On April 18, 2026, AI chipmaker Cerebras Systems filed paperwork with the US SEC to go public, reviving an IPO attempt that fell apart two years earlier. A company spokesperson told TechCrunch the offering is planned for mid-May; the raise amount, exchange, and underwriter lineup remain undisclosed.
What makes this filing different is what the company is holding: $510 million of 2025 revenue, a reported OpenAI compute contract worth more than $10 billion, and an AWS agreement that puts its silicon inside Amazon data centers. The Wall Street Journal broke the filing first.
It is also a test the whole second tier of AI silicon is watching. Since the Nvidia supply squeeze began, investors have funnelled money into every credible alternative accelerator vendor, but almost none of those bets has reached the public tape. If Cerebras prices and trades, it becomes the reference asset for pricing AI compute startups — and the first real public read on how much a non-GPU inference business is worth.
The Numbers in the Filing
Per the SEC filing, Cerebras booked $510 million of revenue in 2025. Profitability has two readings: GAAP net income of $237.8 million, but a non-GAAP net loss of $75.7 million once one-time items are excluded. In plain terms, the accounting profit comes largely from one-off items; the operating business is still in investment mode. For IPO buyers, that gap is the number to interrogate: the question at pricing time will be how durable the underlying inference business is once the one-offs roll off.
CEO Andrew Feldman’s positioning has not changed: Cerebras builds “the fastest AI hardware for training and inference” — a wafer-scale architecture that treats an entire wafer as one chip, a fundamentally different path from clusters of GPUs. The engineering bet is that a huge single die, with memory sitting directly beside the compute fabric, moves tokens faster and with less communication overhead than thousands of networked accelerators. Speed-of-light latency is the product; that is why the company leads its story with inference.
The 2024 Setback and Two Private Bridges
Cerebras first filed for an IPO in 2024. That attempt stalled when the investment from Abu Dhabi-based G42 drew a federal review, and the company eventually withdrew. The episode mattered beyond timing: a foreign-investor review hanging over the cap table is exactly the kind of overhang public-market investors refuse to price, and it forced the company back into private markets to keep funding its roadmap.
Private markets bridged the gap: a $1.1 billion Series G in September 2025, then a $1 billion Series H in February 2026 at a valuation of roughly $23 billion, per the WSJ. Refiling now amounts to a declaration that the regulatory and ownership questions have been settled enough to survive public-market scrutiny a second time — and that management believes the open window matters more than perfect timing.
Two Golden Tickets: OpenAI and AWS
The customer list is the skeleton of the valuation story. The WSJ reports that Cerebras’ computing partnership with OpenAI is worth more than $10 billion. Feldman was blunt in an interview with the paper: “Obviously, [Nvidia] didn’t want to lose the fast inference business at OpenAI, and we took that from them.”
The second ticket is AWS: the two have an agreement to bring Cerebras chips into Amazon data centers. That is a striking purchase from a company whose own silicon franchise (Graviton, Trainium, and friends) already crosses $20 billion in annual revenue. A hyperscaler buying outside wafer-scale inference capacity is a public admission that the approach wins on certain workloads — and it gives Cerebras a second, diversified demand channel beyond a single frontier lab.
What It Means for the Chip Market and Developers
Four observations. First, the inference battlefield has fully formed: with OpenAI and AWS both adopting Cerebras, a “non-Nvidia inference route” now carries two hyperscale endorsements for the first time. Second, the capitalization window for second-tier AI chipmakers is opening: a company that is still unprofitable on a non-GAAP basis feels safe filing, which says the market’s appetite for compute-supply diversification far outweighs its demand for near-term profit. Third, the deal structure is the story to watch: contracts of this size increasingly tie compensation to compute purchases, and how the OpenAI and AWS agreements allocate that risk will be visible as the filing is amended.
Fourth, for developers the effects land on speed and cost: if Cerebras capacity scales through AWS and OpenAI deployments, frontier-model latency and per-token pricing could find a new anchor — worth leaving room for in model selection and cost planning. An inference option that is meaningfully faster at comparable quality changes batch-versus-latency tradeoffs in agent pipelines more than it changes chat applications, and that is where architects should model it first.
Sources
- AI chip startup Cerebras files for IPO — TechCrunch
- Cerebras IPO filing — SEC EDGAR
- Chip Startup Cerebras Files for Initial Public Offering — The Wall Street Journal
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.