Cerebras Systems Inc. (CBRS)

Cerebras makes one very large processor, the Wafer-Scale Engine. Chipmakers normally cut a silicon wafer into many small chips; Cerebras keeps the whole wafer as one, which puts 900,000 compute cores and 44 gigabytes of fast memory on a single piece of silicon. Each engine sits in a CS-3 system with its own cooling, power delivery and interconnect, and many CS-3s link into what it calls AI supercomputers, used both to train AI models and to run them, which the industry calls inference. 9101112

How it operates

Cerebras makes one very large processor, the Wafer-Scale Engine. Chipmakers normally cut a silicon wafer into many small chips; Cerebras keeps the whole wafer as one, which puts 900,000 compute cores and 44 gigabytes of fast memory on a single piece of silicon. Each engine sits in a CS-3 system with its own cooling, power delivery and interconnect, and many CS-3s link into what it calls AI supercomputers, used both to train AI models and to run them, which the industry calls inference. 22232425

It earns money in two ways, reported as two lines. Hardware: systems sold for a customer's own data center, each with a renewable software subscription. Cloud and other services: computing time on Cerebras Cloud or through partner marketplaces including AWS Marketplace, Microsoft Marketplace, IBM watsonx Model Gateway, OpenRouter, Hugging Face and Vercel AI Gateway, bought by the token (the unit of text a model reads or writes), by the week or month for training, or as dedicated capacity on take-or-pay terms, plus software support agreements of one to five years. 2627282930

So far its buyers have been few, and mostly in Abu Dhabi. In 2025 the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) took 62.0% of revenue and G42, an Abu Dhabi technology group, 24.0%; in 2024 G42 alone took 85.0%. The prospectus treats the two as related parties to each other and says a significant majority of 2025 revenue came from customers headquartered in the United Arab Emirates. By the customer's billing location, the region of Europe, the Middle East and Africa brought in most of 2025 revenue, after almost none in 2024. 313233343536

Two new relationships change the list. In December 2025 OpenAI, which had chosen Cerebras for fast inference behind its Codex-Spark coding tool, signed a Master Relationship Agreement to buy 750 megawatts of inference computing capacity, delivered in tranches from 2026 through 2028, each for three or four years, with an option on 1.25 gigawatts more by the end of 2030. OpenAI lent Cerebras about 1.0 billion dollars to build it and holds a warrant for up to 33,445,026 non-voting shares; no revenue from the deal was recognized in 2025. At the end of 2025 the company counted 24.6 billion dollars of contracted revenue not yet recognized, a significant amount of it from OpenAI. In March 2026 AWS signed a term sheet, binding on pricing, exclusivity and minimum capacity, to become the first hyperscaler, meaning one of the giant cloud providers, to deploy Cerebras systems in its own data centers. 373839404142434445

Cerebras owns no chip factory. It depends on one foundry, TSMC, to make its processor, and buys servers, programmable chips, network switches, cooling units, interconnects and chassis from other suppliers, some of them sole or single sources. Apart from some assembly, packaging and testing steps it does itself, contract manufacturers build the systems, and it generally has no long-term capacity commitments from its suppliers: substantially all orders go purchase order by purchase order. 46474849

The specifics

Its founding bet was to make a chip the size of a whole wafer work. By its account the largest chip before it was about 840 square millimeters; it invented a way to connect the separate dies across the full 46,225 square millimeters of a wafer, and an architecture that shuts down and routes around the flaws every wafer carries. It has delivered three generations, at 16, 7 and 5 nanometers, and says it has run commercial wafer-scale systems since 2020. As of March 2026 it owned 96 issued patents, 50 of them in the United States, where they expire between 2038 and 2041. By its own comparison the WSE-3 is 58 times the size of NVIDIA's B200 chip, and its inference runs up to 15 times faster than leading GPU systems on leading open-source models. 505152535455565758

It names its competitors: NVIDIA, AMD and Intel; AI chips designed by hyperscalers and private companies; the full-service clouds of Amazon's AWS, Microsoft's Azure, Alphabet's Google Cloud Platform and Oracle; and specialized AI clouds such as CoreWeave. AWS is thus both a rival and, under the new term sheet, a customer. 5945

Export controls and national security review shape its biggest relationships. Since October 2023 US rules have required licenses for exports of certain chips and components to countries including the United Arab Emirates, home to G42 and MBZUAI; Cerebras holds licenses for its CS-2, CS-3 and future CS-4 systems for both, on condition of rigorous security and compliance duties against diversion. In 2024 an affiliate of G42 agreed to buy 335 million dollars of preferred stock, and the two filed with CFIUS, the committee that reviews foreign investment in the United States. After engaging with CFIUS they agreed to remove G42 from the purchase, CFIUS let them withdraw the filing on March 27, 2025, the purchase agreement was restated without G42, and a separate G42 agreement from May 2024 was terminated. 60616263646566

The prospectus reports material weaknesses in internal control over financial reporting, most recently for 2025 and 2024: too few staff with the expertise to review account reconciliations and apply US accounting rules in revenue recognition, inventory, data center assets and equity, and inadequate IT controls, including poor separation of duties. It says it is hiring accounting and finance staff to fix them. The document read here is the prospectus for its initial public offering of 30,000,000 Class A shares, to be listed on Nasdaq, filed in May 2026. 6768697071

What drives growth, line by line

Revenue rose 76% in 2025 to 510.0 million dollars, after 290.3 million in 2024, 78.7 million in 2023 and 24.6 million in 2022. The 2025 net income of 237.8 million dollars did not come from operations, which lost money in both years: it rests on a gain of 363.3 million dollars from settling a forward contract liability, an accounting item tied to investors' commitments to buy its preferred shares. On its own non-GAAP measure, which strips out that gain and stock-based pay, it lost 75.7 million dollars. 727374757677

Hardware. The larger line, about seven tenths of 2025 revenue: up 69%, an increase of 146.5 million dollars, which the filing puts down mainly to demand from MBZUAI for systems in its own data centers. 7879

Cloud and other services. About three tenths of revenue, up 94%, an increase of 73.3 million dollars: more support and services revenue from the larger base of systems sold to G42 and MBZUAI, and growing demand for its cloud inference in the second half of 2025. Its cost grew faster, by 252%, and gross margin for the whole company fell to 39.0% from 42.3% on higher data-center costs for the cloud service. 80818283

What comes next, in the company's own words: it expects the OpenAI agreement to make cloud and other services a significantly higher share of revenue; it expects gross margin to be significantly lower in the near term, from warrants granted to customers that it books as reductions of revenue from the first quarter of 2026, data-center costs passed through to a customer, and start-up costs for new cloud capacity; and it expects to recognize about 15% of its 24.6 billion dollars of contracted revenue by the end of 2027. 8485868743

The read's section on the people who run it is on the People tab.

What the desk infers Inferred

1. Inferred

   G42's purchases in 2024 were billed to a location in the United States, which suggests the systems it bought that year were installed there rather than in Abu Dhabi.

   Because

   • G42 took 85.0% of 2024 revenue. 31
   • The prospectus assigns revenue to regions by the customer's billing location, and in 2024 the United States accounted for 282,685 thousand dollars of revenue while Europe, the Middle East and Africa accounted for 7,567 thousand. 363531
   • In 2025, when MBZUAI, an Abu Dhabi university, drove hardware demand for its own premises, revenue billed in Europe, the Middle East and Africa rose to 322,231 thousand dollars. 793235
   • Exports of its systems to the United Arab Emirates need US licenses that carry rigorous security and anti-diversion conditions. 6061
   Confidence 60 of 100Wrong if A filing or disclosure showing that the systems G42 bought in 2024 were delivered to sites outside the United States.
2. Inferred

   Cerebras is turning from a seller of machines into an operator that rents out computing time from data centers, which ties up more of its own money in equipment and, by the company's own account, weighs on its gross margin.

   Because

   • Its cloud cost of revenue rose 252% in 2025 while cloud and other services revenue rose 94%. 808281
   • Gross margin fell to 39.0% from 42.3% on higher data-center costs for the cloud inference service. 83
   • It spent 382.7 million dollars on property and equipment in 2025, mainly systems to deliver Cerebras Cloud services. 88
   • It expects the OpenAI agreement to make cloud and other services a significantly higher share of revenue, and gross margin to be significantly lower in the near term. 8485
   Confidence 75 of 100Wrong if A year in which cloud and other services rise as a share of revenue while gross margin rises and spending on its own equipment falls.
3. Inferred

   Securing data-center space and power on schedule has become as central to Cerebras as designing chips, because its largest contract is sold in megawatts with delivery deadlines.

   Because

   • OpenAI committed to buy 750 megawatts of capacity, deployed in tranches from 2026 through 2028. 3839
   • If Cerebras misses the delivery timelines or a set level of service, OpenAI may terminate part or all of the agreement. 89
   • The company says it must find and keep long-term data-center leases for both power and space to keep pace with demand. 90
   • OpenAI's deployment choices decide how much of the capacity is booked as hardware and how much as cloud services. 91
   Confidence 70 of 100Wrong if A filing showing that most of the OpenAI capacity is delivered as hardware installed in OpenAI's own data centers rather than as capacity Cerebras hosts.

We are building the fastest AI infrastructure in the world.

OpenAI, the world’s leading foundation model lab, selected us to be its fast inference solution.

Amazon Web Services (“AWS”), the world’s leading hyperscale cloud, has signed a binding term sheet with us to become the first hyperscaler to deploy Cerebras in its own data centers, providing massive distribution to a broad base of enterprise customers.

What it warns about, in its own words

• We may not sustain our growth rate, and we may not be able to manage future growth effectively.
• We have a history of generating net losses, and if we are unable to achieve adequate revenue growth while our expenses increase, we may not achieve and maintain profitability in the future.
• A substantial portion of our revenue has been, and is expected to continue to be, driven by a limited number of customers.
• We identified material weaknesses in our internal control over financial reporting.
• Our offerings are subject to U.S. export controls and may be exported outside the United States only with the required export license or through a license exception.

Past

Cerebras listed in May 2026 , so almost no past exists on file. Revenue in the first quarter of 2025 was 99.5 million dollars and rose to 180.1 million dollars in the second quarter of 2026 Revenue rose 76% in 2025 to 510.0 million dollars, after 290.3 million in 2024 Insiders have sold 268 times since listing The filings since 2024 show no warning events: no bankruptcy or receivership notice, no debt default notice, no exchange listing notice, no restatement warning, no auditor change, no material cyber incident and no late-report notice 12345

Present

The latest quarterly report is for the period ended June 30, 2026 Leverage is 21 percent, meaning 21 cents owed per dollar of assets Track record is an SI estimate at 5.1 of 10 with a confidence of 20 of 100 because fewer than three filed years of revenue are on file Financial health is an SI estimate at 4.8 of 10 with a confidence of 20 of 100 because no balance sheet is on file Leadership is an SI estimate at 9.9 of 10 with a confidence of 35 of 100, read from officer and director changes on Form 8-K over three years, which rank in the top 1 percent of the sector Andrew D. Feldman co-founded Cerebras in 2016 and has been its chief executive, president and a director since April 2016 6253

Insiders have sold 408.4 million dollars of their own shares and bought zero since the company began filing. By its own prospectus, Cerebras is building the fastest AI infrastructure in the world, with OpenAI as an inference customer and a binding term sheet with AWS The prospectus discloses material weaknesses in internal control over financial reporting 74

Future

The quarterly report for the quarter ending September 30, 2026 is due by November 14, 2026, under the SEC's rules, not a prediction of when it will file The annual report for the fiscal year ending December 31, 2026 is due by March 31, 2027, under the SEC's rules A first full annual report is filed by March 31, 2027, with confidence 85 in 100, graded by that filing date, because the filing calendar, not the business, drives this one Room to expand is an SI estimate at 5.1 of 10 with a confidence of 20 of 100 because the record does not carry enough parts Growth drivers is an SI estimate at 5.0 of 10 with a confidence of 20 of 100 because the record does not carry three filed years of revenue 685

The filing names tariffs or export controls among its risk factors, with confidence 90 in 100, graded by the first 10-K, because nearly every chip company's annual report does 8

How the company operates From the company read

Cerebras makes one very large processor, the Wafer-Scale Engine. Chipmakers normally cut a silicon wafer into many small chips; Cerebras keeps the whole wafer as one, which puts 900,000 compute cores and 44 gigabytes of fast memory on a single piece of silicon. Each engine sits in a CS-3 system with its own cooling, power delivery and interconnect, and many CS-3s link into what it calls AI supercomputers, used both to train AI models and to run them, which the industry calls inference. 9101112

It earns money in two ways, reported as two lines. Hardware: systems sold for a customer's own data center, each with a renewable software subscription. Cloud and other services: computing time on Cerebras Cloud or through partner marketplaces including AWS Marketplace, Microsoft Marketplace, IBM watsonx Model Gateway, OpenRouter, Hugging Face and Vercel AI Gateway, bought by the token (the unit of text a model reads or writes), by the week or month for training, or as dedicated capacity on take-or-pay terms, plus software support agreements of one to five years. 1314151617

So far its buyers have been few, and mostly in Abu Dhabi. In 2025 the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) took 62.0% of revenue and G42, an Abu Dhabi technology group, 24.0%; in 2024 G42 alone took 85.0%. The prospectus treats the two as related parties to each other and says a significant majority of 2025 revenue came from customers headquartered in the United Arab Emirates. By the customer's billing location, the region of Europe, the Middle East and Africa brought in most of 2025 revenue, after almost none in 2024. 181920212223

Two new relationships change the list. In December 2025 OpenAI, which had chosen Cerebras for fast inference behind its Codex-Spark coding tool, signed a Master Relationship Agreement to buy 750 megawatts of inference computing capacity, delivered in tranches from 2026 through 2028, each for three or four years, with an option on 1.25 gigawatts more by the end of 2030. OpenAI lent Cerebras about 1.0 billion dollars to build it and holds a warrant for up to 33,445,026 non-voting shares; no revenue from the deal was recognized in 2025. At the end of 2025 the company counted 24.6 billion dollars of contracted revenue not yet recognized, a significant amount of it from OpenAI. In March 2026 AWS signed a term sheet, binding on pricing, exclusivity and minimum capacity, to become the first hyperscaler, meaning one of the giant cloud providers, to deploy Cerebras systems in its own data centers. 242526272829303132

Cerebras owns no chip factory. It depends on one foundry, TSMC, to make its processor, and buys servers, programmable chips, network switches, cooling units, interconnects and chassis from other suppliers, some of them sole or single sources. Apart from some assembly, packaging and testing steps it does itself, contract manufacturers build the systems, and it generally has no long-term capacity commitments from its suppliers: substantially all orders go purchase order by purchase order. 33343536

New moves

On August 12, 2026, Cerebras filed its quarterly report on Form 10-Q, a current report on Form 8-K, and a post-effective amendment to its employee stock plan registration A 25 percent tariff on certain imported semiconductors took effect on January 15, 2026; imports for uses that expand the US technology supply chain, including data centers, are exempt; Commerce was asked to report on the data-center chip market by July 1, 2026 so the tariff could be modified Cerebras processors train artificial intelligence in data centers, an exempt use today 8