Cerebras Grows Faster Than the Market Can Forgive

Cerebras posted a 103 percent jump in core revenue and a $25.4 billion backlog in its second quarter as a public company, yet a $450 million loss sent its shares sharply lower.

Key Takeaways

• Core revenue reached $209.9 million, up 103 percent year over year, as cloud and inference revenue nearly quadrupled to $127.7 million against a $25.4 billion order backlog.
• A $450.5 million GAAP net loss, driven mostly by $377 million in IPO-related stock compensation, sent Cerebras shares down roughly 14 percent in after-hours trading.
• Hardware revenue fell 23 percent year over year even as cloud surged, evidence of a deliberate shift from selling chips to leasing inference capacity.

A Quarter of Contradictions

Cerebras Systems has spent two years arguing that speed, not scale, is the more interesting bet in artificial intelligence hardware. Its second earnings report as a public company, released on August 12, gave investors reason to believe it. Core revenue for the quarter ended June 30 reached $209.9 million, more than double the $103.3 million reported a year earlier, and the fast inference cloud business at the center of Chief Executive Andrew Feldman’s pitch brought in $127.7 million on a core basis, up 287 percent. Remaining performance obligations, the contracted revenue not yet recognized, stood at $25.4 billion, dwarfing anything the company has previously reported and underscoring, in management’s telling, a backlog that is not merely large but accelerating.

None of that stopped the stock from falling roughly 14 percent in after-hours trading, and by the next session it had shed closer to 15 percent from its prior close. The gap between what the numbers said and how the market responded is the real story here, revealing as much about the mechanics of a newly public, richly valued stock as about the underlying business. Cerebras posted a net loss of $450.5 million under GAAP, its largest since listing, even as the metrics it uses to describe its own performance improved on nearly every measure. Untangling why investors weighted one set of figures over the other is where the quarter becomes genuinely instructive.

The Ledger Behind the Headline

The most consequential number in the release sat not in the headline but in the segment detail. Hardware revenue, the CS-3 systems Cerebras has traditionally sold outright, fell 23 percent year over year to $54.1 million, even as cloud and other services revenue climbed 281 percent to $126.0 million on a GAAP basis. That inversion, a shrinking hardware book beside an exploding cloud one, is no accident. It is the clearest evidence yet of a transition management signaled more than a year ago: from selling expensive machines to renting extremely fast compute by the hour. The economics of that shift, recurring and higher margin against upfront and thin, are exactly what a bullish investor wants, provided it does not outrun the hardware pipeline feeding the cloud business.

Profitability is where the quarter turns genuinely difficult to parse. GAAP gross margin compressed to 14.2 percent from 31.1 percent a year earlier, largely because of pass-through data center costs that generate almost no margin of their own. Strip those out, along with stock compensation and warrant amortization, and core gross margin actually rose to roughly 41 percent, an improvement of about 940 basis points. The net loss follows the same pattern: $377.0 million of the $450.5 million GAAP shortfall was stock-based compensation tied to vesting triggered by the May IPO, a mechanical event rather than a recurring drag. Core net loss narrowed to just $6.9 million from $40.5 million a year earlier. Last year’s apparent GAAP profit was itself something of a mirage, the product of a $363.3 million non-cash gain on the extinguishment of a forward contract rather than of operating income.

Backlog, Capacity, and Concentration

The $25.4 billion backlog matters only if Cerebras can build capacity to deliver against it, and the quarter offered evidence that it can. More than 600 megawatts of data center capacity are now live or under contract for delivery by the end of 2027, with a pipeline running into the gigawatts. Manufacturing capacity is set to expand more than tenfold this year through new production lines at Flex, Sanmina, and Rocket EMS, with wafer supply secured from TSMC well ahead of need. The company also cites a structural advantage: its wafer scale architecture sidesteps the high bandwidth memory, advanced packaging, and leading edge fabrication nodes that leave rival chipmakers competing for the same constrained inputs.

That buildout also targets a concentration problem that nearly kept Cerebras private for good. The company first filed to go public in September 2024, then withdrew the registration after a national security review scrutinized its relationship with G42, the Abu Dhabi based investor and customer that once accounted for the overwhelming majority of its revenue. The Committee on Foreign Investment in the United States cleared the arrangement in March 2025, once G42 agreed to hold only non-voting shares, but by the time Cerebras returned to market in May 2026, the risk had simply relocated. G42’s share of 2025 revenue fell to 24 percent, while Mohamed bin Zayed University of Artificial Intelligence, also UAE based, rose to 62 percent. A December 2025 agreement with OpenAI for 750 megawatts of capacity, valued at more than $20 billion, along with newer commitments from AWS, AMD, and CrowdStrike, marks genuine progress toward a broader customer base, though a single geography still anchors most of what Cerebras has under contract.

The Mechanics of the Sell-Off

Part of the market’s reaction traces to a distinction that got lost somewhere between the boardroom and the trading floor. Cerebras had guided to approximately $194 million in core revenue for the quarter, and core revenue came in at $209.9 million, a clear beat. But GAAP total revenue landed at $180.1 million, below that same $194 million figure, and several data providers built their consensus estimates around the GAAP number rather than the core one Cerebras had guided to. The result was a release that Chief Financial Officer Bob Komin described as exceeding guidance “across all core business metrics,” and that much of the financial press nonetheless characterized as a miss. Both descriptions are defensible. Neither is complete without the other.

Layered onto that confusion were figures with less ambiguity: the size of the loss, a working capital loan carrying an outstanding balance of $918 million understood to trace to OpenAI’s financing support, and operating cash flow that stayed negative through the first half, though narrower than a year earlier. Add $548.9 million of first-half capital spending and a stock still trading well below the roughly $350 level at which it opened in May against a $185 offering price, and the sell-off looks less like a verdict on the business than a recalibration of how much uncertainty a young, volatile stock can absorb in one print. Wall Street’s positioning reflects that split: Needham’s Quinn Bolton holds a $300 target, consistent with a Street average near $290 to $298 and a broad Strong Buy consensus, while Freedom Capital’s Paul Meeks carries the Street’s lowest target at $209, doubting how durable the current margin trajectory is.

The Guidance and the Test Ahead

Management’s response to a punishing stock reaction was not caution but its opposite. Third-quarter core revenue guidance was set at $214 million to $216 million, and full-year core revenue guidance was raised to $880 million to $890 million, up from $855 million to $865 million, with core gross margin guidance improving to a range of 41 to 43 percent. Komin told investors the company expects to more than triple revenue in 2027, a target that would require capacity and manufacturing expansion to hold roughly on schedule, and that treats the $25.4 billion backlog less as an asset to admire than as a commitment to execute against.

What the quarter ultimately demonstrates is a familiar tension for capital-intensive infrastructure companies in the earliest stages of public life. Non-cash compensation charges and pass-through costs can make a GAAP income statement look considerably worse than the business actually is, and a backlog measured in the tens of billions reveals very little, on its own, about near-term cash conversion. The more instructive signal from Cerebras’s second quarter is not the double-digit decline that followed it, but the direction of travel underneath: cloud revenue nearly quadrupling, core margins expanding by hundreds of basis points, and a customer base still concentrated yet visibly widening. Whether that trajectory survives a tripling of 2027 revenue will determine whether this quarter is remembered as an entry point or as the first sign of strain in a story the market has, until now, been eager to believe.