Cerebras Systems
Hybrid: hardware sales (CS-3 wafer-scale systems & supercomputers) + cloud/services (per-token fast-inference API and managed compute), increasingly tilting to recurring cloud revenue.
The thesis on this name
State of AI Compute
Cerebras Systems builds wafer-scale AI processors (the WSE-3, an entire TSMC silicon wafer as one chip) and sells them as CS-3 systems and a high-speed inference cloud; freshly IPO'd on Nasdaq (CBRS) May 14 2026.
Earnings, margins, COGS & capex
Hyper-growth, still loss-making AI-chip upstart. Q1 FY26 core revenue $191.3M (+92% YoY), split hardware $111.6M (+60%) and cloud/services $79.8M (+167%) — the cloud mix-shift is the structural story. Core gross margin 47% but guided to crater to 36-38% in Q2 as Cerebras rents third-party data-center capacity to serve a $20B+ OpenAI contract before owned facilities come online; management frames the dip as a transitory 'cost of speed'. FY25 revenue $510M was 86% concentrated in UAE-linked entities (MBZUAI 62%, G42 24%). $3.3B cash post-IPO funds the buildout.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~53¢ is cost of goods and ~21¢ operating expense, leaving ~26¢ of operating profit.
Revenue trend
Margins
down — Q2 guide 36-38%, FY26 38-41%
hardware 41% / cloud 49%
FY26 guide (28)-(32)%
compressing ~10-15pts on rented capacity
COGS structure
Dominated by wafer-scale silicon (sole-source TSMC 5nm wafers), system assembly/packaging of the 46,225mm2 die, and — increasingly — rented third-party data-center power/capacity to bridge demand; the rental cost is the proximate cause of the Q2 gross-margin guide-down.
Capex
Stepping up materially to build owned inference/data-center capacity (the 750MW OpenAI deployment); interim reliance on rented capacity is explicitly a margin drag until owned facilities come online.
Latest earnings
Revenue BEAT (core $191.3M, +92%); stock FELL hard (~ -20% next day, ~ -47% off highs) on Q2 gross-margin guidance of 36-38% vs 47% prior — a ~1,000bps drop spooked the market.
Q2 FY26 core revenue ~$194.0M (+88% YoY), core gross margin 36-38%, core operating margin (30)-(32)%. FY26 core revenue $855-865M (+69%), core gross margin 38-41%, core operating margin (28)-(32)%.
- Core revenue
- $191.3M (+92% YoY)
- GAAP net loss
- $14.0M (narrowed from $23.9M)
- Core net loss
- $2.5M
- Cash & ST investments
- $3.3B
- OpenAI RPO
- $25.0B remaining performance obligations
- Q2 core gross-margin guide
- 36-38% (down from 47%)
Growth drivers
- OpenAI $20B+ multi-year Master Relationship Agreement (750MW inference) — $25.0B remaining performance obligations
- Cloud/fast-inference mix shift (+167% YoY) as agentic-coding/inference demand explodes
- AWS partnership (Bedrock/Marketplace disaggregated inference, Mar 2026) broadening reach
- New customers beyond UAE: Mistral, Cognition, and others diversifying the concentration risk
- Wafer-scale speed advantage (claimed ~3,000 tok/s on gpt-oss-120B vs GPU)
Bull & bear
A scarce pure-play on AI inference with a genuine architectural edge, an anchor OpenAI contract worth $25B in RPO, and a freshly recapitalized balance sheet to build out capacity — the margin dip is a transitory cost-of-speed, not a broken model.
- OpenAI MRA ($20B+, 750MW, $25B RPO) gives revenue visibility almost no other AI-hardware upstart has
- Wafer-scale speed advantage is structurally hard to replicate and most valuable in the booming inference/agentic-AI segment
- Cloud/services +167% YoY signals a durable mix-shift toward recurring, higher-margin revenue
- $3.3B cash + ~$8B+ total raised funds the owned-data-center buildout that should restore margins once rented capacity rolls off
- Customer base is visibly diversifying beyond the UAE (AWS, Mistral, Cognition) — the #1 bear point is actively eroding
A ~$40B market cap on $510M of FY25 revenue — 86% of it from two UAE-linked customers — that just guided gross margin down ~1,000bps, is GAAP-unprofitable, single-sources its chip from TSMC, and faces Nvidia plus every hyperscaler's in-house silicon.
- Customer concentration is existential: 86% of FY25 revenue from MBZUAI + G42 (UAE) carries geopolitical, export-control, and single-counterparty risk
- Gross margin guide-down to 36-38% (from 47%) undercuts the high-margin AI-chip thesis and triggered a ~47% drawdown
- Still losing money on a GAAP basis with negative operating margins guided through FY26 amid a heavy capex ramp
- Sole-source TSMC 5nm dependence — no second-source for a 46,225mm2 wafer-scale die
- Valuation (~46x FY26 revenue at the high) prices flawless execution against Nvidia, AMD, Groq, SambaNova, and hyperscaler ASICs
What it is worth
Reverse-DCF / revenue-multiple sanity check vs AI-semi peers
Margin compression proves structural, OpenAI buildout slips, UAE concentration triggers export-control or counterparty shock -> multiple compresses toward 10-20x sales, downside to the $100-130 region or below.
Strong (~60-70%) growth but margins stay structurally lower (low-40s%) and OpenAI realization is back-end loaded -> stock range-bound around current ~$180 as growth offsets margin/concentration discount.
OpenAI RPO realizes on schedule + margins recover to 50%+ + customer diversification continues -> revenue compounds to multi-$B, multiple re-rates; upside well above current price (Street avg PT ~$299).
At ~$40B market cap on FY26E core revenue of ~$860M, CBRS trades ~46x forward sales — a premium even by AI-semiconductor standards (NVDA/AMD trade at far lower sales multiples on far larger, profitable bases). The price implies sustained ~60-70%+ revenue CAGR for several years AND a recovery of gross margin back toward 50%+ once owned capacity replaces rented capacity. Reverse-DCF: to justify ~$40B, the market needs Cerebras to scale to multi-billion revenue (OpenAI RPO realization) at 50%+ gross and eventual positive operating margin — i.e. flawless execution against Nvidia. The 47%->36-38% margin guide and 86% UAE concentration are the two assumptions most likely to break the bull case.
SWOT
Strengths
- Differentiated wafer-scale architecture (WSE-3 — 4T transistors, 900k cores, 44GB on-chip SRAM) that sidesteps the GPU memory-bandwidth wall — class-leading inference latency/throughput
- Marquee $20B+ OpenAI anchor contract + $25B RPO providing multi-year revenue visibility
- Fortress balance sheet post-IPO (~$3.3B cash, ~$6.4B raised) to fund the data-center buildout
- Cloud/services growing +167% YoY — higher-margin, recurring, less lumpy than hardware
- Real, named blue-chip demand: OpenAI, Mistral, Cognition, AWS partnership
Weaknesses
- Severe customer concentration: ~86% of FY25 revenue from two UAE-linked entities (MBZUAI 62%, G42 24%)
- Gross margin guided down ~1,000bps (47%->36-38%) — profitability path is now murkier
- Still GAAP loss-making with negative operating margins guided through FY26
- Sole-source dependence on TSMC 5nm — single point of failure for the entire product
- Heavy capex/financing needs; rented-capacity stopgap actively destroys near-term margin
Opportunities
- Inference is the larger, faster-growing AI-compute TAM vs training — Cerebras's speed edge is most valuable here
- Customer diversification away from UAE concentration (OpenAI, AWS, Mistral, Cognition) de-risks the franchise
- Agentic AI / reasoning models reward low-latency inference, Cerebras's core advantage
- Sovereign-AI and hyperscaler demand for non-Nvidia second sources
- AWS Bedrock/Marketplace distribution scales reach without owning the customer relationship
Threats
- Nvidia's overwhelming scale, CUDA software moat, and Rubin-generation roadmap
- Hyperscaler in-house silicon (AWS Trainium/Inferentia, Google TPU) commoditizing inference
- Well-funded specialist rivals Groq and SambaNova chasing the same speed niche
- Geopolitical/export-control exposure on UAE (G42/MBZUAI) revenue base
- Margin-compression narrative + IPO-lockup supply could keep the stock under pressure
Moats, dependencies & bottlenecks
Moats
WSE-3 single-wafer design (4T transistors, 900k cores, 44GB on-chip SRAM) eliminates the GPU memory-bandwidth wall — hard, capital- and know-how-intensive to copy.
$25B RPO and 750MW deployment create deep integration, but a single counterparty is also a concentration risk.
Claimed ~3,000 tok/s leadership is real today but contested by Groq, SambaNova, and Nvidia's roadmap.
No CUDA-equivalent moat; faces Nvidia's entrenched developer ecosystem.
Dependencies
Only foundry able to manufacture the wafer-scale die — single point of failure for the entire product line.
$20B+ MRA / $25B RPO is the growth engine but also concentrates future revenue in one counterparty; realization is multi-year and capacity-gated.
~86% of FY25 revenue; geopolitical and export-control exposure.
Interim capacity to serve OpenAI is the direct cause of the Q2 margin guide-down; rolls off as owned facilities come online.
Bedrock/Marketplace distribution (Mar 2026) extends reach but AWS also sells competing Trainium/Inferentia silicon.
Advantages
- Best-in-class inference latency/throughput from wafer-scale design
- Scarce public pure-play on AI inference hardware
- Anchor OpenAI contract with multi-year RPO visibility
- Strong post-IPO liquidity (~$3.3B cash) to fund the buildout
Weaknesses
- 86% UAE customer concentration (FY25)
- Gross margin guided down ~1,000bps
- GAAP-unprofitable with negative operating margins
- Single-source TSMC dependence
- Premium revenue multiple pricing in flawless execution
Bottlenecks
- Owned data-center capacity not yet online — forcing margin-dilutive rentals
- TSMC wafer allocation as the single manufacturing chokepoint
- Capital intensity of the 750MW OpenAI buildout
- Customer-diversification pace vs. UAE concentration
Top signals & trends
Top signals
~1,000bps drop; market read it as structural, drove ~47% drawdown — mgmt insists transitory.
Rare multi-year revenue visibility for an AI-hardware upstart.
Higher-margin recurring mix shift accelerating.
Directly attacks the UAE-concentration bear case.
De-rating could be opportunity or correct re-pricing of margin/concentration risk.
Trends
Plays directly to Cerebras's latency advantage.
Speed premium where Cerebras leads.
Commoditizes inference; pressures merchant accelerators.
Default-buy gravity and software moat.
Demand for alternative accelerators (also the UAE origin story).
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Sole foundry for the wafer-scale 5nm die — the critical single-source dependency.
Networking/interconnect silicon used across AI-systems supply chain (general ecosystem supplier).
$20B+ MRA, 750MW inference, $25B RPO — the anchor customer.
~62% of FY25 revenue; analysis only, not a US-buyable name.
~24% of FY25 revenue; named as analysis, not a buy/own recommendation.
Inference-cloud customer; diversification beyond UAE.
Agentic-coding workloads on Cerebras inference.
Dominant AI-compute incumbent; CUDA moat, scale, Rubin roadmap. The benchmark Cerebras must beat on inference economics.
Instinct MI-series GPUs + ROCm; the credible #2 merchant GPU and a price/perf pressure on inference.
LPU inference specialist chasing the same ultra-low-latency niche; well-funded direct rival.
Reconfigurable dataflow (RDU) systems for training+inference; overlapping enterprise/sovereign accounts.
Custom AI ASIC/XPU partner to hyperscalers (e.g. Google TPU); enables the in-house-silicon threat.