Cerebras (CBRS) Q1 2026 earnings review
Explosive Growth, Vanishing Margins
Cerebras' first quarter as a public company delivered record core revenue of $191.3M, up 92% YoY, powered by a $20B+ OpenAI deal and a new AWS partnership. But the headline 46.5% gross margin is the high point, not the run rate: management's own guidance calls for it to fall to 36-38% next quarter as one-time pricing benefits roll off and the company rents back its own systems to meet demand. The business is still loss-making (GAAP net loss $14.0M, EPS -$0.22), and full-year guidance implies revenue growth decelerating to 69%.
🐂 Bull Case
Management repeatedly stated the only constraint on growth is data center capacity, not demand or wafer supply. A $20B+ OpenAI agreement went from signature to production in 35 days, and an AWS definitive agreement adds a second hyperscaler channel.
Core operating margin improved from -19% a year ago to -2%, and Adjusted EBITDA flipped positive to $12.7M from -$15.4M. Core operating loss is now near breakeven at -$3.5M.
🐻 Bear Case
Q1's 46.5% core gross margin benefited from one-time hardware incentive pricing and preceded the rent-back cost drag. Guidance cuts it to 36-38% in Q2, with cloud margin alone falling 10-15 points. Hardware margin is guided down from 42% to the low 30s.
Revenue leans on a handful of names (OpenAI, G42, MBZUAI, AWS). OpenAI is simultaneously the largest customer and a lender: a working-capital loan totaling ~$983M now sits on the balance sheet.
⚖️ Verdict: ⚪
Neutral, leaning constructive. Growth and strategic wins are real and large, but the business still loses money, the headline margin deteriorates sharply on guidance, and concentration plus a capacity-driven cash build-out keep this from being a clean bull. Execution on data center buildout is everything.
Key Themes
Q1 Gross Margin Flatters the Reality
The 46.5% core gross margin is not a run rate. Hardware margin of 42% (vs 30.6% a year ago) was lifted by incremental performance-based incentive pricing recognized prospectively on unshipped systems; management guides it back to the low 30s. Cloud margin of 52.9% is about to drop 10-15 points as Cerebras temporarily rents its own systems back from a customer (G42) to free up capacity for the OpenAI ramp. Net effect: core gross margin is guided down to 36-38% in Q2 and 38-41% for the year. This is the clearest data point cutting against the growth narrative.
OpenAI: $20B+ Anchor, Backlog Not Yet Revenue
Cerebras signed a definitive agreement with OpenAI on Dec 24, 2025 for 750MW of inference compute worth more than $20B over several years, reaching production in 35 days. GPT-5.4 now runs on Cerebras for OpenAI engineers and select customers, with GPT-5.5 next. The revenue is back-half weighted in 2026 as cloud capacity comes online; most of the $20B is future backlog. OpenAI also retains the option to take future committed capacity in its own data centers rather than Cerebras' cloud, which would remove the associated pass-through revenue.
AWS Partnership Opens Disaggregated Inference
A definitive agreement with AWS pairs Amazon's Trainium 3 (prefill) with Cerebras CS-3 (decode) in a disaggregated architecture. Management framed decode as where GPUs struggle with sequential generation and Cerebras excels. Importantly, AWS impact is a 2027 story, not 2026, and management explicitly flagged opportunity for additional decode partnerships with other GPU operators.
Hardware Revenue Set to Decline as Mix Shifts to Cloud
Core hardware revenue of $111.6M grew 60% YoY, but management guides it down sequentially for the next few quarters as existing purchase orders are delivered and production shifts toward systems deployed in Cerebras Cloud rather than sold outright. Cloud and services revenue (up 167% YoY to $79.8M) becomes the growth engine. This shift improves long-term recurring economics but concentrates near-term revenue on capacity that is still being built.
Customer and Lender Concentration in One Name
Cerebras' own risk disclosures name OpenAI, G42, MBZUAI and AWS as dependencies. The entanglement runs deeper than sales: OpenAI extended a working-capital loan that now sits on the balance sheet as 'loan from customer' totaling roughly $983M (~$621M current, ~$362M non-current). The same customer Cerebras is renting capacity back from to serve OpenAI is G42. This concentrated web magnifies the impact of any single relationship souring.
Wafer-Scale Sidesteps the HBM Bottleneck
Management's supply-chain argument is specific: the industry's binding constraint is HBM memory, which is scarce and expensive. Cerebras uses on-wafer SRAM instead, printed on its logic wafer, and builds at TSMC's less-contested 5nm node rather than 3nm. Chips are 58x larger than the largest competitor. The company manufactures CS-3 systems exclusively in the US and added San Mina as a second contract manufacturer alongside Flextronics.
Data Center Capacity Is the Real Constraint
Asked directly what limits growth, management was blunt: not demand, not wafer supply, but data center capacity. Cerebras is adding sites across the US, Canada, Europe (France, the Nordics) with early discussions in Israel, UAE, Australia, Singapore, India and Indonesia, plus a 120MW Bell Canada facility. The capacity build is heavily weighted to the back half of 2026 and into 2027, which is why full-year revenue and margins are back-end loaded.
The 'Fast Wins' TAM Thesis
Management argues the entire inference market is addressable because fast tokens command a premium and no technology has let slow own a market over time (search, broadband as analogies). This is a directional bet, not a proven share: fast tokens cost more, and management conceded both their view and the GPU camp's are 'self-interested.' Nvidia has publicly framed fast inference as a minority (~25%) of the market. Treat the full-TAM framing as ambition rather than established fact.
Other KPIs
Core operating margin improved steadily from -19% in 25Q1 to roughly -10% in 25Q4 to -2% in 26Q1 (Reversing toward breakeven). But guidance reverses that progress: -30% to -32% in Q2 and -28% to -32% for the full year, driven by the rent-back cost drag and a step-up in public-company G&A. The Q1 reading is the cleanest the company expects to look until capacity normalizes.
Operating cash flow turned positive ($12.3M) versus -$54.9M a year ago, even as net loss remained negative at -$14.0M. The gap is bridged by $18.2M depreciation, $18.9M non-cash interest, and deferred revenue and customer deposit inflows. Note the working-capital build: inventories rose ~40% to $89.0M and receivables ~24% to $62.6M ahead of the ramp. Against this, CapEx was $132M in the quarter, so the company is funding its buildout from financing, not operations.
Cash, equivalents, restricted cash and short-term investments ended at $3.3B. This precedes the $6.4B IPO (largest semiconductor IPO ever, closed in Q2) and an $850M revolving credit facility added in April. Combined with the OpenAI working-capital loan, Cerebras is heavily capitalized to fund data center expansion ahead of revenue. Total deferred revenue grew to $244M (from $167M), a forward-demand signal.
Core operating expenses grew 51% YoY, roughly half the 92% revenue growth rate, showing operating leverage. R&D dominates at $69.8M (up ~48%), reflecting the push to stay at the silicon and model frontier. G&A of $9.9M is guided to step up meaningfully next quarter on public-company costs.
Guidance
Up 88% YoY and roughly flat sequentially (+1.4% vs Q1's $191.3M). Decelerating versus Q1's 92% YoY, with management explicitly noting growth is back-half loaded as cloud capacity for OpenAI comes online later in the year. Near-flat sequential revenue reflects hardware revenue declining while cloud ramps.
Decelerating sharply from 46.5% in Q1. The drop reflects hardware incentive pricing normalizing to the low 30s and a 10-15 point hit to cloud margin from renting back third-party systems to bridge capacity. Management calls this temporary and reaffirmed a long-term 60%+ gross margin target.
A sharp regression from -2% in Q1, driven by the same rent-back gross margin drag plus a step-up in G&A for public-company costs. Management characterized this as a near-term trough that will persist for a few quarters before improving as owned capacity comes online.
Up 69% YoY at the $860M midpoint. Implies H2 core revenue of ~$475M versus ~$385M in H1 (+23% sequentially), consistent with the back-half cloud ramp. The 69% full-year rate is a deceleration from Q1's 92%, partly base effects and partly the timing of capacity deployments. Achievability hinges almost entirely on bringing data center capacity online on schedule.
Below Q1's 46.5%, embedding the rent-back drag across the year. The full-year range sitting above the Q2 trough (36-38%) implies margin recovery in H2 as owned capacity replaces rented systems. Still well short of the 60%+ long-term target.
Full-year losses deepen versus Q1's -2% as the company invests ahead of demand. Management framed this as deliberate: spending aggressively on capacity and R&D to capture market share while demand exceeds supply. Profitability is explicitly a medium-to-long-term goal, not a 2026 outcome.
Key Questions
Path Back From the Margin Trough
Core gross margin falls from 46.5% to a 36-38% trough and recovers within the 38-41% full-year range. What specific milestones (owned-capacity MW online, end of rent-back) gate the recovery, and what's the risk the rent-back extends beyond 2026?
Customer Concentration Disclosure
Management declined to disclose top-customer concentration on the call, deferring to filings. With OpenAI as both largest customer and ~$983M lender, what share of revenue and backlog do the top one and top three customers represent, and how is the lending relationship structured if the commercial one changes?
OpenAI Cloud vs Own-Data-Center Election
OpenAI can elect to take future committed capacity in its own data centers, removing pass-through revenue. How much of the $20B+ and the FY26 guide assumes Cerebras-cloud deployment, and what is the revenue impact if OpenAI shifts to its own facilities?
AWS Deal Size and 2027 Contribution
Management confirmed AWS impact is a 2027 event but declined to size it (megawatts or dollars). What scale should investors model, and is the deal committed volume or capacity-dependent?
Cash Burn vs Buildout Funding
CapEx was $132M against $12M operating cash flow in Q1, with a much larger buildout ahead. How many quarters of buildout does the post-IPO cash plus credit facility fund, and at what point does the company need to return to markets?