AI Chips · Brand guide

Cerebras: The Complete Guide

Cerebras makes the biggest chip in the world: a whole silicon wafer turned into one AI processor. What it is, why it is so fast at answering questions, who builds it, what it costs to use, the OpenAI deal, and the facts buried in its stock-market filings. Three reading levels, every number sourced.

Stuck at any point? Ask an AI about this page →
Cerebras 16 sections · 3 levels 41 linked sources Checked September 2026
How to read this page. Every brand in our AI Chips section uses the same 16 sections in the same order. Each section starts Simple, then goes Deeper, then Expert: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where a company has not published something, we say so rather than estimate.

1.At a glance

SimpleStart here

Most chips are cut out of a round silicon wafer, dozens at a time. Cerebras keeps the whole wafer as one giant chip, the Wafer-Scale Engine 1. Its current chip, WSE-3, has 4 trillion transistors and 900,000 AI cores 1. Because the memory sits on the chip itself, it can answer AI questions very quickly 2.

What it isA single AI processor made from a whole silicon wafer 1
Latest systemCS-4, with three WSE-3 Turbo wafers, shipping from Q3 2026 3
Who builds itCerebras designs; TSMC makes the wafers 1
Can you use one?Yes, through Cerebras's own cloud; systems are also sold to large customers 4 5
DeeperThe detail

Cerebras started by selling whole systems, mainly to G42 of Abu Dhabi, and now earns more from renting its chips as a cloud service: cloud and services revenue grew 281% to $126.0 million in the June 2026 quarter 5. In January 2026 OpenAI signed for 750 megawatts of Cerebras inference compute in a deal Cerebras values at more than $20 billion 6 7.

ExpertFor specialists

Cerebras listed on Nasdaq in May 2026 at $185 a share 8. Its June 2026 quarter: revenue $180.1 million, up 74%; GAAP gross margin 14%; GAAP net loss $450.5 million 5. It reports $25.4 billion of contracted future revenue (remaining performance obligations) and $8.6 billion of cash 5.

2.Company card

SimpleStart here

Cerebras was founded by five engineers led by Andrew Feldman, who is chief executive 9. Its website says 2015; its stock-market filing says it was incorporated in April 2016 9 10.

Legal nameCerebras Systems Inc. (Delaware) 10
Headquarters1237 E. Arques Avenue, Sunnyvale, California 94085, USA 10 (map)
FoundersAndrew Feldman, Gary Lauterbach, Michael James, Sean Lie, Jean-Philippe Fricker 9
Chief ExecutiveAndrew Feldman 9
Chief Technology OfficerSean Lie 9
Chief Financial OfficerBob Komin 9
StockCBRS, Nasdaq Global Select Market 8
Investor siteinvestors.cerebras.ai
DeeperThe detail

The listing took two tries. Cerebras first filed in 2024 but withdrew in October 2025, days after raising $1.1 billion, press reports say 11. A US security review (CFIUS) of G42's investment ended in March 2025 with G42 limited to non-voting shares, according to press reports 12. Cerebras refiled and priced its shares on 13 May 2026, above its planned $150 to $160 range 8 10.

ExpertFor specialists

The IPO sold 30 million shares at $185, raising about $5.55 billion before the underwriters' option 8. The shares closed their first day at $311.07, up about 68%, press reports say 13. After the IPO, holders of Class B shares control about 99.2% of the votes 10. An analyst reading of the filing puts staff at 708 at the end of 2025 14.

3.The story

SimpleStart here
  • 2016 Cerebras incorporated 10.
  • 2021 WSE-2: 2.6 trillion transistors, 850,000 cores 15.
  • 2023 Condor Galaxy 1 supercomputer built with G42 16.
  • 2024 WSE-3 (March); Cerebras Inference cloud launches (August) 1 17.
  • 2025 Meta's Llama API; six new data centres; $1.1 billion Series G 18 19 20.
  • 2026 OpenAI deal, AWS deal, Nasdaq listing, CS-4 6 21 8 3.
DeeperThe detail

Cerebras's value rose fast: $8.1 billion after the September 2025 funding round, $23 billion after February 2026 20 22, and about $70 billion at the end of the first trading day, press reports say 13. The first wafer, WSE-1, was shown in 2019 and the CS-1 system followed 9.

ExpertFor specialists

The business turned from hardware to cloud in 2024. At launch, Cerebras Inference served Llama 3.1 8B at 1,800 tokens a second and Llama 3.1 70B at 450 tokens a second 17. By 2026 it was serving OpenAI's GPT-5.6 Sol at up to 750 tokens a second 23, and AWS agreed to put CS-3 systems in its own data centres, with Trainium handling one stage of each request and Cerebras the other 21.

4.The lineup

SimpleStart here
ProductWhat it isStatus
WSE-2 / CS-2Second wafer and its system2021 15
WSE-3 / CS-3Third wafer; 15U system up to 23 kW2024; in service 1 24
WSE-3 TurboFaster version of WSE-3: 250 petaflopsAugust 2026 3
CS-4Three WSE-3 Turbo wafers in one systemShipping from Q3 2026 3
Cerebras InferenceCloud service, pay per tokenSince August 2024 17
Cerebras CodeMonthly coding plansSince August 2025 25
DeeperThe detail

CS-3 systems can be joined in clusters of up to 2,048, which Cerebras says reach 256 exaflops 1 26. For training very large models, an external memory unit (MemoryX) holds the model and streams it to the wafer 26 16. Cerebras says CS-4 is twice as fast as CS-3 and delivers up to 10 times more throughput per watt 3.

ExpertFor specialists

The CS-4 was redesigned for building at scale: 50% fewer parts, 60% more automated manufacturing, and power converted much closer to the chips 3. Cerebras says CS-4 can serve "more than 1,000 tokens per second on models exceeding 50 trillion parameters" 27. At Hot Chips 2026 Cerebras previewed a CS-5 for 2027 and a CS-6 with stacked memory, press reports say 28.

5.How it’s made

SimpleStart here

Cerebras designs the wafer; TSMC manufactures it 1.

  1. Design: Cerebras, in Sunnyvale and elsewhere, including an office in Bengaluru 29.
  2. Wafer: TSMC, 7 nm for WSE-2 and 5 nm for WSE-3 15 1.
  3. System assembly: Flex, Sanmina and Rocket EMS 5.
  4. Cooling: custom water cooling 19.
DeeperThe detail

A whole wafer always has some faulty spots. Cerebras builds 970,000 tiny cores and uses 900,000, and its on-chip network routes around faulty ones; it calls its chip "about 100x more fault tolerant" than a GPU 30. Cerebras says its manufacturing capacity will grow more than tenfold in 2026 5.

ExpertFor specialists

Cerebras's filing warns: "We depend on third-party suppliers, including certain sole sources" 10. An analyst reading of the filing says TSMC is its only chip factory and there is no long-term supply agreement 14. Cerebras does not name which TSMC plant makes its wafers or who packages them 1.

6.Where it’s made: factories and addresses

SimpleStart here

Cerebras's chip factory is TSMC; Cerebras does not name the plant, so there is no factory address to give 1. Its own data centres, from official announcements:

SiteLocationWhat we know
Condor Galaxy 1Santa Clara, California (Colovore)64 CS-2 systems, built with G42 16
Condor Galaxy 3Dallas, Texas64 CS-3 systems 31
Oklahoma CityOklahoma, USA (Scale Datacenter)More than 300 CS-3s, planned online June 2025 19
MontrealQuebec, Canada (Enovum)Planned online July 2025 19
MikkeliFinland (Compute Nordic)165 MW at full size; first 50 MW being built 32

Cerebras publishes towns but not street addresses for its data centres 19.

DeeperThe detail

In March 2025 Cerebras said Santa Clara, Stockton and Dallas were already online, with Minneapolis, Oklahoma City, Montreal and sites in the US and Europe to follow, and 85% of capacity in the US 19. In July 2026 it announced 200 MW in France, Norway and Finland by the end of 2027 33.

ExpertFor specialists

Cerebras says it has more than 600 MW of capacity live or under contract, to be delivered by the end of 2027 5. The OpenAI deal alone is 750 MW, rolled out in stages from 2026 6.

7.How it works

SimpleStart here

A graphics chip keeps the AI model in separate memory chips beside it and must fetch it for every word it writes. Cerebras puts 44 GB of fast memory (SRAM) right next to the cores on the wafer, so there is almost no fetching 1 23. Cerebras says its memory bandwidth is 21 petabytes a second, "7,000x that of an H100" 2.

DeeperThe detail

Big models do not fit on one wafer, so Cerebras splits them by layer across several systems; each wafer holds its layers and passes the result to the next 23. At launch, a 70-billion-parameter model needed "as few as four systems" 2. For training, the model is streamed through the wafer "one layer at a time" 16.

ExpertFor specialists

Each core is about 0.05 mm², against about 6 mm² for one processing block of an NVIDIA H100, which is why a fault costs Cerebras so little of the wafer; Cerebras says 93% of the silicon is usable 30. Cerebras says a cluster "looks and programs like a single chip" 26.

8.Specs explained

SimpleStart here

Cerebras's own figures:

ChipTransistorsCoresOn-chip memoryAI computeProcess
WSE-2 (2021)2.6 trillion850,000——TSMC 7 nm 15
WSE-3 (2024)4 trillion900,00044 GB125 petaflopsTSMC 5 nm 1
WSE-3 Turbo (2026)4 trillion900,00044 GB250 petaflopsNot stated 3

The wafer measures 46,225 mm², against 815 mm² for the largest GPU when WSE-2 launched 15.

DeeperThe detail
SystemWafersAI computeMemory bandwidthSize / power
CS-31 WSE-3125 petaflops21 PB/s15U, up to 23 kW 24
CS-43 WSE-3 Turbo750 petaflops129.6 PB/sNot published 3
ExpertFor specialists

WSE-3 Turbo's fabric (on-wafer network) bandwidth is 53.5 PB/s and its off-chip links total 2.4 Tb/s 3. Cerebras claims 19 times the transistors and 56 times the compute of an NVIDIA B300 34; that is a whole wafer against one chip, so compare systems, not chips.

9.Official pricing

SimpleStart here

Cerebras's official cloud prices:

ProductPrice
gpt-oss-120b, input$0.35 per million tokens 35
gpt-oss-120b, output$0.75 per million tokens 35
Free trial$5 of credit 4
Developer tierPay per token, $10 minimum deposit 4
Cerebras Code Pro / Max (at launch)$50 / $200 a month 25
CS-3 or CS-4 systemNot published by the company 3
DeeperThe detail

At its 2024 launch Cerebras priced Llama 3.1 8B at 10 cents and Llama 3.1 70B at 60 cents per million tokens 17. The docs list gpt-oss-120b at about 3,000 tokens a second, with 65,000 tokens of context on the free tier and 131,000 on paid tiers 35.

ExpertFor specialists

The independent tester Artificial Analysis measured Cerebras's gpt-oss-120b at about 1,750 tokens a second, with a blended price of $0.39 per million tokens (third-party measurement) 36. That is well below the "about 3,000" in Cerebras's docs, a reminder that speeds depend on the test 35. For GPU cloud prices see our NVIDIA guide.

10.The software lock-in

SimpleStart here

Speed is Cerebras's selling point. Putting memory on the wafer lets it write answers many times faster than GPU clouds for the same model; Cerebras claims CS-4 is up to 30 times faster than GPU solutions on the gpt-oss-120b model 3. Fast answers matter for coding assistants and AI agents that make many calls in a row 23.

DeeperThe detail

Few companies have tried wafer-scale chips because faults ruin big chips. Cerebras's answer, tiny cores plus spare ones, is its core engineering know-how 30. OpenAI's 750 MW deal is the strongest outside endorsement 6.

ExpertFor specialists

OpenAI is tied in closely: besides the compute deal, it lent Cerebras $1 billion in January 2026 7 ($918.2 million outstanding at 30 June, per a summary of the quarterly filing) 37, and received a warrant for 33,445,026 shares at an exercise price of $0.00001 38. It exercised 10.0 million of them in July 2026; about 23.4 million remain subject to milestones 37.

11.Who buys it and why

SimpleStart here

Customers Cerebras names include OpenAI, AWS, Meta, IBM, Mistral, Notion, AlphaSense, GSK, Mayo Clinic, CrowdStrike, Figma and the US Department of Energy 20 5. Developers can also reach it through Hugging Face, where Cerebras says it is the number one inference provider 20.

DeeperThe detail

For years one customer dominated: G42 of Abu Dhabi was 83% of Cerebras's 2023 revenue and 87% in the first half of 2024, according to press reports on the 2024 filing 39. An analyst reading of the 2026 filing says MBZUAI, an Abu Dhabi university, and G42 together were about 86% of 2025 revenue 14.

ExpertFor specialists

Cerebras's filing names its significant customers as OpenAI, G42, MBZUAI and AWS 10. OpenAI contributed no revenue in 2025, the analyst reading says, so the shift away from Abu Dhabi shows up from 2026 14.

12.Hidden in plain sight

SimpleStart here

Public facts buried in Cerebras's filings and releases. Each links to the exact document.

  1. Its customer is also its lender: OpenAI lent Cerebras $1 billion; $918.2 million was still owed at 30 June 2026 7 37.
  2. OpenAI got a warrant for 33.4 million shares at $0.00001 each, and exercised 10.0 million of them in July 2026 38 37.
  3. 2025's profit was an accounting gain: $237.8 million of net income came mainly from a one-off non-cash gain on ending a forward contract, not from operations 38 10.
  4. Abu Dhabi customers were most of the revenue for years: G42, then MBZUAI and G42 39 14.
  5. Class B holders keep about 99.2% of the votes after the IPO 10.
DeeperThe detail
  1. Gross margin swung from 45% to 14% between the March and June 2026 quarters on the GAAP measure 7 5.
  2. $25.4 billion of contracted revenue against $180.1 million in the quarter 5.
  3. The founding year differs: 2015 on the website, April 2016 in the filing 9 10.
  4. The G42 investment was cleared only with non-voting shares, press reports say 12.
  5. Cerebras's own tests and outside tests differ: about 3,000 against about 1,750 tokens a second for the same model 35 36.
ExpertFor specialists
  1. GAAP loss of $450.5 million against a "core" loss of $6.9 million in the June 2026 quarter 5.
  2. Most contracted revenue is years away: an analyst reading of the IPO filing put 42% of the then $24.6 billion after four years 14.
  3. Named system builders: Flex, Sanmina and Rocket EMS 5.
  4. The IPO priced above its range, at $185 against $150 to $160 8 10.
  5. The first Condor Galaxy supercomputer sits in Santa Clara, not Abu Dhabi 16.

13.Weak spots

SimpleStart here

Limits, from Cerebras's own filings and statements:

  • Losses: the filing warns of "a history of generating net losses" 10.
  • Few customers: "A substantial portion of our revenue has been, and is expected to continue to be, driven by a limited number of customers" 38.
  • Memory per wafer: 44 GB, so big models need several systems 1 23.
  • Sole-source suppliers: the filing names "certain sole sources"; an analyst reading says TSMC is the only chip factory 10 14.
DeeperThe detail

Cerebras's own 2024 comparison puts one NVIDIA B200 GPU at 192 GB of HBM memory, far more than one wafer's 44 GB; Cerebras wins on speed, not capacity 24. The public model list is short: two models in the docs catalogue 40.

ExpertFor specialists

The filing names OpenAI, G42, MBZUAI and AWS and warns that losing any of them, or failing to meet its obligations to OpenAI, "would harm our business" 10. An analyst reading reports accounting control weaknesses flagged in the filing 14.

14.India angle

SimpleStart here

Cerebras opened an engineering office in Bengaluru in August 2022, for research and customer support 29. Lakshmi Ramachandran is its India Country Lead 9.

DeeperThe detail

In February 2026, at the AI Impact Summit in New Delhi, G42, MBZUAI and Cerebras announced an 8-exaflop AI supercomputer with India's C-DAC, hosted in India under the IndiaAI Mission 41.

ExpertFor specialists

No Cerebras data centre in India has been announced, and India is not mentioned in the June 2026 quarterly results 5. See our India's semiconductor push guide.

15.What’s next

SimpleStart here

Only what Cerebras has announced:

  • CS-4: first shipments in Q3 2026 3.
  • With AMD: a combined inference product in Q4 2026 5.
  • AWS: Cerebras in Amazon Bedrock; announced in March 2026 for later in 2026 21, with the August results pointing to early 2027 5.
  • Capacity: more than 600 MW live or under contract, deliverable by the end of 2027 5.
DeeperThe detail

Guidance for 2026: "core" revenue of $880 million to $890 million, raised in August 5. The first European data centre is due by the end of 2026 33.

ExpertFor specialists

Press coverage of Hot Chips 2026 reports a CS-5 in 2027 aimed at 10,000 tokens a second per user; Cerebras has not issued a release on it 28. The test is whether OpenAI's 750 MW turns into revenue on schedule 6 10.

16.Sources

41 sources, all checked September 2026. Official = the company or organisation’s own page; Filing = a document filed with a regulator; Press = news coverage, used only where no official source exists and labelled in the text.

  1. Cerebras: Cerebras announces third-generation Wafer Scale Engine (13 Mar 2024)Official
  2. Cerebras Blog: Introducing Cerebras Inference: AI at instant speed (Aug 2024)Official
  3. Cerebras Investor Relations: Cerebras unveils CS-4, 30 times faster than GPU-based solutions (18 Aug 2026)Official
  4. Cerebras: Inference product page and plansOfficial
  5. Cerebras Investor Relations: Q2 2026 results: fast inference cloud business nearly quadruples (12 Aug 2026)Official
  6. Cerebras Blog: OpenAI partners with Cerebras to bring high-speed inference to the mainstream (14 Jan 2026)Official
  7. Cerebras Investor Relations: Q1 2026 results (23 Jun 2026, PDF)Official
  8. Cerebras: Cerebras Systems announces pricing of initial public offering (13 May 2026)Official
  9. Cerebras: Company page: founders and leadershipOfficial
  10. Cerebras Systems (SEC): Form S-1/A registration statement (May 2026)Filing
  11. Tech Startups (press): Cerebras revives IPO plans after $1.1B raise and CFIUS clearance (19 Dec 2025)Press
  12. The Register (press): Cerebras clears CFIUS hurdle on its path to IPO (31 Mar 2025)Press
  13. Yahoo Finance (press): Cerebras stock slides after near-70% surge in biggest IPO of 2026Press
  14. Mostly Metrics (analyst): Cerebras IPO: S-1 breakdownResearch
  15. Cerebras: Second-generation Wafer Scale Engine, 2.6 trillion transistors (Apr 2021)Official
  16. Cerebras Blog: Introducing Condor Galaxy 1: a 4 exaFLOP supercomputer (20 Jul 2023)Official
  17. Cerebras: Cerebras launches the world's fastest AI inference (27 Aug 2024)Official
  18. Cerebras: Meta collaborates with Cerebras for fast inference in the new Llama API (29 Apr 2025)Official
  19. Cerebras: Cerebras announces six new AI data centres across North America and Europe (11 Mar 2025)Official
  20. Cerebras: Cerebras raises $1.1 billion Series G (30 Sep 2025)Official
  21. Cerebras: AWS collaboration: CS-3 in AWS data centres (13 Mar 2026)Official
  22. Cerebras: Cerebras raises $1 billion Series H (3 Feb 2026)Official
  23. Cerebras Blog: How Cerebras serves GPT-5.6 Sol at up to 750 tokens per second (27 Aug 2026)Official
  24. Cerebras Blog: Cerebras CS-3 vs NVIDIA B200: 2024 AI accelerators compared (Apr 2024)Official
  25. Cerebras Blog: Introducing Cerebras Code (1 Aug 2025)Official
  26. Cerebras Blog: Cerebras CS-3: the world's fastest and most scalable AI acceleratorOfficial
  27. Cerebras: System page: CS-4Official
  28. Wccftech (press): Cerebras at Hot Chips 2026: CS-5 in 2027, CS-6 with 3D-stacked SRAMPress
  29. Cerebras: Cerebras accelerates global growth with new India office (29 Aug 2022)Official
  30. Cerebras Blog: 100x defect tolerance: how Cerebras solved the yield problem (13 Jan 2025)Official
  31. Cerebras: Cerebras and G42 announce Condor Galaxy 3 (13 Mar 2024)Official
  32. Cerebras Investor Relations: Cerebras and Compute Nordic Finland announce 165 MW AI data centre (1 Sep 2026)Official
  33. Cerebras Investor Relations: Cerebras accelerates European expansion, 200 MW of AI compute (9 Jul 2026)Official
  34. Cerebras: Chip page: Wafer-Scale EngineOfficial
  35. Cerebras Inference Docs: OpenAI GPT OSS: price, speed and contextOfficial
  36. Artificial Analysis (third-party): Cerebras provider benchmarksResearch
  37. StockTitan (press): Summary of Cerebras Q2 2026 Form 10-QPress
  38. Cerebras Systems (SEC): Form S-1 registration statement (April 2026)Filing
  39. EE Times (press): Cerebras IPO paperwork sheds light on relationship with G42 (2024)Press
  40. Cerebras Inference Docs: Models overviewOfficial
  41. Abu Dhabi Media Office: G42, MBZUAI and Cerebras partner with C-DAC on an AI supercomputer in India (20 Feb 2026)Government

On this page

1. At a glance2. Company card3. The story4. The lineup5. How it’s made6. Where it’s made: factories and addresses7. How it works8. Specs explained9. Official pricing10. The software lock-in11. Who buys it and why12. Hidden in plain sight13. Weak spots14. India angle15. What’s next16. Sources

Quick facts

HeadquartersSunnyvale, California
Founded2015 (website); incorporated 2016
CEOAndrew Feldman
Makes its own chips?Designs them; TSMC manufactures
StockCBRS, Nasdaq (listed May 2026)
Latest quarter revenue$180.1 million

AI Chips

All chip brandsAmazon TrainiumAMDAppleGoogle TPUGroqHuawei AscendIntelMeta MTIAMicrosoft MaiaNVIDIAAI memory (HBM) and why it is scarceExport rules and the US–China chip contestHow a chip is made: sand to serverHow AI chips workIndia’s semiconductor pushHow to read a chip spec sheetToday’s AI news

In the news

Newest stories mentioning Cerebras, straight from the companies’ own newsrooms and the wider press. Links open the original.

All of today’s AI news →

Ask an AI about this page

Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.

ChatGPTClaudeGeminiPerplexityGrok

Nothing is sent from here. The link carries only this page’s title and address.