S&P 500 7,743.41 +0.51%Nasdaq 27,068.72 +0.48%Dow 51,828.62 +0.93%Russell 2000 2,837.55 +0.07%as of 2026-09-25 close
◈ Frontier Tech Wire
Quantum, AI and frontier-tech small caps — on the wire
Chips & Hardware

Cerebras closed up 12%, then gave it back after hours as two versions of its revenue diverged

The wafer-scale AI chipmaker reported $180.1 million of GAAP revenue and $209.9 million of what it calls core revenue for the second quarter. The $30 million gap, most of it a non-cash customer warrant charge, is why the same print looked like a beat and a miss at the same time.
Illustrative photograph: computer server and electronics hardware.

Cerebras Systems reported second-quarter results after Wednesday's close, and the reaction depended almost entirely on which revenue line an investor was reading. The company said GAAP revenue was $180.1 million, up 74% from a year earlier, while what it labels core revenue was $209.9 million, up 103%. Shares had closed the regular session at $262.06, up 11.63% on the day, according to StockAnalysis market data. In extended trading they fell to $220.50 as of 6:11 p.m. ET, roughly 16% below the close. Published estimates of the after-hours decline varied with the timestamp: GuruFocus put the drop at 11.5%, and CNBC headlined a 14% plunge.

The gap between the two revenue figures is not a rounding convention. In its release, Cerebras said core revenue strips out $14.5 million of pass-through revenue — data center startup and recurring costs billed to customers at minimal margin — and adds back $44.3 million of customer warrant amortization, a non-cash charge that lands as contra-revenue under GAAP. The balance sheet carries $1.13 billion of customer warrants, split between $167.8 million current and $960.6 million non-current, so that quarterly drag is scheduled to keep reducing reported revenue for years. Start at $180.1 million, subtract the pass-through, add back the warrants, and you arrive at $209.9 million.

That distinction is what produced the split scorecard. GuruFocus reported that the $180.1 million GAAP figure came in below an analyst expectation near $194 million, and set the GAAP diluted loss of $2.98 a share against a consensus loss of roughly $0.17. Quiver Quantitative, measuring against the core line, reported revenue of $209.9 million versus an expectation of about $198 million and a loss of $0.04 a share against the same $0.17 estimate. The revenue comparisons are internally consistent on their respective bases, but the earnings comparisons are not: both outlets used the same $0.17 consensus, a figure built on the core basis, so pairing it with the $2.98 GAAP loss measures a non-GAAP estimate against a GAAP result. Held to one basis at a time, the quarter beat on core and missed on GAAP revenue. The after-hours tape suggests the GAAP version got the last word Wednesday evening.

The GAAP loss itself is dominated by items tied to going public. Cerebras disclosed $377 million of stock-based compensation in the quarter, excluding $22.8 million of IPO-related payroll taxes, against a GAAP net loss of $450.5 million. On the core basis the company reported a net loss of $6.9 million. The company priced its IPO at $185 a share in May, raising about $5.55 billion, according to Quartz. Built In reported it was the largest US listing of 2026 to that point, and Wednesday's report was only Cerebras's second as a public company.

Underneath the accounting, the mix shift is the operating story. Cloud and services revenue reached $126.0 million on a GAAP basis, up 281% year over year, while hardware contributed $54.1 million. Chief executive Andrew Feldman said in the release that "core revenue more than doubled to $210 million, and our cloud business nearly quadrupled year-over-year," and framed the company's pitch around inference latency: "Speed changes what AI can do. It makes AI more useful, more productive, and opens entirely new markets." Chief financial officer Bob Komin said in the release that "our quarterly results exceeded our guidance across all core business metrics."

Cerebras also put unusually large forward numbers on the table. It reported remaining performance obligations of $25.4 billion as of June 30 and said it now has 600 megawatts of data center capacity under contract for delivery by the end of 2027. For a company whose trailing quarterly revenue is measured in the low hundreds of millions, a backlog figure two orders of magnitude larger is the single most important disclosure to watch convert — or not — in coming quarters. Liquidity is not the constraint: the company reported $6.7 billion in cash and equivalents, $8.6 billion in total liquidity including investments, and an $850 million revolving credit facility.

Guidance moved higher on the core basis. For the third quarter, Cerebras guided to core revenue of $214 million to $216 million with core gross margin of 38% to 40% and core operating margin of negative 25% to negative 23%. For the full year it raised the core revenue range to $880 million to $890 million, with core gross margin of 41% to 43% and core operating margin of negative 19% to negative 17%. Core gross margin in the quarter was 40.6%, up 940 basis points year over year; GAAP gross margin was 14.2%.

The concentration risk is disclosed and material. Cerebras named OpenAI, Group 42 Holding Ltd, Mohamed bin Zayed University of Artificial Intelligence and AWS among a limited number of significant customers in its forward-looking statements. Those are the same relationships that generate the warrant charges dragging on GAAP revenue, which makes the accounting quirk and the customer-concentration question two faces of one arrangement rather than separate issues.

Cerebras remains a loss-making business by every measure it reports, with operating margins guided to stay negative through year-end and a cash burn profile that only its balance sheet makes tolerable. Wednesday's move was an after-hours move, not a full session's verdict — the stock has not yet traded a regular day against these numbers. Thursday's open will be the first time the market prices the print with the whole book participating.

This article is for general information only and is not investment advice. Figures are as reported by the cited sources at time of writing.

Related coverage

Illustrative photograph: computer server and electronics hardware.
Chips & Hardware

Aeluma Closes Fiscal 2026 With $582,000 of Fourth-Quarter Revenue, $56.0 Million of Cash and a CHIPS Letter of Intent That Would Hand Commerce Stock Worth the Whole Award

The Goleta photonics company reported on Sept. 16 that quarterly revenue fell to $582,000 from $1.3 million a year earlier while the net loss widened to $4.0 million, as it raised $20.1 million through its at-the-market programme. Its 10-K, filed the same day, says $183,000 of the year's $4.5 million revenue came from anything other than government contracts, and that the up-to-$30 million CHIPS award still needs definitive documents.

Karl Lindström · September 25, 2026
Illustrative photograph: an industrial manufacturing facility.
Chips & Hardware

Commerce Signs Its $1 Billion CHIPS R&D Award With IBM's Anderon. The Government's Release Runs to Two Sentences, and Neither Mentions the Equity Stake Promised in May

The Department of Commerce said on Sept. 16 it had signed a final CHIPS R&D award of up to $1 billion with Anderon, the IBM subsidiary building a 300mm quantum wafer foundry in Albany, N.Y. IBM says it is putting in a further $1 billion of its own. Neither document discloses milestones, a disbursement schedule, or how the 'minority, non-controlling equity stake' the Department attached to every quantum award in May applies to a wholly owned IBM unit.

Karl Lindström · September 25, 2026
Illustrative photograph: computer server and electronics hardware.
Chips & Hardware

Quantinuum preprint reports a fault-tolerant architecture validated on its 98-qubit Helios processor

A preprint posted to arXiv on Sept. 2 and written up by Quantum Computing Report on Sept. 9 describes the [[20,2,6]] C4-Helix code running on Quantinuum's 98-qubit trapped-ion machine, with a repeated error-correction figure of 4.6x10^-5 per logical qubit per cycle. The results are error-rate and fidelity gains, not speed gains, and the paper has not been peer reviewed.

Karl Lindström · September 10, 2026