Cerebras Systems Inc. (CBRS) | The Buildout — AI Infrastructure
The Verdict
Cerebras sells speed. It builds a wafer-scale AI processor — one large piece of silicon that uses on-chip SRAM instead of the HBM memory most accelerators rely on — packs it into systems, and sells access through its own fast-inference cloud. It also supplies the decode half of a split inference pipeline, pairing its hardware with a partner's prefill silicon; AWS and AMD are the named partners. Management describes Cerebras as one of four companies that builds processors, systems, data centers and an AI cloud service. The bet underneath is that latency, not raw throughput, is the next thing AI customers pay for.
| Market Cap | — |
| Revenue (TTM) | $681M |
| Revenue Growth | +90.9% |
| EBITDA Margin (TTM) | -74.9% |
| Net Cash | $7.1B |
| Earnings Beats | 1 of 1 |
| P/E (TTM) | — |
| EV/EBITDA (TTM) | — |
What We Like
- Remaining performance obligations were $25.4B at June 30, 2026, up from $25.0B at March 31, 2026, and management says the total includes no backlog from AWS or any other hyperscaler.
- Q2 core revenue beat the guide and management raised the full year on all three lines: FY2026 core revenue to $880-890M, core gross margin to 41-43%, and core operating margin to -19% to -17%.
- Cloud and other services revenue rose 287% year over year to $127.7M, against hardware at $82.1M, up 17%.
- More than 600 MW of data-center capacity is live or contracted for delivery by end-2027, with a pipeline management describes as measured in gigawatts.
- Cash, cash equivalents, restricted cash and marketable securities exceeded $8.6B at June 30, 2026, with an $850M revolver unused.
What We’re Watching
- Concentration: the filed record shows the largest customer at 63% of revenue in the March 2026 quarter, up from 24% a year earlier, with both large customers unnamed related parties. The OpenAI share of RPO is calculated at 79% or more.
- The margin path: Q3 is guided to 38-40% gross margin and -25% to -23% operating margin, both below Q2, and the Q4 recovery depends on owned data centers replacing rented systems on schedule.
- Data-center space is the company's own named bottleneck, and Arm's head-node CPU demand is stated at more than $2B against $1B of secured manufacturing capacity — a supply gate outside Cerebras's own silicon. Neighbors describe tightness across memory, substrates, test equipment and networking.
- Securities-fraud investigation notices from eight law firms appeared between 2026-06-25 and 2026-07-02, six to seven weeks after the May 14, 2026 IPO; no filed complaint detail is available in the source material.
The operating thesis is intact and the disclosure became more specific: a beat-and-raise quarter, 600+ MW secured, dated milestones for AWS, AMD and CS-5. The structural question is unresolved. Concentration rose in the filed record to 63%, AWS contributes nothing to the backlog, and the reduction in concentration is forecast rather than demonstrated. The open question is whether OpenAI takes future committed capacity in Cerebras's cloud or its own data centers, and whether a second scaled customer is ever named.
Earnings Beat
In the quarter ended June 30, 2026, reported revenue was $180.1M, with gross margin of 14.2% and EBITDA of -$452.9M. On the management-defined core basis the company emphasizes, core revenue was $209.9M, up 103% year over year, and core gross margin was 40.6%. Management says the quarter beat its own guide on all three guided lines. Core operating loss was $33.6M, a -16% margin.
| Metric | Q2 FY2026 | Q1 FY2026 | Q2 FY2025 | YoY |
|---|---|---|---|---|
| Revenue | $180M | $193M | $103M | +74.3% |
| Gross margin | 14.2% | 44.6% | 31.1% | -1690bps |
| EBITDA | −$453M | $3M | −$52M | +774.3% |
| EPS | $-2.98 | $-0.06 | $1.44 | −307.4% |
| Core cloud & other services revenue | $127.7M | $79.8M | n/a | +287% YoY |
| Remaining performance obligations | $25.4B | $25.0B | n/a | — |
our $25 billion in RPO does not reflect any backlog of business from AWS or any other hyperscaler at this time.— Feldman, 2026-08-12
Management tone: Management's commentary moved from qualitative bottleneck talk on the June 2026 call to quantified commitments on the August 2026 call: 600+ MW of secured capacity, manufacturing capacity 4x above the first half of 2025 and targeted above 10x for 2026, roughly 500 basis points of rent-back drag, and dates for AWS, AMD and CS-5. Management named Q3 as the gross-margin low point rather than burying the sequential decline, and reaffirmed the 60%+ longer-term target. It again declined to quantify customer concentration or the scale of the AWS agreement.
Management Guidance
With Q2, management raised FY2026 core revenue to $880-890M from $855-865M, core gross margin to 41-43% from 38-41%, and core operating margin to -19% to -17% from -28% to -32%. For Q3 2026 it guided to $214-216M core revenue, 38-40% core gross margin and -25% to -23% core operating margin. Management calls Q3 the low point for core gross margin, expects significant improvement in Q4 as more data centers fill with lower-cost Cerebras-owned systems, and reaffirmed a 60%+ medium-to-long-term gross margin target.
Trajectory
On management's core basis, growth accelerated: core revenue rose 92% year over year in the March 2026 quarter and 103% in the June 2026 quarter, to $209.9M. The cloud line is doing the work, going from +167% to +287% year over year, while hardware went from +60% to +17% and fell sequentially from $111.6M to $82.1M — a shift management flagged in advance. Margins moved the other way. Core gross margin fell from 46.5% to 40.6% sequentially and core operating margin from -2% to -16%, with Q3 guided lower on both. Management attributes the compression to renting back systems at higher cost and says owned capacity replaces them. On the reported spine figures, revenue slipped to $180.1M from $193.4M and gross margin fell to 14.2% from 44.6%, reflecting pass-through revenue carried at a 3% markup and non-cash warrant amortization recorded as a revenue reduction.
The Model
The model projects FY+1 revenue of $890M and EBITDA of -$324M (-36.4% margin), and FY+2 revenue of $2,950.0M and EBITDA of $245M (8.3% margin). Revenue dispersion across the five runs is 12% in FY+1 (minimum $798M, maximum $905M) and 14% in FY+2 (minimum $2,700M, maximum $3,100M). The near term is anchored on the company's own FY2026 core revenue guide of $880-890M and a guided core operating margin of -19% to -17%. FY+2 implies a step-change above a tripling of that base, which is where management's stated expectation to more than triple core revenue in 2027 and its 60%+ gross-margin target point. The EBITDA swing depends on the owned-data-center mix shift and whether the throughput roadmap is delivered rather than described.
| Metric | FY2025 | Next FY (E) | Following FY (E) |
|---|---|---|---|
| Revenue | $510M | $890M | $3.0B |
| YoY Growth | — | +74.5% | +231.5% |
| EBITDA | −$137M | −$324M | $245M |
| EBITDA Margin | -26.8% | -36.4% | 8.3% |
Projections are the median of 5 independent model runs.
With Q2, management raised FY2026 core revenue to $880-890M from $855-865M, core gross margin to 41-43% from 38-41%, and core operating margin to -19% to -17% from -28% to -32%. For Q3 2026 it guided to $214-216M core revenue, 38-40% core gross margin and -25% to -23% core operating margin. Management calls Q3 the low point for core gross margin, expects significant improvement in Q4 as more data centers fill with lower-cost Cerebras-owned systems, and reaffirmed a 60%+ medium-to-long-term gross margin target.
What Could Go Right — and Wrong
- Owned data centers energize on schedule, the rent-back rolls off, and gross margin steps back up in Q4 2026 toward the 60%+ medium-term target.
- AWS becomes generally available through Bedrock in Q1 2027 and produces disclosed revenue, giving the cloud line a second scaled source.
- A named other hyperscaler begins paying from mid-2027 and ramps through 2028, so revenue concentration finally falls.
- Disaggregated GPU inference ships in Q4 2026 with named buyers, extending the decode position into the AWS and AMD stacks.
- The manufacturing ramp — 4x above the first half of 2025, above 10x targeted for 2026, and 3x to 4x more contracted for 2027 — converts the backlog into revenue on schedule.
- Concentration proves structural rather than a phase: the largest filed customer sits at 63% of revenue and more than $20B of backlog traces to one counterparty.
- OpenAI elects to take future committed capacity in its own data centers, removing the corresponding pass-through revenue and shifting the mix.
- Rent-back persists past Q4 2026, the 60%+ target slips, and the back-end-loaded FY2026 guide is at risk.
- Component tightness outside wafer-scale silicon — head-node CPU allocation, memory, substrates, test equipment, networking — gates shipments even with 5nm wafers secured.
- Data-center and power delivery slips, or the post-IPO securities-fraud investigations mature into litigation and change the disclosure posture.
Looking Ahead
The next four quarters turn on capacity execution rather than demand. Q3 2026 tests the guided margin trough and Q4 2026 tests the step-up as owned data centers replace rented systems; disaggregated GPU inference is also slated to be available in Q4 2026. Full-year 2026 tests the above-10x manufacturing expansion. In Q1 2027 AWS is scheduled for general availability through Bedrock, and management expects first revenue from other hyperscalers from mid-2027, ramping through 2028. The CS-5 is targeted for the second half of 2027, with 600+ MW live or delivered by the end of 2027.
- Q3 2026Gross-margin low point — Whether core gross margin lands in the guided 38-40% band.
- Q4 2026Margin step-up test — Whether owned systems replace rented capacity and lift gross margin.
- Q4 2026GPU disaggregation ships — Whether disaggregated GPU inference is available with named buyers.
- Q1 2027AWS Bedrock GA — Whether AWS begins producing revenue through Bedrock.
- Mid-2027Other hyperscaler revenue — Whether a second scaled customer begins paying, easing concentration.
- End 2027600+ MW conversion — Whether contracted megawatts become live capacity.
Financials
Annual Summary
| Metric | FY2024 | FY2025 | TTM | YoY |
|---|---|---|---|---|
| Revenue | $290M | $510M | $681M | +75.7% |
| Gross Margin | 42.3% | 38.6% | 34.9% | 365bps |
| EBITDA | −$97M | −$137M | −$510M | -40.5% |
| EBITDA Margin | -33.5% | -26.8% | -74.9% | +670bps |
| Net Income | −$482M | $238M | −$512M | +149.4% |
| Free Cash Flow | $300M | −$309M | −$596M | — |
| Net Cash | — | — | — | — |
Key Ratios (Trailing)
- P/E TTM—
- EV/EBITDA TTM—
- EV/Revenue TTM—
- Price/FCF TTM—
- Gross Margin (TTM)34.9%
- EBITDA Margin (TTM)-74.9%
- Net Margin (TTM)-75.3%
- ROIC-21.4%
- SBC / Revenue60.1%
The Company
Cerebras builds a wafer-scale AI processor — one large piece of silicon that uses on-chip SRAM instead of HBM memory — and sells it three ways: as hardware systems, as token generation through its own fast-inference cloud, and as the decode half of a disaggregated pipeline where a partner supplies prefill. The CS-3 ships today, the CS-4 was unveiled on 2026-08-18, and the CS-5 is targeted for the second half of 2027.
The company reports one segment and steers readers to management-defined 'core' figures, which exclude pass-through data-center revenue and cost carried at a 3% markup to OpenAI, non-cash warrant amortization recorded as a revenue reduction, stock-based compensation and one-time items. Manufacturing runs through Flex and Sanmina; wafers come from TSMC's 5nm node, and management says it uses no HBM, no CoWoS packaging and no 3nm capacity. More than 600 MW of data-center capacity is live or contracted for delivery by end-2027 across named sites in the U.S., Canada and Europe.
Business Segments
Competitive Landscape
Cerebras competes for inference workloads against GPU-based systems and against other low-latency accelerators. Management's framing is that its chips are 58x larger than the largest competitor, which lets it use SRAM while rivals use HBM, and that it is one of two hardware vendors currently serving OpenAI models. Management also concedes the arguments are self-interested: asked about a rival's share estimate, it said "both of our arguments are, I think, in some way self-interested." The most unusual feature of the competitive set is AMD, which appears both as a listed competitor and as the announced Helios disaggregation partner.
- Listed in the relationship graph as a competitor on AI processors (GPUs versus wafer-scale). Management says serving frontier models opened strategic advantage "previously only available to NVIDIA."
- AMDListed as a competitor on AI GPUs and simultaneously the announced Helios disaggregation partner. Management says it does not see AMD's acquired inference team's first application as data center inference.
- GroqNamed in the relationship graph as a low-latency inference LPU maker; not discussed elsewhere.
- GraphcoreNamed in the relationship graph as an IPU maker; not discussed elsewhere.
- SambaNovaNamed in the relationship graph for a reconfigurable dataflow architecture; not discussed elsewhere.
Supply Chain
Cerebras buys 5nm wafers from TSMC, has systems built by Flex and Sanmina, colocates in third-party data centers, then sells inference through its own cloud. Of its supply-chain neighbors, only Arm and WhiteFiber mention Cerebras by name.
More on CBRS: Earnings recap