Summary: Cerebras filed IPO paperwork with the US SEC on April 18, 2026 and targeted a mid-May listing. Its filing reported $510 million of 2025 revenue, GAAP net income of $237.8 million, and a non-GAAP net loss of $75.7 million. The company also cited a reported OpenAI deal above $10 billion and an AWS arrangement to place its silicon in Amazon data centers.
Cerebras Systems has taken its second shot at the public markets: the AI chipmaker submitted IPO paperwork to the US SEC on April 18, 2026, picking up a listing push that collapsed two years ago. Speaking to TechCrunch, a company spokesperson set the target window for the listing at mid-May, while the size of the raise, the choice of exchange, and the underwriting lineup all remain under wraps.
What separates this filing from the last attempt is what the company now holds: revenue of $510 million for 2025, an OpenAI compute deal reported to be worth more than $10 billion, and an AWS arrangement that will place Cerebras silicon inside Amazon data centers. The Wall Street Journal was first to report the filing.
The entire second rank of AI silicon vendors has a stake in watching this one. Ever since Nvidia’s supply squeeze began, capital has poured into nearly every credible alternative accelerator maker, yet hardly any of those positions has reached public markets. Should Cerebras price and start trading, it becomes the reference asset for valuing AI compute startups — and the first genuine public measurement of what a non-GPU inference business is actually worth.
What the Filing Reveals About the Business
The SEC document puts Cerebras’ 2025 revenue at $510 million. Profitability reads two ways: on a GAAP basis the company netted $237.8 million, yet strip out one-time items and a non-GAAP net loss of $75.7 million surfaces. The takeaway is straightforward — accounting profit rests mostly on one-off items, and the core operation remains in build-out mode. That gap is exactly what IPO buyers should press on: when the stock prices, the live question will be how sturdy the underlying inference business proves once the one-offs roll off.
CEO Andrew Feldman still pitches the product the same way: Cerebras makes “the fastest AI hardware for training and inference”, built on a wafer-scale architecture that treats a whole wafer as one chip — a route that departs entirely from clusters of GPUs. The engineering wager behind it: one enormous die with memory placed directly beside the compute fabric moves tokens faster, with less communication overhead, than thousands of linked accelerators manage. The product, at bottom, is latency at the speed of light, and that is why inference sits at the front of the company’s story.
A Failed First Attempt, Then Two Private Lifelines
The company’s first IPO run began in 2024 and stalled after the investment from G42, an Abu Dhabi-based backer, drew a federal review; Cerebras eventually pulled the filing. The episode carried weight beyond the calendar: public-market investors will not price a cap table shadowed by a foreign-investor review, so the company had little choice but to return to private capital to keep funding its roadmap.
Private money carried it across: Series G brought in $1.1 billion in September 2025, and Series H added $1 billion in February 2026 at a valuation of roughly $23 billion, per the WSJ. Coming back now amounts to a public statement that the regulatory and ownership entanglements have been worked through well enough to face public-market scrutiny a second time — and that leadership values the currently open window over flawless timing.
OpenAI and AWS: The Two Marquee Customers
What anchors the valuation narrative is the customer roster. Per the WSJ, the computing partnership between Cerebras and OpenAI is reported to be worth more than $10 billion. Interviewed by the paper, Feldman put it bluntly: “Obviously, [Nvidia] didn’t want to lose the fast inference business at OpenAI, and we took that from them.”
AWS is the other marquee name: under their agreement, Cerebras chips will run inside Amazon data centers. Coming from a company whose in-house silicon line (Graviton, Trainium, and friends) already tops $20 billion in annual revenue, that outside purchase stands out — a hyperscaler paying for external wafer-scale inference capacity is a public acknowledgment that the approach wins on certain workloads, and it hands Cerebras a second, diversified demand channel beyond any single frontier lab.
The Takeaways for the Chip Market and Developers
Read it four ways. One: the inference battlefield has taken its final shape — with OpenAI and AWS both on board, the “non-Nvidia inference route” now holds two hyperscale endorsements at once, a first. Two: the window for capitalizing second-tier AI chipmakers is swinging open — a company still losing money on a non-GAAP basis judges it safe to file, a sign that the market’s hunger for diversified compute supply dwarfs any insistence on near-term profit. Three: the deal structures deserve the closest watch — agreements at this scale increasingly tie compensation to actual compute purchases, and amendments to the filing will show how the OpenAI and AWS contracts divide up that risk.
Four: developers will feel this through speed and price. If Cerebras capacity grows alongside the AWS and OpenAI deployments, frontier-model latency and per-token pricing may settle on a new anchor — reason enough to keep the option on the table in model selection and cost planning. An inference offering that is noticeably quicker at comparable quality reshapes batch-versus-latency tradeoffs in agent pipelines far more than it reshapes chat applications, and agent architectures are where architects should model it first.