Cerebras Systems Inc. (CBRS) | The Buildout — AI Infrastructure
The Verdict
Cerebras makes the CS‑3 supercomputer, built around its wafer‑scale engine — a single chip the size of a 300 mm silicon wafer — that delivers extreme inference speed. It sells systems directly to AI labs, cloud providers, and governments, and operates a cloud service for inference. Its wafer‑scale approach avoids the HBM and advanced packaging bottlenecks that constrain GPU makers, giving it a supply‑chain edge. The company’s role is to be the high‑throughput, low‑latency inference layer in the AI buildout, both for proprietary models and in disaggregated inference with partners.
| Market Cap | — |
| Revenue (TTM) | $604M |
| Revenue Growth | +87.8% |
| EBITDA Margin (TTM) | -18.9% |
| Net Cash | $1.9B |
| Earnings Beats | 1 of 1 |
| P/E (TTM) | — |
| EV/EBITDA (TTM) | — |
What We Like
- $25.0 billion remaining performance obligation as of March 31, 2026, time‑phased to deliver ~$4 billion in the next two years.
- Core cloud‑services gross margin reached 52.9% in Q1. Management targets 60%+ overall gross margin long term.
- Manufactures in the US with dual contract sources; avoids HBM, CoWoS, and 3 nm constraints.
- AWS signed a definitive agreement for disaggregated inference, adding a second hyperscale relationship beyond OpenAI.
- Raised $6.2 billion net in the largest semiconductor IPO, plus $3.3 billion in pre‑IPO cash and investments.
What We’re Watching
- Data‑center capacity additions are the binding constraint; any delays in utility approvals or construction would push revenue out.
- Arm AGI CPU bottleneck: Arm’s CEO says the chip is sold out; Cerebras has not disclosed its allocation, risking a production halt.
- Securities‑fraud investigations by at least seven law firms could lead to class‑action litigation or regulatory action.
- Full‑year core operating margin guided to (28)%–(32)%, implying heavy cash burn until 2027 or later.
The investment thesis has strengthened on extraordinary demand visibility but is tempered by acute execution risk and a credibility cloud. The $25 billion backlog, now publicly filed, confirms that Cerebras has secured massive AI inference demand. Yet the company’s ability to deliver depends on the speed of its data‑center rollout, and the sudden wave of fraud investigations raises questions about pre‑IPO disclosures. The thesis hangs on execution over the next 12–18 months. The key open question is whether Cerebras can bring enough megawatts online before its cash burn and legal overhang undermine confidence.
Earnings Beat
Cerebras reported its first public quarter with GAAP revenue of $193.4 million, up 92% year‑over‑year, and positive EBITDA of $3.1 million. Cloud‑services revenue jumped 167% as the OpenAI ramp went live in February, while hardware revenue grew more slowly. The remaining performance obligation soared to $25.0 billion.
| Metric | Q1 FY2026 | Q4 FY2025 | Q1 FY2025 | YoY |
|---|---|---|---|---|
| Revenue | $193M | $171M | $100M | +94.4% |
| Gross margin | 44.6% | 41.0% | 41.8% | +280bps |
| EBITDA | $3M | −$34M | −$28M | −110.9% |
| EPS | $-0.22 | $-0.12 | $-0.11 | +100.7% |
| Remaining Performance Obligation | $25.0B | n/a | n/a | — |
We signed a definitive agreement … on December 24, 2025. … By February 1, we were in production, running a model we’ve never before seen — 35 days from signature to production deployment.— Andrew Feldman, CEO, June 23, 2026
Management tone: On the company’s first public call, CEO Andrew Feldman was confident and expansive, emphasizing a 13 × speed advantage over GPUs and declaring that “demand is not the constraint. The constraint is data centres.” CFO Bob Komin was more measured, calling the full‑year guide “conservative” and detailing the temporary margin hit from renting back systems. The call occurred two days before the first securities‑fraud investigations were announced, so management did not address any legal issues.
Management Guidance
For Q2, management guided to core revenue of ~$194 million, core gross margin of 36%–38%, and core operating margin of (30)%–(32)%. For the full year, it expects core revenue of $855–$865 million, core gross margin of 38%–41%, and core operating margin of (28)%–(32)%. The margin pressure reflects the cost of renting back systems from a customer to meet near‑term demand while the company’s own data centers are built.
Trajectory
Revenue is shifting rapidly from hardware to cloud services, which grew 167% in Q1 and will dominate going forward. Total revenue growth of 92% YoY is set to accelerate in the second half as new data‑center capacity comes online. Gross margins, however, will dip to the mid‑30s% temporarily as the company rents capacity, before recovering toward a 60% target.
The Model
The model projects FY+1 (2026) revenue of $865 million and EBITDA of -$227 million, implying a margin of -26.3%, consistent with management’s full‑year guidance for heavy operating losses. FY+2 (2027) revenue jumps to $2,776.0 million with EBITDA of $611 million (22%), anchored by the $25 billion backlog conversion and the expected scale‑out of AWS and other cloud deployments.
| Metric | FY2025 | Next FY (E) | Following FY (E) |
|---|---|---|---|
| Revenue | $510M | $865M | $2.8B |
| YoY Growth | — | +69.6% | +220.9% |
| EBITDA | −$146M | −$227M | $611M |
| EBITDA Margin | -28.6% | -26.3% | 22.0% |
Projections are the median of 5 independent model runs.
For Q2, management guided to core revenue of ~$194 million, core gross margin of 36%–38%, and core operating margin of (30)%–(32)%. For the full year, it expects core revenue of $855–$865 million, core gross margin of 38%–41%, and core operating margin of (28)%–(32)%. The margin pressure reflects the cost of renting back systems from a customer to meet near‑term demand while the company’s own data centers are built.
What Could Go Right — and Wrong
- Data‑center capacity comes online faster than forecast, accelerating cloud revenue and margin recovery.
- AWS deployment scales and a third hyperscaler signs a disaggregated inference deal, diversifying revenue.
- The securities‑fraud investigations are dismissed or settled quickly with no material impact.
- Arm allocates sufficient AGI CPUs, removing a hidden production bottleneck.
- OpenAI exercises hardware purchase options, pulling some revenue forward while cloud continues to grow.
- OpenAI renegotiates or cancels its contract; the $25 billion backlog shrinks dramatically.
- Data‑center delays extend, pushing revenue out by years and deepening cash burn.
- The fraud investigations uncover material misstatements, leading to litigation, restatements, or capital‑access issues.
- Arm CPU shortage limits CS‑3 production, undercutting the “supply is not the constraint” claim.
- Inference speed gap closes to 2–3×, eroding pricing premiums and margin ambitions.
Looking Ahead
The next 12 months will test Cerebras’s ability to execute on its data‑center build‑out and to manage the fallout from the legal investigations. The company expects new facilities to go live each quarter, which should drive the cloud revenue ramp. At the same time, the AWS disaggregated inference partnership is slated to begin having an impact in 2027, while multiple hardware partners may emerge in the second half of 2026.
- Q3 FY2026Next data‑center additions — New capacity comes online; critical for cloud revenue ramp in H2 2026.
- 2027Disaggregated inference partners — AWS disaggregated inference expected to begin impact; other partners may follow.
- 2027AWS deployment begins — Cerebras decode at scale in AWS data centers; revenue impact expected.
- UnknownFraud investigation resolution — Dismissal or legal filing will clarify risk; major overhang for stock.
- ShortlyGPT‑5.5 on Cerebras — OpenAI deepening partnership; validates architecture for next‑gen models.
Financials
Annual Summary
| Metric | FY2025 | TTM |
|---|---|---|
| Revenue | $510M | $604M |
| Gross Margin | 38.6% | 40.4% |
| EBITDA | −$146M | −$221M |
| EBITDA Margin | -28.6% | -18.9% |
| Net Income | $238M | $248M |
| Free Cash Flow | — | $30M |
| Net Cash | — | — |
Key Ratios (Trailing)
- P/E TTM—
- EV/EBITDA TTM—
- EV/Revenue TTM—
- Price/FCF TTM—
- Gross Margin (TTM)40.4%
- EBITDA Margin (TTM)-18.9%
- Net Margin (TTM)41.0%
- ROIC-12.3%
- SBC / Revenue7.7%
The Company
Cerebras Systems designs and sells the CS‑3 supercomputer, an AI compute system built around its Wafer‑Scale Engine (WSE‑3) — the largest chip in production, fabricated on TSMC’s 5 nm process. The architecture delivers what the company describes as industry‑leading inference speed, demonstrated at 13 times faster than a leading GPU‑based cloud on a trillion‑parameter model. This speed is the core economic proposition: faster tokens command a premium in AI inference.
The company sells hardware directly to AI labs, cloud providers, enterprises, and governments, and offers compute through its Cerebras Inference Cloud, which is consumption‑based. It also supplies an Integrated‑AI‑rack through OEMs like Dell, HPE, and Supermicro. Systems are assembled in the US by Flex and Sanmina. Cerebras operates data centers in multiple regions, with a large planned site in Saskatchewan (300 MW) subject to approvals, and has committed $3.9 billion in future data‑center lease payments.
Business Segments
Competitive Landscape
Cerebras competes in AI accelerators against Nvidia’s dominant GPU ecosystem, AMD, and a growing field of custom ASICs from hyperscalers. It differentiates by offering what it claims is leading speed in the decode phase of inference, and by partnering with some competitors through disaggregated inference architectures — for example, AWS Trainium for prefill and Cerebras for decode.
- NVIDIA (NVDA)Dominant with GPUs and CUDA; noted in competitors list. Cerebras claims 13 × speed advantage on a large model.
- AMD (AMD)Competes with MI300X/MI400 AI GPUs.
- AI startups (Graphcore, Groq, SambaNova)Alternative architectures; listed as competitors in company intel.
- Amazon (AWS) TrainiumCustom ASIC; now a partner for disaggregated inference but also a potential long‑term competitor.
- Google TPU, Meta custom acceleratorHyperscaler internal ASICs that reduce external dependency over time.
Supply Chain
Cerebras sits atop a global supply chain with sole‑source dependencies on TSMC and likely on Arm, but its wafer‑scale design sidesteps the industry’s most constrained inputs — HBM memory and TSMC’s CoWoS packaging.