The wafer-scale AI processor company
HQ: Sunnyvale, CA, United States | Founded: 2015 | Employees: 700-800 (708 in FY2025, ~774 as of March 2026) | Stage: Public — Nasdaq: CBRS, IPO May 2026 | Website: https://www.cerebras.ai
Cerebras exists because of a specific constraint in chip manufacturing. Lithography exposes a pattern onto a die that must fit inside a reticle of roughly 815 square millimetres. That limit is why a modern GPU is a slab of silicon dozens of times smaller than a wafer, and why the bandwidth needed to feed a large model has to come from off-package high-bandwidth memory stacked beside the compute die. Cerebras was founded in 2015 by Andrew Feldman and Gary Lauterbach — who had previously built SeaMicro, which AMD acquired in 2012 — along with Michael James, Sean Lie and Jean-Philippe Fricker, and decided to ignore the reticle limit. Fabricate a whole 300mm wafer and etch it as one processor.
The engineering problem is that whole wafers contain defects, and yield is the economics of semiconductors. Cerebras's answer is redundancy at extremely fine granularity: fault-tolerant core granularity of about 0.05 mm², against 6.2 mm² for an H100, so the maximum area lost to defects is 2.2 mm² rather than 361 mm². The rest is interconnect — on-die nanowires plus a two-dimensional network-on-chip fabric, later extended with a programmable wafer I/O module using RDMA over Ethernet and switchless direct wafer links at two microseconds.
The generations came fast. WSE-1 in August 2018 had 1.2 trillion transistors, 850,000 cores and 8.5GB of SRAM on 15nm. WSE-2 in November 2019 doubled SRAM to 40GB at 7nm. WSE-3, launched in March 2024, went to 4 trillion transistors, 900,000 cores and 44GB of SRAM on TSMC 5nm, with 125 PFLOPS of FP16 and 21 petabytes per second of memory bandwidth. There has been no WSE-4; instead the company built WSE-3 Turbo, which doubles per-wafer power to reach 250 PFLOPS and 43.2 petabytes per second, and combined three of them into the CS-4 Nexus rack announced in August 2026 — 750 PFLOPS and 129.6 petabytes per second per rack, and a claimed 4,400 tokens per second per user on a 120-billion-parameter open-weight model.
A deliberate and important decision runs through all of it: no high-bandwidth memory. Model weights live in on-wafer SRAM. That insulates Cerebras from the HBM allocation war that constrains everyone else, and it chose TSMC 5nm over 3nm deliberately, because 5nm capacity is less contended and because High-NA EUV would shrink the reticle and therefore reduce total transistors per wafer. The trade-off is that SRAM is fixed at 44GB per wafer, so a 120-billion-parameter model has to span many wafers and racks, pushing the bottleneck onto interconnect.
The software is the other half. Cerebras pioneered weight streaming — streaming weights one layer at a time across the wafer fabric in pure data-parallel fashion, so a 100-billion-parameter model uses the same code as a one-billion-parameter one, with no pipeline parallelism, no model parallelism and no sharding. Its PyTorch layer is native, and the company claims a GPT-3-scale model in 565 lines of code, 97% less than an equivalent GPU implementation. For inference, Cerebras Inference offers an OpenAI-compatible API — two lines of code to migrate — which is how it reached developers quickly and how it serves OpenAI's own models, including GPT-5.6 Sol at 750 tokens per second.
The business changed character in 2025-26. Legacy revenue came overwhelmingly from G42 and MBZUAI in the UAE and from Condor Galaxy supercomputers. In January 2026 OpenAI signed the master relationship agreement, having reportedly passed over Nvidia silicon. AWS then became the first hyperscaler channel, with Trainium3 handling prefill and CS-3 handling decode on Bedrock, and AMD Helios following. Management now runs more than 600 megawatts of contracted capacity across thirteen named metros, is building a first-party cloud at a ten-times manufacturing scale-up, and holds an $850 million revolving credit facility on top of more than $8.6 billion in liquidity.
Cerebras is the only company shipping a wafer-scale processor, and after five years of being called a science project it is now a public company with a $25.4 billion contracted backlog. The engineering bet was simple and audacious: stop trying to shrink a chip to fit inside a lithography reticle and instead fabricate an entire 300mm wafer and wire it together as one device. The result is 4 trillion transistors and 44GB of on-wafer SRAM, which is why Cerebras can claim roughly 7,000 times the memory bandwidth of an H100 and why its inference cloud produces thousands of tokens per second per user.
The commercial proof arrived in January 2026, when OpenAI signed a master relationship agreement for up to 750 megawatts of compute — reported at more than $20 billion — expandable to two gigawatts, with warrants for up to roughly 10% of the company. That single contract, alongside the legacy G42 and MBZUAI relationship, is what turned a hardware company into an infrastructure company. Q2 2026 core revenue was $209.9 million, up 103% year over year, with cloud revenue up 287%.
Two things complicate the story. The first is concentration. In FY2025, 86% of revenue came from two UAE-linked entities, and MBZUAI alone owed the company 77.9% of year-end receivables. The OpenAI agreement reduces that dependence in principle and increases it in practice. The second is the quality of reported earnings. FY2025 GAAP net income of $237.8 million included a $363.3 million one-time non-cash gain tied to G42; underneath it was a $145.9 million operating loss and a $75.7 million non-GAAP net loss. Q2 2026's core-to-GAAP gap was driven by $44.3 million of customer-warrant amortisation — customers being paid partly in equity, with warrants on the balance sheet at $1.13 billion.
The structural question is whether the software ecosystem catches up. Cerebras has no CUDA. It answers that with native PyTorch, an OpenAI-compatible API and a 565-line implementation of a GPT-3-scale model, but it competes against the most entrenched software moat in the industry while several of its largest customers are also its chip-design rivals.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki