Cerebras Systems CBRS
Technology · Semiconductors · Synthos Deep Dive · 2026-07-20
The Overview
Cerebras builds AI chips the size of a dinner plate — an entire silicon wafer as one chip, instead of cutting it into hundreds of small ones the way Nvidia does. Keeping everything on one huge chip means the memory sits right next to the compute, so when an AI model answers you ("inference"), it can run 15–20x faster than a GPU — that's the company's claim, and independent benchmarks do show it running open models about 3–6x faster than its closest specialist rival, Groq.
The business case is unusually concrete for a young company: customers have already signed contracts worth $25 billion — that's a real number from the company's own SEC filing, not a press-release flourish — mostly one giant deal in which OpenAI will pay over $20 billion through 2028 to run its models on Cerebras machines. Against that, Cerebras sold only about $600 million of stuff in the last twelve months. If the contracts convert to revenue on schedule, revenue roughly triples in 2027 and triples again in 2028.
So why aren't we buyers today? Three reasons, in order:
- The price already assumes success. At ~$52 billion (counting all shares that will exist), you're paying about 5x the sales the company hopes to have in 2028. If everything goes right you make maybe +40%; if the plan slips you lose a third; if it breaks you lose 80%. That's a lopsided bet — against you.
- The customer list is dangerously short. About 86% of current revenue comes from just two customers in the United Arab Emirates, and the $25B backlog is mostly one customer (OpenAI). If any one of them renegotiates, the story changes overnight.
- A flood of insider shares is coming. Early investors and employees were barred from selling after the May IPO — but that ban starts lifting in stages from August through mid-November 2026, freeing roughly 5–6 times as many shares as currently trade. More sellers than buyers usually means lower prices, and the stock has already fallen 45% from its first-day close.
Here's what our three scores mean in everyday terms:
- Downside Risk 8/10 (high). Lots of cash in the bank, but everything else — customers, contracts, share supply, valuation — is concentrated and fragile.
- Growth Quality 6/10 (good, not clean). Growth is explosive and margins are improving fast, but it comes from very few customers and requires burning billions on data centers.
- Exponential Potential 9/10 (very high). This is one of the few companies whose "exponential" is written into signed contracts, not just a story.
The one big worry: the backlog and the concentration are the same thing. The $25 billion that makes the bull case is mostly one customer's signature. Our stance: don't buy at $172 — put it on the watchlist and let the lockup flood do its work. Below roughly $110 this becomes the genuinely asymmetric bet the thesis wants; below $70 it would be exceptional.
Putting a number on it: our fair-value estimate is $115 against a current price of $172.51 — consistent with our call to stay away or wait for a better setup.
Our summary metrics
What does “fair value” mean?
Fair value is Synthos’s estimate of what one share is worth today, based on our model of the company’s future cash generation and the risks to it. It is not a price target or a prediction of where the stock trades next quarter — it’s the price where we think risk and reward are balanced. Above it, you’re paying for outcomes better than our base case; below it, the market is offering a margin of safety.
The model’s inputs and weightings are proprietary — they’re the product. What we publish is the output, We don’t publish a reverse-DCF cross-check for pre-profit companies — negative or missing earnings break that math — so take this number on our modeling alone.
What do the 5 tiers mean? (Core · Tactical · Watch · Hold · Avoid)
The Road Ahead
What we expect to matter in each window, and the evidence that would prove us wrong.
Short term 0–6 months
No differentiated view- Driver
- A phased lockup structure releases share supply equal to roughly 5-6x the 30M-share IPO float between August and mid-November 2026, into a stock already 45% below its day-one close, 2% off its all-time low, RSI 33, and below its $185 IPO price. Supply overwhelms narrative in this window.
- What we’re watching
- Q2 print (~Aug 11) — OpenAI revenue actually appearing in the mix, RPO holding at/above $25B, gross margin holding ≥40%; then insider behavior as each lockup tranche opens (Benchmark already converted and distributed 2.16M Class B shares on 2026-06-29).
- Confidence
- Medium
Medium term 6–24 months
Tailwind- Driver
- The contracted OpenAI ramp (up to 750MW through 2028, >$20B) plus the 10x manufacturing scale-up converts backlog into reported revenue — consensus has 2027 revenue tripling to ~$2.8B and 2028 reaching ~$7.2B with positive EPS. If delivery tracks, the stock's problem shifts from supply to valuation, a much better problem.
- What we’re watching
- Backlog conversion rate (RPO drawdown vs revenue recognized), diversification beyond UAE + OpenAI, gross margin under Nvidia/Groq price pressure, and whether the 10x manufacturing claim (Feldman's own test) is met.
- Confidence
- Medium
Long term 2+ years
Tailwind- Driver
- Inference is the growth segment of AI compute and wafer-scale memory-next-to-compute is a real architectural edge for latency-sensitive, reasoning-heavy workloads — if Cerebras holds a 3-6x speed lead while Moore-rate doubling slows, it earns a durable specialist franchise beside Nvidia rather than under it.
- What we’re watching
- Nvidia's Rubin-generation response (reportedly integrating Groq-style scheduling — secondhand, unverified), hyperscaler ASICs commoditizing inference, and the 2025-vintage architecture critique (models outgrowing on-chip memory) reasserting itself at the frontier.
- Confidence
- Low
Exponential Potential
What could take this further than the base case — and where the market may be underpricing it. Full forward-growth and acceleration math in Deeper analysis, §4 below.
Deeper analysis
Technicals, fundamentals, valuation, and the full expert-claim evidence panel — the detail behind the numbers above.
Reference table
| Street consensus | $297.25 (high $325 / low $273; 5 Buy · 0 Hold · 0 Sell) — 5 analysts, weeks after they underwrote the IPO; we discount this lane heavily |
| Valuation | EV/S ~62x TTM (~$604M) · ~44x 2026E ($861M) · ~13.6x 2027E ($2.77B) · ~5.3x 2028E ($7.18B) — the multiple is a bet on backlog conversion, nothing else |
| The backlog (the whole story) | $25.0B remaining performance obligations — company-disclosed in the Q1 10-Q, of which >$20B is one counterparty: the OpenAI agreement (up to 750MW through 2028) |
| Technicals | Broken parabola — IPO'd 2026-05-14 at $185, opened $350, day-one close $311; now $172.51, −45% from day-one close, 2% above the all-time low ($168.52 EOD), below the IPO price, RSI 33, MACD −13.8, no 50/200-DMA (only ~45 trading days) |
| Conviction | Low–Moderate — 9 entity-tagged KB claims / 3 thinkers, net conviction +48, but the bull lane is dominated by the CEO's own relays (half-weighted); independent voices are split |
| Position sizing | 0% today. In the $80–110 asymmetric zone: satellite 1–2% max (concentration + lockup risk caps size even there) |
What the experts actually said 1 traceable claims on CBRS · showing the highest-conviction voices
“Cerebras's dinner-plate-sized wafer chip puts fast memory next to compute; when OpenAI uses them they run 15-18x faster than a GPU, which real-time AI demands.”
Every claim reconciles to a real claim_id in the Synthos knowledge base — this is the evidence the verdict is built on, not vibes. Management (the company itself) is shown but half-weighted; one cautionary voice is included on purpose.
Price & moving averages 12 months · 50 & 200-day averages · 52-week range
Solid line = price · dashed line = 50-day average · dotted line = 200-day average · the two thin horizontal lines mark the 52-week high and low. Price above both averages is an uptrend.
Bollinger Bands 20-day average ± 2 standard deviations
The shaded band widens when the stock gets more volatile. Riding the upper edge = strong momentum (sometimes stretched); the lower edge = weak / potentially oversold.
Data summary: price $172.73 is currently inside the band (band $158–$234).
RSI (14) momentum gauge · 0–100
Above 70 (overbought zone, shaded) = overbought, below 30 (oversold zone, shaded) = oversold. Currently 38.
MACD 12 / 26 / 9 · trend & momentum
The MACD line crossing above the signal line (bars flip to the up color) = momentum turning up; crossing below (bars flip to the down color) = turning down. Bar height = the size of that gap.
Data summary: MACD is currently below its signal line by 0.53, negative momentum.
Relative performance vs S&P 500 & its sector (XLK (sector)), set to 100 a year ago
Solid = CBRS · dashed = S&P 500 · dotted = XLK (sector). A rising line means it is beating that benchmark — the sector line shows whether it is a leader or laggard within its own group.
Forward revenue & earnings actual → estimate · "FY" = fiscal year, "E" = estimate
Darker bars = actual results, brighter = analyst estimates. Taller bars to the right = expected growth.
Key stats an RIA wants
1. What it is
Cerebras Systems (Nasdaq: CBRS, IPO 2026-05-14) designs the Wafer-Scale Engine (WSE) — a single AI processor built from an entire 300mm silicon wafer, currently in its third generation (WSE-3), sold inside rack-scale CS systems and, increasingly, consumed as a cloud inference service from Cerebras-operated data centers. The architectural bet: keeping enormous SRAM directly on-die next to compute removes the memory-bandwidth bottleneck that throttles GPU inference, yielding token-generation speeds the company claims at 15–20x GPUs and independent testing pegs at ~2,100 tokens/sec on Llama-class 70B models versus ~750 for Groq's LPU (verticalapi comparison). Founded 2015, Sunnyvale CA, ~784 employees, founder-CEO Andrew Feldman (previously sold SeaMicro to AMD). Dual-class share structure (Class B super-voting held by insiders/VCs).
Revenue mix (from filings and press summaries of the Q1 2026 10-Q):
- By offering: hardware systems sales plus a fast-growing services/cloud line (the OpenAI agreement is structured as cloud services — Cerebras builds or leases the data centers and OpenAI pays for inference capacity, per DCD). FMP carries no segment tables yet (recent IPO; seg_prod/seg_geo empty).
- By customer (the number that matters): press summaries of the Q1 10-Q put ~86% of Q1 2026 revenue with two UAE-linked customers — ~62% MBZUAI (Mohamed bin Zayed University of AI) and ~24% G42 (TECHi). This is the successor to the pre-IPO disclosure that G42 alone was ~87% of H1-2024 revenue (CNBC). OpenAI revenue is contracted but barely flowing yet. One oddity worth naming: reports note Cerebras has been renting capacity back from G42 to meet deployment schedules — a related-party circularity we flag rather than fully evaluate.
The structural story: inference — not training — is becoming the dominant AI-compute workload; latency and tokens-per-second are becoming product features (reasoning models "think" in tokens); and a specialist architecture with a real speed lead can carve a durable franchise beside Nvidia. The counter-story, from our own KB's bear lane, is that first-wave AI-hardware architectures already got burned once betting on on-chip memory when models outgrew it.
2. The expert thesis — why the panel leans bullish (traceable, with a big caveat)
KB coverage: 9 entity-tagged claims across 3 thinkers (all_in ×4, no_priors ×4, latent_space ×1), net conviction ~+48, plus 5 mention-lane claims (dylan_patel, jensen_huang, compound_and_friends, bill_gurley, ml_street_talk) where Cerebras appears in the thesis text without an entity tag. The critical caveat: 5 of the 9 entity-tagged claims record CEO Andrew Feldman as the speaker — podcast relays of the vendor's own numbers. House rules half-weight the management lane; we do so explicitly below.
Bull lane (vendor-relayed — Feldman as speaker, half-weight):
- all_in, 2026-07-10, bullish/80: "Cerebras has a $25 billion backlog; AI-compute demand is already booked and way outstrips the ability to build data centers and fill them with hardware." (This one is corroborated: the $25.0B RPO figure appears in the company's Q1 10-Q — see §5.)
- all_in, 2026-07-10, bullish/75: "Cerebras broke Moore's law on inference; over the next 18 months its performance will be way over 2x, versus GPUs' traditional ~18-month doubling." (Falsifiable test on the clock: >2x by ~January 2028.)
- no_priors, 2026-05-21, bullish/80: "Cerebras builds the fastest AI inference computers — 15, 18, 20x faster than GPUs across the board: big and small, US and Chinese, 1B to trillion-parameter models."
- no_priors, 2026-05-21, bullish/65: "Cerebras will try to increase manufacturing 10x this year — about as fast as anybody in the history of hardware — against a $20B+ backlog." (Second falsifiable test: 10x manufacturing in 2026.)
- all_in, 2026-06-06, bullish/70: wafer-scale puts fast memory next to compute; "when OpenAI uses them they run 15-18x faster than a GPU, which real-time AI demands."
Bull lane (independent):
- all_in, 2026-01-17, bullish/65 (fact relay): Cerebras "won a $10B+ OpenAI compute deal and [is] heading to IPO near a $22B valuation." (Deal verified by CNBC/Bloomberg/TechCrunch; note the IPO ultimately priced at ~$39.8B basic / $56.4B fully diluted — nearly triple the valuation the panel heard in January.)
- latent_space, 2025-07-08, bullish/75: specialized accelerators (Cerebras, SambaNova, Groq) push token-output speed far beyond GPU baselines.
- Mention lane, bill_gurley pod (Sunny Madra, 2024-10-13, bearish-NVDA/60): Nvidia's inference moat is weaker than its training moat — CUDA isn't required for inference and "Groq/Cerebras/SambaNova already lead on performance and price."
Bear / skeptic lane (independent — this is what keeps conviction at Low–Moderate):
- no_priors, 2025-08-14, bearish/75: "First-wave AI-hardware firms bet on more on-chip memory/low off-chip bandwidth; models grew too big to fit, invalidating the architecture bet." (Note the tension: the same show hosted Feldman bullishly nine months later — the 2026 inference turnaround, where speed matters more than fitting frontier training runs, is the rebuttal. But the critique isn't dead: if frontier models keep growing, the constraint can reassert.)
- no_priors, 2024-09-12, neutral/58: the new-chip wave's hard problem is "committing years ahead to an architecture while matching Nvidia/AMD price-performance."
- dylan_patel, 2026-02-03, bullish/55 (the caveat is the claim): "Cerebras makes sense for price-insensitive customers who'll pay 10x to complete 10x faster; but the OpenAI deal is only 750MW — modest capacity, and product exposure to users is unsolved." (The sharpest single sentence in the file: the best-informed semis analyst in our KB sizes the flagship deal as a niche, not a platform.)
- jensen_huang, 2026-05-21, bullish-NVDA/80 (management lane, interested party): "competitors like AMD and Cerebras will have to work very hard to catch us."
- compound_and_friends (Kai Wu), 2026-05-15, neutral/52: accelerated computing is historically cyclical; Nvidia's competition (including Cerebras) makes earnings sustainability the real question.
Net read: the panel's genuinely independent signal is moderately bullish on the technology and pointedly unresolved on the business. The loudest numbers in the bull case ($25B backlog, 15-20x, 10x manufacturing) all originate from one voice — the CEO's — even when they arrive through three different podcasts. One of those numbers (the backlog) is now SEC-corroborated, which upgrades it from vendor claim to disclosed fact; the performance and manufacturing claims remain vendor claims with clocks attached.
3. Synthos scores & the Bull / Base / Bear cases
| Score | 0–10 | The read |
|---|---|---|
| Downside Risk (lower = safer) | 8 · High | The cushion: ~$1.1B cash+investments at year-end 2025 plus ~$5.5B gross IPO proceeds (May 2026) — call it ~$6B+, no conventional debt. Everything else is tail risk: ~86% of revenue from two UAE customers; a $25B backlog that is >80% one counterparty; ~$2.3B of non-cancelable data-center leases; FY25 FCF −$393M and going more negative as the 750MW buildout accelerates; a phased lockup releasing ~5-6x the float by mid-November; dual-class governance; and a stock with 45 trading days of history, down 45% from its first close. |
| Growth Quality | 6 · Good, not clean | Revenue $24.6M (FY22) → $78.7M (FY23) → $290.3M (FY24) → $509.99M (FY25, +76%); Q1 2026 $193.4M, +94% y/y; gross margin 39.0% (FY25) → 44.6% (Q1 2026); Q1 operating loss narrowed to −$15.0M (net −$14.0M; adjusted EPS −$0.04 vs −$0.16 expected). Real, accelerating, margin-expanding. Marked down for concentration, prepayment-funded working capital ($485M current deferred revenue), and the capital intensity of the cloud pivot (FY25 capex $383M, 21.9% of revenue). |
| Exponential Potential | 9 · Very High | $25.0B disclosed RPO vs ~$604M TTM revenue (~41x) — the largest backlog-to-revenue ratio we have scored. Consensus: $861M (2026E) → $2.77B (2027E, +222%) → $7.18B (2028E, +159%), EPS turning positive 2027E ($0.96) → $4.68 (2028E). Plus two falsifiable acceleration claims (>2x perf/18mo; 10x manufacturing/2026). Held below 10 because the exponential has a single point of failure (OpenAI) and thin, underwriter-adjacent estimate coverage (5–9 analysts). |
The three cases (our own scenario model — assumptions stated; each target is a ~12–18-month fair value on ~290M fully-diluted shares; we deliberately attach no probabilities). FV method note: this is a scenario-based backlog-conversion × margin × multiple model, labeled honestly as NOT DCF-defensible — a two-month-old company with one dominant contract does not support a defensible DCF; anyone who shows you one is decorating.
| Case | Key assumptions | Fair value |
|---|---|---|
| Bull | OpenAI ramp lands on schedule; 2028 revenue ~$7.5B (consensus $7.18B), gross margin ~45%, net margin scaling toward mid-teens (~$1.0-1.2B NI, consensus 2028 EPS $4.68); market pays ~8-9x forward sales / ~50x EPS for a still-hypergrowth specialist → EV ~$65-70B. | ~$240 (+39%) |
| Base (our anchor) | Conversion slips ~25-30% (data-center buildouts always slip): 2027 revenue ~$2.2B, 2028 ~$5B; margin gains partly competed away by Nvidia/Groq pricing (GM ~40%); 6-7x 2028E sales → EV ~$30-35B + ~$4B net cash. | ~$115 (−33%) |
| Bear | OpenAI second-sources or renegotiates as Rubin-generation parts close the speed gap; UAE anchor wobbles; revenue plateaus at a ~$1.5-2B run-rate against $2.3B of lease obligations and heavy burn; 3-4x sales → EV ~$6-9B + remaining cash. | ~$35 (−80%) |
Synthos fair value = the base case, ~$115 (−33% from $172.51), full range $35–$240. Read that carefully: at today's price the base case is negative and the bull case (+39%) is smaller in magnitude than the base-case loss. We sit ~61% below the Street's $297.25 — but that consensus is five analysts, all Buy, publishing weeks after their banks priced the IPO; the underwriter-coverage lane gets the same skepticism the vendor lane does. This is a tracked call — the Forecaster Scorecard grades it once it matures.
The asymmetry map (the founder's question, answered in bands):
| Entry zone | Approx. FD valuation | Payoff at our scenarios (bull / base / bear) | Verdict |
|---|---|---|---|
| >$150 (incl. today's $172.51) | >$44B | +39% to +60% / negative / ~−80% | Not asymmetric — skew is against you. Avoid new money. |
| $110–150 | $32–44B | ~+60-120% / ~flat / −68-77% | Symmetric at best; watch, don't chase. |
| $80–110 | $23–32B (≈ ~8-12x 2027E sales; at $95 the FD cap ~$27B sits below the $22B-heading-to-IPO panel figure adjusted for the $5.5B raise) | +120-200% / +5-45% / −55-68% | The genuinely asymmetric zone — IF backlog and OpenAI terms remain intact. Starter buys earn their risk here. |
| <$70 | <$20B (below the pre-IPO private valuation; <1x disclosed RPO) | +240%+ / +65% / ~−50% | Table-pound asymmetric absent a thesis-breaking event — a price this low would itself imply the market believes the backlog broke, so verify the kill conditions (§10) before acting. |
Honesty note on the bands: the bear case is deep (−50% or worse) at every entry price, because the bear case is the thesis breaking, not a multiple compressing. The asymmetry you buy at lower prices is bull-magnitude versus entry — you can never buy this name "safely," only cheaply enough that being right pays for the times you're wrong.
4. Exponential Potential
Synthos separates compounders from exponentials. Cerebras is the clearest exponential candidate in our coverage — with a single point of failure:
- The booked curve: $25.0B RPO against ~$604M TTM revenue. Consensus converts this to +69% (2026E), +222% (2027E), +159% (2028E) — a ~189% two-year forward CAGR. No other name we cover has the exponential contractually pre-sold.
- Acceleration (2nd derivative) is positive: quarterly revenue $99.5M → $103.3M → $135.7M → $171.4M → $193.4M over five quarters, with gross margin rising through the ramp (39% → 44.6%) — accelerating growth with improving unit economics, the exact profile the flagship philosophy hunts.
- Performance compounding claim: Feldman's ">2x every 18 months vs GPUs' ~2x" (all_in, 2026-07-10, vendor claim, testable by early 2028). If true, the speed moat widens mechanically. WSE-4 is expected to lean into low-precision (FP4/FP8) throughput per trade-press roadmaps (hashrateindex).
- Room to run: inference is the growing share of AI compute, and Dylan Patel's own framing — customers who'll "pay 10x to complete 10x faster" — describes a genuine premium niche even in the skeptical reading. The bull re-rating comes if that niche turns out to be the whole reasoning-model market.
- The single point of failure: ~$20B+ of the $25B curve is one signature (OpenAI, through 2028). This is an exponential on a leash.
Exponential Potential: Very High (9/10). The reason this dive exists — and the reason the price bands in §3, not enthusiasm, must govern entry.
5. Financials (real numbers — FMP statements + Q1 2026 10-Q; recent-IPO data-quality flags at bottom)
- Revenue: FY25 $509.99M, +75.7% (FY24 $290.3M, FY23 $78.7M, FY22 $24.6M). Q1 2026 (Mar): $193.4M total / $191.3M core, +94% y/y, beating the ~$181M estimate. TTM ≈ $604M.
- Quarterly ramp: Q1'25 $99.5M → Q2 $103.3M → Q3 $135.7M → Q4 $171.4M → Q1'26 $193.4M — five consecutive sequential increases through the IPO window.
- Margins: FY25 gross margin 39.0%; Q1 2026 44.6% (headline ~46.5% incl. non-core, per earnings-call summaries). FY25 operating loss −$145.3M; Q1 2026 operating loss just −$15.0M — approaching operating breakeven far ahead of typical hardware-IPO vintage.
- Reported net income needs adjustment: FY25 GAAP shows +$237.8M net income, but this includes ~$390M of non-operating gains (preferred/warrant remeasurement around the IPO). The operating loss is the truer FY25 read. Q1 2026: net −$14.0M, GAAP EPS −$0.22, adjusted EPS −$0.04.
- The backlog (company-disclosed, not vendor-relayed): $25.0 billion of remaining performance obligations per the Q1 2026 10-Q (StockTitan filing summary), overwhelmingly the OpenAI Master Relationship Agreement (>$20B). Deferred revenue $485.5M current + $35.8M non-current at FY25 — customers pay ahead.
- Cash flow (the cost of the pivot): FY25 operating CF −$10.1M, capex −$382.7M (21.9% of revenue), FCF −$392.8M. FY24's +$429M FCF was prepayment-driven (G42), not operational — income quality is volatile by construction. Non-cancelable data-center leases ~$2.3B (Q1 10-Q, per press summaries) sit on top of capex.
- Balance sheet: FY25-end cash + short-term investments $1.108B, no conventional debt ($262M of lease obligations); plus ~$5.55B gross IPO proceeds (May 2026) → pro-forma ~$6B+ liquidity. GAAP equity was negative (−$579M) at FY25-end due to redeemable preferred, which converted at IPO — post-IPO equity is substantially positive.
- Concentration: ~62% MBZUAI + ~24% G42 of Q1 2026 revenue (press summaries of the 10-Q); pre-IPO, G42 was ~87% of H1-2024 revenue. The OpenAI ramp is the diversification plan — which replaces UAE concentration with OpenAI concentration.
Data-quality flags (recent IPO, FMP feed): quarterly opex line-items are zeroed/merged in several 2024-25 quarters; share-count fields are unreliable (215M vs 62.8M weighted in adjacent rows — we use market-cap-implied ~227M basic and ~305M fully diluted from IPO prospectus math); FMP's estimate rows show internally inconsistent EBITDA (deeply negative alongside positive net income — discarded); 50/200-DMA fields are null/artifactual (insufficient history). All flagged figures are excluded from our math.
6. Valuation — priced in or room?
At $172.51 the market is not paying for what Cerebras is — it is paying for roughly half of what Cerebras promises to become by 2028:
- EV ~$37.7B on the basic count (FMP TTM); on the fully-diluted ~305M shares, FD cap ~$52.6B. EV/S: ~62x TTM → ~44x 2026E → ~13.6x 2027E → ~5.3x 2028E. Every multiple before 2028 is decorative; the stock only "works" if the 2027-28 consensus ramp lands.
- Reference points: Nvidia in its hypergrowth years rarely sustained >25-30x forward sales; Astera Labs (FMP's nearest real peer, $54B cap) trades rich on delivered hyperscaler revenue. Cerebras at ~44x current-year sales with 86% two-customer concentration is priced like the conversion is done.
- The IPO tape is the tell: priced $185 (above two raised ranges: $115-125 → $150-160 → $185), opened $350, touched $385, closed day one $311.07, and has since fallen to $172.51 — below the IPO price and −45% from the first close, in nine weeks (IPOScoop, CNBC). The market has been repricing the promise downward continuously since the first print.
- Street targets (heavily discounted lane): consensus $297.25 (high $325 / low $273), 5 Buy, 0 Hold, 0 Sell — a unanimous, tightly-clustered, post-IPO underwriter cohort. FMP's own composite rating is a C+ (DCF and P/E screens score 1-2/5). We treat the $297 the way we treat management guidance: an interested party's number, stated for the record.
- FMP's PEG of 0.25x is the one screen that captures the bull case: if the growth is real, the multiple compresses explosively (5.3x 2028E sales, ~37x 2028E EPS of $4.68). That is a fine 2028 entry multiple — it is just being charged in 2026, before delivery, concentration-adjusted.
Bottom line: at spot, fully priced for success; the valuation only becomes an asymmetric offer in the $80–110 band (~8-12x 2027E sales), which the lockup calendar may well deliver.
7. Technicals (from the tech block; 45 trading days of history — read with humility)
- Trend: there isn't one yet — no 50-DMA or 200-DMA exists (nulls in the feed; the single populated average is an artifact). What exists is a nine-week waterfall: $350 open → $311 first close → $172.51.
- Location: −44.7% from the EOD peak ($311.07; intraday high $385-386) and +2.1% above the all-time EOD low ($168.52) — the stock is sitting ON its low, below the $185 IPO price. There is no technical support below; the only references underneath are round numbers and our valuation bands.
- Momentum: RSI(14) 33 — oversold-adjacent but not washed out; MACD −13.8, deeply negative and not yet curling. Sellers remain in control.
- Relative strength: n/a (no 3/6/12-month returns exist). Context: SPY +5.0%/3mo, QQQ +8.0%/3mo while CBRS fell ~45% — the underperformance is total.
- The house parabola lens: our attention-parabola work (43% of parabolas round-trip; NVTS demonstrated it live) exists for exactly this chart. CBRS is a post-IPO attention parabola — priced above two raised ranges on AI-tsunami sentiment, +89% intraday at the open, then a 45% nine-week bleed. The round-trip below the IPO price has already happened from the open/close prints; the remaining question is whether the lockup wave extends it into our $80-110 band. Parabola history says: do not catch this knife on chart grounds; let supply exhaust itself (main unlock ~November), then buy the valuation band, not the bounce.
8. Moat & competitive position
The moat is real, narrow, and time-boxed:
1. Architecture lead (the core): wafer-scale SRAM-next-to-compute delivers measured ~2,100 tok/s on 70B-class open models vs ~750 for Groq and far less for GPU serving — independent testing puts Cerebras at 3-6x Groq on identical models (verticalapi). For reasoning models that "think" in tokens, speed is product quality — this is why OpenAI signed.
2. A decade of process learning: yield-managing a wafer-sized die is a manufacturing art competitors cannot copy quickly; Cerebras has been at it since 2015 (The Register).
3. Contracted demand: $25B RPO is itself a moat of sorts — capacity pre-sold through 2028 funds the roadmap.
Against it, the erosion vectors are serious:
- Nvidia's response: Rubin-generation parts (late 2026) are expected to adopt on-chip-memory features; trade press reports Nvidia acquiring Groq (~$20B reported) and integrating its deterministic scheduling into the Rubin platform, announced March 2026 (Silicon Report, hashrateindex) — secondhand trade-press claims we could not verify against primary filings; label accordingly, but Jensen's own KB quote ("competitors like AMD and Cerebras will have to work very hard to catch us") confirms Cerebras is now on Nvidia's named-threat list.
- Price competition is live: on Llama 3.3 70B, Groq lists ~$0.59/$0.79 per 1M input/output tokens vs Cerebras ~$0.85/$1.20 — Cerebras charges a per-token premium justified only by speed (digitalapplied pricing matrix). Hyperscaler ASICs (TPU, Trainium) compress the floor from below.
- The architecture critique (KB bear lane): no_priors' 2025 claim — models outgrew on-chip memory once already — is the structural version of this risk. The 2026 inference boom rescued the bet; a jump in frontier-model working-set size could un-rescue it.
- Dylan Patel's ceiling: if wafer-scale only "makes sense for price-insensitive customers," the moat guards a lucrative niche, not a platform — and 750MW is niche-sized against multi-gigawatt GPU buildouts.
Peer set: no clean public comp. Astera Labs (ALAB, $54B) for AI-infrastructure multiple context; Groq (now reportedly inside Nvidia) was the direct specialist rival; SambaNova/Tenstorrent private; Nvidia the ecosystem incumbent. FMP's listed "peers" (Fervo, X-Energy, Voyager…) are IPO-vintage cohort noise, not comps — ignored.
9. Management, capital allocation & guidance
- Founder-CEO Andrew Feldman — serial founder (SeaMicro → AMD), a decade on one contrarian bet, and demonstrably the company's chief evangelist: five of our nine KB claims are him on podcasts. The claims are vivid, numerical, and — to his credit — falsifiable (>2x/18mo performance; 10x manufacturing in 2026). We will grade both.
- Capital allocation: the company just pivoted from selling boxes to operating AI cloud capacity — FY25 capex $383M (up 16x y/y), ~$2.3B of non-cancelable DC leases, and ~$5.5B of IPO proceeds largely earmarked for the OpenAI buildout. This is a leveraged operational bet: brilliant if utilization follows the contract, brutal if the contract wobbles (the leases don't cancel).
- Governance/alignment flags: dual-class structure (insider Class B supervoting). Insider tape post-IPO: COO Dhiraj Mallick sold 10,000 shares at $206.51 on 2026-06-30 (small, likely programmatic); Benchmark Capital converted and distributed 2.16M Class B shares on 2026-06-29 — consistent with early lockup-release mechanics already operating (see §10), and worth watching as a template for what the VC cap table does at each gate.
- Guidance (self-interested — half-weight): first public quarter beat (rev $191.3M core vs ~$181M est; adj EPS −$0.04 vs −$0.16). Management frames the backlog as demand "already booked" outstripping buildout capacity. Fine — but note what management does not control: OpenAI's own funding cadence, and UAE counterparties' budgets. The guide inherits its customers' balance sheets.
10. Catalysts & what to watch
- ~2026-08-11: Q2 2026 earnings (FMP also lists a 9/22 date — calendar noise; treat 8/11 as the print). Watch: first material OpenAI revenue in mix; RPO held at/above $25.0B; gross margin ≥40%; concentration percentages moving down.
- THE LOCKUP WAVE (the dominant near-term supply event — verified, and it is NOT a single cliff): the S-1 ties release to the earlier of two trading days after the Q3-2026 (September-quarter) report or ~180 days post-prospectus (~2026-11-10), with early-release tranches around the earnings windows before that — secondary analyses estimate 60M+ shares freeing around the Q2 print and ~171M shares (≈5-6x the 30M-share IPO float) unrestricted at the main release; one social-media breakdown claims ~84M Class B eligible by end-August and ~87M more in Sept-Oct (Seeking Alpha "IPO Lockup Comes Fast", QuantAbundancia, AInvest; tranche share-counts are secondary-source estimates, not verified against the S-1 by us). Benchmark's June 29 conversion/distribution suggests the mechanism is already live. For the asymmetry thesis this is the gift: maximum share supply into a stock at its lows is precisely how a quality name reaches an irrational price. The buy zone likely opens between August and November.
- The two Feldman clocks: 10x manufacturing by end-2026; >2x inference performance by ~January 2028. Both graded claims in our scorecard.
- OpenAI tranche deliveries: capacity comes online "in multiple tranches through 2028" (DCD) — each confirmed tranche de-risks the RPO; each slip is a tell.
- Nvidia Rubin / Rubin-CPX-class launches (late 2026): the first head-to-head benchmarks against next-gen Nvidia inference parts will either validate or vaporize the speed premium that justifies Cerebras's per-token pricing.
- Diversification prints: AWS partnership mentions and any new named non-UAE, non-OpenAI customer >5% of revenue would be the single most thesis-strengthening datapoint possible.
Kill conditions (thesis tripwires — any one moves us from Watch toward Avoid, and invalidates the buy bands):
1. OpenAI materially renegotiates, delays beyond 2028, or second-sources the contracted capacity (the RPO is only as good as its counterparty's intent).
2. RPO declines sequentially without corresponding revenue recognition (backlog evaporation, not conversion).
3. Gross margin back below ~35% (the speed premium failing against Groq/Rubin pricing).
4. MBZUAI/G42 payment or lease-back irregularities deepening (the related-party circularity turning load-bearing).
5. The 10x manufacturing ramp missing badly (Feldman's own test — a miss impeaches the vendor lane wholesale).
6. Founder/executive selling at scale into the lockup gates (beyond the COO's 10k programmatic-scale sale).
11. Key risks
- The backlog IS the concentration (the dominant risk): >$20B of the $25B RPO is OpenAI — a customer that is itself burning cash, raising continuously, and famous for multi-sourcing compute (Nvidia, AMD, Broadcom ASICs, now Cerebras). The bull case and the single-counterparty risk are the same number.
- UAE dependency, present tense: ~86% of current revenue is MBZUAI + G42 — geopolitically exposed, related-party-entangled (Cerebras rents capacity back from G42), and historically the reason the 2024 IPO attempt drew CFIUS attention.
- Lockup/dilution mechanics: ~5-6x the float unlocking Aug–Nov 2026; fully-diluted count (~305M) is 35% above the basic count — per-share math must use FD.
- Architecture-bet risk: the KB's own bear lane (no_priors 2025) documents this exact class of company being invalidated once before by model growth; Rubin-generation parts adopting on-chip memory features could compress the speed gap from above while Groq pricing compresses economics from below.
- Capital-intensity trap: $2.3B of non-cancelable leases + multi-hundred-million quarterly capex against a contracted-but-undelivered revenue ramp — if utilization slips, the leases don't.
- Recent-IPO information asymmetry: 45 trading days of price history, one public quarter, 5 underwriter-adjacent analysts, and a promotional founder — every number that isn't in an SEC filing deserves the vendor-claim discount we've applied.
- Valuation: ~44x current-year sales leaves zero room for any of the above.
12. Verdict, position sizing & monitoring
Watch — Asymmetric Only Lower. The founder asked the precise question: is this a genuinely asymmetric bet, and at what price? Answer: yes — and not at this price. Cerebras has the most contractually documented exponential we cover (a company-disclosed $25.0B backlog, 41x trailing revenue, with independent verification of the >$20B OpenAI anchor), a real and measured speed moat, and near-breakeven operations far ahead of hardware-IPO vintage. But at $172.51 (~$52.6B fully diluted) the payoff map is inverted: bull +39%, base −33%, bear −80%. Exciting, yes. Asymmetric, no — the skew currently favors the seller.
The market is actively manufacturing the asymmetric entry: the stock sits 2% off its all-time low, below its $185 IPO price, with a phased lockup releasing ~5-6x the float between August and mid-November 2026. That supply wave is not a reason to despair; it is the mechanism by which the $80–110 band becomes reachable.
- Sizing: 0% today. In the $80–110 zone (FD ~$23-32B, ~8-12x 2027E sales): satellite 1-2% maximum — the concentration and counterparty risk cap size even at good prices. Below ~$70 (FD <$20B, below the pre-IPO private mark, <1x RPO): the table-pound zone, contingent on kill conditions 1-4 being clean — a price that low without broken news is the asymmetry; with broken news it is a trap.
- Monitoring: the §10 kill conditions every print; the lockup gates (post-Q2 tranche ~August, main release ~2 days post-Q3 report or ~2026-11-10); RPO drawdown vs revenue each 10-Q; the two Feldman clocks (10x manufacturing end-2026, >2x performance ~Jan-2028) as graded scorecard claims. Formal re-score on any ±25% move, the Q2 print, or any OpenAI/UAE contract news. This verdict is logged as a tracked Synthos call as of 2026-07-20 at $172.51.
- Single biggest risk: one counterparty. If OpenAI's signature weakens — renegotiation, delay, second-sourcing — the backlog, the estimates, the moat premium, and the buy bands all reprice at once. The bear case ($35) is what that looks like.
Provenance & disclosures
- Traceability: 9 entity-tagged KB claims, breadth 3 thinkers (all_in ×4, no_priors ×4, latent_space ×1), net conviction ~+48, plus 5 mention-lane claims (dylan_patel 2026-02-03, jensen_huang 2026-05-21, compound_and_friends 2026-05-15, bill_gurley/Sunny Madra 2024-10-13, ml_street_talk 2025-05-26) where Cerebras appears in thesis text without an entity tag. Claim IDs on file in
data/knowledge_base.jsonl(e.g.,all_in-y2NeAef6d30:9ac5598bb4,no_priors-vGhlJqnECd0:36840d1ecd,latent_space-sRpqPgKeXNk:423fcd90b3). Speaker-lane disclosure: 5 of the 9 entity-tagged claims record CEO Andrew Feldman as speaker — vendor lane, half-weighted throughout. The Jensen Huang claim is management-lane for NVDA (an interested competitor). Street targets are a 5-analyst, all-Buy, post-IPO underwriter cohort — discounted accordingly. - Data as-of: fundamentals 2026-03-31 (Q1 2026 10-Q, filed 2026-06-24); estimates and prices 2026-07-20 (FMP pull, $172.51 close basis); latest KB claim 2026-07-10. Forward figures are analyst consensus (5-9 analysts — thin) or management/vendor claims, labeled as such.
- FV method: scenario-based backlog-conversion × margin × multiple on ~290M fully-diluted shares, explicitly labeled not-DCF-defensible — a 45-trading-day-old company with one dominant contract cannot support a defensible DCF. No probability weights attached (the base case is the expected path; a blend would restate it with false precision).
- Recent-IPO data-quality flags: FMP FY25 GAAP net income (+$237.8M) includes ~$390M of non-operating preferred/warrant remeasurement — operating loss (−$145.3M) is the truer read; share-count fields are unreliable (we use market-cap-implied ~227M basic, ~305M FD from prospectus math); several quarterly opex rows are zeroed/merged; estimate-row EBITDA is internally inconsistent (discarded); DMA/return fields are null or artifactual on 45 days of history; FMP's "peers" list is IPO-cohort noise (ignored); duplicate earnings dates (8/11 and 9/22) flagged.
- Vendor/secondhand labels: the $25.0B backlog is company-disclosed (10-Q RPO) — upgraded from vendor claim to disclosed fact; the 15-20x performance, >2x/18mo, and 10x-manufacturing claims remain vendor claims (Feldman, podcast relays) with falsifiable tests logged; lockup tranche share-counts are secondary-source estimates not verified against the S-1; the Nvidia-Groq acquisition/Rubin-integration reports are trade-press claims we could not verify against primary filings; the 62%/24% concentration split is from press summaries of the 10-Q.
- Web sources: Cerebras IPO pricing press release · CNBC — IPO priced above range · IPOScoop — $185 pricing, $350 open, $311.07 close · Quartz — $5.55B raised · CNBC — OpenAI deal >$10B · DCD — 750MW through 2028, cloud-services structure · Bloomberg — OpenAI deal · StockTitan — Q1 10-Q summary, $25.0B RPO · TECHi — Q1 concentration (62% MBZUAI / 24% G42) · Seeking Alpha — lockup structure · QuantAbundancia — AI IPO lockup calendar · AInvest — lockup analysis · The Register — wafer-scale history/IPO · verticalapi — Groq vs Cerebras benchmarks · digitalapplied — Q2 2026 inference pricing matrix · Silicon Report — non-GPU inference bet · hashrateindex — independent AI chip companies
- Management caveat: Cerebras guidance and all Feldman podcast claims are management's own book, half-weighted by design; the honest-numbers standard applies double to a nine-week-old IPO with a promotional founder.
- Not investment advice. Independent research, educational and informational only, never personalized. Hypothetical/forward figures are labeled; the only performance numbers Synthos will headline are the live, real-money Flagship's.
- Version: 2026-07-20. Prior versions available via the deep-dive version dropdown ("based on the info at the time").