06-reference

mostlymetrics cerebras ipo s1 breakdown

2026-05-12·reference·source: Mostly Metrics·by CJ Gustafson
cerebrasai-infrastructurechipsipos1-breakdowncustomer-concentrationopenaiwafer-scaleinnermost-loop

Cerebras IPO: S1 Breakdown - @CJ Gustafson

Why this is in the vault

Cerebras is filing to IPO this week at ~$33B implied market cap on $510M of 2025 revenue (76% Y/Y growth), a $24.6B remaining performance obligation backlog that is almost entirely OpenAI, and an 86%-of-revenue concentration on two related Abu Dhabi entities (MBZUAI + G42). This is the cleanest public-comp data point yet for the Layer 2 (chips) tier of the [[01-projects/investing/theses/2026-05-12-innermost-loop-ai-infrastructure|Innermost Loop AI Infrastructure thesis]] filed earlier today. It is also the single document that resolves whether wafer-scale is a Layer 2 position candidate or a structural-risk passing curiosity. CJ's S1 breakdowns are signal-dense and source-cited.

Issue contents

  1. Sponsor block - Brex agentic finance pitch (declared, not part of editorial).
  2. What does Cerebras do - wafer-scale engine (WSE) explainer: kept the silicon wafer intact instead of slicing it; chip is 58x the size of NVIDIA B200, dinner-plate-sized; defect-routing built in; manufactured with TSMC on a custom process.
  3. Why now - Modern reasoning models do most of their thinking during inference, not training. Inference became the bottleneck in early 2025. Cerebras built for that bottleneck four years before it existed. Revenue: $25M (2022) -> $79M (2023) -> $290M (2024) -> $510M (2025). 20x in 3 years.
  4. Key stats scorecard - Revenue, gross margin (39%, down from 42%), GAAP net income ($238M but distorted by a $363M one-time forward-contract gain), operating cash flow (-$10M), RPO ($24.6B = 48x last year's revenue), employees (708), target valuation ($33B at $155 midpoint), Series H step-up (74% in four months).
  5. How they make money - Hardware (70% of revenue, 43% GM) vs Cloud (30% of revenue, 30% GM). Cloud is the growth line and the lower-margin line. Quarterly cloud GM in 2025: 68% -> 26% -> 16% -> 21%. Inverse of the standard hardware-to-SaaS pivot story.
  6. Customer concentration - MBZUAI 62% of 2025 revenue, G42 24%. AR concentration even worse: MBZUAI 78% of AR at year-end 2025. G42 prepaid $640M in 2024 (which is what made 2024 OCF look healthy) and received warrants worth ~$544M for ~$35,000 in exercise cost.
  7. OpenAI warrant - 33.4M shares at $0.00001/share = ~$5.18B in equity for ~$334 total exercise cost. Plus a $1B working capital loan, 6% interest, secured. Vesting in three tranches, one of which auto-triggers at $40B market cap (i.e., 21% above IPO price).
  8. Financials detail - R&D 48% of revenue ($243M, +54% Y/Y); S&M +237% Y/Y building cloud GTM on top of hardware GTM; G&A down 31% Y/Y on lower legal costs; operating loss ($146M); $701M cash at year-end, $1B OpenAI loan on top (but clawback-able).
  9. Red flags - Material weakness disclosed in revenue recognition and IT general controls; single-foundry TSMC dependency with no long-term supply commitment; UAE geopolitical concentration; cloud business 18 months old and they're betting on it.
  10. Cap table + Series H - Founders 8.3% combined; Series G $36.23 (Sept 2025) -> Series H $89.02 (Jan 2026) -> IPO $155 midpoint = 4.3x in a year, almost all on the OpenAI deal signing.
  11. Voting structure - Class B (insiders) holds 99.2% of voting power post-IPO; public Class A holders get 14% economic share and 0.8% of voting power.
  12. Valuation - ~10-13x EV/forward revenue. Premium to every peer except NVIDIA. CoreWeave trades ~7x; AMD ~6-7x. NVIDIA-style multiple (~20x = $50B) is what the OpenAI narrative will be selling.
  13. CEO history - Andrew Feldman had a 2007 SEC settlement and DOJ guilty plea for circumventing accounting controls at Riverstone Networks; disclosed in S1; standing next to a present-day material weakness on revenue recognition makes it more notable than it would otherwise be.

Mapping against Ray Data Co

(a) Innermost Loop thesis - does this change Layer 2 sizing?

Filed thesis ([[01-projects/investing/theses/2026-05-12-innermost-loop-ai-infrastructure]]) calls Layer 2 (chips) "lowest forward leverage, most picked-over" and recommends NVDA + AVGO as structural holders, not new entries. The Cerebras S1 reinforces that call rather than reopening it. Five reasons:

  1. Customer concentration is disqualifying for an investing thesis. 86% of revenue from two related entities under one sovereign AI program is not a public-equities position; it's an Abu Dhabi geopolitics position with chip optionality. If US export framework on advanced semis tightens further, a meaningful chunk of the backlog gets stranded. That's not a Layer 2 candidate; that's an event-driven trade.

  2. The cap-table is upside-down. OpenAI gets $5.2B in equity for $334 in exercise cost. G42 got $544M for $35,000. New IPO buyers pay full $155, get one vote per share against insiders' 20-vote Class B, and watch the largest customer get auto-tranches of equity as the stock works higher. The economic-vs-voting structure is the opposite of what you want as a public minority holder.

  3. The cloud margin profile is the inverse of the standard pivot. Hardware GM 43%, cloud GM 30%, and the OpenAI backlog converts mostly to cloud revenue. Mix shift is margin-dilutive going forward, not accretive. Bull case (utilization recovers as backlog burns) is plausible but unproven; bear case (structurally lower than hyperscaler GM forever) is also live.

  4. Material weakness in revenue-recognition controls at a $33B IPO with a 48x-of-revenue backlog under a complex MRA is exactly the disclosure footnote that goes in a textbook five years later. Read alongside the CEO's prior plea on circumventing accounting controls, the audit-quality risk is real even if the audit committee is on it.

  5. Single-foundry TSMC dependency with no long-term supply commitment is a strictly worse version of the supply-chain risk NVDA already carries. NVDA also relies on TSMC, but at a scale and priority level Cerebras cannot match. If TSMC reallocates, Cerebras has no second source for a custom wafer-scale process.

Conclusion: Cerebras is NOT added to candidates/. Layer 2 candidate list stays NVDA + AVGO. The S1 is filed as the canonical "why we don't chase the wafer-scale story" reference; if the founder wants Layer 2 chip exposure with the inference-shift thesis baked in, NVDA's substrate position remains the cleaner play. The narrower TSM bet (the foundry layer underneath both NVDA and Cerebras) is the next position to consider, not Cerebras itself.

The interesting tactical note: the Tranche 2 OpenAI warrant vests at $40B market cap on a 30-day rolling average. From a $33B open, that's a ~21% pop, which is one good IPO week. If it triggers, ~$864M of stock auto-transfers to the largest customer. That is information that should flow into any decision to short or hedge post-pop, but it is not a fundamental case for going long.

(b) Harness-engineering thesis cluster - compute-supply economics data point

The compute-substrate question Thompson raised in [[06-reference/2026-05-11-stratechery-inference-shift-agentic]] - that agentic inference cares about memory hierarchy and capacity, not on-chip bandwidth - gets a financial-side data point here. Cerebras built the bandwidth-optimized substrate for answer-inference (44GB on-chip SRAM at 21 PB/s) and is now pivoting most of its IPO narrative to a cloud business serving a workload (OpenAI agentic) that, per Thompson, plays to a different chip profile than the one Cerebras built. The Q3 2025 cloud GM crater (16%) is what shows up in financials when you build capacity ahead of a workload regime that may or may not actually need your specific hardware advantage.

This is corroborating evidence that the wafer-scale wager is structurally a single-regime bet (answer-inference dominance), and the regime may be ending just as the IPO prices. Not a thesis-changing input, but a useful "watch this trade unwind in public" data point as the harness-engineering thesis matures.

(c) MAC framework

Weak overlap, as expected. The only MAC-relevant signal is the disclosure pattern itself: an S1 that names a material weakness in revenue-recognition controls and a CEO with a prior accounting-controls plea, going public at a premium-to-peers multiple, is a small-bet-relevant case study in "what disclosures actually mean when the market is excited" - which lives more in the founder's general business-judgment training set than in any active MAC project.

Strength: STRONG. Resolves the Layer 2-Cerebras candidacy question, corroborates the inference-shift harness thesis, and provides a canonical S1 to cite when this topic resurfaces. No decision-required.

Related