through 2028 — a component the company has described as worth more than $10 billion. This single relationship explains most of the $25.4 billion remaining performance obligation figure and most of the warrant accounting that separates GAAP from core revenue.

Historically, Cerebras' revenue was heavily concentrated in G42, the Abu Dhabi-based technology group, a concentration that was widely discussed during the company's earlier attempt to go public. The OpenAI agreement changes the composition of that concentration without eliminating the underlying characteristic: this remains a business where a small number of very large agreements determine the outcome.

Strengths and key questions

Leadership

Andrew Feldman is co-founder and chief executive, and previously co-founded SeaMicro, which AMD acquired in 2012. His stake was reported at roughly $3.2–3.4 billion following the IPO. Sean Lie, also a co-founder, serves as chief technology officer and is the most visible public explainer of the wafer-scale architecture. The founding team's prior experience building and selling a server company is directly relevant, because Cerebras' product is a systems business rather than a component business.

What to watch

• Capacity delivery milestones. Progress against the 750 MW commitment is the clearest test of whether backlog converts to revenue.
• Customer diversification. Whether AWS or another hyperscaler appears in backlog, which management explicitly noted is not currently included.
• GAAP revenue convergence. Whether the gap between GAAP and core revenue narrows as warrant amortisation works through.
• Independent latency benchmarks. Third-party measurements across model sizes, context lengths and concurrency levels, not single-stream demonstrations.
• CS-4 shipping and power. Cerebras has not disclosed CS-4 power consumption; independent estimates put the WSE-3 Turbo near 54kW, which would place real constraints on where it can be sited. Availability dates were not given at launch.
• WSE-4. Any disclosure on a genuinely new wafer-scale generation and its process node, as distinct from further overclocking of the WSE-3 design.
• The 30× claim. Cerebras launched the CS-4 under a headline of being up to 30 times faster than GPU-based solutions. The comparison basis, model and batch configuration matter enormously and should be checked against independent measurement.

Sources

1. CNBC — Cerebras Q2 2026 earnings report (12 August 2026). GAAP and core revenue, guidance raise and share reaction.
2. Cerebras — CS-4 launch announcement (18 August 2026). Company statement of CS-4 positioning and performance claims.
3. ServeTheHome — WSE-3 Turbo and rack-scale CS-4 (19 August 2026). Independent specification detail and power analysis.
4. The Next Platform — Cerebras overclocks WSE-3 for the Nexus CS-4. Clock speed and architecture analysis.
5. Cerebras Investor Relations — First quarter 2026 results. Company-reported quarterly detail.
6. CNBC — Cerebras IPO (14 May 2026). Offering price, proceeds, first-day performance.
7. Forbes — Andrew Feldman after the Cerebras IPO. Founder stake and background.
8. Seeking Alpha — FY2026 core revenue guidance and margin target. Guidance range and gross margin target.
9. Cerebras — $1B Series H (February 2026). Final private round before listing.
10. Tech Times — Cerebras cloud growth and OpenAI warrant accounting. Explanation of the GAAP-to-core difference.

As a listed company, Cerebras' filings on SEC EDGAR are the authoritative record and supersede any figure reported here. See the editorial methodology for the source hierarchy used on this site.

Related pages