Cerebras Goes Public: The $23B Wafer-Scale Bet, the OpenAI Lifeline, and the G42 Concentration Cliff

On April 17, Cerebras Systems filed its Form S-1 with the SEC, targeting a Nasdaq listing under the ticker CBRS at an approximate $23 billion valuation. The roadshow extended into May, with secondary-market quotes on Forge, Hiive, and Notice recently bracketing $102–$107 per share — an implied $26–$28 billion private mark heading into pricing. For pre-IPO investors who have spent the last two years watching the AI infrastructure trade run almost entirely through one ticker, Cerebras is the first venue-priced, pure-play alternative to NVIDIA on a public tape.

It is also one of the more contradictory listings of the cycle. The S-1 includes a $10 billion+ OpenAI compute agreement, a 245% revenue growth trajectory, a Series H led by Tiger Global with strategic participation from AMD, and a wafer-scale chip that benchmarks at multiples of NVIDIA Blackwell on inference. It also includes 86% historical revenue concentration to entities tied to the United Arab Emirates, a CFIUS file that took two years to clear, and a preferred-stock conversion mechanic that quietly reprices the deal as the offering moves up the price range.

The S-1 in plain language

Cerebras refiled in late February 2026 after withdrawing its 2024 paperwork the prior October. The company closed a $1.01 billion Series H on February 4, 2026 at $89.02 per share, anchoring a $23 billion post-money. Investors of record include 1789 Capital, Alpha Wave Global, Altimeter, AMD, Atreides, Benchmark, Coatue, Fidelity, and Tiger Global. Morgan Stanley, Citigroup, Barclays, and UBS are listed as underwriters, with the company targeting a roughly $2 billion raise.

The financial story is unusual for a chip IPO. Estimated 2024 revenue, annualized from the original S-1 disclosures, was approximately $272 million. The OpenAI agreement alone — up to 750 megawatts of inference compute through 2028, valued at more than $10 billion — is large enough to materially restate the revenue mix in 2026 and beyond. If the deployment hits its scheduled ramp.

The OpenAI deal is the entire story — and also the entire question

The January 14 announcement that OpenAI had committed to $10 billion of compute over three years was the single largest non-NVIDIA AI infrastructure contract ever signed. Reuters and CNBC put the cap on the deal at 750 megawatts; CEO Andrew Feldman confirmed the term sheet was signed before Thanksgiving 2025. That contract is what moved Cerebras from an $8.1 billion valuation in September to $23 billion in February.

Two questions matter for the public market.

First, the timing. 750 megawatts of inference capacity is enormous — roughly the load of a mid-sized U.S. city — and that capacity has to be physically built. Cerebras' CS-3 systems draw approximately 23 kilowatts each, are manufactured at TSMC on a 5nm process, and require purpose-built data center sites with power agreements, cooling, and grid interconnect. Slippage in any one of those gates compresses revenue recognition.

Second, the dependency. OpenAI is a company that is itself reportedly preparing for an IPO at up to a $1 trillion valuation, and that has signed multiple parallel infrastructure deals — with Oracle, with NVIDIA, and as of late 2025 with NVIDIA-acquired Groq. A Cerebras public valuation that prices in full OpenAI delivery is, structurally, also a position on OpenAI's capital plan and on the political stability of cross-border AI infrastructure.

The G42 cliff and the F2 strike-price problem

The harder issue in the S-1 is concentration. Group 42 Holdings, an Abu Dhabi-based AI company in which Microsoft holds a stake, accounted for 83–87% of Cerebras' revenue across 2023 and the first half of 2024. The company reports that G42's 2025 share dropped to 24%, but that the gap was filled by Mohamed bin Zayed University of Artificial Intelligence at 62% — another UAE-linked counterparty. In CFIUS terms, the customer base is still concentrated; the names just rotated.

Layered on top is the Series F-2 preferred mechanic. G42 holds the right to buy 23 million shares of Series F-2 convertible preferred at $14.66 per share. At a $23 billion IPO price, that strike is deep in-the-money; at a $25–$28 billion print, the dilution to public-market investors gets worse the better the deal does. Practical Venture Capital's Dave McClure flagged this on April 26 as the reason he expects Cerebras to broaden the customer story before pricing — either through a publicly disclosed Amazon, Oracle, or hyperscaler agreement, or by trimming the F-2 conversion through negotiation.

Why this matters for the wider pre-IPO trade

Cerebras is the pricing test case for three things that have been building underneath the secondary market for eighteen months.

• Inference is the new training trade. Spending on running models has now overtaken spending on training them. Cerebras' wafer-scale architecture — 4 trillion transistors, 900,000 cores, 21 PB/s on-chip memory bandwidth — is purpose-built for the latency profile that agentic AI workloads demand. The IPO is the public market's first chance to price an inference-native chip company directly against an NVIDIA comp.
• The "anti-CUDA" premium is real but bounded. Roughly 30x forward sales is what NVIDIA trades at; AMD trades around 10x; Arm sits below 40x. A $23 billion mark on roughly $272 million of 2024 revenue prices Cerebras at ~85x. That spread is the public-market value of being the credible non-NVIDIA option. It is not a price that survives a single missed quarter.
• Foreign-customer concentration is now an underwriting variable. CFIUS approved the G42 share sale in March 2025. A future administration, or a future incident, can re-open it. Public-market investors price geopolitics the way private investors price founder risk — expensively, and only after something breaks.

The secondary-market read

The pricing tell, before the deal is locked, is the secondary tape. Forge, Hiive, and Notice quotes around $102–$107 per share imply a $26–$28 billion private valuation — meaningfully above the $89 Series H clearing price and roughly in line with where speculative pre-IPO money believes Cerebras will open. Subsequent strategic disclosures (Oracle confirmed Cerebras availability in its catalog during March's earnings; Mayo Clinic and Department of Defense engagements have been disclosed publicly) suggest underwriters are working on the customer-diversification narrative the F-2 problem requires.

The structural setup — concentrated customer base, dilutive preferred mechanic, transformational anchor contract, and the only liquid public name in the inference category — is the kind of profile that resolves either as a step-change re-rating or as a 12-month grinding correction back toward the $14.66 strike. There is very little room in the middle.

Education takeaway

Cerebras sits at the intersection of two of our standing research verticals: AI & Robotics infrastructure, and Pre-IPO Markets. The questions a pre-IPO investor needs to be able to answer before pricing are not "can I get an allocation" — they are structural. How is the F-2 conversion priced into the cap table? What happens to revenue recognition if OpenAI's Stargate cadence slips by a quarter? What is the public-market multiple compression risk if G42 share-of-revenue does not continue to fall? Those are the questions our weekly research is built to walk through. Cerebras is the cleanest case study to do it on this year.