Test report DSG-2512 · Rev A · tested November 14, 2026

AI Datacenter InfrastructureDevice under test

Cerebras lands Microsoft as third hyperscaler customer

Microsoft becomes Cerebras' third hyperscaler customer after AWS and Oracle. Q3 revenue hit $231M, up 70 percent YoY, but capacity limits and OpenAI delays persist.

Read
3 min
Words
670
Node
28nm
Operator
Priya Raman

Spec summary

  1. Microsoft is the third hyperscaler to sign with Cerebras, after AWS and Oracle; deal details deferred to 2027
  2. Q3 2026 revenue was $231 million, up 70 percent year-on-year; full-year guidance raised to $940M–$960M
  3. Remaining performance obligations stand at $26.1 billion, excluding Oracle and Microsoft deals
  4. Roughly 5 percent of the OpenAI contract value is exposed to a 90-day delay cancellation clause
  5. Over 75 percent of planned 2027 data center capacity is outside the US
Cerebras secures Microsoft as a customer
Fig. ACerebras secures Microsoft as a customer — TECNIC Bioprocess Solutions / unsplash

Microsoft has signed on to deploy Cerebras AI chips, making it the third hyperscaler customer after AWS and Oracle. The company disclosed the deal on its Q3 2026 earnings call. It disclosed almost nothing else about it.

CEO Andrew Feldman told investors that details of the Microsoft contract will arrive in 2027. Until then, the announcement functions as a name, not a number. Notably, the $26.1 billion in remaining performance obligations as of September 30, 2026 excludes both the Oracle and Microsoft agreements. The backlog, in other words, is larger than reported.

What drove the hyperscaler signings?

CFO Bob Komin credited the company's own cloud service, launched in early 2025. He said: "We chose to ramp a cloud offering in early 2025 because in this rapidly advancing AI landscape, AI natives want to scale compute quickly and don't want to own data centers."

His explanation of the mechanism is more revealing than the headline. Hyperscalers allocate data center compute years ahead. New hardware often has no near-term slot in their facilities. Cerebras therefore used its private cloud as a proof-of-scale venue. Komin said: "After having the opportunity to thoroughly evaluate our solution in our private cloud, these customers can then choose to buy our hardware or purchase our cloud offering." The cloud offering is a sales instrument. The strategy has now converted three times.

What did Q3 deliver?

Revenue reached $231 million, up 10 percent sequentially and 70 percent year-on-year, beating expectations. The composition:

  • AI cloud: $145 million, expected to stay roughly flat in Q4
  • Core hardware: $86 million, which excludes the Oracle deal
  • Gross margin: 41 percent, guided to 48 percent next quarter, 50 percent near term, 60 percent long term
  • Core net income: positive at $30 million; adjusted core EBITDA positive year-to-date at $2 million
  • Full-year 2026 core revenue guidance raised to $940 million–$960 million

The Q4 cloud plateau is a capacity problem, not a demand problem. Data center capacity will not begin ramping until December, and larger portions of the planned 700MW are under contract for mid-2027 deployment. The company expects to triple revenue in 2027 on that basis. That forecast rests on capacity that does not yet exist.

Feldman identified the bottlenecks as land, power, and data center shells. "Data center availability remains a constant constraint across the industry, and we are no exception," he said. He added that US policy could worsen domestic development. Consequently, more than 75 percent of planned 2027 capacity sits outside the US, in Canada, France, and Norway. The quarter added contracts in Norway and Finland, with a São Paulo facility in final contracting, the company's first in South America.

On silicon: CS-5 production starts summer 2025, CS-3 to CS-4 crossover lands in Q1 2027, and Cerebras expects to exit 2026 with ten times its beginning-of-year manufacturing capacity.

What about the OpenAI delays?

Feldman spent much of the call on speculation. The OpenAI agreement expanded from Q2, and he called the response to the "Ultrafast GPT-5.6 Sol preview" overwhelming. He denied that Cerebras hardware cannot technically support OpenAI's Astra model. He conceded capacity fell short: "Cerebras didn't have sufficient capacity to support Ultrafast in a broad rollout, including for Astra."

Delays are real. Feldman said the company is "running behind schedule on the initial tranches." If delays exceed 90 days, either party can cancel the affected portion. Cerebras estimates roughly 5 percent of contract value is exposed. No cancellations have occurred.

Is there a yield problem?

Culper Research, a short-selling firm, disclosed a short position claiming Cerebras wafers fail after roughly two years of use, which it said "undermines the entire business." Feldman attributed low hardware margins to the manufacturing ramp and final shipments under special pricing for an early customer, not yields. He described fabrication methods that compensate for performance variation across the wafer. No independent data settled the dispute. Cerebras shares trade at $97.75.

via datacenterdynamics.com (Source)

Filed under

  • cerebras
  • microsoft
  • ai-accelerators
  • hyperscalers
  • ai-cloud
Share this article:

More from Priya Raman

Priya Raman

Show full bio

Correspondent covering money, funding and market shifts around hardware layer of artificial intelligence at Die Signal.

573 articles

Same lot · LOT-362A

« Previous articleNext article »