CBRSCerebras Systems Inc.

CBRS — dossier, 2026-09-21

Slot names-ai-capex-3. Model-drafted unattended by the cloud research routine; nothing here is human-verified. Sources and locators are in relationships.yaml; figures below carry theirs inline. Anything from background knowledge is marked [background].

Price reference: $208.315 close, 2026-09-21 (ROIC.ai get_latest_stock_price NASDAQ:CBRS; volume 5,000,300 shares, so roughly $1.04bn traded that day). IPO priced at $185.00 on 2026-05-14 and closed 2026-05-15 with the over-allotment exercised in full (34,500,000 Class A shares); the stock is +12.6% on its issue price four months later.

Shares at 2026-08-05 (10-Q cover): 112,247,109 Class A + 111,601,424 Class B + 13,715,508 Class N = 237,564,041, so ~$49.5bn of market capitalisation. Add the two customer warrants (33,445,026 Class N at $0.00001, 2,696,678 at $100.00) and it is ~273.7m shares, ~$57.0bn.

This name is a recent IPO under the skill's §3 rule, so the prospectus risk factors and use of proceeds replace the 10-K read. There is no 10-K, no proxy and no DEF 14A to skim: the entire EDGAR history is two S-1s, one 424B4, two 10-Qs and three 8-Ks.

S1 — Source

watchlist (owner, Apple Stocks, 2026-09-20), cluster ai_semis, coverage sec_domestic, cik 2021728, carrying the watchlist note "recent IPO; S-1 is the primer". No Tier 0 promotion.

S2 — Kill test

Annual figures as the company states them in the prospectus MD&A overview; quarterly from the XBRL company facts (data.sec.gov/api/xbrl/companyfacts, CIK 0002021728).

The FY2025 GAAP profit is not earnings power. The company's own reconciliation removes stock-based compensation and the "change in fair value (extinguishment) of forward contract liability" to get to a non-GAAP net loss of $75.7m. Cerebras has never made money on operations.

Quarterly (10-Q XBRL):

Q2 2026 looks like a catastrophe and mostly is not. Stock-based compensation in the quarter was $377.0m ($15.4m in cost of revenue, $222.1m in R&D, $71.1m in sales and marketing, $68.5m in G&A) — the IPO triggered the liquidity condition on RSUs granted under the 2016 Plan. A further $44.3m was charged against revenue as amortisation of the OpenAI customer warrant asset. Adding both back: gross revenue ~$224.4m, cost of revenue ex-SBC ~$139.2m, gross margin ~38% — in line with FY2025's 39% — and an operating loss nearer ~$100m than $477m. These adjustments are derived here, not reported by the company; the company gives no non-GAAP figures in the 10-Q.

The real deterioration is sequential revenue: $193.4m → $180.1m, down 7% quarter on quarter, in a company the market prices for triple-digit growth. Part of that is the warrant contra-revenue starting to bite ($2.0m in Q1, $44.3m in Q2). Gross of it, revenue rose ~15% sequentially.

Balance sheet at 2026-06-30: cash $6,742.2m plus available-for-sale securities $1,179.4m = $7.92bn, against the OpenAI Working Capital Loan of $736.0m current and $182.2m long-term. Net cash ~$7.0bn. Total assets $11.63bn; stockholders' equity swung from −$578.7m at 2025-12-31 to +$9.15bn after the IPO and the conversion of preferred into 124,652,775 Class B shares. Equibles GetGoingConcernStatus: no flag in either examined 10-Q.

Cash runway: not the binding constraint. H1 2026 operating cash flow was −$47.5m and the company holds $7.9bn. What the $7.9bn is for is the constraint — this is now a data-centre business, and the prospectus says the cloud offering requires "significant data center capacity and capital investments for which we expect to require significant additional capital".

Share count: no five-year history exists — the company has been public for four months. Diluted weighted-average shares were 150,968,000 for Q2 2026. The dilution that matters is ahead: 36.1m warrant shares to two customers, a 42,650,268-share 2026 Plan reserve, and a 3,554,189-share ESPP that grows 1% a year.

Analyst coverage: at least five firms asked questions on the first call — UBS, TD Cowen, Barclays and two others in the turns read. No count obtained.

Verdict — survives on the balance sheet, not on the business. $7.9bn of cash against a ~$100m quarterly cash operating loss is years of runway, and there is a $25.4bn contracted backlog behind it. It does not die of funding. It dies, if it dies, of one customer.

S3 — Filings, gate-minimum (recent-IPO variant)

Read this run: the 424B4 prospectus (risk factor summary, the customer-concentration, internal- control and export-control risk factors, use of proceeds, MD&A overview, cover-page capital structure), the Q2 2026 10-Q (revenue note, Working Capital Loan note, warrants note, stock-based compensation note, balance sheet), and the Q2 2026 earnings call (turns 2–11 of 31). Not read: the September 2024 S-1, the Q1 2026 10-Q in full, the prepared remarks of the call (the transcript returns the CEO's opening as a single short turn and then jumps to Q&A — the guidance discussed below is the analyst's restatement of remarks this run did not see).

The three biggest risks management itself discloses

1. The revenue is three counterparties, and one of them is also the lender. From the risk factors: "Group 42 Holding Ltd (together with its affiliates, "G42") accounted for 24.0% and 85.0% of our total revenue for the years ended December 31, 2025 and 2024, respectively" and "in the year ended December 31, 2025, Mohamed bin Zayed University of Artificial Intelligence ("MBZUAI") accounted for 62.0% of our total revenue." Those two are 86% of FY2025 revenue and the company states "G42 and MBZUAI are considered related parties with respect to each other" — so it is one Abu Dhabi relationship, not two customers. Receivables are worse: MBZUAI was 77.9% of the balance at 2025-12-31, G42 91.0% at 2024-12-31. Going forward the concentration transfers rather than dissolves: the December 2025 OpenAI MRA "represents a substantial portion of our projected revenues over the next several years".

This independently confirms, from the primary document, the third-party teardown figures the cluster pass recorded as search-summary evidence (_clusters/cluster-ai-capex-2026-09-21.md, §S6 row for CBRS: "MBZUAI 62.0% + G42 24.0% = 86% of 2025 revenue"). Those numbers are right.

1. Material weaknesses in internal control over financial reporting, unremediated. "we identified certain material weaknesses in our internal control over financial reporting, including most recently for the years ended December 31, 2025 and 2024." They relate to "inadequate or missing resources who possess an appropriate level of expertise to timely review account reconciliations and identify, select, and apply U.S. generally accepted accounting principles ... including revenue recognition, inventory management and costing, data center assets accounting, and equity administration". Every one of those four areas is load-bearing for the numbers in S2 above, and the company is an emerging growth company taking the reduced reporting accommodations.

1. Export controls on a business whose historical customers are in the UAE. The offerings are subject to US export controls and "may be exported outside the United States only with the required export license or through a license exception". The prospectus records licences held for export to G42 and MBZUAI with conditions attached. A significant majority of FY2025 revenue came from customers headquartered in the United Arab Emirates, and the risk factors separately name the war in the Middle East, "where our strategic partners G42 and MBZUAI are headquartered".

Use of proceeds

Net proceeds ~$5,408.5m at $185.00 (up to $6,221.4m with the over-allotment, which was taken). Purpose: "general corporate purposes, including working capital, operating expenses and capital expenditures" — no debt to repay, no acquisition named. Of that, ~$416.1m goes straight back out as tax withholding on ~4,799,565 shares of RSUs vesting on the IPO. So roughly 8% of the raise funds the tax bill on its own compensation event.

Latest quarter

Covered in S2. The one forward number obtainable this run came from the Q&A, where a UBS analyst restated management's guide — "you did say that revenue would be up more than 3x next year" — and the CEO did not dispute it, adding that capacity will "end the year well over 10x our manufacturing capacity, and we already have contracted facilities three or four times more for growth in 2027". Treat the 3x as thin: it is an analyst's paraphrase of prepared remarks this run did not read.

Insider pattern

3spread: 173 Form 4 filings since 2025-09-21, essentially all post-IPO, and 11 Form 144 proposed sales since 2026-03-25, dated from 2026-08-21 and 2026-09-04. The IPO closed 2026-05-15; a customary 90-day lock-up would have expired around 2026-08-13 [background], and the first Form 144 lands eight days later. Insiders began filing to sell at the first legal opportunity. Amounts and sellers were not itemised this run.

Red-flag sweep

• Going concern: no flag (Equibles, 10-Q filed 2026-08-12).
• Auditor change / restatement: no NT 10-K, no Item 4.01 8-K in an EDGAR history four months long. The efts.sec.gov full-text sweep was not run (time).
• Executive changes: not checked (Equibles budget reserved for ASML).
• Customer concentration: read from the prospectus directly rather than from the Equibles tool, because the company has no 10-K for the tool to tag.
• Material weaknesses: flagged above. This is the red flag; it is disclosed, not hidden, and it is unremediated.

S5 — Ownership

13F, 2026-06-30 — the first quarter in which anyone reported CBRS at all. 297 filers, 67.2m shares. Every line is a new position by construction.

Two market-makers (Jane Street, Susquehanna) and a multi-strategy (Citadel) in the top six is what a newly listed, heavily optioned name looks like; it is inventory, not conviction. The conviction names are the pre-IPO venture holder (Eclipse) and the crossover funds (Coatue, Tiger, Atreides).

13D/G: six 13Gs filed 2026-08-14 for the 2026-06-30 period, plus a 13D/A filed 2026-08-18 by Eclipse Continuity GP I, LLC — the pre-IPO venture holder's control-group filing, not an activist position.

Insider ownership: no proxy exists yet, so no clean percentage. What the cover page gives instead is control: Class B carries 20 votes and "will represent approximately 99.2% of the voting power of our outstanding capital stock immediately following this offering." Public shareholders own economics, not votes.

Short interest (Equibles, FINRA): 8.14m (2026-06-15) → 13.53m → 10.42m → 12.90m → 15.84m (2026-08-14) → 13.33m (2026-08-31), days to cover 1.4–2.3. Against 112.2m Class A shares — and a tradeable float much smaller than that, since the IPO sold 34.5m — a low-double-digit-million short position is a large fraction of the free float, even though days-to-cover looks benign on the elevated post-IPO volume. The tool's 2026-09-15 estimate (~13.0m) is a model output, not FINRA data.

Shelf / ATM / buyback: none. No S-3 is on file (the company is not yet S-3 eligible [background]), no ATM, no repurchase authority, no dividend. The dilution here does not come from a shelf — it comes from 36.1m warrant shares owed to two customers and a 42.65m-share 2026 Plan reserve.

Who is on the other side: the pre-IPO venture and crossover holders who are now free to sell, and a short base sized against a thin float. The insiders' own Form 144s start eight days after lock-up expiry.

S8 — Valuation and sizing (proposed only; owner-only items OPEN)

At $208.315: market cap ~$49.5bn on 237.6m shares, ~$57.0bn fully diluted for the two customer warrants. Net cash ~$7.0bn → EV ~$42.5bn / ~$50.0bn diluted.

LTM revenue (FY2025 $510.0m − H1 2025 $202.8m + H1 2026 $373.5m) = $680.7m. EV/LTM revenue ≈ 62x basic, ≈ 73x diluted. There is no earnings multiple to compute.

The only forward anchor the filings give is the $25.4bn remaining performance obligation at 2026-06-30 — 37x LTM revenue — of which the OpenAI MRA is the bulk, and against which the unamortised customer-warrant asset (~$1.1bn already, up to ~$2.74bn at the OpenAI warrant's $82.02 grant value) will be charged as contra-revenue as it is recognised.

Out-year model. Revenue is built off the capacity commitments rather than a growth rate, because the growth rate is a backlog conversion question, not a demand question.

The base case is below the traded price. That is the honest read of this model: at $208 the market is already paying for the bull path, and the bear case is not a slow derating but a step-change tied to one contract. The spread between $55 and $300 is the widest of any name in this slot and it is the right shape for the risk.

⏳ Buy price and position size are owner-only and stay OPEN (user). Liquidity is ample today (~$1.04bn traded on 2026-09-21) but that is post-IPO volume on a thin float and should not be assumed durable.

S9 — Monitoring

Next earnings: Q3 2026 results, expected in the second week of November 2026 — derived from the two data points that exist (Q1 10-Q filed 2026-06-24, Q2 10-Q and call 2026-08-12), not from a calendar feed.

Filing feed: the Tier 0 EDGAR sweep covers cik 2021728. For this name the sweep matters more than for most: with no 10-K and no proxy, 8-Ks are the disclosure channel, and the first S-1/S-3 or lock-up-related filing will arrive there.

The leading indicator: the remaining performance obligation, quarter on quarter. $25.0bn at 2026-03-31 → $25.4bn at 2026-06-30. This single number is the OpenAI relationship made visible. It is disclosed in the revenue note every quarter, it leads revenue by years, and it is the only figure that would show an MRA renegotiation before revenue did. Secondary: the customer-warrant asset amortisation charged against revenue (which tells you the rate at which capacity is actually being delivered) and the current portion of the Working Capital Loan.

Trigger that converts this to a buy: two consecutive quarters in which the RPO holds or grows and a customer other than OpenAI, G42 or MBZUAI is disclosed above 10% of revenue — i.e. the backlog is real and the concentration is actually breaking, rather than transferring from Abu Dhabi to San Francisco. AWS revenue appearing as a disclosed line is the specific form that takes. Until then this is a single-contract instrument priced as a platform. The price and the size remain the owner's.