Cerebras Systems: Disaggregated Inference Partnerships With AMD and AWS Set Up Tripling of 2027 Revenue as $25.4 Billion Backlog Anchors Guidance
Q2 2026 earnings call, held August 12, 2026, first results since IPO
Cerebras Systems delivered its first earnings report as a public company on Wednesday, beating guidance across core revenue, gross margin and operating margin, while laying out a growth thesis for 2027 that rests heavily on a strategy the company calls disaggregated inference. Core revenue came in at $209.9 million, up 103% year-over-year, with core cloud and other services revenue nearly quadrupling to $127.7 million, up 287%. Management raised full-year 2026 core revenue guidance to $880 million to $890 million and reiterated that 2027 core revenue should more than triple, a target CEO Andrew Feldman said is underpinned by the $25 billion of remaining performance obligations already on the books, which grew to $25.4 billion by June 30.
Disaggregation with AMD and AWS is the new margin and throughput lever
The most consequential new disclosure was the mechanics of Cerebras's disaggregated inference partnerships with AMD and AWS, which the company frames as a structural unlock for both throughput and gross margin rather than a one-off product announcement. Feldman explained that inference splits into two stages, prefill and decode, and that GPU-based HBM architectures are well suited to the parallelizable prefill stage while Cerebras's SRAM-based wafer-scale engine is better suited to the sequential, memory-bandwidth-intensive decode stage. "GPU solutions can support high throughput, but only at low speeds. When configured to support even moderate speeds, GPU throughput drops precipitously," he said, adding that this trade-off is not unique to GPUs but applies to "all solutions that use HBM."
By pairing AMD's Helios racks for prefill with Cerebras systems for decode, the company says it can maintain its industry-leading speed while lifting throughput by 5 times, which management argues directly improves gross margin, power efficiency and the economic value of every data center megawatt deployed. CFO Bob Komin noted that up to 5 times more tokens can now be produced per CS system, reducing power and cost per token. The AWS relationship, which pairs Trainium chips with Cerebras systems, is expected to go generally available in the first quarter of 2027 through AWS's Bedrock platform, giving Cerebras hyperscale distribution it previously lacked. Feldman noted the two partnerships together mean Cerebras has "tied up about half the leading chip makers," and left the door open to eventually hosting AWS's Trainium hardware inside Cerebras's own cloud, saying that approach is "very much on the table for the future with AWS."
Gross margin dip in Q3 is a temporary function of renting back systems
Investors should note a mechanical wrinkle in the margin trajectory. Core gross margin came in at 40.6% in Q2, down sequentially from 46.5% in Q1, because Cerebras has been temporarily renting back its own systems from cloud customers to meet overwhelming inference demand sooner. Komin said that absent this dynamic, core gross margin would have been roughly 500 basis points higher and more in line with the prior quarter. Management guided Q3 core gross margin to 38% to 40%, calling it the low point for the year, with improvement expected in the fourth quarter as more Cerebras-owned systems come online in the private cloud, and further progress toward a long-term target above 60% in 2027 as higher-cost rented capacity rolls off and new pricing on hardware and cloud deals is recognized.
Customer concentration remains a live question but is expected to diversify
UBS analyst Timothy Arcuri pressed management on customer concentration heading into 2027, asking whether OpenAI and AWS together could represent two-thirds of revenue. Feldman pushed back on the framing by pointing to Cerebras's history of expanding beyond concentrated relationships, from early government contracts to a sovereign cloud deal to its frontier lab partnership with OpenAI. "OpenAI is an enormous customer, and they're an enormous part of not just our business, but of everybody's business in the sector," he said, adding that OpenAI will remain "a meaningful portion" of revenue next year even as AWS and other categories such as coding and security grow and OpenAI shrinks as a percentage of the mix over time. Notably, management reiterated that the $25.4 billion RPO figure does not include any backlog from AWS or other hyperscalers, meaning the disclosed backlog understates the potential scale of future bookings.
Capacity build-out: 600 megawatts secured, manufacturing scaling more than 10 times
Data center capacity remains the binding constraint industry-wide, and Cerebras detailed an aggressive seven-month push that secured more than 600 megawatts of capacity either live or contracted for delivery by the end of 2027, spanning sites in Alabama, Dallas, Denver, Minneapolis, Santa Clara, Stockton, and international locations including France, Finland, Manitoba, Montreal, Norway, Saskatchewan and Toronto. Feldman said the company's pipeline of new opportunities is "now measured in gigawatts." On manufacturing, Komin said capacity is already 4 times above first-half 2025 levels and will finish 2026 more than 10 times higher, with facilities already contracted for another 3 to 4 times growth in 2027. The company continues to emphasize that its supply chain avoids the tightest industry bottlenecks, including HBM memory, CoWoS packaging and 3-nanometer fab capacity, since Cerebras runs on TSMC's 5-nanometer node.
Frontier model support and new verticals: coding, agentic workflows and security
Cerebras confirmed it now serves OpenAI's GPT-5.6 Sol model at 10 times the speed of competing infrastructure, which Feldman said "lays to rest any of the remaining concerns regarding our ability to support large frontier models." Beyond OpenAI, the company signed six new deals worth more than $30 million each in the quarter, with wins at Figma, Cognition and Lovable in coding, and Block, AlphaSense and GSK in agentic workflows. A new vertical, cybersecurity, emerged via a partnership with CrowdStrike, which Feldman described as "an application that only exists if AI is fast," enabling large language models to inspect enterprise traffic in real time without introducing latency. Management believes this type of inline security inspection will become standard practice as enterprises look to secure growing volumes of AI-generated and AI-processed traffic.
Balance sheet and roadmap: CS4 launch next week, CS5 targeted for second half 2027
Cerebras ended the quarter with more than $8.6 billion in cash, cash equivalents, restricted cash and marketable securities, plus an undrawn $850 million revolving credit facility, giving it a materially stronger liquidity position following its IPO. Komin also highlighted that Cerebras's capital intensity per megawatt is lower than most AI cloud peers because it deploys its own hardware at bill-of-materials cost rather than paying third-party margins, and because a meaningful portion of data center fit-out costs are reimbursed by its largest customer. On the product side, the company will unveil its fourth-generation CS4 system next week at its Supernova conference, with the CS5 targeted for the second half of 2027, alongside a roadmap that calls for doubling speed annually and increasing throughput more than 20 times over the next 18 months.