Cerebras Systems, Inc. (NASDAQ: CBRS) — One Customer, One Contract, One Hundred Times Sales

Independent equity research note Report date: 2026-06-10 · Price reference: ~$237 (IPO 2026-05-14 at $185.00) · Fresh coverage

The institutional analysis in §1–§15 below carries no investment recommendation and no price target, per firm policy. The sole exception is the clearly-labeled “Claude’s Take” block immediately below, which is a subjective view fenced off from the house framework.

⚡ Claude’s Take

This block is the author’s own subjective opinion and general information only — not investment advice. The analysis that follows takes no position and sets no price target.

Verdict: AVOID at ~$237 — a remarkable engineering story priced as a fait-accompli, but not investable here. Explicitly NOT a short. “Great story, impossible price.” Directional zone where the risk/reward starts to interest me: the low-$100s to ~$130 — roughly 3–4x the Series H private mark ($89, Jan 2026) and a still-rich ~25–40x forward sales against a credible (not heroic) OpenAI ramp. I would need a materially lower entry, or proof the OpenAI capacity is converting to profitable recognized revenue, before the math works.

The market is pricing Cerebras as though its $24.6B remaining performance obligation (RPO) — essentially one OpenAI contract — is already de-risked, already diversified, and already at semiconductor-grade margins. None of those are true today. The company recognized $0 of OpenAI revenue in 2025; its gross margin is falling (42.3% → 39.0%); its non-GAAP loss is widening (−$21.8M → −$75.7M); free cash flow is −$393M; and 86% of 2025 revenue came from two related UAE entities (MBZUAI 62%, G42 24%). The headline +$237.8M GAAP “profit” — and the P/E of ~500, the “46% margin,” the “38% ROE” the data screens show — is a non-cash accounting gain on a forward-contract liability, not earnings. Strip it out and this is a cash-burning, single-customer hardware company at ~90–125x sales versus NVIDIA at ~23x (with 75% margins and ~$200B of FCF). The framing is late-cycle momentum + IPO scarcity (14% float, 20:1 founder super-vote, sell-side initiating unanimously Buy at $250–$340) — the same signature seen on ARM, where mean targets sat below spot.

Why not a short, and why only medium conviction: the RPO is real and contracted, the OpenAI relationship is genuine (a $1B loan and a ~$8B penny-warrant equity grant are real skin in the game), wafer-scale inference may be durably fast, and a 14%-float momentum name can squeeze violently. Conviction: medium. The single piece of evidence that flips me bullish: two to three quarters of OpenAI capacity converting to recognized revenue at expanding gross margin with a second non-UAE, non-OpenAI anchor customer signed — i.e., concentration genuinely breaking. The single piece that flips me outright bearish: any OpenAI de-scope/slip, a margin print below ~35%, or the early lock-up release (up to ~171M shares) hitting the 30M float. Tag: “the whole thesis fits on one contract.”

1. Executive Summary

Cerebras Systems designs and sells wafer-scale AI compute — the Wafer-Scale Engine (WSE-3), a single processor cut from an entire TSMC silicon wafer (46,225 mm², ~4 trillion transistors, 900,000 cores, 44 GB on-chip SRAM), housed in CS-3 systems and lashed into AI supercomputers. The architectural bet is that keeping compute and memory on one piece of silicon eliminates the off-chip data movement that bottlenecks GPU clusters, delivering AI inference up to ~15x faster than leading GPU solutions. The company IPO’d on 2026-05-14 at $185/share (30M Class A shares, ~$5.4B net proceeds) after a ~20-month delay driven by CFIUS review of its UAE ownership/customer ties. It trades at ~$237 today (~$51B basic market cap, ~$69B economically diluted).

The business has grown revenue impressively — $24.6M (2022) → $78.7M (2023) → $290.3M (2024) → $510.0M (2025, +76%) — but three facts dominate the thesis. First, concentration is extreme and related-party: MBZUAI (62%) and G42 (24%) — two affiliated Abu Dhabi entities — were 86% of 2025 revenue; G42 was 85% in 2024. Second, the forward story is one customer: the entire $24.6B RPO is predominantly a single Master Relationship Agreement with OpenAI (750 MW, signed December 2025), of which zero was recognized in 2025, layered with a $1.0B OpenAI working-capital loan and a ~33.4M-share penny warrant (~$8B of equity granted to the customer). Third, the economics do not yet work: gross margin fell to 39.0%, the operating loss widened to −$145.9M, the non-GAAP net loss widened to −$75.7M, and FCF was −$393M as the company spent $383M of capex pivoting into owned cloud capacity. The GAAP profit is an artifact of a $363.3M non-cash forward-contract extinguishment gain tied to the unwinding of G42’s CFIUS-blocked equity stake.

On competitive position, Cerebras possesses real, hard-won engineering differentiation but no durable moat in Greenwald’s sense. It sells a premium point-solution (inference speed) into the most fragmentable, most contested slice of AI compute, against a ~90%-share, CUDA-locked incumbent (NVIDIA), better-funded merchant rivals (AMD, Groq, SambaNova), and its own largest customers’ in-house ASICs (Google TPU, AWS Trainium). Crucially, it sits outside the CUDA ecosystem — it must persuade customers to leave their lock-in, the inverse of a switching-cost moat. Governance is founder-controlled (Class B = 20 votes, ~99.2% of voting power; ~14% public float) with a 9M-share founder PRSU mega-grant vesting on $75B/$150B/$250B market-cap hurdles. Verdict across the framework: a structurally attractive end-market, a structurally brutal competitive position for a sub-scale entrant, exceptional concentration risk, unproven unit economics, and a valuation that already underwrites flawless execution of a single mega-contract. This memo takes no position; the analysis is laid out below.

2. Business Overview

What Cerebras does. Cerebras Systems, founded in 2015 and headquartered in Sunnyvale, California, is a fabless AI-infrastructure company. Its defining product is the Wafer-Scale Engine (WSE-3) — rather than dicing a silicon wafer into hundreds of small GPU dies (NVIDIA’s approach) and reconnecting them with networking, Cerebras keeps the entire wafer as one monolithic processor. The WSE-3 is fabricated by TSMC on a 5nm process and, per the prospectus, spans 46,225 mm² with roughly 4 trillion transistors, 900,000 AI cores, and 44 GB of on-chip SRAM — the company claims it is ~58x larger than NVIDIA’s B200 GPU die. The chip is packaged into the CS-3 system, and multiple CS-3s are interconnected into “Cerebras AI supercomputers” (the Condor Galaxy installations built for G42 are the flagship reference deployments). With only ~340 full-time employees, this is an extraordinarily small company relative to its revenue and valuation.

How it makes money — three revenue streams (FY2025, reconciled to the 424B4 financial statements):

1. On-premises hardware — outright sales of CS-3 systems to customers who want to own and control their compute (e.g., sovereign data-center build-outs). Recognized at a point in time on delivery/acceptance, which makes hardware revenue inherently lumpy quarter to quarter and gives a single large shipment outsized influence on any period.
2. Cloud & inference services — the “Cerebras Cloud,” sold both as dedicated reserved capacity (recognized over time) and on-demand/by-the-token inference, distributed partly through partner marketplaces (AWS, Microsoft, IBM watsonx, Hugging Face, OpenRouter, Vercel). This is the strategic growth vector and the reason for the 2025 capex surge, but its gross margin collapsed from 61.4% to 29.9% in 2025 as the company absorbed data-center start-up costs.
3. Custom AI model / professional services — training and fine-tuning work; a small, services-like contributor.

Recurring vs. non-recurring. Today the model is mostly non-recurring: 70% of revenue is point-in-time hardware, and even the “cloud” revenue is concentrated in a handful of large reserved-capacity contracts rather than a diversified, sticky subscription base. Management states it expects cloud/services to grow as a share of the mix over time — a transition that, if achieved, would improve revenue visibility but currently carries lower gross margin than hardware.

End markets and customers. Cerebras serves (in its own words) hyperscalers, foundation-model labs, AI-native businesses, enterprises, and Sovereign AI initiatives. In practice the realized customer base has been overwhelmingly UAE-sovereign-adjacent: G42 (an Abu Dhabi AI holding company) and MBZUAI (the Mohamed bin Zayed University of Artificial Intelligence, related to G42) together were 86% of 2025 revenue. The forward pivot is toward OpenAI (the December 2025 MRA) and AWS (a partner/term-sheet relationship). The customer set is thus not just concentrated but rotating — from one sovereign anchor to one frontier-lab anchor — rather than diversifying.

The technical wager, in plain terms. A modern AI model is a pile of matrix multiplications whose weights and activations must constantly move between compute units and memory. On a GPU cluster, that movement crosses chip boundaries — die-to-die, GPU-to-GPU, server-to-server — over interconnects (NVLink, InfiniBand) that are fast but far slower than on-die SRAM. NVIDIA’s answer is to make the interconnect ever better and to hide latency behind massive parallelism; Cerebras’s answer is to remove the boundary entirely by keeping the whole model resident on one wafer’s worth of SRAM. For latency-sensitive inference of a single large model — generating tokens one after another, where each token depends on the last — that architecture can be genuinely faster, because there is no inter-chip hop in the critical path. This is the kernel of truth in the “up to ~15x faster” claim. The weakness is equally structural: a wafer holds a fixed amount of SRAM (44 GB), so very large models must be split across multiple CS-3s anyway, and the economics of dedicating an entire wafer to one workload are unforgiving unless utilization is high and the customer values speed enough to pay a premium. Wafer-scale is therefore not a general-purpose GPU replacement; it is a specialized bet that a valuable slice of inference will pay up for latency.

Software and distribution. Cerebras ships its own software stack — a compiler and graph-mapping toolchain that takes standard PyTorch models and lays them across the wafer, explicitly pitched as “eliminating the need for CUDA.” It distributes inference both directly (Cerebras Cloud / Cerebras Inference API) and through third-party marketplaces (AWS, Microsoft Azure/Foundry, IBM watsonx, Hugging Face, OpenRouter, Vercel), which broadens reach but cedes the customer relationship and a margin slice to the platform. The Condor Galaxy supercomputers built for G42 are the flagship reference installations and the template for the sovereign and OpenAI build-outs.

Verdict (Business Overview): a genuinely novel, single-architecture hardware company with a nascent, low-margin cloud overlay, selling mostly lumpy hardware to a tiny number of very large, very concentrated customers, on the bet that latency-sensitive inference will pay a premium for wafer-scale speed. The model’s quality hinges entirely on whether the cloud/inference transition can convert the OpenAI relationship into diversified, higher-margin recurring revenue — none of which is yet visible in the financials.

3. Industry Dynamics

A genuinely enormous, fast-growing end-market. There is no debate that AI compute is one of the great secular demand stories of the decade. Cerebras cites Dell’Oro that worldwide data-center infrastructure capex grows from $679B (2025) to $1.7T (2030) — a ~21% CAGR — and splits the AI-silicon pool into training (~$185B → ~$380B by 2029, ~20% CAGR) and inference (its addressable slice ~$66B → ~$292B by 2029, ~45% CAGR). Independent industry estimates corroborate the order of magnitude, putting hyperscaler capex near ~$1T in calendar 2026 (top-five ~$700B). The directional claim that inference will be the larger long-run pool is credible and corroborated — inference scales with end-user adoption toward the global internet base, and token volumes have exploded (Google reported serving ~1.3 quadrillion tokens/month by late 2025). On the demand side, the bull case is real.

But the framing is self-serving. Cerebras’s “$251B → $672B (2029), 28% CAGR” headline number cherry-picks training plus a sub-slice of inference — not the full $1.7T capex pool — engineered so that inference (where Cerebras claims its speed edge) looks like the dominant pool the company is levered to. Treat the company’s TAM math as marketing, not data.

Competitive structure — hostile for a sub-scale merchant entrant. This is the crux. The prospectus’s own risk factors concede Cerebras operates “in a market with a dominant incumbent” — NVIDIA, “a dominant market leader” — that holds an estimated ~90%+ of merchant data-center accelerators behind a CUDA software-ecosystem moat, full-stack co-design, and R&D scale widely judged to be one of the most durable franchises in technology. Around NVIDIA sit: better-funded merchant challengers (AMD/ROCm, Intel Gaudi), direct fast-inference specialists (Groq’s LPU, SambaNova), and — most structurally threatening — the in-house ASIC programs of Cerebras’s own largest potential customers: Google TPU, AWS Trainium/Inferentia, Meta MTIA, Microsoft Maia, designed via Broadcom and Marvell and amortized over captive hyperscaler volume. Note the uncomfortable duality: AWS is simultaneously a marquee Cerebras partner and an in-house-silicon competitor; the filing admits “certain of our competitors are also current or prospective customers.”

Through Greenwald’s lens there is no genuine barrier to entry protecting Cerebras’s position. It has none of the three durable advantages: no software-ecosystem captivity (it sits outside CUDA), no economies of scale (it is a rounding-error wafer buyer), and no customer captivity (it sells a premium point-solution into a market the prospectus itself warns “may become increasingly commoditized”). Wafer-scale integration is a real supply/technology edge — and Greenwald is explicit that proprietary-technology advantages are the weakest and most transient category (“in the long run everything is a toaster”).

Training vs. inference economics — why the pivot is rational but contested. Training is a small number of very large, capital-intensive runs concentrated among a handful of frontier labs and hyperscalers; inference is the ongoing cost of serving trained models to end users, scaling with adoption rather than model count. As the installed base of AI applications grows, inference becomes the larger and more durable compute pool — a widely-held industry thesis. Cerebras’s pivot toward inference is therefore strategically sound directionally. The contested part is the structure of inference spend: it fragments into latency-critical (chat, agents, code), throughput-batched (embeddings, bulk generation), and edge workloads, each with different optimal silicon. Cerebras’s wafer-scale edge is sharpest in the latency-critical slice, which is real and growing but is also where every challenger and every hyperscaler ASIC concentrates. “Inference is bigger” does not automatically mean “Cerebras’s addressable, defensible inference profit pool is bigger.”

Capital cycle (Marathon) — IPO-ing into the supply peak. Every diagnostic of a late-boom capital cycle is lit: enormous lumpy/lagged capacity build-out (OpenAI 750 MW, Stargate, sovereign data centers), a rash of AI IPOs (Cerebras, CoreWeave, IREN, Nebius), unanimous investment-bank cheerleading, and AI weighting at index extremes. High incumbent returns (NVIDIA’s ~75% gross margin) are simultaneously attracting capital from competitors, customers building their own silicon, and sovereign states — the textbook supply flood that mean-reverts merchant economics. Because supply additions are lumpy and lagged (multi-year lead times), the excess only becomes visible after the training-demand pulse normalizes. A small, cash-burning entrant has the least cushion when digestion hits — and Cerebras is the marginal supplier IPO-ing as the cycle peaks.

Value chain & dependencies — minimal bargaining power. Cerebras is fully dependent on TSMC (“we are currently dependent on TSMC to produce all of the wafers we use… we have no formalized long-term supply or allocation commitments”), and TSMC also fabricates for “certain of our competitors, many of whom are significantly larger.” It is thus a tiny, unallocated buyer queued behind NVIDIA, AMD, and Apple at a near-monopoly foundry, equally exposed to the industry-wide constraints in HBM/memory, advanced packaging (CoWoS-class), and data-center power.

Regulation — export controls over a single-geography revenue base. Cerebras’s exposure is acute precisely because it overlaps the concentration. UAE customers dominate revenue; the UAE sits in EAR Country Group D:4; Cerebras holds BIS export licenses for its systems to G42/MBZUAI but under “rigorous security and compliance obligations to prevent diversion,” revocable even post-manufacture. Most tellingly, CFIUS forced the unwinding of G42’s planned $335M equity investment (joint notice July 2024, withdrawal granted March 2025) — the state permitted the customer relationship but blocked foreign ownership, a live national-security ceiling over the company’s single largest historical demand source. The episode is the single most important precedent for the thesis: it demonstrates that the US government will intervene directly in Cerebras’s most important commercial relationships when national-security concerns arise, and that the UAE-AI nexus is squarely in scope. The export-license regime compounds this — Cerebras’s licenses to ship CS-class systems to G42/MBZUAI carry diversion-prevention obligations and are revocable even after a system is manufactured, meaning inventory could in principle be stranded by a policy shift. The broader US policy posture toward advanced-AI-chip exports to the Middle East has oscillated, and a tightening (or a deterioration in US–UAE relations, or regional conflict) would strike at the base that still represents the bulk of recognized revenue. The pivot to OpenAI partially diversifies away from this specific exposure, but it does so by concentrating into a different single counterparty — and OpenAI’s own compute ambitions are themselves entangled with sovereign and hyperscaler politics. Geopolitical risk here is not a tail footnote; it is woven into the revenue base.

Verdict (Industry): a structurally attractive industry to consume from and a structurally brutal one to enter and compete in as a sub-scale merchant supplier — a ~90%-share CUDA-locked incumbent above it, its best-funded customers building competing silicon, a monopoly foundry above it with no allocation guarantee, and a capital cycle at peak euphoria. Great demand, terrible competitive structure for this player.

4. Competitive Position

The moat question, answered bluntly: there is no durable moat. Wafer-scale integration is a genuine, hard-to-replicate feat of engineering — Cerebras is the only company to ship a commercial wafer-scale processor, now in its third generation. But a hard-to-replicate product is not a moat unless it produces customer captivity or scale economics that competitors cannot match, and Cerebras has neither. Pressure-testing each axis:

(a) Is the inference-speed advantage defensible? The speed lead appears real today (benchmarked by third party Artificial Analysis), but inference is the least defensible battlefield in AI compute — the most fragmentable workload (smaller models, latency-driven, amenable to fixed-function silicon), precisely where CUDA captivity is weakest and where the most competitors converge. Groq (LPU), SambaNova, AMD, Google TPU, AWS Inferentia, and Intel Gaudi all attack the identical speed/$ axis. A single benchmark on select open-source models is not durability across model generations.

(b) Switching costs — the killer point, and it runs against Cerebras. NVIDIA’s moat is CUDA: ~20 years of libraries, kernels, and developer muscle memory that lock customers in. Cerebras’s own pitch is that its compiler “eliminates the need for CUDA” — i.e., it sits outside the dominant ecosystem and must convince customers to leave their lock-in. That is a customer-acquisition hurdle, not a retention moat. Cerebras has the inverse of the industry’s defining switching cost.

© Scale. Cerebras is rounding-error small — ~340 employees and $510M revenue against NVIDIA’s $200B+. Marathon’s maxim that “market growth is the enemy of scale advantages” cuts against it: the exploding TAM helps the incumbent’s scale economics, not the sub-scale entrant’s. It cannot reach economies-of-scale-plus-captivity, the only truly durable advantage type.

(d) Does the moat show up in the financials? No — definitively. A real franchise shows pricing power, high and stable gross margins, high ROIC, and customer stability. Cerebras shows ~39% and falling gross margin (vs. NVIDIA ~75%), no pricing power (the filing concedes “we sell a premium solution, and our competitors’ solutions may be less expensive” and flags “price erosion”), operating losses every year, and zero customer diversification. The moat claim cannot be tied to a financial outcome that would deteriorate without it — because the favorable outcomes are not there to begin with.

Head-to-head positioning:

Customer concentration is itself a competitive-position liability. A franchise with pricing power diversifies its customers; Cerebras’s 86%-from-two-related-UAE-entities (and forward 100%-from-OpenAI RPO) is the opposite — it signals a product that has won a few large, captive, often strategically-motivated buyers (a sovereign AI program; a frontier lab seeking a second source to NVIDIA) rather than broad market pull.

The Groq problem and the “second-source” thesis. The most direct refutation of Cerebras’s differentiation is Groq, a fellow fast-inference specialist whose LPU architecture attacks the identical “fastest tokens/second” value proposition with a different (deterministic, SRAM-heavy ASIC) approach. When two well-funded startups both claim the inference-speed crown on overlapping benchmarks, neither has a moat — they have a feature that the other can match and that NVIDIA can erode with each generation (Blackwell and Rubin both heavily target inference). The more durable version of the Cerebras bull case is therefore not “we are permanently fastest” but “the market wants a credible second source to NVIDIA, and large buyers will allocate a slice of spend to us for supply diversification and negotiating leverage.” That thesis has real merit — OpenAI’s interest is plausibly driven as much by second-sourcing away from NVIDIA/Microsoft as by raw speed — but it is a commodity-supplier thesis, not a franchise thesis: a second source competes on price and availability, which is precisely why Cerebras’s gross margin is 39% and falling rather than NVIDIA’s 75%. Being the designated alternative to a monopolist is a real business; it is not a moat, and it caps margins structurally.

Why the speed edge may not convert to economics. Even granting a durable latency advantage, three things blunt its financial value: (1) customers capture much of the benefit (they pay for tokens, not for the architecture, and competition sets the token price); (2) wafer-scale’s fixed-SRAM constraint means utilization economics are brutal on anything but high-throughput, well-batched workloads; and (3) the largest, most sophisticated buyers — the hyperscalers — are precisely the ones building their own inference ASICs (TPU, Inferentia, MTIA, Maia), so the segment most able to pay is the segment most likely to self-supply. The addressable, pays-a-premium, can’t-self-supply slice may be narrower than the $292B inference-TAM headline implies.

Verdict (Competitive Position): a crowded, fast-commoditizing inference market with weak structural differentiation and no customer captivity — the precise opposite of the CUDA moat that defines the category leader. The most defensible framing is “credible second source to NVIDIA,” which is a viable business but a margin-capped, commodity-supplier one, not a franchise. The differentiation is real but of the most transient (proprietary-technology) type, deployed in the most contested workload, by the smallest-scale player.

5. Growth History and Forward Opportunities

Historical growth is real but decelerating and concentration-driven. Revenue compounded from $24.6M (2022) to $510.0M (2025) — a ~174% three-year CAGR — but the year-over-year rate decelerated from +269% (2024) to +76% (2025), and essentially all of it traces to a handful of sovereign-linked contracts (G42/Condor Galaxy build-outs, then MBZUAI). This is low-quality growth: it is concentrated, lumpy, related-party, and not yet accompanied by improving economics (margins fell, losses widened). Growth that requires giving a customer a $1B loan and ~$8B of equity (OpenAI) to secure is growth bought, not earned through broad demand.

Forward opportunities — large but conditional:

• OpenAI MRA (the whole forward story). Signed December 2025: a 750 MW commitment, >$20B headline value, and the dominant component of the $24.6B RPO. Recognition is back-end-loaded (~15% by 12/31/27, ~43% in months 25–48, remainder later) and milestone-gated and terminable — Cerebras must deliver capacity tranches across data centers on time, or OpenAI can terminate part/all and potentially call the $1B working-capital loan. This is the single largest growth lever and the single largest risk, simultaneously.
• Cloud/inference platform. The pivot to owned cloud capacity (the $383M 2025 capex) aims to convert one-time hardware buyers into recurring, by-the-token inference customers and to broaden distribution through AWS/Microsoft/IBM/Hugging Face marketplaces. Strategically sensible; economically unproven (cloud GM collapsed to 30%).
• Sovereign AI expansion. Genuine new demand vector industry-wide, but for Cerebras it has so far meant single-geography (UAE) concentration under export-control and CFIUS risk rather than diversification.
• Diversification into enterprises / new geographies. Necessary to de-risk the thesis; not yet evident in the numbers.

The deceleration that matters. Headline growth is still high, but the second derivative turned: +269% (2024) decelerated to +76% (2025), even with the cloud build ramping and a full year of MBZUAI. The market is underwriting a re-acceleration off the OpenAI ramp from 2027 onward — the RPO schedule (~15% recognized by end-2027, ~43% in months 25–48) implies a step-up in 2028–2029 toward multiple-billion annual revenue. That is plausible if capacity is delivered on time and OpenAI does not de-scope, but it is the opposite of a smooth compounding curve: it is a back-end-loaded, milestone-gated, single-customer step-function. Growth-quality investors should note that the years between now and the step-up are the cash-burning, margin-pressured, dilution-heavy build years, and that the re-acceleration is a forecast, not a trend.

Verdict (Growth): high headline growth of low quality — concentrated, lumpy, related-party, margin-dilutive, decelerating before a forecast OpenAI-driven re-acceleration, and forward-dependent on one conditional mega-contract. The growth that matters (diversified, profitable, recurring) is entirely prospective.

6. Financial Quality

The single most important point: the headline GAAP profitability is an accounting mirage. Cerebras reported +$237.8M GAAP net income for 2025 — the source of the ~500x P/E, “46% profit margin,” and “38% ROE” on data screens. That figure is driven entirely by a non-cash, non-operating $363.3M gain on the extinguishment of a forward-contract liability (the mirror image of a $401.3M loss on the same instrument in 2024, which produced the −$481.6M 2024 net loss). The instrument was a freestanding derivative tied to the G42-linked redeemable convertible preferred stack; it was remeasured to fair value each period and extinguished in 2025 when the CFIUS-forced unwinding of G42’s equity stake completed. It has nothing to do with operations or cash. The clean picture:

The truth: on +76% revenue growth, the non-GAAP net loss widened ~3.5x, from −$21.8M to −$75.7M.

Reconstructed P&L:

Margins are going the wrong way under scale. Gross margin fell 330 bps in 2025 (42.3% → 39.0%), driven by cloud GM collapsing from 61% to 30% as the company ate data-center start-up costs; management explicitly guides gross profit dollars to fall near-term. S&M tripled (to $70.6M) and R&D rose 54% (to $243.3M, ~48% of revenue). This is the opposite of operating leverage — opex is outrunning revenue and the operating loss widened. A separate one-time IPO-triggered SBC charge of ~$150.5M (RSU/PRSU liquidity vesting) hits post-IPO quarters.

Cash-flow quality — the “2024 FCF” was a prepayment mirage. Reported operating cash flow was +$452.0M in 2024 but −$10.1M in 2025. The 2024 figure was not earnings — it was manufactured by a +$640.3M increase in customer deposits (G42/MBZUAI hardware prepayments) plus deferred-revenue growth. When that working-capital timing reversed in 2025 (deposits −$285.9M), operating cash went negative. Capex then exploded to $382.7M (from $23.4M) to build owned cloud capacity, taking FCF to ~−$393M (vs. the prepayment-flattered ~+$429M in 2024). Deferred revenue ended 2025 at $166.9M (from $57.4M) and customer deposits at $354.5M — confirming the cash engine here is customer prepayments, not profit.

Balance sheet — strong post-IPO, but built on a complex stack. At year-end 2025: liquid assets ~$1,108M (cash $701.7M + investments $406.5M), plus $228.7M restricted cash; A/R $50.4M; inventory $63.6M. Total stockholders’ equity was −$578.7M — a mezzanine-preferred artifact (the $1,933M redeemable convertible preferred sat outside permanent equity, with an accumulated deficit of −$905.3M), not insolvency; actual total liabilities were only $971.3M. The IPO ($5.4B net) plus the automatic preferred conversion reverses the deficit and leaves pro-forma post-IPO cash of roughly $6.5–7B (existing $1.1B + IPO $5.4B + the $1.0B OpenAI loan received January 2026). Debt is the OpenAI $1.0B working-capital loan (6%, repayable in cash or compute/hardware/services, maturity ≤2032) and a $250M revolver (cash-collateralized, restricts dividends). Runway is not a near-term concern despite ~$75M annual non-GAAP burn — the constraint is the capital intensity of the cloud build, not survival.

RPO is the entire valuation case — and it is one customer. Remaining performance obligations were $24.6B at 12/31/25 (~48x FY25 revenue), predominantly the OpenAI 750 MW MRA, with $0 OpenAI revenue recognized in 2025. Backlog of this magnitude on a $510M-revenue company is unprecedented and is the literal foundation of the multiple — but it is forward, conditional, back-end-loaded, and single-customer.

The forward-contract liability, explained. Because this single item swung reported net income by ~$770M across two years, it deserves a plain-English account. As part of the G42 arrangement, Cerebras issued instruments (tied to the redeemable convertible preferred stack and a forward purchase commitment) that were classified as liabilities rather than equity, because the underlying shares were redeemable on events outside the company’s control. GAAP requires such liabilities to be marked to fair value every reporting period, with the change running through the income statement. As Cerebras’s implied equity value rose into 2024, the liability grew — generating a $401.3M non-cash remeasurement loss that drove the −$481.6M 2024 net loss. When the CFIUS-forced unwinding removed G42 as an equity party in 2025, the instrument was extinguished, reversing the accumulated liability and booking a $363.3M non-cash gain — the sole reason 2025 GAAP net income printed positive. No cash moved; no product was sold; the “profitability” is an accounting echo of a financing dispute. Any analyst or screen quoting Cerebras’s P/E, profit margin, or ROE is quoting noise.

Deferred revenue and the prepayment engine. The mechanism that makes Cerebras look cash-generative in good years is customer prepayment. Sovereign and frontier-lab customers pay substantial deposits ahead of hardware delivery; those deposits land in “customer deposits” and “deferred revenue” on the balance sheet and in operating cash flow before the matching revenue or cost is recognized. In 2024 that inflow was enormous (+$640.3M in customer deposits) and turned a −$481.6M GAAP loss into +$452M of operating cash; in 2025 the timing reversed (−$285.9M) and operating cash flipped negative. Deferred revenue still grew to $166.9M (from $57.4M), and customer deposits ended at $354.5M — healthy forward signals, but they confirm the cash engine is the timing of customer money, not the conversion of profit. A concentrated customer base makes this lumpier still: one large deposit or one delayed milestone can swing a quarter’s reported cash flow by hundreds of millions.

Stock-based compensation and the IPO charge. SBC was $49.8M in 2025 (R&D $32.2M, S&M $10.0M, G&A $6.8M) — already a meaningful drag on the non-GAAP loss — and a one-time ~$150.5M SBC charge from IPO-triggered RSU/PRSU liquidity vesting hits post-IPO quarters, alongside the ~$416M of cash used to sell-to-cover RSU tax withholding. Investors should expect reported GAAP results in the first public quarters to be heavily distorted by these one-time equity charges on top of the operating loss.

Verdict (Financial Quality): do economics improve with scale? No — not demonstrated. On 76% revenue growth, gross margin compressed, cloud margin collapsed, the operating loss widened, and the non-GAAP loss widened 3.5x. The only positive operating signal was hardware GM (+770 bps to 42.9%), swamped by cloud drag and opex. All operating leverage is prospective — it lives in the $24.6B OpenAI RPO, not in any realized trend. The demonstrated record is a deepening margin-and-cash burn funded by customer prepayments and a $5.4B IPO, dressed up by a one-time non-cash gain.

7. Capital Allocation

Funding history — a serial preferred-raiser on an escalating mark. Cerebras has never funded itself from operations. The pre-IPO ladder: Series F-1 (May 2024, $14.66/sh), Series G (Sep–Oct 2025, $36.23/sh, ~$1.1B; Fidelity $700M, Alpha Wave, Benchmark), Series H (Jan–Feb 2026, $89.02/sh, ~$1.0B; Benchmark $225M, Alpha Wave $100M, Fidelity $100M), and the IPO at $185. The mark walked $14.66 → $36.23 → $89.02 → $185 in under two years — a private-to-public quadrupling that itself signals late-cycle exuberance. The IPO was primary-only (no selling stockholders), raising ~$5,408.5M net on 30.0M Class A shares, with ~$416M carved out to satisfy RSU tax withholding (sell-to-cover). There has been no M&A.

Where the capital goes — R&D and a large, unhedged cloud-capex bet. 2025 spend went two places: R&D $243.3M (48% of revenue, up from $158.2M) and a capex explosion to $382.7M (from $23.4M) to build Cerebras Cloud data-center capacity — a large bet placed before the cloud unit economics are demonstrated (cloud GM fell to 30%). The company also signed a $2.2B, 10-year Canada data-center lease (March 2026), a sizeable off-balance-sheet commitment.

The customer-funded growth model — and its entanglement. This is the defining feature of Cerebras’s capital structure: it finances its build-out with customer money and customer equity. The $1.0B OpenAI working-capital loan; substantial G42/MBZUAI advance payments (G42 was 91% of A/R at YE2024); and large warrant grants to anchor customers. The line between “customer,” “lender,” and “owner” is essentially erased — OpenAI is simultaneously the largest contracted customer, a $1B lender, and a ~13% potential equity holder.

The OpenAI penny warrant — ~$8B of equity to a customer. The OpenAI Warrant grants 33,445,026 Class N shares at $0.00001 (effectively free). Vesting is tied to OpenAI milestones: a tranche on receipt of the $1B loan, a tranche on the earlier of a $40B market cap or fee payments, and ~23.4M shares across committed compute-delivery dates (full vesting requires OpenAI to buy all 2 GW of Additional Capacity). At ~$237 this is ~$7.9B of equity handed to a customer — economically a contra-revenue customer-incentive instrument dressed as a warrant, and ~13% potential dilution. Class N is non-voting, so it dilutes economics without diluting founder control. (Add a 2.7M Class N AWS warrant at $100 and ~3.5M Class N G42 warrants.)

Governance — founder-controlled, outside-shareholder-hostile. Three classes: Class A = 1 vote, Class B = 20 votes, Class N = 0 votes. Post-IPO, Class B holds ~99.2% of voting power; the 20:1 ratio lets insiders retain control down to ~5% economic ownership. It is not technically a “controlled company” (no single >50% holder), but directors/officers/5%-holders together control ~50.8% of the vote. Public Class A float is ~30M of ~215M shares ≈ 14% — tiny, amplifying volatility and the lock-up overhang, and the multi-class structure risks index exclusion.

The founder PRSU mega-grant. In February 2026 the board granted 9,000,000 market-cap-based PRSUs (Feldman 5.7M, Lie 3.3M, settled in super-voting Class B), vesting in thirds on 90-day trailing market caps of $75B / $150B / $250B. These are pure share-price hurdles, not operational/ROIC metrics — they reward market-cap inflation, not capital discipline — layered on make-whole and annual RSUs, with a consultant engaged to lift founders into the “top 10%” of peers. A classic IPO-timed founder mega-grant red flag.

Insider activity / lock-up. The IPO is primary-only with no insider selling; IPO-window Form 4s show only RSU tax-withholding (code F) and J entries — no open-market buys (no conviction signal) and no discretionary sells. The lock-up ends the earlier of two trading days after Q3-2026 earnings or 180 days, but early-release provisions free up to ~171.1M shares (including ~15.0M held by Section 16 insiders) — a large latent supply overhang against a 30M float. Separately, a December 2025 tender bought back 2.16M Class B at $36.23 ($78.1M) from employees (including $1.0M each from the CFO and COO) — an insider liquidity event, not a return of capital.

Principal holders: Fidelity 11.0%, Benchmark 9.5%, Foundation Capital 8.3%, Eclipse 7.3%, Alpha Wave 6.5% (Abu Dhabi-linked via Lunate/Chimera); founders Feldman 5.4% / Lie 2.9% economic. NEO 2025 pay is ~92% equity (Feldman $11.75M, Lie $11.57M) — “alignment” runs through share price, consistent with the PRSU design.

The customer-as-capital-provider model, assessed. There is a sophisticated-bull reading of the entanglement: by financing its build-out with customer loans (OpenAI’s $1B), customer prepayments (G42/MBZUAI deposits), and customer equity (the penny warrant), Cerebras de-risks demand and conserves its own balance sheet — the customer has put real money down before the capacity exists. That is genuinely better than building speculative capacity into a hope. But the same structure is a double-edged sword: it means the customer holds the leverage. OpenAI can direct the bank to freeze the $1B loan and demand repayment on a termination/trigger event; the penny warrant only vests if OpenAI keeps buying, so Cerebras is incentivized to keep one customer happy at almost any cost; and the prepayment model makes reported cash flow a function of one or two customers’ payment timing. Customer-funded growth lowers speculative risk but raises dependency risk — it trades the risk of building too much for the risk of being captured by the buyer. For a company whose entire forward book is one customer, that trade compounds the central vulnerability rather than offsetting it.

Verdict (Capital Allocation & Governance): negative on governance, unproven on capital allocation. Governance is structurally hostile to Class A holders (20:1 super-vote, ~99.2% Class B voting power, ~14% float, a 9M market-cap-based founder PRSU grant, index-exclusion risk). Capital allocation has no track record to judge: the company is pre-FCF, funded by ever-higher preferred rounds and customer loans/prepayments, and is giving away billions in equity to its own customers in a structure that fuses revenue, financing, and dilution into one concentrated, entangled web. Founder-controlled, customer-entangled, heavily-diluting — alignment is to share-price escalation, not disciplined capital returns.

8. Changes and Headwinds — Last Two Years

The last ~24 months reshaped the company:

• The CFIUS saga (2024–2025). Cerebras’s planned $335M G42 equity investment triggered a CFIUS joint voluntary notice (July 2024), the IPO was shelved, and the parties ultimately removed G42 as an equity party and terminated the purchase option (withdrawal granted March 2025), producing the $363.3M 2025 forward-contract extinguishment gain. The customer relationship survived; the foreign-ownership stake did not.
• The IPO (May 2026). After withdrawing the 2024 attempt (RW filed October 2025) and refiling, Cerebras priced at $185 on 2026-05-14 and closed 2026-05-15. The stock spiked (52-week high $386.34) before settling near $237.
• The OpenAI pivot (December 2025–January 2026). The MRA, the $1B loan, and the penny warrant fundamentally re-anchored the forward revenue story from UAE-sovereign to OpenAI — the dominant change to the thesis.
• The cloud-capex pivot (2025). Capex 16x’d to $383M as the company shifted from selling hardware to operating owned inference capacity — strategically rational, but margin-dilutive so far.
• Sell-side initiation wave (June 8, 2026). Ten brokers initiated coverage post-quiet-period, unanimously Buy/Overweight, PTs $250–$340.

Headwinds: decelerating growth, falling gross margin, widening losses, negative FCF, single-customer forward dependence, export-control/CFIUS overhang on the UAE base, a TSMC-allocation dependency, a peaking AI-capex capital cycle, and a near-term lock-up/early-release supply overhang against a 14% float.

The IPO journey is itself a data point. Cerebras first filed publicly to go public in September 2024 and could not complete the offering for roughly twenty months — an unusually long gestation that reflected the CFIUS review of the G42 ties and, implicitly, investor wariness of the concentration. The company bridged the gap with two large, escalating private rounds (Series G at $36.23 in late 2025, Series H at $89.02 in early 2026) and only listed once the OpenAI MRA gave it a non-UAE forward narrative to sell. In other words, the OpenAI deal was arguably the precondition for the IPO — which underscores how central that single relationship is to the entire equity story. The post-IPO price action (a spike to $386 then a settling to ~$237, roughly 28% above the $185 issue) is consistent with a thin-float, high-momentum name finding a level well above its private marks but below its first-day euphoria; it tells us about supply/demand technicals and sentiment, not about intrinsic value.

Verdict (Changes): the period’s changes shifted the concentration from UAE to OpenAI rather than reducing it, monetized a balance-sheet (IPO cash) without yet fixing the income statement, and added a large unhedged cloud-capex bet. On balance these developments raise the stakes of the single-contract thesis rather than de-risking it.

9. Risk Analysis

The risk matrix skews high-likelihood / high-impact across the items that matter most — concentration, margins, competition, and valuation — which is unusual and itself a flag. In most equities the dominant risks are medium-likelihood or medium-impact, leaving a margin of safety in the diversification of risk; here the top risks are correlated and concentrated in the same root cause. Concentration risk, OpenAI-execution risk, and valuation risk are not independent — they are three faces of one fact: the price embeds the flawless conversion of a single customer’s conditional contract. If OpenAI slips, the concentration risk and the valuation risk fire together, with no offsetting diversification to cushion the equity. That correlation is why the bear case is a large-drawdown case rather than a modest-derating case, and why the near-term lock-up release (up to ~171M shares into a 30M float) is dangerous: it would add a technical supply shock precisely when the fundamental risks are most likely to be re-priced (around the first one or two public quarters). The one genuine mitigant is the balance sheet — ~$6.5B of post-IPO cash makes a solvency failure remote and buys years of runway to execute or diversify. The risk here is overwhelmingly to the multiple and the per-share value, not to corporate survival; an investor’s loss function is dominated by valuation de-rating, not bankruptcy.

10. Valuation Discussion (Embedded Expectations)

No price target; embedded-expectations and scenario framing only.

Share count and EV. Basic post-IPO shares are 215.1M (30.0M Class A + 185.1M Class B). But the economic count is much higher: the 33.4M OpenAI penny warrant alone adds ~13% once vested, before options (28.4M @ $4.97), RSUs (~24.7M), founder PRSUs (9M), and other warrants. At ~$237:

EV/EBITDA is N/M (operating loss −$145.9M). The basic headline understates true economic dilution by ~25–40%.

Comp table (peer multiples, public data):

Cerebras trades at 3.5–5x the EV/sales of profitable hyperscale-AI peers and ~1.3–1.8x even ARM — the market’s richest large-cap semi comp — despite the lowest margins and the only operating losses in the set. No comp justifies the multiple on trailing economics. ARM is the closest structural analog: a richly-valued, thin-float scarcity story where the sell-side mean target sits below spot — the same momentum signature now visible in CBRS’s $250–340 initiations.

Embedded expectations (reverse-DCF). The valuation is 100% a forward construct. To justify the ~$64B economically-diluted EV at a mature ~25x FCF multiple and ~12% discount rate, Cerebras must generate roughly $5–6B of FCF ~6 years out (~2032) — which at a 20–25% FCF margin (it runs −39% today) implies ~$22–28B of revenue. Mapped onto the backlog, the $24.6B RPO recognized over ~2026–2033 averages ~$3–4B/year — meaning the full RPO is already the base-to-bull case, not upside. The price additionally requires capacity revenue beyond OpenAI and a gross-margin reversal from 39% toward semi-like 50–60%. Four things must all go right: (1) capacity delivery on committed dates (cash-hungry); (2) OpenAI does not terminate/de-scope; (3) margins reverse their slide; (4) diversification de-risks concentration.

Scenario analysis (value relative to today — no target):

The skew is the inverse of a typical post-IPO setup: base ≈ today, bull is the only meaningful upside, and the bear carries large downside because the thesis rests on one customer and one contract while the cash build is front-loaded — a thin margin of safety.

The dilution waterfall the headline hides. A buyer paying ~$237 on the 215M basic count is paying ~$51B; the economic claim on the business is meaningfully smaller per share once the full instrument stack is counted. On top of basic: ~28.4M options (WAEP $4.97 — deeply in-the-money, near-certain to exercise), ~24.7M service/liquidity RSUs, 9M founder PRSUs (vest on $75B/$150B/$250B market caps), the 33.4M OpenAI penny warrant (vests on capacity delivery — i.e., as the bull case materializes), a 2.7M AWS warrant, ~3.5M G42 warrants, plus 42.65M shares reserved under the 2026 Plan and 3.55M under the ESPP, both with automatic annual increases. The economically-diluted count is ~292M and the fully-loaded count approaches ~308–354M — 35–65% above the basic headline. Critically, the penny warrant vests precisely in the bull scenario, so the very success that justifies the multiple also dilutes the per-share claim — a structural headwind the $300 consensus targets largely ignore.

Why the own-history valuation percentile is useless here. Standard practice would anchor a multiple against the stock’s own trailing range; Cerebras IPO’d on 2026-05-14, so there is no trailing band — every “percentile” is undefined and any screen reporting one is fabricating it. The only valid anchors are (a) the private-round ladder ($14.66 → $36.23 → $89.02 → $185 IPO), which shows a ~13x mark expansion in under two years, and (b) the public peer set, against which Cerebras is the most expensive name on the weakest economics. Both anchors point the same direction: the price embeds extraordinary forward success.

What the market is pricing correctly: the OpenAI RPO is real, contracted, and enormous (~48x revenue); wafer-scale inference could be genuinely differentiated; ~$6.5B post-IPO cash funds the near-term build; inference is the larger long-run pool. What it is pricing incorrectly / under-weighting: single-customer concentration; falling margins while the multiple assumes expansion; deeply negative FCF; widening non-GAAP losses; the GAAP “profit” that is a non-cash artifact; 35–65% hidden dilution above the basic headline (vesting into the bull case); and sell-side initiations that extrapolate the bull ramp as the base case.

11. Variant Perception

Consensus. Unambiguously bullish. Ten brokers initiated June 8, 2026, all Buy/Overweight, PTs $250–$340 (Citi $340, Craig-Hallum $325, UBS/Mizuho/Rosenblatt/Needham $300, Barclays $280, TD Cowen $275, Wedbush $270, Morgan Stanley $250). The consensus narrative: a differentiated wafer-scale architecture, a generational $24.6B backlog anchored by OpenAI, and exposure to the fastest-growing slice (inference) of the largest secular capex wave in tech.

The ARM analogy, and why it is the right one. The closest market template is ARM Holdings: a richly-valued, high-multiple, thin-float, scarcity-driven IPO into a structurally important franchise, where the sell-side initiated bullishly and the mean target ultimately sat below spot for an extended period as the multiple did the work the fundamentals could not yet justify. CBRS rhymes — ~14% float, ~90–125x sales, unanimous Buy initiations clustered just above spot — but with two adverse differences: ARM enjoys ~92.5% gross margins and a genuine licensing moat (the entire mobile ecosystem is built on its ISA), whereas Cerebras has 39% margins, no ecosystem lock-in, and one customer. If ARM is the bullish template, Cerebras is the lower-quality, higher-concentration version of it — which argues for more multiple risk, not less.

Strongest bull case. Cerebras is the only company shipping commercial wafer-scale compute, with a real inference-speed edge in the workload that will dominate long-run AI demand. The OpenAI MRA (>$20B, 750 MW) is a frontier-lab validation that simultaneously funds the build (the $1B loan), aligns incentives (the penny warrant), and provides a multi-year revenue runway. AWS as a partner and the cloud pivot open a path to recurring, diversified revenue. With ~$6.5B of post-IPO cash, the company can fund the capacity build to inflection. If the RPO converts and margins scale, today’s multiple is justified by 2030 earnings power.

Strongest bear case. The entire valuation rests on one customer and one contract, recognized $0 to date, milestone-terminable, against a balance sheet that hands that same customer ~$8B of equity and owes it $1B. Beneath the backlog, the demonstrated business is a sub-scale hardware company with falling margins, widening losses, negative FCF, no moat (it sits outside CUDA), and 86% related-party UAE concentration, IPO-ing into a peaking AI-capex capital cycle at ~90–125x sales — 4–5x peers on the weakest economics in the group. Governance is founder-controlled with a market-cap-based PRSU grant, and a 14% float faces a large near-term lock-up release.

The 3–5 assumptions that matter most: (1) the OpenAI RPO converts to recognized revenue on schedule and OpenAI does not terminate; (2) gross margin reverses its decline toward semi-like levels with scale; (3) the company diversifies beyond OpenAI/UAE; (4) wafer-scale stays competitive against NVIDIA’s next architectures and Groq/hyperscaler ASICs; (5) the AI-capex cycle does not roll over before the build pays off.

What would falsify each side. Bull falsified by: any OpenAI de-scope/slip, a sub-35% margin print, continued FCF burn with no diversification, or a competitor matching wafer-scale inference economics. Bear falsified by: two to three quarters of OpenAI capacity converting at expanding margin plus a signed second non-UAE/non-OpenAI anchor — i.e., concentration genuinely breaking and operating leverage finally appearing.

12. Fact vs. Interpretation Table

13. Open Questions

1. OpenAI MRA recognition cadence and termination triggers — exact capacity-delivery milestones, service-level thresholds, and the precise conditions under which OpenAI can de-scope/terminate and call the $1B loan.
2. Path of gross margin — does the cloud build reach scale economics that reverse the 39% slide, and by when?
3. Diversification — is there any signed or pipeline non-UAE, non-OpenAI anchor customer of size?
4. TSMC allocation — what wafer allocation, if any, is contractually secured for the OpenAI ramp, and on what node?
5. Lock-up early-release mechanics — what triggers free the ~171M shares, and over what window relative to the 30M float?
6. Capex intensity — total capital required to deliver 750 MW (and the 2 GW Additional Capacity), and the financing mix (cash vs. more customer loans vs. equity)?
7. Penny-warrant accounting — how the OpenAI warrant is treated (contra-revenue vs. SBC vs. other) and its drag on reported gross margin as it vests.
8. First public quarter (Q2/Q3 2026) — does any OpenAI revenue begin to recognize, and what does the un-prepayment-flattered cash flow look like?

14. What Must Be True

Bull case — what must be true: OpenAI converts the $24.6B RPO to recognized revenue roughly on schedule without de-scoping; gross margin reverses toward 50%+ as the cloud build scales; the company signs at least one large non-UAE/non-OpenAI customer; wafer-scale stays ahead of NVIDIA’s next architectures and Groq/hyperscaler ASICs in inference; and the AI-capex cycle holds long enough for the build to pay off. Falsification test: if, over the next 3–4 quarters, OpenAI revenue fails to begin recognizing on plan, or gross margin prints below ~35%, or no diversifying anchor customer is signed, the bull thesis is broken.

Bear case — what must be true: the single-contract concentration bites (OpenAI slips, de-scopes, or terminates); margins stay sub-40% as competition commoditizes inference pricing; FCF stays deeply negative through the build; and the ~90–125x multiple compresses toward peer levels as the IPO scarcity premium fades and the lock-up releases supply. Falsification test: if, over the next 3–4 quarters, OpenAI capacity begins converting to recognized revenue at expanding gross margin and a credible second anchor customer is signed, the bear thesis is broken.

The two falsification tests are deliberately symmetric and near-term-observable: the first two to three public quarters will largely resolve which case is right.

15. Source Appendix

See the Source Appendix below for the full citation list. Primary sources: Cerebras Systems Form 424B4 prospectus (filed 2026-05-14, SEC CIK 0002021728); Form 8-K closing (2026-05-15); Form S-1/S-1-A series (2024-09-30 through 2026-05-11); Forms 3/4 and Schedules 13D/13G (May–June 2026). Quantitative cross-checks via SEC EDGAR XBRL and public market data; peer multiples from public filings of NVDA, AMD, AVGO, ARM, MRVL, TSM, QCOM. Consensus from sell-side initiations (June 8, 2026). All figures reconciled to the 424B4 where material.

This article is independent research and general information only — not investment advice. The main analysis carries no recommendation and no price target; the “Claude’s Take” block is a clearly-labeled subjective exception.

APPENDIX A — Standard Diligence Questionnaire

Cerebras Systems, Inc. (NASDAQ: CBRS) — as of 2026-06-10

Supplemental diligence appendix. Labels: Fact / Interpretation / Assumption. Where a question does not map to the business model, the correct analog is given.

General

What thoughtful questions have other investors asked about this company? The recurring institutional questions cluster on five points: (1) How real and how terminable is the OpenAI $24.6B RPO? — recognition cadence, milestone triggers, and de-scope rights. (2) Is the customer concentration improving or just rotating from UAE-sovereign (G42/MBZUAI, 86% of 2025 revenue) to OpenAI? (3) Why is gross margin falling (42.3%→39.0%) and when does the cloud build show operating leverage? (4) What is the true economic share count given the OpenAI penny warrant and PRSU dilution? (5) Is the GAAP profit real? (No — it is a non-cash forward-contract gain.) Bears additionally ask whether wafer-scale is a durable advantage or a niche architecture against NVIDIA/CUDA.

Cyclicality & Earnings Nature

Are earnings at a cyclical high or low? (Interpretation) There are no operating earnings — the company runs operating and non-GAAP losses. Revenue is at an all-time high but is driven by a peaking AI-capex capital cycle and a few large contracts, so the revenue base is arguably cyclically and idiosyncratically elevated.

Driven by the external environment or internal actions? (Interpretation) Both — the AI-capex boom is the external tailwind; internally, revenue is the product of a small number of negotiated mega-deals (G42, MBZUAI, now OpenAI) rather than broad market demand.

How stable are revenues? (Fact/Interpretation) Unstable and lumpy. 70% is point-in-time hardware; 86% comes from two related customers; year-over-year growth swings (+269% then +76%). A single shipment timing can move a quarter materially.

Outlook for products/services? (Interpretation) Demand outlook (inference) is strong; competitive outlook is hostile. The forward book is enormous ($24.6B RPO) but concentrated and conditional.

How big will this market be? (Fact, company-cited) Inference addressable ~$66B → ~$292B by 2029 (~45% CAGR); total data-center capex ~$679B → ~$1.7T by 2030. Growing, global. (Interpretation) The market is large; Cerebras’s capturable, profitable share is the open question.

Business Quality & Competitive Moat

Is the industry getting more or less competitive? (Interpretation) More. Merchant challengers (AMD, Groq, SambaNova), hyperscaler in-house ASICs (TPU, Trainium, MTIA, Maia), and NVIDIA’s accelerating cadence all converge on inference.

How profitable is the business (ROIC, ROE)? (Fact) Not profitable operationally — operating margin −28.6%; non-GAAP net loss −$75.7M. Reported ROE (38%) and the GAAP “profit” are artifacts of a non-cash forward-contract gain and a negative-equity denominator; both are meaningless here. ROIC is negative.

How profitable is the industry — competitors, barriers to entry? (Fact/Interpretation) The incumbent is extraordinarily profitable (NVIDIA ~75% GM, ~60% op margin) behind the CUDA barrier; merchant challengers are mostly sub-scale and unprofitable. Barriers protect NVIDIA, not Cerebras.

Can the business be easily understood? (Interpretation) The product is simple to state; the financials (forward-contract liability, mezzanine preferred, penny warrants, three share classes, customer-as-lender-as-owner) are unusually complex and obscure the underlying loss-making operations.

Can it be undermined by foreign low-cost labor? (Interpretation) Not the relevant threat — the threats are foundry dependence (TSMC) and better-capitalized domestic/hyperscaler competition, not labor arbitrage.

Do brands matter? (Interpretation) Marginally — “Cerebras” has technical cachet, but purchasing is driven by performance/$, ecosystem, and supply, not brand.

Nature of competition? (Interpretation) Performance/price, software ecosystem, capital access, and foundry allocation. Cerebras competes mainly on raw inference latency — the narrowest and most contestable axis.

Customers’ switching costs? (Interpretation) Run against Cerebras. The dominant switching cost in the industry is CUDA lock-in, which benefits NVIDIA; Cerebras sits outside it and must persuade customers to leave their incumbent stack.

Financial Condition & Balance Sheet

Assets not fully recognized on the balance sheet? (Interpretation) The $24.6B RPO is an off-balance-sheet future asset; the OpenAI relationship’s value is not capitalized. Conversely, IP/know-how is expensed via R&D.

Off-balance-sheet liabilities? (Fact) A $2.2B, 10-year Canada data-center lease (March 2026) and an AWS warrant commitment; large purchase/capacity-delivery obligations under the OpenAI MRA.

How conservative is the accounting? (Interpretation) Mixed. Revenue recognition is standard (point-in-time hardware, over-time cloud), but the GAAP net income is heavily distorted by fair-value remeasurement of the forward contract, and the prepayment-driven 2024 cash flow flatters the cash narrative. The reporting is technically compliant but easy for screens to misread.

How CapEx-hungry is the business? (Fact) Very, and increasingly so — capex 16x’d to $382.7M in 2025 as it builds owned cloud capacity; the 750 MW / 2 GW build implies substantial further capital intensity.

Capital Allocation & Management

How much FCF does the business generate; how is it used; philosophy? (Fact) Negative FCF (−$393M in 2025). No FCF to allocate; the company consumes capital, funded by preferred raises, the IPO, and customer loans/prepayments. Philosophy is growth-at-all-costs build-out.

Significant acquisitions recently? (Fact) None.

Buying back shares? (Fact) No buybacks of public stock; a December 2025 pre-IPO employee tender ($78.1M at $36.23) provided insider liquidity but is not a return of capital.

Issuing large amounts of new shares to insiders? (Fact) Yes — a 9M-share founder PRSU mega-grant (market-cap hurdles $75B/$150B/$250B), make-whole and annual RSUs, plus ~$8B of penny warrants to OpenAI. Heavy dilution is a defining feature.

Compensation policy of directors/management? (Fact) ~92% equity-weighted NEO pay (Feldman $11.75M, Lie $11.57M, 2025); incentives keyed to share price/market cap, not operating returns — a misalignment with disciplined capital allocation.

Motivations of management? (Interpretation) Founder-controlled (20:1 super-vote, ~99.2% voting power) with comp tied to market-cap escalation; motivated to maximize valuation and retain control, not necessarily to optimize per-share economics for outside Class A holders.

Valuation & Market Data

Is the stock an ADR, MLP, or K-1 issuer? (Fact) No — a US C-corp common stock (Class A, NASDAQ: CBRS). No K-1.

Dividend policy? (Fact) None; the revolver restricts dividends; none anticipated.

How profitable is the business? (Fact) Unprofitable operationally (see above).

Is net income diverging from cash from operations? (Fact) Yes, dramatically and in both directions. 2024: GAAP net loss −$481.6M vs. OCF +$452.0M (prepayment-driven). 2025: GAAP net income +$237.8M (non-cash gain) vs. OCF −$10.1M. Neither year’s net income tracks cash — a textbook quality-of-earnings divergence.

Risks & Downside

What factors would cause the stock to decline? (Interpretation) OpenAI de-scope/slip/termination; a sub-35% margin print; continued FCF burn with no diversification; multiple compression as IPO scarcity fades; the lock-up early-release flooding the 14% float; a competitor matching wafer-scale inference; an AI-capex cycle roll-over; or any CFIUS/export-control action on the UAE base.

Risk of a catastrophic loss? (Interpretation) Plausible in the bear case — a single-contract thesis where the contract is terminable. The ~$6.5B post-IPO cash mitigates insolvency risk, but the equity could de-rate severely (the multiple is 4–5x peers).

Chance of a total loss? (Interpretation) Low in the near term — ample cash and a real (if concentrated) revenue base make a zero unlikely soon. Total loss would require OpenAI collapse and failure to diversify and exhaustion of the cash war chest — a multi-year tail, not a base case.

Recent News & Events

Has the business environment changed recently? (Fact) Yes — the OpenAI MRA (Dec 2025), the $1B loan and penny warrant, the $185 IPO (May 2026), the cloud-capex pivot, and the CFIUS-driven unwinding of G42’s equity stake (March 2025) all reshaped the company within ~18 months.

Significant acquisitions? (Fact) None.

Change in accounting policies? (Fact) None material flagged beyond the extinguishment of the forward-contract liability (an event, not a policy change) and standard IPO-related re-classifications (preferred conversion, RSU settlement).

Recent changes — new markets, facilities, management? (Fact) New owned-cloud facilities (incl. the $2.2B Canada lease); a pivot toward OpenAI/AWS and the inference-cloud market; an IPO-timed founder PRSU grant; and a board with three VC-affiliated directors (Benchmark, Eclipse, Foundation).

APPENDIX B — Source Appendix

Cerebras Systems, Inc. (NASDAQ: CBRS) — Research Sources, as of 2026-06-10

Primary sources prioritized. All material figures reconciled to the Form 424B4 prospectus. URLs accessed 2026-06-10.

Primary — SEC Filings (CIK 0002021728)

1. Form 424B4 — Final IPO Prospectus (filed 2026-05-14). The principal source for this report: business description, risk factors, MD&A, audited 2022–2025 financial statements, customer concentration, RPO, capitalization/dilution, executive compensation, principal stockholders, related-party transactions, OpenAI MRA, G42/CFIUS history, share classes/voting. https://www.sec.gov/Archives/edgar/data/2021728/000162828026035214/cerebras-424b4.htm
2. Form 8-K — IPO Closing (filed 2026-05-15). https://www.sec.gov/Archives/edgar/data/2021728/000162828026035605/closing8-k.htm
3. Form S-1/A (2026-05-11) and S-1/A (2026-05-04) — pre-pricing registration amendments. https://www.sec.gov/Archives/edgar/data/2021728/000162828026033143/cerebras-sx1a2.htm
4. Form S-1 (2026-04-17) — re-filed registration statement. https://www.sec.gov/Archives/edgar/data/2021728/000162828026025762/cerebras-sx1april2026.htm
5. Form S-1 (2024-09-30) — original (delayed) IPO registration; Form RW withdrawal (2025-10-03). https://www.sec.gov/Archives/edgar/data/2021728/000162828024041596/cerebras-sx1.htm
6. Form 8-A12B (2026-05-11) — Exchange Act registration of Class A common stock.
7. Forms 3 / 4 (May 2026) — initial insider ownership and IPO-window transactions (codes F/J; no open-market buys/sells).
8. Schedules 13D / 13G (2026-05-22 to 2026-06-05) — institutional/insider beneficial-ownership stakes.
9. Forms S-8 (2026-05-14) — registration of shares under the 2026 Incentive Award Plan / ESPP.

EDGAR filing index: https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0002021728&type=&dateb=&owner=include&count=100

Primary — Quantitative cross-checks

10. SEC EDGAR XBRL company facts (CIK 0002021728) — authoritative US-filer financial concepts.
11. yfinance / Yahoo Finance — live price (~$237), market cap, 52-week range ($196.73–$386.34), shares (Class A float). Unofficial; reconciled to the 424B4 for all material figures.
12. Multi-period financial statements (income statement, balance sheet, cash flow), 2022–2025 — reconciled to the 424B4 audited statements.

Secondary — Consensus & Market Color

13. Sell-side initiations, 2026-06-08 (post-IPO quiet-period expiry): Citigroup ($340, Buy), Craig-Hallum ($325, Buy), UBS ($300, Buy), Mizuho ($300, Outperform), Rosenblatt ($300, Buy), Needham ($300, Buy), Barclays ($280, Overweight), TD Cowen ($275, Buy), Wedbush ($270, Outperform), Morgan Stanley ($250, Overweight). Used for consensus/variant-perception framing only — not as a price-target basis.

Peer / Industry Cross-Reference

14. Public filings and disclosures of peer companies (NVDA, AMD, AVGO, ARM, MRVL, TSM, QCOM) used for comparative multiples, hyperscaler-capex figures, and competitive/capital-cycle framing.

Analytical Frameworks

16. Competition Demystified (Greenwald & Kahn) and Capital Returns (Marathon Asset Management / Chancellor), via the repository’s investment-research-frameworks skill — applied to the moat-type, barriers-to-entry, and capital-cycle analysis.

Company filings (424B4, S-1, 8-K, Forms 3/4, 13D/G) are the source of record; public aggregator data was used for orientation and reconciled to filings. Sell-side targets are reported as consensus color only and do not inform the author’s view, which carries no price target.