Cerebras Systems designs wafer-scale AI processors — the WSE-3, a single chip the size of a dinner plate with ~900,000 cores and enough on-chip memory to hold large models without external DRAM. It sells on-prem CS-3 systems and a hosted Cerebras Inference cloud, competing with NVIDIA on inference latency. Founded 2015; went public on Nasdaq in May 2026 under ticker CBRS.
Research
The verdict
Real wafer-scale inference moat priced at ~92x sales on demand that is circular (OpenAI is customer + lender + shareholder) and concentrated — the IPO swapped UAE concentration for OpenAI concentration. WATCHING, bear-leaning on valuation.
Cerebras designs and sells wafer-scale AI compute — the WSE-3, a single chip the size of a dinner plate (~900,000 cores, massive on-chip SRAM) — as on-prem CS-3 systems and as a hosted Cerebras Inference cloud. The pitch is latency: by keeping a whole model resident in on-chip memory it skips the GPU's external-DRAM and inter-chip networking penalty, claiming order-of-magnitude faster token generation. FY2025 revenue $510.0M (+76% YoY) split $358.4M hardware / $151.6M cloud & services. Two business models in one: lumpy hardware sales + a nascent recurring inference cloud.
Fabless: the WSE is built on TSMC (5nm-class), but as a single wafer rather than diced die — a genuinely differentiated process flow with packaging/yield/thermals all bespoke to Cerebras. Upstream concentration on TSMC is the same chokepoint every AI-silicon name shares; downstream the company increasingly operates the compute itself (its own + partner datacenters — six new sites, ~40M tokens/sec aggregate ), so it is moving from "sell the box" toward "sell the tokens." CDMO-equivalent risk: any wafer-scale yield problem has no second source.
The moat is architectural, not yet economic. Wafer-scale is hard to copy and delivers a real, benchmarked inference-latency edge — Cerebras claims ~2.4–6x Groq and faster-than-Blackwell on several models. But it sits inside NVIDIA's CUDA gravity well; the switching costs run against Cerebras. The durable question: is "fastest tokens/sec" a moat customers pay a premium for, or a spec that commoditizes as everyone optimizes inference? Today the edge is real; its defensibility is unproven.
Hardware 70% / cloud-&-services 30% of FY25 revenue. The strategic tell is geography: 86% of FY25 revenue is UAE-sourced, and US-billed revenue shrank 34% YoY — the cloud and US-commercial lines are early and not yet offsetting the customer-concentration risk. Mix is shifting toward cloud/inference, which is the right direction but small.
No 10-Q yet (IPO 2026-05-14). The S-1 is the anchor: $510M revenue, +76%, and — notably for a hyper-growth AI name — reportedly profitable ("files again, with profit this time" ); confirm net income against the S-1/first 10-Q before relying on it. The IPO raised $5.55B (30M shares at $185, upsized).
No earnings calls yet; the narrative arc is the roadshow → first-day pop → fade. Management messaging has pivoted hard from "G42 partner" (the CFIUS liability) to "OpenAI + AWS + Meta anchor customers" — a deliberate reframing of the concentration story (see Lens 10/13).
At ~$214 / ~$47B market cap (2026-06-15; prev close $226.55), on $510M TTM revenue that is **~92x sales ** — versus NVIDIA's far lower multiple on vastly larger, diversified, profitable revenue. Mechanism comps are the inference field: NVIDIA (Blackwell, the default), Groq (LPU), SambaNova, AMD (MI-series), Google TPU, AWS Trainium (now also a partner), Broadcom ASICs. Cerebras is the fastest on paper and the most expensive per dollar of revenue.
IPO $185 → +68% to $331 first-day close (~$95B cap) → faded ~35% to ~$47B in five weeks. The tape already encodes the debate: euphoria on the inference-vs-NVIDIA story, then a valuation/concentration hangover. Forward catalysts: first 10-Q (does OpenAI revenue show up, does US revenue inflect?), lock-up expiry, any OpenAI-deal documentation.
Co-founder/CEO Andrew Feldman (sold SeaMicro to AMD for ~$334M) and a long-tenured founding team — a decade-plus betting on wafer-scale when the consensus said diced-die. Credible, technical, mission-committed operators with real skin in the game. Capital allocation has been aggressive growth (eight private rounds, ~$1.9B raised, into self-operated datacenters). Watch for post-IPO insider selling and the governance optics of the OpenAI equity entanglement.
This is the heart of the bear case. (1) Customer concentration: 86% of FY25 revenue from two UAE-linked entities — MBZUAI 62% + G42 24% (G42 itself was 85% of 2024 revenue). (2) The OpenAI entanglement: OpenAI committed >$10B (Jan 2026, 750MW), reportedly up to $20B over 3 years for ~11% equity — making OpenAI simultaneously customer, lender, and shareholder. That is textbook circular AI financing and a related-party flag: the "we solved concentration" story replaces UAE concentration with OpenAI concentration and adds a conflict of interest. (3) US revenue −34% YoY undercuts the organic-diversification narrative. CFIUS cleared the G42 stake (Mar 2025, G42 → non-voting, later divested) so the national-security overhang is resolved, but the concentration economics remain.
CFIUS review of the G42 partnership withdrew the first (2024) IPO; cleared 2025-03-31 after G42 restructured to non-voting and divested Chinese holdings. No SEC enforcement (newly public). Standard newly-IPO'd litigation risk; revisit after the first 10-K. Verified via web + S-1 narrative as of 2026-06-15 (no EDGAR EFTS ingest this run).
Bull path: OpenAI/AWS/Meta convert the inference-speed edge into multi-hundred-MW recurring cloud revenue → revenue could multiply off $510M, justifying some of the multiple. Base path: hardware stays lumpy/UAE-tied while cloud ramps slowly; revenue ~$0.8–1.1B FY26 — still ~45–60x sales. Bear path: OpenAI terms prove low-margin or circular, US demand stays soft, the multiple compresses to a hardware-cyclical 8–15x. Brier-scoreable forecast to log: "CBRS FY2026 revenue ≥ $1.0B," p≈0.45 (resolves ~2027-03).
Bull: a real, defensible architectural edge in the growth workload (inference, not training); anchor logos (OpenAI, AWS, Meta, Mistral, Perplexity) that validate the tech; founder-led conviction; the only credible non-NVIDIA wafer-scale story. Bear: ~92x sales on revenue that is 86% concentrated and increasingly circular (OpenAI funds the customer that funds the revenue); US organic demand shrank; NVIDIA's CUDA moat + Blackwell cadence; inference is the most contested, fastest-commoditizing layer. Pre-mortem (18mo): the OpenAI deal is restructured or slips, a 10-Q shows UAE still dominant, the multiple halves. Contrarian read the market is underpricing: "diversification achieved" is a forward projection sold as a current fact.
The short thesis writes itself: a single-customer business that swapped one whale (G42) for another (OpenAI) and called it diversification — while the related-party who is also a shareholder and lender books the revenue. Strip the OpenAI forward commitment and you have a ~$510M, UAE-dependent hardware company at a ~$47B cap with shrinking US sales. What permanently impairs it: NVIDIA matches "good-enough" inference latency at lower TCO and the architectural edge stops commanding a premium. Most dangerous competitor bulls underrate: AWS Trainium — Cerebras's own new partner, which can learn the workload and in-source it.
The number-one HBM supplier and the chokepoint NVIDIA depends on.