Cerebras Systems is the most compelling pure-play AI infrastructure company approaching public markets. In six months, the company has transformed from a single-customer chip startup into a multi-hyperscaler platform with $24.6B in remaining performance obligations and a $20B+ OpenAI contract, validated by OpenAI, Amazon/AWS, Meta, Oracle, the U.S. Department of Energy, and the Government of India. We initiate coverage with an OUTPERFORM rating and a 2-year market cap target of $70B, with significant upside to $150B on a 5-year horizon.
| SNAPSHOT | |
| Current valuation | $49B (IPO range, May 2026). Offering 28M shares at $185. Book 2.5x+ oversubscribed ($10B+ IOIs). Series H was $23B. |
| Total capital raised | ~$4B+ across 8 rounds |
| IPO target | May 13, 2026 | Nasdaq: CBRS | $3.5B raise | Morgan Stanley lead |
| FY2024E revenue | $290.3M actual (269% YoY growth) — S-1 filed April 17, 2026 |
| FY2026E revenue (est.) | $2.05B (revised down; anchored on $510M FY2025 actual) |
| Contracted backlog | $24.6B RPO as of Dec 31, 2025. OpenAI deal valued at $20B+ through 2030 |
| Key customers | OpenAI, Amazon/AWS, Meta, Oracle, IBM, Mistral, Perplexity, DOE, India |
| Inference capacity | 100+ exaflops of deployed compute, 85% U.S.-based |
| Core technology | WSE-3: 900K cores, 44GB on-chip SRAM, up to 15x faster than leading GPU-based solutions (S-1). 58x larger than Nvidia B200, 2,625x more memory bandwidth |
40 days that changed everything
Between January and March 2026, Cerebras announced six transformative developments that collectively rewrote the investment thesis:
| Date | Catalyst |
| Jan 14 | $20B+ deal for 750MW of inference compute through 2030, with expansion rights to 2 gigawatts. OpenAI also provided a $1B loan at 6% and received warrants for 33.4M shares |
| Feb 4 | Series H: Raised $1B at $23B valuation (184% increase in 5 months). Tiger Global led; AMD participated as strategic investor |
| Feb 12 | OpenAI Codex-Spark: First production AI model on non-Nvidia hardware. GPT-5.3-Codex-Spark runs at 1,000+ tok/s on WSE-3. 1M+ weekly active Codex users |
| Feb 20 | India: 8 exaflop national AI supercomputer with G42/MBZUAI/C-DAC under the India AI Mission. One of the largest AI infra investments in Asia |
| Mar 10 | Oracle: CEO Clay Magouyrk named Cerebras alongside Nvidia and AMD on Q3 FY26 earnings call. Oracle RPO quadrupled to $553B |
| Mar 13 | Amazon/AWS: Disaggregated inference (Trainium prefill + CS-3 decode) deployed on Amazon Bedrock. EXCLUSIVE to Bedrock. Amazon Nova models on Cerebras. AWS is first hyperscaler to commit |
The Amazon/AWS deal is particularly significant: AWS VP David Brown stated the result will be “inference that’s an order of magnitude faster and higher performance than what’s available today.” The solution is built on the AWS Nitro System, with Amazon Nova models running on Cerebras hardware, and Cerebras’ disaggregated inference architecture is exclusive to Amazon Bedrock. This embeds Cerebras into the distribution fabric of the world’s largest cloud platform.
Investment thesis: the five pillars
1. The inference TAM is exploding — and Cerebras owns the speed crown
Deloitte estimates inference will account for ~66% of all AI computation in 2026, up from ~50% in 2025. McKinsey projects 80% by 2027. The combined market for AI training and addressable inference is estimated at $251B in 2025, growing to $672B by 2029 — a 28% CAGR (Bloomberg Intelligence, per S-1). AI inference will grow more than twice as fast as training through 2029. Cerebras delivers 2,600+ tokens/second on Llama 4 Scout versus ~130 for ChatGPT — up to 15x faster than leading GPU-based solutions (per S-1), an advantage that is architecturally fundamental. This speed enables new categories: real-time voice agents, sub-second reasoning, interactive code generation, and agentic workflows.
2. OpenAI deal transforms the business — and is already in production
The $20B+ commitment through 2030 is now producing live product. GPT-5.3-Codex-Spark, launched February 12, is OpenAI’s first-ever production deployment on non-Nvidia hardware. Running at 1,000+ tok/s, it reduced per-client round-trip overhead by 80% and time-to-first-token by 50% — improvements that benefit ALL OpenAI models. Sam Altman personally teased the launch. Codex has 1M+ weekly active users.
The S-1 revealed the deal is far deeper than a compute purchase. OpenAI provided a $1 billion loan to Cerebras at 6% interest for datacenter buildout. OpenAI received warrants for 33.4 million shares of non-voting Class N stock. Most significantly, OpenAI and Cerebras signed a co-design agreement to build future models optimized for future Cerebras hardware. The initial commitment is 750MW with expansion rights to 2 gigawatts through 2030.
3. Six hyperscaler relationships in 40 days
Cerebras now has chip-level or infrastructure-level relationships with OpenAI ($20B+ deal), Amazon/AWS (Bedrock exclusive, Trainium+CS-3 disaggregated inference), Meta (Llama API backend), Oracle (OCI infrastructure), IBM (watsonx), plus deep partnerships with Mistral, Perplexity, Cognition, and 15+ additional named customers across enterprise, government, and life sciences.
4. Sovereign AI creates a massive new revenue stream
Cerebras for Nations (launched Nov 2025) positions the company as the infrastructure partner for sovereign AI. The India 8 exaflop supercomputer (announced Feb 20) is one of the largest AI infra investments in Asia. The UK AI Minister personally endorsed Cerebras. India’s government is targeting $200B in total AI infrastructure investment over the next two years. The Condor Galaxy network with G42 was the prototype; now it’s scaling globally.
5. Architectural moat is real and widening
The WSE-3 uses the entire 300mm silicon wafer as a single processor: 4 trillion transistors, 900K AI cores, 44GB on-chip SRAM with 21 petabytes/second bandwidth — 2,625x more memory bandwidth than Nvidia’s B200 package (per S-1). Cerebras maintains a 5x speed lead over Nvidia Blackwell on GPT-OSS-120B. The DARPA “Fuse” project ($45M) with Ranovus for photonic interconnects could extend this lead. Groq, the most direct inference competitor, was acquired by Nvidia for $20B in December 2025 — removing it from the independent market.
Financial model
Revenue buildup by segment
| Segment | FY24 act. | FY25 act. | FY26E | FY27E | FY28E | Notes |
| OpenAI | — | $50M | $1,200M | $4,000M | $6,000M | $20B+/thru 2030 |
| AWS / Bedrock | — | $10M | $150M | $500M | $1,000M | Binding term sheet |
| Meta | — | $30M | $100M | $250M | $400M | Llama API |
| Oracle (OCI) | — | — | $50M | $100M | $200M | OCI infra |
| G42 / MBZUAI | $220M | $280M | $200M | $150M | $100M | 62%+24% FY25 |
| Sovereign AI | $10M | $20M | $100M | $250M | $400M | India+UK |
| Startups/ent/gov | $60M | $120M | $250M | $400M | $600M | |
| TOTAL | $290M | $510M | $2,050M | $5,650M | $8,700M |
Revenue diversification transformation
In H1 2024, G42 represented 87% of revenue. By FY2025 (per the S-1), G42 dropped to 24% — but MBZUAI rose to 62%. Revenue concentration shifted between UAE partners rather than fully diversifying. By FY2027E, OpenAI’s $20B+ contract will dominate at ~71%, with AWS, Meta/Oracle, and Sovereign AI creating meaningful secondary channels.
Profitability and margin trajectory
| Metric | FY25E | FY26E | FY27E | FY28E |
| Revenue | $510M (actual) | $2,050M | $5,650M | $8,700M |
| Gross margin | 42–45% | 48–52% | 55–60% | 58–63% |
| Operating margin | 0–5% | 20–28% | 33–40% | 40–46% |
| Est. adj. EBITDA | ~-$76M (non-GAAP) | $400–600M | $1.8–$2.3B | $3.5–$4.0B |
Margin expansion thesis: As revenue shifts from hardware (36% GM) to cloud inference-as-a-service (60–70%+ GM), profitability inflects rapidly. The AWS Bedrock integration is especially accretive — Cerebras provides silicon, AWS handles sales, billing, and customer acquisition. Oracle OCI follows a similar model. FY2025 GAAP net income was $237.8M, but non-GAAP net loss was $75.7M after excluding stock-based compensation and fair value adjustments.
Valuation: where could the stock go?
| Bear | Base | Bull | |
| 2-year (Q2 2028) | $35B | $70B | $110B |
| 5-year (Q2 2031) | $50B | $150B | $250B+ |
| vs. $26.6B IPO | -$0 / +$88% | +163% / +464% | +314% / +840%+ |
| Key assumption | OpenAI delays; Nvidia closes gap | Full $20B+ ramp; AWS + Oracle scale | De facto inference standard; new hyperscalers |
Comparable company analysis
| Company | Mkt Cap | LTM Rev | EV/Rev | Growth |
| Nvidia | $3.4T | $115B | ~30x | ~95% |
| Arm | $190B | $4B | ~48x | ~25% |
| AMD | $180B | $26B | ~7x | ~15% |
| Marvell | $80B | $6B | ~13x | ~25% |
| Cerebras (FY26E) | $49B IPO | $2.05B FY26E | ~13x (FY26) / ~4.7x (FY27) | ~302% |
At $49B on $2.05B FY2026E revenue, Cerebras trades at ~13x — below Arm (48x) and Nvidia (30x) while growing 3–20x faster. On FY2027E revenue of $5.65B, the multiple compresses to just 4.7x. The $24.6B in RPO and $20B+ OpenAI contract provide contracted visibility that justifies the near-term premium. If the market assigns 15x on FY2027E, the implied market cap is $85B — a 220% return from IPO valuation.
Key risks
Customer concentration 2.0: OpenAI could represent 60–65% of FY2026 revenue. Mitigant: AWS Bedrock exclusive + Oracle OCI now create two additional mega-channels, materially reducing single-customer dependency vs. the G42 era.
Nvidia competitive response: Blackwell and Rubin GPUs will narrow the inference gap. Nvidia acquired Groq for $20B. Cerebras maintains a 5x speed lead over Blackwell on GPT-OSS-120B today, but this must be defended.
TSMC single-source: WSE-3 is fabricated exclusively on TSMC 5nm. No alternative exists for wafer-scale manufacturing. Any TSMC disruption would halt production.
Execution on buildout: 750MW for OpenAI + AWS Bedrock integration requires massive parallel capex. Cerebras has raised $2.1B in 5 months partly to fund this, but delays would reduce near-term revenue recognition.
Software ecosystem: Nvidia’s CUDA remains deeply entrenched. CSoft and the Cerebras developer platform are growing (#1 on HuggingFace, 5M+ monthly requests) but are still nascent vs. CUDA’s 20-year head start.
Conclusion: the asymmetric opportunity
Cerebras represents a rare asymmetric opportunity. The downside is bounded by $10B+ in contracted OpenAI backlog, exclusive AWS Bedrock integration, and validated technology (20x faster inference, #1 on HuggingFace). The upside is substantial if the company becomes the de facto inference standard for a market projected to reach $255B by 2030.
The key insight the market may be underpricing: inference is not a commodity where Nvidia automatically wins. Unlike training (where CUDA lock-in is extreme), inference is workload-by-workload, latency-sensitive, and architecturally favors single-chip solutions. The AWS disaggregated architecture (Trainium for prefill, CS-3 for decode) may become the industry standard — and every Bedrock customer becomes a potential Cerebras revenue source.
At $23B on forward FY2026 revenue of $3.9B, Cerebras trades at 5.9x — a fraction of comparable AI infrastructure multiples. With six hyperscaler relationships, the strongest pre-IPO customer roster of any AI hardware company in history, and the only independent high-speed inference platform left standing after Nvidia’s Groq acquisition, the path to $65–140B+ market cap within 2–5 years is credible.
⚠️ Disclaimer: This report is for informational purposes only and does not constitute investment advice, a recommendation, or a solicitation to buy or sell any security. All revenue figures, projections, and price targets are based on assumptions and estimates that may prove materially incorrect. Past performance is not indicative of future results. The author holds a small position in Cerebras purchased through secondary markets and may have a financial interest in the subject matter discussed. Do your own diligence before making any investment decisions.
