Cerebras Systems

Snapshot

• Ticker: NASDAQ: CBRS
• Price: $185 IPO (+68% debut, May 14 2026)
• Market cap: ~$56bn
• Revenue: $510m FY2025 (+76% YoY)
• Growth: +76% YoY
• Profitability: net income $237.8m (likely one-off-flattered)
• Valuation: ~100x sales
• Founded / HQ: 2015 / Sunnyvale, CA
• CEO: Andrew Feldman (co-founder)
• Top competitors: NVIDIA, Groq, SambaNova, AMD, Google TPU
• Key customers: OpenAI, AWS, Meta, G42, MBZUAI
• Key suppliers: TSMC
• Verdict: Strongest AI-compute challenger; concentration- & valuation-rich
• Confidence: 0.57

Executive summary

Cerebras builds the Wafer-Scale Engine (WSE-3) — a single dinner-plate-sized chip cut from an entire silicon wafer (~4 trillion transistors, ~900,000 cores) — sold as CS-3 systems and “Condor Galaxy” supercomputers, and increasingly delivered as a record-speed AI-inference cloud. It is the most credible non-GPU challenger of the cohort: revenue grew from $24.6m (2022) to $290m (2024) to $510m in 2025 (+76%), it reported a swing to net income of $237.8m, and it completed the biggest US tech IPO since Snowflake in May 2026 (Nasdaq: CBRS, +68% on debut) [S1][S2][S10]. It has begun landing marquee, non-UAE customers — a $10bn+ OpenAI inference agreement, an AWS deployment, and Meta’s Llama API [S2][S4]. The defining risk is quality of revenue: concentration merely rotated within Abu Dhabi — from G42 (85% of 2024 sales) to G42 24% + MBZUAI 62% in 2025 (~86% combined) — while it faces NVIDIA’s CUDA moat, inference specialists like Groq, wafer-scale economics, and an extreme (~100x-sales) valuation [S2][S7][S8].

Verdict: the strongest name in the cohort — genuine NVIDIA-alternative traction (record growth, reported profitability, a landmark IPO, OpenAI/AWS/Meta wins) — but still ~86% Abu-Dhabi-concentrated, up against the CUDA moat and inference rivals, with wafer-scale supply/economic risk, at a valuation that already prices a platform outcome. Confidence: 0.57

Company overview

Founded in 2015 (incorporated 2016) and based in Sunnyvale, CA, Cerebras set out to beat GPU clusters by putting an entire AI system on one wafer. It nearly failed early (reportedly burning ~$8m/month before product-market fit), then scaled via systems sales and the Condor Galaxy supercomputers for G42 (CG-1 ~4 exaflops, CG-3 ~8 exaflops) and a fast-inference cloud [S5][S10]. After filing to IPO in 2024, the listing was delayed by a CFIUS review of G42’s stake; the review concluded in October 2025 (G42’s holding converted to non-voting), clearing the May 2026 IPO [S3][S5].

Management & founders

Cerebras is founder-led by Andrew Feldman (co-founder & CEO), with Sean Lie as CTO/chief technology; the other co-founders are Gary Lauterbach, Michael James and Jean-Philippe Fricker [S5]. The team’s pedigree is notable: the same group built and sold SeaMicro to AMD in 2012 for $334m, so this is a proven systems/silicon group on its second act [S5]. The May-2026 IPO reportedly minted two billionaires, underscoring meaningful founder ownership and alignment — a positive — though concentrated founder/insider control and a recently public governance structure are worth monitoring [S10].

Business model & products

The WSE-3 keeps an entire model resident in fast on-wafer SRAM, eliminating the chip-to-chip communication that bottlenecks GPU clusters — yielding headline records (Meta’s Llama API reportedly ~18x faster than GPU alternatives) and very fast single-model inference [S2][S5]. Revenue has two engines: hardware/systems (CS-3, Condor Galaxy) at ~70% of 2025 sales ($358.4m), and cloud/inference services at ~30% ($151.6m) and growing [S4]. The model’s elegance is also its constraint: on-wafer SRAM is fast but low-capacity, so very large models must stream weights off-wafer (the MemoryX/SwarmX scheme), reintroducing bandwidth complexity, and a full-wafer device is bespoke to manufacture, power and cool.

Financial analysis

A rare combination for this cohort: explosive growth and reported profitability — though the profit figure warrants scrutiny.

Revenue compounded ~4.5x from 2022 to 2024 and grew another 76% in 2025 to $510m [S2][S4]. The swing to +$237.8m net income (from a −$481.6m loss) is striking, but a ~47% net margin on hardware-heavy revenue almost certainly reflects sizable non-operating / one-off items (fair-value or warrant remeasurement tied to the financing/IPO and large customer deals) rather than clean operating profitability — a key item to verify in the filings. Cash is bolstered by the ~$5.55bn IPO raise [S1]. Net: the top line and customer wins are real; the durability and operating quality of the profit are the open questions.

Customers & suppliers

Customers: revenue remains highly concentrated — in 2025, MBZUAI ~62% and G42 ~24% (both Abu-Dhabi-linked, ~86% combined), with G42 also a (now non-voting) shareholder — so most revenue is effectively related-party [S2][S3]. The strategic priority is diversification, now underway via OpenAI (a $10bn+ inference agreement), AWS (first hyperscaler to host Cerebras), and Meta (Llama API), plus research/enterprise users [S2][S4]. Suppliers: Cerebras depends on TSMC to fabricate its wafer-scale chips but reportedly has no formalized long-term supply or allocation commitment, and TSMC also serves Cerebras’s competitors — a genuine single-source/allocation risk — alongside a bespoke power, cooling and packaging supply chain [S5].

Sector & market (TAM)

Cerebras plays in AI accelerators, with a wedge in fast inference — a large, fast-growing market NVIDIA dominates:

• AI accelerator market: set to exceed US$600bn by 2033 (Bloomberg Intelligence), driven by hyperscale spend and ASIC adoption [S9]; the broader AI-chip market is forecast around US$295bn by 2030 (~33% CAGR) [S2].
• AI inference chips: ~US$31–105bn (2024–25) scaling to ~US$170–570bn by 2032–33 at ~24–28% CAGR; inference is the fastest-growing slice — the part Cerebras targets [S6].
• NVIDIA share: ~80–90% of AI accelerators (training >90%, inference ~60–75%), expected to hold ~70–75% through 2030 [S7].

Cerebras’s served market is the non-NVIDIA, non-hyperscaler-captive merchant slice, especially latency-sensitive inference — enormous, but tightly held by the incumbent and shrinking as hyperscalers build in-house silicon.

Competitive landscape

Cerebras fights NVIDIA above, hyperscaler in-house silicon laterally, and inference specialists head-on.

The core competitive fact is the CUDA moat: a non-GPU architecture must continuously map new models/kernels onto an exotic fabric, so Cerebras can “win the benchmark but lose the deployment” if new architectures land on GPUs first [S7][S8]. Its wedge — fastest single-model/inference throughput — is real and is what won OpenAI, Meta and AWS interest, but it is contested directly by Groq and SambaNova and indirectly by ever-cheaper GPU inference.

Growth drivers & catalysts

• OpenAI agreement — $10bn+ multi-year, 750 MW of inference capacity (expandable to ~2 GW by 2030); the largest commitment in company history and a major diversification beyond UAE [S2][S4].
• AWS deployment — first hyperscaler to host Cerebras (Bedrock), a powerful validation and channel [S4].
• Meta Llama API — production inference at record speeds, showcasing the wedge [S2].
• Inference-cloud mix shift — recurring, higher-quality revenue growing toward ~30%+ of sales [S4].
• IPO war chest — ~$5.55bn raised funds capacity and diversification [S1].

Recent news

• May 14, 2026 — IPO: debuted on Nasdaq (CBRS), shares +68% on day one; ~$5.55bn raised, the biggest US tech IPO since Snowflake; minted two billionaires [S1][S10].
• Mar 2026 — AWS: signed a term sheet making AWS the first hyperscaler to deploy Cerebras in its own data centers [S4].
• Jan 2026 — OpenAI: multi-year inference agreement, 750 MW, $10bn+ at signing [S2].
• Oct 2025 — CFIUS: review of G42’s stake concluded; stake converted to non-voting, clearing the IPO path [S3].

Headwinds & key risks

• Customer concentration (~86% Abu-Dhabi-linked): G42 (24%) + MBZUAI (62%) in 2025; concentration rotated rather than diversified, with related-party/geopolitical exposure [S2][S3].
• CUDA moat / ecosystem gap: no CUDA-class software; new models land on GPUs first; benchmark-wins-not-deployments risk [S7][S8].
• Inference competition: Groq and SambaNova attack the same wedge; GPU inference costs keep falling.
• Wafer-scale economics & supply: defect-tolerant full-wafer devices, bespoke power/cooling, low-capacity SRAM, and single-source TSMC dependence with no long-term commitment [S5].
• Profit quality: 2025 net income likely embeds large one-offs; recurring operating profitability unproven.
• Valuation: ~$56bn vs $510m revenue (~100x sales) prices a platform outcome; high beta to the AI-capex cycle.

Valuation

Cerebras priced one of the largest AI IPOs to date and rose +68% on its May-14-2026 debut, for a market value reported near ~$56bn, raising ~$5.55bn [S1][S10]. On $510m of 2025 revenue that is on the order of ~100x sales — extreme even for hyper-growth AI compute, underwritten by the OpenAI/AWS ramp and the “credible NVIDIA alternative” narrative. The valuation requires Cerebras to (1) convert the OpenAI/AWS/Meta wins into large recurring inference revenue, (2) genuinely diversify away from the ~86% Abu-Dhabi base, and (3) defend its speed wedge against Groq and falling GPU costs — while sustaining real (not one-off) profitability. Any stumble implies large downside from here.

Verdict & what to watch

Cerebras is the standout of the names analyzed: a genuine technical breakthrough that has converted into explosive revenue ($510m, +76%), reported profitability, a landmark IPO, and the kind of marquee customers (OpenAI, AWS, Meta) that begin to validate a real NVIDIA-alternative platform. But the bull case is not yet proven where it most matters — revenue is still ~86% from two Abu-Dhabi-linked entities, the CUDA moat and inference rivals are formidable, wafer-scale carries unique economic/supply risk, the 2025 profit likely flatters with one-offs, and the ~100x-sales valuation prices much of the platform outcome in advance. Verdict: the most compelling AI-compute challenger here, but concentration- and valuation-rich — confidence 0.57.

Decision boundaries (what would change the view):

• Non-UAE revenue (OpenAI/AWS/Meta) driving combined G42+MBZUAI below ~50% of sales, with absolute dollars growing → materially more positive (+).
• Durable, growing recurring inference revenue at clearly operating margins → more positive (+).
• Day-one model/ecosystem support (new models run on Cerebras without long porting) → more positive (+).
• Concentration staying ~80%+ Abu-Dhabi-linked, or OpenAI/AWS deals slipping → more negative (−).
• Profit revealed as one-off-driven with operating losses persisting → more negative (−).
• AI-capex normalization or multiple compression from ~100x sales → more negative (−).

Open questions (highest-leverage unknowns):

• How much of 2025 net income is operating vs non-operating/one-off?
• Pace and dollar size of non-UAE revenue recognition vs the ~86% concentration.
• Segment (hardware vs inference) gross margins and inference-cloud utilization/ARR.
• WSE yield/cost economics and TSMC allocation; production scalability to many customers.
• Durability of the speed wedge vs Groq/SambaNova and falling GPU inference costs.