Cerebras’s contract book reached US$26.1bn, 124 times its US$210.8m quarter

01Record revenue, and a loss larger than it

Revenue was US$210.8m against US$135.7m a year earlier, up 55%, and US$30.7m more than in April–June. Gross profit fell US$13.0m to US$42.1m, and the gross margin halved to 20% from 41%. Operating expenses rose to US$266.2m from US$81.6m, so the operating loss was US$224.1m, 106% of revenue. Interest on the IPO cash helped: other income was US$63.9m, and the net loss was US$161.2m.¹²

Most of the gap is pay in shares. Restricted stock units granted over years vest only once the company lists; the IPO met that condition in May, and stock-based compensation was US$155.1m in the quarter against US$11.3m a year earlier. General and administrative costs rose to US$55.6m from US$3.5m.²

Cerebras also reports “core” results, which it defines as excluding stock-based compensation, the non-cash amortisation of warrants it gave customers, and data-centre costs it passes through to a customer. Core revenue takes the US$210.8m, removes US$16.4m of pass-through revenue and adds back US$36.4m of warrant amortisation, which accounting rules deduct from revenue: US$230.9m, up 70%. The core operating loss was US$33.1m, or 14% of core revenue, against a guided loss of 23–25%; core revenue beat the US$214–216m guide.¹⁴ The core gross margin was 41%, against 40.7% a year earlier; the 10-Q credits more cloud revenue and better cloud margins, offset by lower hardware margins.²

02From selling wafer-sized chips to renting them out

The product is the Wafer-Scale Engine, which the 10-Q describes as “a chip encompassing an entire silicon wafer”. The current WSE-3, made by TSMC on a 5nm process, has 4 trillion transistors, 900,000 cores and 44 GB of memory on the chip itself, which Cerebras says is 58 times the size of Nvidia’s B200. Its argument is that keeping memory on the chip removes the bottleneck in serving AI models: moving data on and off a GPU.²⁵ The WSE-3 sits inside the CS-3 system. Cerebras sells those systems to run in a customer’s data centre, or runs them in its own cloud and sells the output: dedicated capacity, “generally structured as take-or-pay”, or tokens paid for as they are used.²

Cloud and other services revenue rose 265% to US$141.9m, 67% of the total; hardware fell 29% to US$68.9m. The OpenAI arrangement alone brought in US$72.4m, net of US$3.2m of warrant amortisation. Hardware margins fell to 11% from 50%, which Cerebras attributes to the warrants, stock pay, higher material costs and “preferential pricing” for a large strategic customer signed in 2025.² Its inference is also sold through AWS Marketplace, Microsoft Marketplace, OpenRouter and Hugging Face, among others. The next systems are on the way: the CS-4, introduced in August and “soon” in full production, and the CS-5, targeted for early summer 2027, whose WSE-5 processor is in system testing.¹²

03A contract book 124 times the quarter

Remaining performance obligations (RPO) are the revenue a company has signed for and not yet delivered, including deferred revenue and amounts still to be invoiced. Cerebras’s was US$26.1bn, and the 10-Q says “a significant amount” of it is owed under the Master Relationship Agreement (MRA) with OpenAI. It does not give OpenAI’s share as a figure. Cerebras’s first-quarter release valued the deal at “more than $20 billion”.²³⁴⁵

The schedule has moved forward as the deal ages. At the end of 2025 Cerebras expected 15% of its book within 24 months; now it expects 33% by September 2028, 44% in the two years after, and the rest later. Timing “may vary”, the filing says, with deliveries.²⁵ The figure also leaves things out: data-centre costs that will be passed on to OpenAI for sites not yet leased are excluded, because the amounts depend on “factors outside the Company’s control”.²

04Two customers, 72% of revenue

The 10-Q’s customer table uses letters; its risk factors give the names. The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) paid 38% of revenue in July–September, against 70% a year earlier. OpenAI paid 34%. G42, which the filing treats as a related party of MBZUAI, paid 7%, against 14%. Two customers also held 80% of what Cerebras was owed on 30 September.²

The OpenAI agreement, signed in December 2025, has three parts. OpenAI is “contractually committed to purchase” 750 MW of inference capacity over three or four years, extendable to five, with an option on 1.25 GW more by the end of 2030. It lent Cerebras about US$1.0bn in January at 6% for the build-out, so far repaid through services rather than cash: US$175.5m had been offset by 30 September, leaving US$829.1m. And it holds a warrant for up to 33.4m shares at a nominal price, vesting as milestones are met; 10.0m had been exercised by 30 September.² If the agreement ends for any reason other than OpenAI’s own uncured breach, the loan can fall due at once.²

05What turns the book into revenue

The contract pays only as capacity arrives. Cerebras must deliver it in tranches, across a set number of data centres, by milestone dates; a late tranche shortens that tranche’s term day for day. The 10-Q is direct about where things stand:

As of September 30, 2026, without adjustment for any excused delays … we delivered less capacity to OpenAI than the quantity contemplated by the MRA.

It expects delays “may continue through a portion of 2027”. Before any adjustment for excused delays, service credits, which reduce OpenAI’s later monthly fees, began accruing on 1 October; a shortfall left uncured for 90 days lets either side end that tranche, and Cerebras expects less than 5% of the committed capacity to be terminable this year.²

The capacity is being lined up. Cerebras says it has secured 700 gross MW of data-centre capacity for delivery by the end of 2027, up from more than 600 MW in August, with new sites in Norway and Finland; it does not say how much of that serves OpenAI’s 750 MW. It spent US$1.05bn on property and equipment in nine months and held US$7.8bn of cash, restricted cash and investments, with an undrawn US$850m credit line.¹²⁴ For October–December it guides core revenue of US$310–330m and a core operating margin of minus 1% to plus 1%; for the year, US$940–960m, up from US$880–890m in August and US$855–865m in June. Chief executive Andrew Feldman said the company expects “to triple revenue in 2027”.¹³⁴