Detailed narrative
The speed thesis: fast tokens are the most valuable tokens
Feldman framed AI as expanding the addressable compute market 'many thousands of times' by giving computers purchase on text, images, video, agents and robotics. His central claim is that inference speed equals productivity — 'fast is productive and slow is unproductive' — and that Cerebras delivers the fastest AI in the world by 'an order of magnitude.' A live demo on the trillion-parameter Kimi K2 open-source model showed Cerebras completing a prompt in 21 seconds versus 4 minutes 37 seconds on a leading GPU, which he characterized as 13x faster.
He also argued speed removes the safety-vs-experience trade-off because guardrails add a compute layer that is far less painful on fast infrastructure.
OpenAI partnership: >$20B commitment and frontier-model access
Cerebras signed a definitive agreement with OpenAI on December 24, 2025 to purchase more than $20 billion of Cerebras compute over the next several years, and was in production running a never-before-seen model by February 1 — 35 days from signature.
GPT 5.4 is now running on Cerebras and available to OpenAI engineers and select OpenAI customers as part of a strategic rollout, with GPT 5.5 being actively brought onto the platform next. Beyond revenue, management stressed the strategic value of direct visibility into frontier-model development.
Cerebras is one of only two hardware vendors currently serving OpenAI models.
AWS and the disaggregated inference opportunity
After a binding term sheet in March, Cerebras completed a definitive agreement with AWS this week to deploy in AWS data centers, combining AWS Trainium 3 (prefill) with Cerebras CS-3 (decode) in a disaggregated solution expected to be an order of magnitude faster at massive throughput.
Management explained the divide-and-conquer logic — prefill is highly parallelizable while decode is strictly sequential, where GPUs struggle and Cerebras excels. Feldman flagged 'multiple partners' for disaggregated solutions, signaling opportunities to provide decode to GPU owners beyond AWS, and confirmed interest even if Amazon sells Trainium hardware externally.
Supply-chain advantage: sidestepping HBM, CoWoS and 3nm
Management argued Cerebras avoids the industry's three binding constraints: it uses on-wafer SRAM instead of scarce, expensive HBM ('supply is approximately infinite'); it does not use TSMC's CoWoS advanced-packaging process; and it manufactures at the 5-nanometer node rather than contended 3nm.
Its salable unit is the CS-3 system (sold on-prem or as cloud time), manufactured exclusively in the U.S. — claimed to be the only accelerator maker doing so. Cerebras added hundreds of thousands of square feet of manufacturing and clean-room space, expanded its Flextronics partnership, and added Sanmina as a second contract manufacturer.
Data-center capacity is the binding constraint
When asked directly what limits growth, Feldman said 'Demand is not the constraint. Supply is not the constraint. The constraint is data centers.' The company has added capacity across the U.S., Canada, Europe (France, the Nordics) and is in early discussions for Israel, the UAE, Australia, Singapore, India and Indonesia, engaging with dozens of data-center owner-operators.
New capacity is coming online across Q3 2026 through Q4 2027. To bridge near-term demand it is temporarily renting its own systems back from an existing customer, accepting a 10-15 point cloud-margin hit while it builds out owned capacity.
New non-GAAP 'core' framework and margin dynamics
As a newly public company, Cerebras introduced 'core' metrics that exclude OpenAI data-center pass-through costs (billed at a 3% markup, reported gross and dilutive to margin), non-cash amortization of customer warrants recorded as a revenue reduction, stock-based compensation and one-time📎 items.
Core gross margin rose to 46.5% (from 42.1% YoY and 41% sequentially), with cloud margin at 52.9% on premium pricing and improved utilization, and hardware margin at 42% aided by prospectively-recognized incentive pricing. Bob Komin cautioned that renting systems back will regress margins and near-breakeven profitability for the next few quarters.
Balance sheet and fundraising
Cerebras ended the quarter with $3.3 billion in cash, equivalents, restricted cash and marketable securities. It raised $1 billion in Series G equity in September 2025, another $1 billion in Series H in February 2026, added a revolving credit facility of up to $850 million in April 2026, and completed 'the largest semiconductor IPO in history,' raising $6.4 billion a few weeks before the call.
Management said this liquidity positions it to accelerate sourcing and deployment of data centers and its supply chain, and to invest ahead of demand to win key customers and gain market share.