Cerebras Books $26.1B Backlog, Targets CS-5 Launch in 2027

Cerebras Systems closed the third quarter of 2026 with a $26.1 billion revenue backlog and $7.36 billion in cash, and CEO Andrew Feldman says the company expects to triple sales in 2027.

The waferscale AI compute maker is juggling two things at once: ramping its cloud business to serve a $10 billion inference contract with OpenAI signed in January 2026, and preparing its fifth-generation system. The OpenAI deal covers 750 megawatts of gear the model builder rents for fast inference, with figures between $20 billion and $30 billion cited for 2026 through 2032.

The larger number apparently hinges on OpenAI exercising an option for Cerebras to build another 1.25 gigawatts of capacity by the end of 2030, with services running to 2032. Provided OpenAI goes public and has the cash, there's no reason it wouldn't stretch the deal.

What happened to Q3 hardware sales?

Hardware revenue fell 28.8 percent year over year to $68.9 million, and gross margin came in at just 11 percent against a long-term target of 60 percent. Two factors did the damage: the factory is still ramping to a new production level, and an early customer — presumably cloud provider G42 — negotiated aggressive pricing years ago on machines now being installed.

Feldman dismissed speculation that Cerebras has a chip yield problem. The argument holds up. The WSE-3 in CS-3 systems and the WSE-3 Turbo in CS-4 systems are built on mature 5-nanometer processes at Taiwan Semiconductor Manufacturing Co — not bleeding-edge nodes like the 2-nanometer process AMD is using for its just-launched "Venice" Epyc 9006 server CPUs.

The WSE-3 Turbo is a stop-gap, double-clocked waferscale processor shipped in CS-4 systems in August. There will be no WSE-4 — the next chip is the WSE-5.

When does the WSE-5 arrive?

Sooner than planned. Feldman told Wall Street analysts on a conference call:

"Our product road map continues to advance rapidly. Our CS-4 will soon be in full production, doubling both performance and throughput. Our next-generation processor, the Wafer Scale Engine 5, which will be deployed in our fifth-generation CS-5 system, is back from the fab ahead of schedule. The WSE-5 taped out on June 15. The chip is back, has been brought up, and is already running models and is undergoing in-chassis testing. Moving from tape-out to system test in less than five months is exceptional execution. The CS-5 will more than double the speed of CS-4. It will enter production in early summer of 2027. It is built on TSMC's 5 nanometer node, where competition for wafers is less intense and wafer costs are much lower."

The CS-4 and CS-5 systems will compete with new GPUs from Nvidia and AMD, LPUs and other XPUs from Grok, SambaNova and Rebellions, plus homegrown hardware from hyperscalers and AI model builders themselves.

At Hot Chips this summer, Cerebras said it would use 3D stacked DRAM with its wafers. Silicon photonics linking wafers and memory servers, and SRAM stacking, are two other plausible additions to the WSE-5 design.

What does this change for the hyperscalers?

Cerebras has now signed deals with Amazon Web Services, Oracle Cloud and Microsoft Azure on top of OpenAI. None of the three cloud deals is in the backlog yet, and none is generating revenue. Details remain vague. But three of the world's largest clouds intend to run Cerebras iron on inference workloads, presumably decode operations — and with 2X the performance of today's WSE-3 Turbo, the CS-5 is almost certainly the machine they want.

Notably, Cerebras uses AMD GPUs for prefill in its own cloud, which makes the economics of the OpenAI contract work: inference prefill on GPUs, decode on waferscale silicon.

On full-year numbers, hardware sales are lumpy at every supercomputer maker, so annual comparisons matter more than quarterly ones. The best estimate puts GAAP revenue for 2026 at around $885 million, up 73.4 percent from 2025. Operating and net income were hit by a spike in R&D spending and stock-based compensation. Margins should improve in 2027 as manufacturing ramps with partners Flex and Sanmina.

Given the pace of token-cost declines, generative AI looks durable. Feldman's 2027 triple is the number to watch.