Cerebras Systems filed an updated IPO prospectus on Monday targeting 28 million shares at $115 to $125 each, a range that would raise up to $3.5 billion and value the wafer-scale AI chip company at up to $26.6 billion, making it the most significant test of public-market appetite for specialised AI hardware since Nvidia's own trajectory became the benchmark every AI infrastructure company is measured against.
The technology is genuinely different. Cerebras builds what it calls the Wafer-Scale Engine, a processor that occupies an entire silicon wafer rather than being cut into individual chips like every conventional processor in commercial production. The WSE-3, its current generation, contains 4 trillion transistors and 900,000 cores on a single surface measuring 46,225 square millimetres, roughly 57 times the physical size of Nvidia's H100. That scale eliminates the inter-chip communication bottleneck that limits throughput in GPU clusters: where an H100 system moves data between discrete chips over high-speed interconnects, the WSE moves it internally across a unified memory space at 21 petabytes per second of on-chip bandwidth. The company claims the CS-3 system delivers 21 times the performance of Nvidia's DGX B200 at one-third the cost and power draw for specific inference workloads. Those claims have not been independently audited at scale, but they are specific enough and the customer roster, which includes Argonne National Laboratory, GlaxoSmithKline, IBM, and OpenAI, is credible enough that they cannot be dismissed as marketing alone.
The financial profile that accompanies those technical claims is where the roadshow conversations will concentrate. Revenue grew 76% year over year to $510 million in 2025, a trajectory that justifies serious attention. The company reported GAAP net income of $237.8 million in 2025 after a $481.6 million net loss in 2024, a swing that reflects the revenue ramp rather than an operating model transformation. The critical number in the S-1 is not the revenue figure. It is the customer concentration figure: approximately 86% of Cerebras's 2025 revenue came from two UAE-based entities, G42 and Mohamed bin Zayed University of Artificial Intelligence. In 2023, G42 alone accounted for 83% of revenue. In the first half of 2024, the combined UAE concentration reached 87%. For a company seeking a $26.6 billion valuation, that is an unusual level of dependence on a handful of large sovereign deployments in a single geography, rather than the diversified, repeatable enterprise revenue base that justifies the kind of multiple Cerebras is seeking.
The OpenAI relationship changes the customer concentration calculus but does not resolve it. Cerebras signed a $20 billion Master Relationship Agreement with OpenAI in early 2026 for 750 megawatts of inference compute capacity, with a $10 billion core commitment and a 2 gigawatt expansion option. The deal is the largest AI infrastructure contract on record for a non-hyperscaler. If OpenAI's inference demand for applications like Codex Spark follows the adoption trajectory of its consumer products, Cerebras's compute could be fully committed well ahead of the MRA's three-to-four-year contemplated timeline. That scenario transforms a concentration risk into a capacity constraint, which is a different problem entirely. It is also the single strongest argument for the valuation. But it substitutes UAE customer concentration for OpenAI customer concentration, which is trading one form of revenue fragility for another, even if OpenAI's trajectory is more predictable than sovereign compute contracts.
CEO Andrew Feldman is not selling shares in the offering, retaining approximately 10.3 million post-IPO at values up to $1.28 billion at the top of the range. That is the right signal for a founder who wants public market investors to read his participation as conviction rather than exit. The company previously withdrew its first IPO filing in 2024 after pivoting from hardware sales to a cloud service model built on its proprietary chips. The revised model, which sells compute access rather than hardware units, has better revenue visibility and lower customer procurement friction. It also means Cerebras now operates a data centre business alongside a chip design business, which adds capital intensity and operational complexity to an already demanding engineering challenge. AMD participated in Cerebras's most recent private funding round, which is a notable strategic signal from a direct competitor about the technical credibility of the wafer-scale approach.
For other AI hardware startups watching this roadshow, Cerebras's IPO pricing carries information that no private fundraise can provide. If the offering prices at the top of the range and the stock holds or rises in early trading, it confirms that public markets are willing to assign premium valuations to AI infrastructure companies with differentiated architectures and large anchor customers, even with visible concentration risk and a software ecosystem that is years behind Nvidia's entrenched CUDA platform. That window, if it opens, changes the calculus for companies like Groq, SambaNova, and a cohort of inference acceleration startups that have been waiting for a public market signal before pursuing their own listings. If Cerebras prices poorly or the stock stumbles, it signals that public investors remain sceptical of anything in AI hardware that cannot demonstrate diversified revenue at scale, which is a tougher standard than most of the waiting cohort currently meets.
The valuation arithmetic is the last thing to stress-test. At $26.6 billion against $510 million in 2025 revenue, Cerebras is asking for a 52x revenue multiple. Nvidia trades at approximately 25x forward revenue, and Nvidia has 80% gross margins, dominant software lock-in through CUDA, and a customer base that includes every major hyperscaler and AI lab in the world. Cerebras has a technically superior architecture for specific inference workloads, two large anchor customers, a 76% revenue growth rate, and a manufacturing dependency on TSMC that it shares with most of the semiconductor industry. The premium over Nvidia's multiple is the market's bet on whether wafer-scale can scale from sovereign and anchor-customer deployments to the kind of broad enterprise adoption that justifies infrastructure valuations. The roadshow will determine whether investors believe that bet is worth making at $26.6 billion.
Also read: Japan's $2,000 cardboard drones are not a novelty and the defense startup implications are more serious than the material suggests • Six tech giants allegedly avoided $278 billion in US taxes while building their AI empires on public infrastructure and government contracts • Utah's VPN liability law for age verification is the clearest sign yet that state-by-state internet compliance is becoming a startup operating cost
