Skip to content
AI · Markets · August 13, 2026

Cerebras Shares Tumble After Hardware Revenue Miss Overshadows a Strong Quarter

Cerebras Systems’ stock did something unusual after its second-quarter numbers hit the wire Wednesday evening, August 12: it climbed as much as 12% during the regular session on pre-earnings optimism, closed near $262, then fell 14% to 17% in after-hours trading once the actual figures landed — settling into a band around $214 to $220 by Thursday morning.

The swing traces to a genuinely confusing set of numbers, because Cerebras reports two different revenue figures that told two different stories this quarter. Under standard U.S. accounting rules (GAAP), Cerebras booked $180,100,000 — up 74% from a year earlier, but short of the roughly $191 million to $194 million analysts expected. Under “core revenue,” the company’s own non-GAAP measure of its underlying AI-chip and cloud business that Cerebras and most Wall Street analysts actually use to track growth, the company brought in $209,900,000, up 103% year over year and comfortably ahead of estimates.

That gap between an accounting miss and a genuinely strong quarter in the business Cerebras actually runs is the story that follows. It is not, on the numbers the company itself reported, a story about growth slowing down.

§ 01 / The Numbers
Two Numbers, One Quarter

GAAP revenue: $180,100,000, up 74% year over year — missed the roughly $191–194 million Wall Street expected.

Core revenue: $209,900,000, up 103% year over year — beat estimates and drove a guidance raise.

GAAP is the figure required under standard accounting rules. Core revenue is Cerebras’ own adjusted measure of the underlying chip-and-cloud business, and it is the number the company, its CFO, and most sell-side analysts point to when describing how fast Cerebras is actually growing.

Cerebras’ GAAP net loss for the quarter was $450,500,000, or $2.98 a share — a sharp reversal from the $309,500,000 in GAAP net income the company reported in the second quarter of 2025, and far worse than the roughly $0.18 a share Wall Street had modeled. But the loss is not primarily a sign of the business deteriorating: roughly $377 million of it is non-cash stock-based compensation tied to equity that vested when Cerebras completed its initial public offering in May — a one-time accounting charge, not cash leaving the company. The year-ago comparison carries its own asterisk, too: that $309,500,000 in Q2 2025 GAAP income was itself driven largely by a one-time, non-cash gain tied to the extinguishment of a forward contract linked to Cerebras’ UAE-based customer G42 — not by underlying operating profit. Neither year’s GAAP bottom line is a clean read on the business alone.

Strip that charge out and the underlying business looks considerably healthier. GAAP cloud and services revenue — the fast-inference business Cerebras sells directly to AI developers — came in at $126,000,000, up 281% year over year. Core gross margin, the company’s preferred profitability measure, was 40.6%, up roughly 940 basis points from a year earlier. Cerebras did not just meet its own targets; CFO Bob Komin told analysts on the call that “our quarterly results exceeded our guidance across all core business metrics.”

The company responded to that underlying strength by raising, not cutting, its outlook. Cerebras lifted its full-year 2026 core revenue guidance to a range of $880 million to $890 million, up from the $855 million to $865 million range it had set in June, and guided to $214 million to $216 million in core revenue for the third quarter. It also raised its full-year core gross margin guidance to a range of 41% to 43%.

This was an outstanding quarter for Cerebras. Core revenue more than doubled to $210 million.

Andrew Feldman · CEO & Co-Founder, Cerebras Systems · Q2 2026 earnings release, August 12, 2026
Cerebras Stock (CBRS) Earnings Call | Q1 2026 — Future Investing
Keep Reading
Alibaba Says Qwen Passed 3 Billion Downloads — Ahead of Meta and Google, by Alibaba's Own Math
AI · Markets
Alibaba Says Qwen Passed 3 Billion Downloads — Ahead of Meta and Google, by Alibaba's Own Math
Read →
§ 02 / Why Hardware Fell

If core revenue and cloud/services revenue were both climbing sharply, what actually dragged the GAAP number below consensus and rattled investors? One specific line item: hardware sales — the systems Cerebras builds and ships to customers who run their own on-premises AI infrastructure — fell to $54,100,000, down 23% from roughly $70.3 million a year earlier. That is the figure Bloomberg’s own headline on the quarter singled out, describing “lumpy demand” for Cerebras hardware.

A 23% drop in one hardware line item, not a slowdown in the underlying business, is what dragged Cerebras' GAAP revenue below estimates — Civic Intelligence illustration

The pattern is not unusual for a company selling large-ticket AI infrastructure: hardware orders arrive in chunky, multi-million-dollar batches tied to when a customer builds out a new data-center deployment, rather than as a smooth monthly revenue stream. A single delayed shipment, or a customer choosing to lean on Cerebras’ cloud offering instead of buying hardware outright, can swing the quarter’s hardware number sharply in either direction. What alarmed some investors more than the hardware dip itself was a separate figure: Cerebras’ backlog of contracted future business — what the company calls remaining performance obligations — grew to just $25,400,000,000, up only 1.6% from $25,000,000,000 the prior quarter. A backlog growing that slowly, next to a 103% core-revenue growth rate, raises the question of whether the pipeline behind that growth is itself keeping pace.

This was also not the first time in 2026 that a Cerebras earnings report produced a stock reaction disconnected from its headline growth rate. When the company reported its first quarter as a public company on June 23, revenue beat expectations, but full-year sales and margin guidance disappointed investors enough to send the stock down as much as 20% the next trading day; by June 25, shares were down roughly 47% from the high they touched on IPO debut day. Wednesday’s reaction continued that pattern rather than establishing a new one.

§ 03 / The Concentration Problem

Analysts who cover Cerebras have circled the same underlying question since before the IPO: how much of the company’s growth depends on a small number of very large customers. In January 2026, Cerebras and OpenAI announced a supply agreement under which Cerebras will provide OpenAI with 750 megawatts of computing capacity through 2028, initially valued at more than $10,000,000,000; by the time of Cerebras’ first public earnings report in June, the disclosed value of the relationship had roughly doubled to more than $20,000,000,000. By one analyst’s estimate, that single relationship now accounts for roughly 24% of Cerebras’ revenue and “most of the backlog” the company reports each quarter.

That concentration is not new for Cerebras — it has simply moved. The company’s own IPO prospectus disclosed that 86% of its fiscal 2025 revenue came from entities tied to the United Arab Emirates, principally G42 and the Mohamed bin Zayed University of Artificial Intelligence. Independent analyst Patrick Moorhead put the shift bluntly: “Concentration did not go away; it rotated.” He added: “The margin path, the cash burn, and the concentration are the questions that matter now.”

CEO Andrew Feldman has argued the concentration will ease over time rather than deepen, telling analysts on the earnings call that “OpenAI will remain important, but its share of revenue should shrink over time as AWS, other hyperscalers and new customer categories grow.” Feldman has also used earnings-adjacent appearances to press the company’s core pitch against market leader Nvidia, whose accelerators hold an estimated 75% to 92% share of the AI-chip market; Cerebras’ WSE-3 chip is roughly 58 times the physical size of Nvidia’s Blackwell B200 and, the company says, runs inference 15 to 20 times faster on certain benchmarks. Asked about competing on price, Feldman told Reuters on August 12: “Nvidia’s prices have gone through the roof because of HBM prices” — high-bandwidth memory, the same component shortage driving up costs industrywide.

Cerebras CEO on Delivering AI Inference at Scale — Bloomberg Live
Keep Reading
Nvidia Taps Wall Street for $500 Billion in AI Infrastructure Financing
AI · Markets
Nvidia Taps Wall Street for $500 Billion in AI Infrastructure Financing
Read →
§ 04 / What Wall Street Thinks Now

Wall Street’s read on Wednesday’s numbers split along familiar lines. Wedbush Securities analyst Matt Bryson maintained the $280 price target and Outperform rating he set in late June, effectively unchanged by Wednesday’s release. Morgan Stanley analyst Joseph Moore has separately argued, in commentary tied to Cerebras’ prior-quarter reaction, that after-hours weakness following a strong quarter can be a buying opportunity — a framing several analysts echoed again this week, though a Wednesday-specific note from Moore could not be independently confirmed.

Cerebras earnings on deck: Here's what to expect (Q1 2026 preview) — CNBC Television
X
LiveSquawk
@LiveSquawk · August 12, 2026

Cerebras Q2: revenue $180.1M vs $194M est., cloud revenue $126.6M vs $116.3M est. (beat), hardware revenue $54.1M vs $73.1M est. (miss). Q3 guide $214-216M.

That real-time post’s cloud-revenue figure is a live wire-speed estimate that differs slightly from the $126,000,000 Cerebras’ own release later confirmed — the kind of small reconciliation that’s normal between in-the-moment earnings coverage and the final numbers.

By Thursday, August 13, Cerebras shares had stabilized in the $214-to-$220 range — down sharply from Wednesday’s regular-session close near $262, but still above where the stock had traded through most of the summer following June’s guidance-driven selloff. The company remains far below its post-IPO peak: Cerebras priced its May 14 offering at $185 a share and closed its debut trading day near $311; Wednesday’s after-hours low left the stock closer to its offering price than to that first-day high.

Bottom Line

Cerebras did not have a bad quarter by the measure the company and most analysts who cover it actually track: core revenue more than doubled year over year to $209,900,000, beating estimates, and the company raised its full-year guidance on the strength of it. What spooked investors was narrower and more specific — a GAAP revenue miss driven substantially by a 23% year-over-year decline in hardware sales, a $450,500,000 GAAP net loss inflated by a one-time, non-cash IPO stock-compensation charge, and a $25,400,000,000 backlog that grew just 1.6% from the prior quarter. Those are legitimate questions about the durability of Cerebras’ pipeline and its dependence on a small number of enormous customers. They are not evidence that the underlying business — the one growing 103% a year — is slowing down.

More From Civic Intelligence
Sources & Methodology · 9 Sources
Methodology: this page reports two distinct revenue figures from Cerebras’ own Q2 2026 earnings release and earnings call and is careful to label which is which throughout. “GAAP revenue” ($180,100,000) is the figure required under standard U.S. accounting rules; “core revenue” ($209,900,000) is Cerebras’ own non-GAAP measure of its underlying AI-chip and cloud business, which the company and most analysts covering it use to track growth quarter to quarter. Both are drawn from the same primary earnings release and call transcript. This is a non-partisan technology and markets story on the site’s AI beat: no party affiliation applies to any named individual, and Truth Social is omitted per the beat’s convention for business and financial coverage. Only one X post specific to Wednesday’s earnings could be independently verified as live at the time of writing; this page relies on that post plus its three video embeds rather than manufacture an additional citation. Two of the three videos are from Cerebras’ Q1 2026 earnings cycle (labeled as such in their titles) and are used for CEO/company background rather than as Q2-specific reaction coverage, since no verified Q2 video existed at the time of writing.