Cerebras' $25 Billion Backlog Is Real. Its Margins Are Now Paying for It.

The company with the biggest order book in AI is also the one investors keep selling. Cerebras Systems signed a multi-year deal valued at more than $20 billion to run fast inference for OpenAI, holds a $25.4 billion backlog of undelivered work, and tells shareholders it expects core revenue to more than triple next year — and still the stock fell roughly 12% since it reported second-quarter 2026 results, slipped on its first quarterly report after the IPOshares drop on earnings debut, with margins, and sits about 40% below the high it set earlier this year.

The demand is not the puzzle. The demand is the point of friction.

What Cerebras actually wins

Cerebras makes "wafer-scale" chips — instead of cutting a silicon wafer into dozens of processor dies, it uses nearly the whole wafer as one enormous chip. That design solves a specific, increasingly valuable problem: quickly generating long, word-by-word output from an AI model. On Nvidia GPUs, that decode phase is the latency bottleneck; Cerebras eliminated most of the memory and interconnect transport that slows it down. As the industry shifts from training models to serving them — the inference transition that determines who earns the next cycle — speed on that single step is a real architectural edge, not a marketing line.

The scale of the bet is what makes the valuation story. The OpenAI agreement, announced in January, commits up to 750 megawatts worth of computing power to OpenAI through 2028. For context, the company's entire expected revenue for 2026 is roughly $885 million. One multi-year contract from one customer is worth more than 20 times a year of sales. Cerebras has also tied into AWS, Cerebras systems to be offered on Amazon Bedrock in Q1 2027, and pairs its fast decode step with GPU prefill from AMD.

Why a strong report still fell

Here is the operating result that the cheerleading around the backlog misses. To serve all that demand, Cerebras does not yet own enough of its own machines, so it rents compute from third parties — and the rental charge lands directly in cost of sales. Core gross margin fell from roughly 47% in the first quarter to about 41% in the second, and management said it would have been about five points higher without the rented systems. That is why, despite cloud revenue up 287% and a raised full-year outlook, the market read the quarter as a miss and sold the stock again.

Look close and the accounting tells the same story. Reported GAAP revenue actually fell, from $193.4 million in Q1 to $180.1 million in Q2, because the hardware line shrank even as cloud billings soared. Reported gross margin was just 14% in the quarter, versus the 41% "core" figure management highlights. The gap is the difference between what a company earns from operations and what it reports after the stock-based compensation and customer-warrant costs that come with paying customers partly in paper — a reported net loss of roughly $450 million in the quarter. For a retail investor, "core" is the metric Cerebras wants you to watch; GAAP is the bill.

What the market is already paying for

At a market value near $50 billion, Cerebras trades at roughly 56 times this year's expected revenue. Even if management's forecast comes true and 2027 revenue triples to about $2.6 billion, the current price is still near 19 times that 2027 number — and it assumes two things that have not happened yet. The first is that the rented-capacity period is temporary, that Cerebras' own lower-cost machines replace it and pull gross margin from the current 40s toward the 60% the company targets long term. The second is that the growth compounds after the triple, because 19 times a single good year is not cheap for a company that just reported a negative core operating margin and a large GAAP loss.

The counterweights are real. If owned systems replace rented ones and margins snap back in the fourth quarter as promised, then a $25 billion backlog — about 29 times this year's revenue — can carry growth for several years. The risk is that the rental crutch lasts longer than planned, or that the two customers holding most of that backlog, OpenAI and G42, throttle their deployments; Cerebras' own filings warn that a reduction in demand from a limited number of significant customers would hit results, and just two years ago G42 accounted for 87% of revenue in the first half of 2024.

This is the dual reading of a supply commitment. The $25.4 billion order book is simultaneously the strongest demand signal in the AI hardware trade and a delivery-and-margin risk, because filling it today costs more than it earns. At $210, you are not buying the growth story — that is already priced, fully. You are buying the margin swing: that Cerebras converts its backlog into revenue faster than it pays to rent the capacity to do it. Demand is not the question here. Whether the buildout outruns its own cost is.