Cerebras WSE-4 and 2026 IPO
Last reviewed
Jun 7, 2026
Sources
18 citations
Review status
Source-backed
Revision
v1 · 2,022 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
18 citations
Review status
Source-backed
Revision
v1 · 2,022 words
Add missing citations, update stale details, or suggest a clearer explanation.
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 has not been officially announced or specified, and the third-generation WSE-3 remains the company's shipping chip. The 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. 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. It valued Cerebras at roughly $56 billion at the offer price, before a first-day surge pushed the market value much higher. 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.
Cerebras Systems, founded in 2015 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. 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 product is the WSE-3, announced on March 13, 2024 and fabricated by TSMC on a 5 nm process. 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 and the 2026 listing; the broader corporate history and the WSE-3 chip itself are treated on the Cerebras and WSE-3 pages.
No official WSE-4 specifications exist as of June 2026. Cerebras has not published transistor counts, core counts, a process node, or a launch date for a fourth-generation part, and the WSE-3 remains the latest announced design. Trade and analyst commentary published around the IPO generally expects a WSE-4 to arrive in roughly 2027, most likely on a TSMC 3 nm class process, 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. None of those WSE-4 figures 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). 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). The company says the WSE-3 doubles the performance of the WSE-2 at the same power and price.
The published Cerebras generations compare as follows:
| Generation | Year | TSMC process | Transistors | AI cores | On-chip SRAM |
|---|---|---|---|---|---|
| WSE-1 | 2019 | 16 nm | 1.2 trillion | 400,000 | 18 GB |
| WSE-2 | 2021 | 7 nm | 2.6 trillion | 850,000 | 40 GB |
| WSE-3 | 2024 | 5 nm | 4 trillion | 900,000 | 44 GB |
| WSE-4 | Not announced (expected ~2027) | reported as likely 3 nm | Not disclosed | Not disclosed | Not disclosed |
There is likewise no announced CS-4 system. A CS-4 would, by the company's naming convention, be the chassis built around a future WSE-4, but Cerebras has not detailed one. The current platform is 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. 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. Until a WSE-4 is unveiled, the CS-3 remains the basis of both Cerebras' on-premises sales and its inference cloud.
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. 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). 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.
In October 2025 Cerebras formally withdrew its registration and raised private bridge capital instead, including a round reported at about $1.1 billion. It refiled confidentially in early 2026 and made its prospectus public on April 17, 2026. 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. The sale of 30 million shares brought in about $5.55 billion and implied a fully diluted valuation near $56.4 billion.
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. 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. 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. The shares gave back part of the gain in the days that followed.
The prospectus also clarified the company's finances. Revenue reached $510 million in 2025, up about 76 percent year over year. 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. The business is therefore growing quickly but is not yet profitable from operations.
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, 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. 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.
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. 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.
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. 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. 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. The arrangement diversifies Cerebras beyond its Gulf customers while binding it tightly to a single, very large US buyer.
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. 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 chip, the still-unannounced WSE-4, 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. The WSE-4 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.