Cerebras WSE-4 and 2026 IPO

The Cerebras WSE-4 (Wafer Scale Engine 4) is the anticipated fourth generation of the dinner-plate-sized wafer-scale AI processor built by Cerebras Systems. As of June 2026 the WSE-4 had not been officially announced or specified, and the third-generation WSE-3 was the company's shipping chip. That was still the case in September 2026: on August 18, 2026 Cerebras introduced a fourth-generation system, the CS-4, but built it on a higher-clocked variant of the existing wafer, the WSE-3 Turbo, rather than on a new WSE-4, and the company's Hot Chips 2026 roadmap places its next-generation wafer in a later system, the CS-5, targeted for 2027, without naming the chip or its process node.^[19][22] The WSE-4 label is most often paired with the larger 2026 milestone for Cerebras: the company's long-delayed initial public offering, which priced on May 13, 2026 and began trading on the Nasdaq the next day under the ticker CBRS.^[1][2] The offering sold 30 million shares at $185 each and raised about $5.55 billion, which CNBC described as the largest US technology IPO since Uber in 2019; with the underwriters' option exercised in full at the May 15 closing, the total came to 34.5 million shares and roughly $6.38 billion in gross proceeds.^[2][6][20] It valued Cerebras at roughly $56 billion on a fully diluted basis at the offer price, before a first-day surge pushed the market value much higher.^[2][6] The IPO had been stalled for more than a year and a half, largely by a national-security review of a minority investment from the Emirati group G42.^[13]

Overview

Cerebras Systems, incorporated in April 2016 (its prospectus says the company was founded in 2016) and led by chief executive Andrew Feldman, designs AI chips that span almost an entire semiconductor wafer rather than being cut into many small dies.^[14] Where companies such as Nvidia slice a 300 mm (12-inch) wafer into hundreds of separate GPUs, Cerebras keeps the wafer whole and operates it as a single processor. The result is a chip with vastly more cores and on-chip memory than a conventional GPU, which Cerebras argues makes it especially well suited to running large neural networks at high speed. Each chip is paired with a system, the CS series, that supplies power, cooling, and data movement.

The current wafer design is the WSE-3, announced on March 13, 2024 and fabricated by TSMC on a 5 nm process.^[8] In August 2026 Cerebras added a higher-clocked version of the same wafer, the WSE-3 Turbo, and a three-wafer rack, the CS-4, built around it.^[19] The 2026 IPO, the OpenAI inference agreement, and a still-heavy reliance on G42 are the developments that have brought the company the most attention. This article covers the prospective WSE-4, the CS-4 as announced, and the 2026 listing; the broader corporate history and the WSE-3 chip itself are treated on the Cerebras and WSE-3 pages.

WSE-4 specs vs WSE-3

No official WSE-4 specifications existed as of June 2026, and none had been published as of September 2026. Cerebras has not published transistor counts, core counts, a process node, or a launch date for a fourth-generation part, and does not use the name "WSE-4" in its CS-4 launch materials or its Hot Chips 2026 roadmap post.^{[19][21][22][25]} Trade and analyst commentary published around the IPO generally expected a fourth-generation wafer to arrive in roughly 2027 on a smaller TSMC node, continuing the company's two-to-three-year generational cadence (WSE-1 in 2019, WSE-2 in 2021, WSE-3 in 2024). Several outlets noted that a portion of the IPO proceeds is expected to fund next-generation research and development. Cerebras' April 2026 prospectus said only that it intends to "expand on-chip memory and memory bandwidth, improve interconnect density, and leverage advancements in process technology to increase transistor counts and reduce power in future WSE generations", and that future WSEs would have faster compute and "more and faster memory and communication onto and off of the wafer".^[14] None of the WSE-4 figures in press commentary should be treated as confirmed.

What is confirmed is the WSE-3 baseline that a WSE-4 would have to surpass. The WSE-3 carries 4 trillion transistors and 900,000 AI-optimized cores on a die of 46,225 mm squared, about 57 times the area of an Nvidia H100 (roughly 814 mm squared).^[8][10] It holds 44 GB of on-chip SRAM with about 21 petabytes per second of memory bandwidth, and Cerebras rates its peak AI performance at 125 petaflops at 16-bit precision (12.5 petaflops dense).^[8][11] The company says the WSE-3 doubles the performance of the WSE-2 at the same power and price.^[8]

The WSE-3 Turbo (August 2026)

The chip Cerebras did announce on August 18, 2026 is the Wafer Scale Engine 3 Turbo (WSE-3T), which the company presents as a faster version of the WSE-3 rather than a new generation. According to the press release and the CS-4 datasheet, the WSE-3T keeps the WSE-3's physical design: four trillion transistors, 900,000 AI-optimized cores, 46,225 mm squared of silicon on TSMC's 5 nm process, and 44 GB of on-wafer SRAM.^[19][25] What changes is the operating point. Cerebras says AI compute doubles to 250 petaflops per wafer (a sparse FP16 figure, per the datasheet footnote), memory bandwidth doubles to 43.2 PB/s, on-chip fabric bandwidth doubles to 53.5 PB/s, off-wafer I/O doubles to 2.4 Tb/s, and I/O latency falls from five microseconds to as low as two.^[19][25] The company attributes the higher operating frequency to a redesigned power path in the CS-4 that moves power conversion from roughly 50 mm from the processor, as on conventional GPU boards, to about 0.5 mm, delivering twice as much power to the wafer.^[19][21] Cerebras has not published clock speeds; The Register and The Next Platform each estimated that the wafer now runs at about 2.8 GHz, up from about 1.4 GHz, and both characterised the WSE-3T as the existing silicon pushed harder rather than a redesign.^[27][28] These are company-published performance figures and outlet estimates, respectively.

What Cerebras has said about the next wafer

The clearest company statement on a post-WSE-3 chip came a week after the CS-4 launch, in a Cerebras blog post on its Hot Chips 2026 presentation. It says the Nexus rack platform introduced with the CS-4 "was also co-designed to support our next-generation WSE, which will debut in CS-5", and that the CS-5 is "targeted for 2027".^[22] The post gives performance targets for the CS-5 rather than chip specifications: up to 10,000 output tokens per second per user on open models such as Gemma 4 31B and gpt-oss-120b, up to 5,000 tokens per second per user on multi-trillion-parameter models, 3 million tokens per second per megawatt, and support for models above 50 trillion parameters.^[22] It also previews a CS-6 that would stack DRAM in 3D on top of wafer-scale SRAM and compute, work the company says it began in 2024.^[22] Reuters reported from the launch briefing that "Cerebras plans another generation of the chip and server in 2027", and quoted Feldman: "We're going to get four times as fast between now and the end of the year, end of 2027, and we're going to get 20 times more throughput."^[26] Whether the next-generation WSE will carry the WSE-4 name, and on what node it will be built, had not been stated by Cerebras as of September 2026. The Register's Tobias Mann wrote that he expects a WSE-4, "which is presumably still coming", to offer only modest FP16 gains while roughly doubling SRAM capacity, and The Next Platform published its own speculative WSE-4 configurations; both are analyst projections, not company disclosures.^[27][28]

The published Cerebras generations compare as follows:

The CS-4 system

As of June 2026 there was likewise no announced CS-4 system. A CS-4 was widely assumed, by the company's naming convention, to be the chassis built around a future WSE-4, but Cerebras had not detailed one. The platform at the time was the CS-3, which houses a single WSE-3. Cerebras pairs the CS-3 with an external memory subsystem it calls MemoryX, offered in capacities reported up to roughly 1.2 petabytes, and a fabric, SwarmX, that lets many systems be linked.^[9] The company says a single CS-3 can train models with up to 24 trillion parameters, and that a cluster of up to 2,048 CS-3 systems reaches about 256 exaflops at 16-bit precision.^[8] Until August 2026 the CS-3 was the basis of both Cerebras' on-premises sales and its inference cloud.

The CS-4 as announced (August 18, 2026)

On August 18, 2026, at its Supernova 2026 event, Cerebras introduced the CS-4.^[19][21][22] It broke the assumed pairing: the CS-4 is not built around a WSE-4 but around three WSE-3 Turbo wafers in a single rack, where the CS-1 through CS-3 each held one wafer in one chassis.^[19][28] The company describes it as the first product on its Nexus rack-scale platform architecture, a modular design that separates compute, power, and I/O into self-contained assemblies. Each wafer sits in a rear-mounted "Wafer-Scale Backpack" that packages power conversion, direct liquid cooling, high-speed I/O, and control electronics around the wafer; Cerebras says the backpack has 50 percent fewer components than the prior-generation system, uses 60 percent more automated manufacturing, and cuts deployment time from days to hours.^[19][21] A new programmable Wafer I/O Module supports standards-based RoCE v2 RDMA over Ethernet for connecting to other vendors' equipment, and a switch-free mode called Direct Wafer Links for connecting wafers within and across racks with latency as low as two microseconds.^[19][25]

All CS-3 and CS-4 figures in the table come from the CS-4 press release and datasheet unless otherwise marked.^[19][25] Cerebras' headline claims are that the CS-4 is up to twice as fast as the CS-3, which it says extends its tokens-per-second-per-user advantage over GPU systems to "up to 30x"; the release cites a head-to-head run of gpt-oss-120B at more than 4,400 tokens per second per user, with a footnote that actual throughput varies by model, context length, precision, and serving configuration.^[19] It also claims up to 10 times more throughput per watt than the CS-3, which the company blog sources to "internal benchmarking and projections", and more than 1,000 tokens per second on models exceeding 10 trillion parameters, labelled an extrapolation from internal benchmarking.^[21][25] Cerebras' own materials are not consistent on the last point: the blog and datasheet say 10 trillion parameters, while the product page on cerebras.ai says 50 trillion.^[21][25] SemiAnalysis founder Dylan Patel is quoted in the release on the system's deployability and networking.^[19] The Register estimated the rack's power draw at 120 to 140 kW, a figure Cerebras has not published.^[27]

The CS-4 is also the first Cerebras system designed from the outset for disaggregated serving, in which a separate prefill engine processes the prompt and hands the model state to Cerebras for decode; the launch materials name AMD Helios and AWS Trainium as prefill platforms it is built to pair with.^[21][25] Chief technology officer Sean Lie told a San Francisco media briefing that the reduced component count would speed data-center construction.^[26] Cerebras said first CS-4 shipments would begin in the third quarter of 2026, and full specifications are in its CS-4 datasheet.^[19][25]

The 2026 IPO

Cerebras first moved toward the public market in 2024, filing a confidential registration with the US Securities and Exchange Commission in September 2024 and targeting a listing as early as that October.^[13] The plan stalled almost immediately. G42, an Abu Dhabi AI conglomerate, had agreed to buy about $335 million of Cerebras stock, a stake of more than 5 percent, and that purchase triggered a review by the Committee on Foreign Investment in the United States (CFIUS).^[13] The concern centered on G42's past ties to Chinese firms, including Huawei. The companies restructured the deal so that G42's shares would be non-voting, and CFIUS cleared the investment in early 2025, but the offering did not proceed amid a soft IPO market.^[12]

In October 2025 Cerebras formally withdrew its registration and raised private bridge capital instead, including a round reported at about $1.1 billion.^[12] It refiled confidentially in early 2026 and made its prospectus public on April 17, 2026.^[14] Demand proved strong: Cerebras twice raised its expected price, from an initial $115 to $125 range to $150 to $160, and finally priced at $185 per share on May 13, 2026, above the range.^[2] The sale of 30 million shares brought in about $5.55 billion and implied a fully diluted valuation near $56.4 billion.^[2][6] Cerebras' pricing release granted the underwriters a 30-day option on a further 4.5 million shares and set the closing for May 15, 2026.^[24] The offering closed that day with the option exercised in full, for an aggregate 34.5 million Class A shares and gross proceeds of approximately $6.38 billion before underwriting discounts and expenses, a figure Cerebras rounds to $6.4 billion in its second-quarter 2026 results release.^[20][23]

Trading opened the next day, May 14, 2026. The stock began at about $350, roughly 89 percent above the offer price, and closed its first session at $311.07, up about 68 percent.^[1] At that close CNBC put the company's market capitalization near $95 billion on a fully diluted basis, while The Register and some other outlets, counting basic shares, cited a figure closer to $66 billion.^[1][7] The debut made paper billionaires of Feldman, whose stake was valued near $1.9 billion, and co-founder and chief technology officer Sean Lie, at roughly $1 billion.^[3] The shares gave back part of the gain in the days that followed.^[4]

The prospectus also clarified the company's finances. Revenue reached $510 million in 2025, up about 76 percent year over year.^[14] Cerebras reported GAAP net income of $237.8 million, but that figure was lifted by a one-time, non-cash gain of about $363.3 million tied to the extinguishment of a forward-contract liability related to G42; on an operating basis the company still ran a loss of roughly $146 million for the year.^[14] The business is therefore growing quickly but is not yet profitable from operations.

Inference business and the G42 and OpenAI context

Cerebras has increasingly positioned itself around AI inference, the work of running already-trained models to answer queries, rather than only model training. It launched a hosted Cerebras Inference service in 2024, later listed it on the AWS Marketplace, and markets the speed advantage of keeping an entire large language model on one chip's SRAM. Independent benchmarks by Artificial Analysis cited in 2026 reported the CS-3 running OpenAI's gpt-oss-120B at more than 2,700 tokens per second versus about 900 on an Nvidia B200,^[15] and in May 2026 Cerebras reported 981 tokens per second on the Kimi K2.6 model, which it said was about 6.7 times the next-fastest GPU cloud.^[16] Cerebras' own materials claim the CS-3 is roughly 21 times faster than an Nvidia DGX B200 at one-third the cost and power for some inference workloads; these are vendor figures and should be read as such.^[15]

Customer concentration remains the company's central risk. In 2024 G42 accounted for the overwhelming majority of revenue, reported at about 85 to 87 percent.^[14] That share fell to 24 percent in 2025 as other customers grew, but the prospectus disclosed that about 86 percent of 2025 revenue still came from two UAE-affiliated entities: the Mohamed bin Zayed University of Artificial Intelligence at 62 percent and G42 at 24 percent.^[14]

The most consequential new relationship is with OpenAI. On January 14, 2026 the two companies announced a multiyear agreement, reported by CNBC as worth over $10 billion, to supply 750 megawatts of inference capacity through 2028.^[5] Cerebras' April 2026 prospectus described a binding Master Relationship Agreement valued at more than $20 billion once expansion options are included, with provisions allowing OpenAI to scale the commitment toward 2 gigawatts by 2030.^[14][17] Reporting around the deal and the IPO also described OpenAI taking an equity position of about 11 percent in Cerebras and extending a roughly $1 billion loan, at about 6 percent interest, to help build the data centers that will house the chips.^[17] The arrangement diversifies Cerebras beyond its Gulf customers while binding it tightly to a single, very large US buyer.

Significance

Cerebras' listing was the first large US technology IPO of 2026 and was widely read as a test of investor appetite for an AI hardware company that competes, in a narrow lane, against Nvidia. Nvidia held an estimated 86 percent of the AI data-center market by revenue in its most recent fiscal year, with revenue above $215 billion, and dominates training.^[18] Cerebras does not challenge that breadth; instead it argues that its single-wafer design wins on latency for high-speed inference, an area analysts have suggested could grow into a market worth hundreds of billions of dollars by 2030 and where buyers want alternatives to a single supplier.

The open questions after the IPO are the same ones the prospectus raised: whether Cerebras can convert headline contracts such as the OpenAI agreement into recurring, diversified revenue; whether it can fund and ship a competitive next wafer, the next-generation WSE it has slated for the CS-5 in 2027 and which was still unnamed and unspecified as of September 2026, against Nvidia's annual cadence; and whether its dependence on a small number of customers, several with foreign-government ties, leaves it exposed to export rules and concentration risk.^[22] The CS-4 answered part of the roadmap question by doubling per-wafer speed without new silicon, although, as The Register and The Next Platform noted, it left SRAM capacity where it has been since the WSE-2 and spent a clock-speed lever that can be pulled only once.^[27][28] The next-generation wafer itself, when and if it is detailed, will be the clearest signal of whether the wafer-scale approach can keep pace at the leading edge.

References

1. ^1 ^2 ^3CNBC, "Cerebras pops 68% in Nasdaq debut, pushing the AI chipmaker's market cap to $95 billion," May 14, 2026. cnbc.com/...cerebras-cbrs-stock-trade-nasdaq-ipo
2. ^1 ^2 ^3 ^4 ^5CNBC, "Cerebras prices IPO above expected range, as Wall Street braces for AI tsunami," May 13, 2026. cnbc.com/...ted-range-wall-street-expects-ai-flood
3. ^CNBC, "Cerebras IPO mints two billionaires, sets stage for potential AI wave," May 14, 2026. cnbc.com/...aires-sets-stage-for-potential-ai-wave
4. ^CNBC, "Cerebras stock falls after blockbuster IPO debut," May 15, 2026. cnbc.com/...cerebras-stock-ipo-debut-ai
5. ^CNBC, "Cerebras scores OpenAI deal worth over $10 billion ahead of AI chipmaker's IPO," January 14, 2026. cnbc.com/...ores-openai-deal-worth-over-10-billion
6. ^1 ^2 ^3TechCrunch, "Cerebras raises $5.5B, kicking off 2026's IPO season with a bang," May 14, 2026. techcrunch.com/...off-2026s-ipo-season-with-a-bang
7. ^The Register, "Cerebras risked it all on dinner plate-sized AI accelerators a decade ago. Today it's worth $66B," May 15, 2026. theregister.com/...5240821
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Cerebras Systems, "Cerebras Announces Third Generation Wafer-Scale Engine" (WSE-3 press release), March 2024. cerebras.ai/...third-generation-wafer-scale-engine
9. ^The Next Platform, "Cerebras Goes Hyperscale With Third Gen Waferscale Supercomputers," March 14, 2024. nextplatform.com/...1642584
10. ^ServeTheHome, "Cerebras WSE-3 AI Chip Launched 56x Larger than NVIDIA H100," March 2024. servethehome.com/...vertiv-supermicro-hpe-qualcomm
11. ^IEEE Spectrum, "Cerebras WSE-3: Third Generation Superchip for AI." spectrum.ieee.org/cerebras-chip-cs3
12. ^1 ^2Tech Startups, "AI chipmaker Cerebras revives IPO plans after $1.1B raise and CFIUS clearance," December 19, 2025. techstartups.com/...1-1b-raise-and-cfius-clearance
13. ^1 ^2 ^3GovCon Wire, "Report: Cerebras Likely to Delay Plans of Going Public Due to National Security Review of G42 Deal." govconwire.com/...onal-security-review-of-g42-deal
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8U.S. Securities and Exchange Commission, Cerebras Systems Inc. Form S-1, April 2026. sec.gov/...cerebras-sx1april2026
15. ^1 ^2Cerebras Systems, "Faster inference from Cerebras, beats Blackwell." cerebras.ai/...blackwell-vs-cerebras
16. ^CryptoBriefing, "Cerebras reports 981 tokens per second on Kimi K2.6 model, 6.7x faster than GPU cloud," 2026. cryptobriefing.com/...981-tokens-kimi-k2-gpu-speed
17. ^1 ^2CryptoBriefing, "OpenAI commits $20B for potential 11% stake in Cerebras," 2026. cryptobriefing.com/openai-20b-cerebras-stake
18. ^The Motley Fool, "Nvidia vs. Cerebras: Better AI Stock to Buy Now," May 17, 2026. fool.com/...vs-cerebras-better-ai-stock-to-buy-now
19. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16Cerebras Systems, "Cerebras Unveils CS-4: Up to 30 Times Faster than GPU-based Solutions" (press release, GlobeNewswire, via Markets Insider), August 18, 2026. markets.businessinsider.com/...olutions-1036472378
20. ^1 ^2Cerebras Systems, "Cerebras Systems Announces Closing of Initial Public Offering," May 15, 2026. cerebras.ai/...-closing-of-initial-public-offering
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7Cerebras Systems (Angela Yeung and Eric Gardner), "Introducing Cerebras CS-4: The Fastest AI Just Got Faster," Cerebras blog, August 18, 2026. cerebras.ai/...introducing-cerebras-cs-4
22. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Cerebras Systems (Jessica Liu), "Ultrafast Frontier Inference: Cerebras Deep Dive at Hot Chips 2026," Cerebras blog, August 25, 2026. cerebras.ai/...erebras-deep-dive-at-hot-chips-2026
23. ^Cerebras Systems, "Cerebras Systems Fast Inference Cloud Business Nearly Quadruples in Second Quarter 2026" (press release, via Nasdaq), August 12, 2026. nasdaq.com/...early-quadruples-second-quarter-2026
24. ^Cerebras Systems, "Cerebras Systems Announces Pricing of Initial Public Offering," May 13, 2026. cerebras.ai/...-pricing-of-initial-public-offering
25. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Cerebras Systems, "Cerebras CS-4 Datasheet" (PDF), August 2026. cerebras.ai/cs4-datasheet
26. ^1 ^2Reuters (Max A. Cherney), "Cerebras launches new server chip and system designed to speed AI chatbots," August 18, 2026 (syndicated copy). lufkindailynews.com/...54c0-5a0d-a50b-729b6481d995
27. ^1 ^2 ^3 ^4 ^5The Register (Tobias Mann), "Cerebras CS-4 rack systems juice chips for every last drop of AI performance," August 19, 2026. theregister.com/...5289286
28. ^1 ^2 ^3 ^4 ^5The Next Platform (Timothy Prickett Morgan), "Cerebras Overclocks WSE-3 Waferscale Engine To Boost Inference Oomph In 'Nexus' CS-4," August 19, 2026. nextplatform.com/...5289400