Cerebras Systems priced its IPO at $185 on May 13, opened at $350 on May 14, hit $386 intraday, and closed its first session at $311. By Friday it was trading near $280. The fade began before the champagne was warm. That is not random noise. That is the market starting to read the prospectus.
The Revenue Story Has One Chapter
In 2025, Cerebras generated $510 million in revenue, up 76% from $290 million in 2024. The growth curve is real and the trajectory from $25 million in 2022 to $510 million in 2025 is genuinely impressive. None of that changes what the revenue actually is.
Two UAE-linked customers accounted for 86% of 2025 sales. G42, the Abu Dhabi AI conglomerate, contributed 24%. The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) contributed 62%. Cerebras’ own prospectus identifies MBZUAI as a G42 related party. The customer concentration did not improve between the first IPO attempt and the second. The labeling did.
This is the same structure that triggered a CFIUS national security review in 2024 and forced Cerebras to withdraw its original filing. The concern then was G42’s historical ties to Huawei and the risk of advanced American AI technology transiting through UAE intermediaries toward China. That review was never formally resolved. It was navigated around. Renaming 62% of your revenue from one UAE entity to a different UAE entity with documented organizational ties to the first UAE entity is a legal restructuring, not a risk reduction.
The Backlog Is a Promise
The number that drove the 20x oversubscription is the $24.6 billion remaining performance obligation — contractually committed revenue not yet recognized. That figure is approximately 48 times 2025 sales and it is the entire bull case compressed into one line. Management’s own guidance is that 15% of that backlog will be recognized across 2026 and 2027 combined, implying annualized forward revenue near $1.85 billion.
The largest component of that backlog is the $20 billion, 750-megawatt compute deal signed with OpenAI in early 2026. OpenAI is a sophisticated counterparty that builds its own infrastructure, negotiates on price as a core competency, and has every structural incentive to use Cerebras as leverage against Nvidia rather than as a primary long-term vendor. A commitment in a backlog is not the same as recognized revenue, and a backlog driven by one customer whose own financial position is not without complexity is not the same as a diversified order book.
The Profitability Gap
The non-GAAP net income figure of $237.8 million circulated widely during the roadshow. The GAAP operating loss was $345 million. The gap between those two numbers is not rounding error. It reflects one-time accounting reversals tied to forward-contract liabilities that will not recur. The business is not operationally profitable at the GAAP line. At $280 per share, investors are paying roughly 55 to 60 times trailing sales for a company losing money by standard accounting, with customer concentration that would be disqualifying in any sector not currently bathed in AI enthusiasm.
The Chip Is Real. The Moat Is Not.
The wafer-scale engine is a legitimate technological achievement. At 57 times the die size of Nvidia’s H100 and purpose-built for inference speed, it solves a real problem for a real class of workloads. This is not a paper company.
It is also not a platform. Nvidia’s advantage in AI infrastructure is not the H100. It is CUDA — fifteen years of developer tooling, software libraries, enterprise integration, and institutional knowledge that makes switching to any alternative a multi-year organizational project. Cerebras competes on hardware specifications against a company that wins on software lock-in. The AWS partnership gives Cerebras distribution. It does not give it an ecosystem.
The Short Thesis
At the current price, the market is pricing Cerebras as though it has already won the inference market, the OpenAI backlog executes on schedule, a third major non-UAE customer materializes within two years, and CFIUS finds no further cause for concern. All four of those conditions need to hold simultaneously.
The timing for a short position is not today. The float is thin, the borrow is expensive, and a single positive catalyst — an OpenAI deployment announcement, a new enterprise customer, an AWS volume disclosure — could push the stock back toward $350 before the fundamental gravity sets in. The 180-day lock-up expiration in November 2026 is the structural inflection point. When insider shares unlock and the float expands, the selling pressure becomes institutional rather than retail-driven.
The first and second earnings reports are the diagnostic tests. If OpenAI recognized revenue remains immaterial and UAE concentration holds above 80%, the narrative cracks on its own without requiring a fight against post-IPO momentum. The short thesis is patient. The balloon does not need to be punctured. It needs to be left alone long enough to find its own ceiling.
The fundamental case is straightforward: 86% revenue dependence on two related UAE entities, a $345 million GAAP operating loss, no durable software moat, and a valuation that requires every optimistic assumption to prove correct. That combination has a name. It is not a $280 stock.