Executive Summary
Cerebras builds the Wafer-Scale Engine, a single 5nm chip the size of a dinner plate, and pairs it with the Cerebras Inference cloud. On raw token speed it is the fastest silicon in the market, a fact independently corroborated on the Artificial Analysis leaderboards. It IPO'd on Nasdaq on May 14, 2026 at $185, popped ~68% to a ~$311 day-1 close, ran to a $386 high, then tumbled to $160.81 after its first post-IPO earnings before recovering toward the offer near $187.53 in roughly ten weeks of public trading. The reported $24.6B remaining performance obligation1 (RPO), about 48x FY25 revenue, is the headline number, but it is back-loaded and driven by a single OpenAI master agreement.
Tactical: CBRS trades at $187.53, ~1.4% above our $185 base-case fair value, which itself sits deliberately below the ~$292 Street average. Blended fair value ~$201 (30/45/25 weights) sits just above spot on a fat but low-probability OpenAI-ramp tail. Rating Hold: own the speed leadership, size for the concentration and margin risk. This is a framework, not a formal rating or target.
Quarterly Revenue ($M)
Backlog (RPO) Growth ($B)
Investment Thesis Summary
Bull Case
- OpenAI 750MW agreement ramps on schedule and expands toward 2GW
- UAE-related-party share falls below ~40% by 2027 as US / cloud diversifies
- Gross margin recovers to 45%+ as one-time pricing effects normalize
- FY28E revenue reaches ~$4.5B on faster backlog conversion
- EV/Sales re-rates to ~19x on durable inference-speed leadership
Base Case
- Real but back-loaded ramp; backlog converts on the disclosed cadence
- UAE-related parties stay ~50-60% of revenue through 2027
- Gross margin holds high-30s / low-40s after the guided-down Q2
- FY28E revenue ~$3.5B; approximate operating breakeven around 2028
- EV/Sales ~15x; roughly fairly valued at spot
Bear Case
- Margin compression persists (Q2 core GM guided to 36-38%)
- OpenAI slips or renegotiates; near-term recognition stays tiny
- Export controls tighten on the UAE entities (CFIUS overhang)
- NVIDIA's CUDA ecosystem blunts share despite the speed edge
- FY28E revenue ~$2.5B; EV/Sales de-rates to ~12x
Overall Rating: Hold · Base fair value $185 (Blended FV ~$201)
We frame Cerebras as a Hold with a $185 base-case fair value. This is a genuine technology leader, the fastest silicon for inference, wrapped in a set of risks that are unusually concentrated: two UAE-linked related parties were ~86% of 2025 revenue, FY25 GAAP profitability is an accounting artifact, and the backlog that anchors the valuation is back-loaded to 2028 and beyond. Weighting Bull / Base / Bear at 30 / 45 / 25 produces a blended fair value of roughly $201, only a few percent above spot. The Street is more bullish: sell-side consensus is a Buy with an average price target near $292, and our Hold with a $185 base fair value sits deliberately below it as a valuation-discipline stance. The reward for owning the speed franchise is real, but the risk/reward at $187.53 is close to balanced, which is why we would own it small rather than chase it. This is a framework, not a formal rating or price target.
1 · The Wafer-Scale Engine
Cerebras's core innovation is refusing to cut the wafer into chips. The WSE-3 is a single TSMC 5nm wafer carrying 4 trillion transistors and 900,000 AI cores, delivering 125 PFLOPS of FP16 compute across roughly 57x the silicon area of the largest GPU.3 The decisive number for inference is memory: 44GB of SRAM sits on the wafer itself, versus tens of megabytes on a conventional GPU. Keeping model weights on-chip is what removes the memory-bandwidth bottleneck that throttles token generation.
On-Chip SRAM (GB): WSE-3 vs GPUs
- CS-3 system: houses one WSE-3; the deployable unit sold to customers and run in the Cerebras cloud.
- MemoryX: adds up to 1.2PB of external memory for weight streaming on the largest models.
- SwarmX fabric: clusters up to 2,048 CS-3s, addressing models up to 24 trillion parameters.
- Condor Galaxy: the network of Cerebras-built AI supercomputers underpinning the training and inference footprint.
WSE-3 vs the Largest GPU
| Metric | WSE-3 | Leading GPU (H100 class) | Cerebras edge |
|---|---|---|---|
| Silicon area | ~46,225 mm² (full wafer) | ~814 mm² | ~57× |
| Transistors | 4 trillion | ~80 billion | ~50× |
| AI cores | 900,000 | ~16,900 (CUDA/Tensor) | Massively parallel |
| On-chip SRAM | 44 GB | ~50 MB | ~880× |
| Peak compute (FP16) | 125 PFLOPS | ~2 PFLOPS (sparse) | Wafer-scale |
GPU figures are approximate, drawn from public specifications and used here for scale. The point is not the exact ratio but the architecture: Cerebras trades manufacturing complexity and yield risk for a memory-bandwidth advantage that a chiplet-based GPU cannot match.
2 · Inference Speed: The Strategic Growth Engine
Because the model weights live in on-wafer SRAM, the Cerebras Inference cloud achieves roughly 7,000x the memory bandwidth of an H100 (a vendor figure).4 Memory bandwidth, not raw training FLOPs, is the binding constraint on token generation, so this is where wafer-scale wins. The claim is independently corroborated by the Artificial Analysis speed leaderboards, which is what separates this from a vendor benchmark.5
Inference Output Speed (tokens/sec)
- >2,200 tokens/sec on Llama 3.3 70B, multiples of a typical GPU cloud.
- 969 tokens/sec on Llama 3.1 405B, a frontier-scale open model.
- ~1,000 tokens/sec on a trillion-parameter model.
- Cerebras claims to beat NVIDIA's Blackwell on token speed; the edge is memory-bound inference, not training.
Why Speed Is the Wedge
Inference is the strategic growth engine, and it is where Cerebras's architecture is most differentiated. Agentic and reasoning workloads generate long token streams, and latency compounds across chained calls, so a several-fold speed advantage is a genuine product differentiator rather than a spec-sheet win. The open question is monetization: the benchmark leadership is not in dispute, but converting it into diversified, non-related-party cloud revenue at healthy gross margins is the crux of the entire thesis, and it has not happened yet.
3 · Customer Concentration: The UAE Problem
The catch to the whole story is on the revenue line. UAE-linked related parties were ~86% of 2025 revenue1: MBZUAI (Mohamed bin Zayed University of AI) at ~62% and G42 at ~24%. MBZUAI alone was ~77.9% of year-end accounts receivable. Critically, concentration did not fall year over year, it rotated from G42 to MBZUAI, a second Abu Dhabi entity. So the diversification the bull case needs has not started; the buyer simply changed names within the same ecosystem.
2025 Revenue by Customer
RPO Recognition Schedule ($B)
The $24.6B Backlog Is Real, But Back-Loaded
The RPO of ~$24.6B is roughly 48x FY25 revenue, an unusually large book for a company this size. It is anchored by the OpenAI cloud-compute agreement, in which Cerebras runs OpenAI's workloads on its own wafer-scale chips rather than selling chips outright: roughly 750MW of capacity for three years, extendable to 2GW, worth about $27B over three years and the majority of the backlog. To join OpenAI's orbit (the FT's "Altman-osphere"), Cerebras handed OpenAI warrants for up to ~10% of the company, roughly $5B at the IPO midpoint, giving back much of the deal's profit.7 Recognition, though, is years out: OpenAI was only $16.9M of Q1 2026 revenue, and only ~15% of the backlog is scheduled to recognize within two years, with ~43% across 2028-29 and ~42% only after 2029. A separate Amazon AWS chip-supply deal will not reach full production until around next year. The backlog underwrites the FY28E anchor in our valuation, not the near-term print.
4 · Competitive Landscape
NVIDIA holds roughly 80-92% of the AI accelerator market behind the CUDA software moat, and that moat is strongest in training. Cerebras leads on raw inference speed5, but the challenger field has thinned: NVIDIA quasi-acquired Groq in late 2025, folding the LPU streaming architecture into its own low-latency inference stack and leaving SambaNova (RDU dataflow, ~$11B) as the main independent inference challenger. Its real disadvantage is not speed, it is ecosystem and scale: the credible, already-scaled non-NVIDIA silicon sits with the hyperscalers.
| Player | Approach | Position vs Cerebras |
|---|---|---|
| Cerebras | Wafer-scale WSE-3, on-chip SRAM | Fastest on inference tokens |
| NVIDIA | GPUs + CUDA ecosystem | ~80-92% share; the moat is software |
| Google TPU v7 "Ironwood" | Custom TPU, Anthropic anchor | Scaled non-NVIDIA silicon |
| AMD MI350X / MI355X | 288GB HBM3E GPUs | Credible memory-rich alternative |
| AWS Trainium | Custom silicon; also a Cerebras prefill partner | Competitor and partner |
| Broadcom-designed ASICs | Hyperscaler custom accelerators | Structural share shift away from merchant silicon |
| Groq | LPU streaming inference | Absorbed by NVIDIA (late 2025) for low-latency inference |
| SambaNova | RDU dataflow | Enterprise inference challenger (~$11B) |
The competitive read is nuanced: Cerebras wins the benchmark but faces a market where the largest buyers are increasingly designing their own silicon and where CUDA compatibility still governs most enterprise deployment decisions. As New Street's Antoine Chkaiban put it, "Cerebras is not going to replace Nvidia," though CEO Andrew Feldman counters that a workload can move from NVIDIA to Cerebras "with ten keystrokes." Speed leadership is necessary but not sufficient.
5 · Financial Health
Annual Revenue ($B)
Revenue vs Operating Loss ($B)
Revenue Trajectory
| ($M unless noted) | 2022 | 2023 | 2024 | 2025 | Q1 2026 |
|---|---|---|---|---|---|
| Revenue | 24.6 | 78.7 | 290.3 | 510 | 193.4 |
| YoY growth | — | +220% | +269% | +76% | +92% |
| Gross margin | — | — | — | 39% | 46.5% core |
| Operating income (loss) | — | — | (101.4) | (145.9) | — |
| GAAP net income | — | — | — | +871 | — |
| Adj. net income (loss) | — | — | — | (~75.7) | — |
Margins, Cash, and Guidance
- Gross margin: 39% for FY25, with Q1'26 core GM at 46.5%, but management guided Q2'26 core GM down to 36-38% as one-time pricing rolls off. Margin direction, not level, is the concern.
- Cash: ~$3.3B at end-Q1'26 (post-IPO plus a ~$1B OpenAI-linked term loan), up from $701.7M at end-2025. Net cash of roughly $2.3B after the loan.
- FY26 guidance: core revenue $855-865M, with operating margin guided to negative 28-32%.2
Material Weaknesses
Disclosed material weaknesses in internal controls over financial reporting.
Single-Source Supply
TSMC 5nm with no disclosed long-term supply agreement for the wafer.
Governance
Super-voting founder shares concentrate control with insiders.
6 · Valuation & Comps
Cerebras is rich on any basis. It trades at roughly 104x trailing FY25 sales (EV ~$52.9B on $510M), about 5x the NVIDIA multiple8; at its IPO-day close it was ~134x sales, higher than any PHLX Semiconductor index name at the time.7 Reported TTM P/E is ~160x on GAAP EPS of $1.15, a figure distorted by the one-time gain. The multiple only makes sense if you underwrite the backlog. Our house EV/Sales scenario model therefore anchors to FY2028E revenue (~$3.5B), the year the $24.6B RPO is scheduled to convert, at 14.9x, plus ~$2.3B of net cash, which reconciles to the $185 base fair value.
Price / Trailing Sales vs Peers (x)
Analyst Targets vs Our Fair Value ($)
Peer Comparison
| Company | Ticker | Price / Trailing Sales | Note |
|---|---|---|---|
| Cerebras | CBRS | ~90x | Speed leader; backlog-anchored |
| Arm | ARM | ~49x | IP licensing, premium multiple |
| Broadcom | AVGO | ~31x | Hyperscaler ASIC design |
| NVIDIA | NVDA | ~27x | Market leader, CUDA moat |
| AMD | AMD | ~19x | Merchant GPU challenger |
6b · EV/Sales Scenario Model
Cerebras is loss-making with negative operating cash flow, so a DCF would rest almost entirely on terminal-value assumptions. Instead we apply an EV/Sales multiple to FY2028E revenue, the backlog-conversion anchor year, add net cash, and divide by shares: (FY28E revenue × EV/Sales) + net cash, ÷ ~294M shares. The EV/Sales multiple already encodes the forward discount, so no separate WACC is layered on outside Reverse mode.
Model Inputs
Implied value roughly in line with the $187.53 price; the base scenario is about fairly valued.
Scenario Assumptions
| Scenario | FY28E Rev | EV/Sales | Net cash | Implied $ |
|---|---|---|---|---|
| Bear | $2.5B | 12.0x | $2.3B | $110 |
| Base | $3.5B | 14.9x | $2.3B | $185 |
| Bull | $4.5B | 19.1x | $2.3B | $300 |
| Blended FV (30/45/25) | — | — | — | $201 |
Sensitivity: EV/Sales × FY28E Revenue
Base-case sensitivity grid.
7 · Scenario Targets
Scenario Fair Value vs Current
| Scenario | Probability | Fair Value | EV/FY28E Sales | Driver |
|---|---|---|---|---|
| Bull | 30% | $300 | ~19x | OpenAI ramps; UAE share <40% by 2027; GM 45%+; FY28E ~$4.5B |
| Base | 45% | $185 | ~15x | Back-loaded ramp; UAE ~50-60%; GM high-30s; FY28E ~$3.5B |
| Bear | 25% | $110 | ~12x | Margin compression; OpenAI slips; controls tighten; FY28E ~$2.5B |
| Blended FV (prob-weighted) | $201 | — | +7% vs $187.53 | |
The 30 / 45 / 25 weighting reflects a fat but low-probability bull tail (a clean OpenAI ramp with genuine diversification) against real, near-term concentration and margin downside. Blended fair value of ~$201 sits only modestly above spot, and still well below the ~$292 Street average, consistent with the Hold framing.
8 · Risk Factors
Customer / Related-Party Concentration
~86% of 2025 revenue from MBZUAI (~62%) and G42 (~24%); MBZUAI ~77.9% of year-end A/R. Concentration rotated, it did not fall.
Margin Compression + Cash Burn
Q2'26 core GM guided down to 36-38%; FY26 operating margin guided negative 28-32%. Adjusted FY25 net loss ~$76M.
CFIUS / Export-Control Overhang
The UAE entities carry national-security exposure. Tightened controls would convert concentration into a revenue cliff.
Back-Loaded, OpenAI-Dependent Backlog
$24.6B RPO is majority-OpenAI, with only $16.9M recognized in Q1'26 and ~42% deferred past 2029.
Single-Supplier TSMC
Sole-sourced 5nm wafers with no disclosed long-term supply agreement.
NVIDIA / CUDA Moat
~80-92% market share behind a software ecosystem that governs most enterprise deployment.
Controls & Governance
Disclosed material weaknesses in internal controls and super-voting founder shares.
9 · Technicals
Price History (IPO → Today)
52-Week Range
Key Levels
With only ~10 weeks of public history, "technicals" here are context, not a tradable structure. The stock priced at $185, popped ~68% to a ~$311 day-1 close, spiked to $386 in the first week, tumbled to $160.81 after its first post-IPO earnings, and has round-tripped to the offer.69
| Level | Type | Note |
|---|---|---|
| $386.34 | High | First-week spike (May 2026) |
| $300 | Bull FV | Clean OpenAI ramp + diversification |
| $187.53 | Spot | Current, back near the offer |
| $185 | IPO / Base FV | Offer price and base fair value cluster |
| $160.81 | 52-wk low | Post-IPO trough |
| $150 | Stop | Tactical stop reference |
| $110 | Bear FV | Margin compression + concentration |
★ Interactive Tools
Price Target Calculator
UAE share below ~45% expands the multiple (concentration risk eases); above ~90% docks it.
Risk / Reward Calculator
Weak R/R, below 1.5.
Sources & Citations
All data verified as of July 27, 2026. Superscripted numbers in the body link to the matching entry below; the ↩ at the end of each entry returns to the citation point. Figures that cannot be tied to a primary filing are labeled illustrative in the prose.
- Cerebras Form S-1 / 424B4 prospectus, SEC EDGAR (April–May 2026) · RPO / backlog, UAE related-party revenue, CFIUS history, one-time G42 gain, GAAP net income, share structure. ↩ ↩ ↩ ↩ ↩ ↩
- Cerebras Q1 2026 results & IPO-pricing press releases · Q1 2026 revenue $193.4M (+92% YoY), FY26 core-revenue guidance $855–865M, operating-margin guide. ↩ ↩
- Cerebras WSE-3 / CS-3 product page · 4 trillion transistors, 900,000 cores, 44GB on-chip SRAM, 125 PFLOPS, wafer-scale specifications. ↩
- Cerebras Inference · On-wafer memory bandwidth (~7,000× an H100, vendor figure) and inference-cloud positioning. ↩
- Artificial Analysis inference speed leaderboards · Independent corroboration of output speed (>2,200 tok/s Llama 3.3 70B; 969 tok/s Llama 3.1 405B). ↩ ↩
- CNBC, Cerebras IPO coverage · IPO priced $185, popped ~68% to a ~$311 day-1 close, first-week high $386, tumble after the first post-IPO earnings, subsequent round-trip toward the offer. ↩
- Financial Times, "Chipmaker Cerebras joins OpenAI's inner circle, for a price" (John Foley, May 12, 2026) · OpenAI cloud-compute agreement, ~10% warrants and the "Altman-osphere," ~134x IPO-close sales, competitive framing. ↩ ↩
- The Wall Street Journal, Heard on the Street, "The Blockbuster Cerebras IPO Is a Huge Bet on Nvidia Fatigue" (Dan Gallagher, May 15, 2026) · Valuation multiple versus NVIDIA, IPO framing. ↩
- WSJ / MarketWatch CBRS quote and Barron's earnings coverage · Real-time price, market cap, 52-week range, analyst consensus (Buy; average $292, median $300, range $209–340), and "Cerebras Stock Tumbles After First Earnings Report Since IPO." ↩ ↩
See the Important Disclaimers in the footer for the full not-investment-advice notice.