MRVLMarvell Technology, Inc.

MRVL — dossier, 2026-09-21

Slot names-ai-capex-2. 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: $244.25 close, 2026-09-18 (ROIC.ai get_latest_stock_price NASDAQ:MRVL; volume 23,549,789 shares, so roughly $5.8bn traded that day — liquidity is not a constraint). Shares outstanding 876.9m at 2026-08-21 (10-Q cover, dei EntityCommonStockSharesOutstanding), plus up to ~21.8m on conversion of the NVIDIA preferred → ~898.7m as-converted, ~$219bn market capitalisation as-converted.

S1 — Source

watchlist (owner, Apple Stocks, 2026-09-20), cluster ai_semis, coverage sec_domestic, cik

1. No Tier 0 promotion; queue.json has never promoted MRVL.

S2 — Kill test

All figures from the XBRL company facts of the SEC filings (data.sec.gov/api/xbrl/companyfacts, CIK 0001835632), fiscal years ending late January.

Quarterly, most recent six (10-Q XBRL):

Survives — but the headline profit number is a trap. FY26 GAAP net income of $2,670m is not earnings power: $1,901m of it landed in Q3 FY26 alone, the quarter of the automotive-ethernet divestiture, against $358m of operating income in the same quarter. The FY26 operating line — $1,323m on $8,195m — is the honest read, and it is the first positive operating year since FY23. GAAP gross margin has climbed five consecutive quarters, 50.2% → 53.2%.

• Growth. Revenue +42% in FY26 and +37% year-on-year in Q2 FY27, "primarily due to a 46% increase in sales from the data center end market which benefited from strong AI-related demand" (10-Q, Item 2, MD&A). Data center was $6,100.3m, 74% of FY26 revenue (10-K, Note 3, end-market table).
• Leverage. $5.0bn of borrowings at 2026-08-01, all senior notes, against $3.93bn cash — roughly $1.1bn net debt, comfortably under 1x any reasonable EBITDA measure. The $1.5bn revolver is undrawn and runs to 2030-06-30. The nearest maturity was the $500.0m 2026 notes, repaid at maturity out of the April 2026 issue.
• Dilution. Diluted weighted-average shares 676m (FY20) → 870m (FY26), +29% in six years, and that is before the ~21.8m as-converted preferred. Buybacks are not offsetting it: $400.0m repurchased in the first half of FY27 against $9.7bn of authority.
• Cash runway. Not applicable; operating cash flow positive every year in the table.
• Analyst coverage count. Not obtained — see "Evidence that was thin".

S3 — Filings, gate-minimum

Read: FY26 10-K (Item 1A risk factors, Item 7 MD&A, Notes 2/3); Q2 FY27 10-Q (Item 2 MD&A, Notes 4, 5, 10, plus Part II Item 1A); DEF 14A filed 2026-05-13 (beneficial ownership); the EDGAR form census since 2023-01-01. No transcript was read — guidance came from Equibles' structured rows off the 2026-08-27 8-K and call, not from the call text.

The three biggest risks management itself discloses

1. Customer and channel concentration, deepening. "during fiscal 2026, there were two customers (one distributor and one direct customer) whose revenues represented 10% or more of total net revenue... net revenue from our ten (10) largest customers... represented 82% of our total net revenue for fiscal 2026" (10-K, Item 1A, "Our sales are concentrated in a few large customers"). The Q2 FY27 table puts Distributor A at 44% of net revenue against 34% a year earlier, and Customer A flat at 16%. Credit follows: "The Company's accounts receivable was concentrated with four customers at January 31, 2026, who represented a total of 73 % of gross accounts receivable" (10-K, Note 2).
2. The AI capex cycle itself, named as a risk in management's own words. "the current level of capital expenditure (capex) on AI infrastructure may not be sustainable over the long term and a significant reduction in AI-related spending will likely harm our financial results" (10-K, Item 1A, "Unfavorable or uncertain conditions in the Data Center and Communications markets..."). With 74% of revenue in data center, this is the whole company.
3. Insourcing by the customers themselves. "AI-driven design tools may lower traditional barriers to entry in the semiconductor industry by enabling new market participants... to develop high-performance, custom semiconductor solutions in-house with reduced reliance on third-party chip suppliers" and "Given their dependence on semiconductor products to operate their data centers... some large customers may begin developing and making their own semiconductor solutions which could result in a loss of business for Marvell" (10-K, Item 1A). Marvell's business is custom silicon for hyperscalers; the risk factor describes its customers becoming its competitors.

Alongside these, geographic concentration rose sharply: sales to customers with operations in Asia were 84% of net revenue in Q2 FY27 against 76% a year earlier (10-Q, Item 2).

The quarter's structural change

Two acquisitions closed in the first ten days of FY27 — Celestial AI for $3.5bn (2026-02-02, photonic fabric for scale-up interconnect, with contingent cash and shares payable through fiscal 2029 on revenue milestones) and XConn for $469.0m (2026-02-10, PCIe/CXL switching). Goodwill went $11.1bn → $13.9bn; $2.5bn of the consideration was non-cash. Marvell then entered a cash-settled forward stock purchase contract with a $300.0m notional to hedge its own share price against the Celestial contingent consideration — the earn-out is payable partly in stock.

Red-flag sweep

• Auditor: Deloitte put to the 2026 AGM for appointment; no 8-K item 4.01 and no NT 10-K in the form census since 2023. No restatement signal found. This is a filing-census read, not the efts.sec.gov full-text sweep the skill specifies — see "Evidence that was thin".
• Going concern: not applicable on these cash flows.
• Related party: nothing surfaced in the sections read. The NVIDIA preferred is disclosed as an equity transaction, not a related-party one.
• Executive changes: none in the period read.

S5 — Ownership

Register (DEF 14A table, as of 2026-01-31, on 847,287,680 shares): FMR LLC 126,698,928 / 14.95%; The Vanguard Group 79,610,185 / 9.40%; BlackRock 60,534,116 / 7.14%.

Insider ownership is negligible. All 13 directors and executive officers as a group held 1,046,798 shares, marked "*" for under one percent — about 0.12%. Matthew Murphy, Chairman and CEO, holds 412,871. Alignment here runs through compensation, not through stock.

13F, report date 2026-06-30 (threespread.py holders MRVL; 1,560 filers, 499.6m shares reported). Largest adds: BlackRock +21,798,190 to 85,641,174; State Street +18,495,866; T. Rowe Price +12,134,409; Cantor Fitzgerald +11,818,810; Susquehanna +9,769,993; Citadel +8,984,768. Sixth Street Partners is new at 5,225,664. Largest cut: Ameriprise −32,577,650 to 9,111,628. ⚠️ The quarter is incompletely ingested. FMR (131.3m shares), both Vanguard entities (78.2m and 33.2m), UBS, Geode, Franklin, FIL, Morgan Stanley and Northern Trust have no 13F in this quarter yet. Their absence is ingestion lag, not an exit, and must not be read as one.

Insider transaction cadence: 143 Form 4 filings in the twelve months to 2026-09-21 and 22 Form 144 proposed-sale notices since 2026-03-25, the most recent 2026-08-17. The tool returns filings, not transaction detail, so this is a cadence reading — heavy and one-directional in the Form 144 sense — not a measured net buy or sell.

Short interest (FINRA via Equibles): 38,175,634 shares at 2026-08-31, 4.4% of shares outstanding, 1.4 days to cover. Nobody is short this in size.

What hangs over it. A live S-3ASR automatic shelf filed 2025-03-12 — the only S-3 on the register since 2023, and the one the April 2026 $1.0bn notes drew on. An S-3ASR carries no stated dollar cap. Against it, $9.7bn of buyback authority (a $5.0bn addition authorised 2025-09-24), of which only $400.0m was used in the first half of FY27 while $2.0bn of preferred and $1.0bn of notes were issued. Capital is coming in, not going out.

And the fact that reframes the register: on 2026-03-31 NVIDIA bought 2.0m shares of Series A Convertible Preferred for $2.0bn cash, convertible into up to ~21.8m common shares at ~$91.8355, voting with the common on an as-converted basis on everything except director elections. At $244.25 that conversion right is worth roughly $3.3bn more than the $2.0bn paid — a gain to NVIDIA larger than Marvell's entire FY26 operating income. Management's own risk factor names the consequence: "because the Series A Preferred Stock was issued to a single significant holder, the concentration of ownership could increase NVIDIA's influence over matters submitted to our stockholders". Marvell's 10-K lists "Nvidia Corporation" among companies that "compete directly with our businesses" (10-K, Item 1, Competition).

S8 — Valuation and sizing ⏳ OPEN (user)

Built from management's own guides (Equibles GetGuidance, rows dated 2026-08-27, sourced to the Q2 FY27 8-K and call): FY27 revenue ~$12B; data-center revenue growth 60% in FY27 and 60% in FY28; FY27 non-GAAP opex $2.55B; Q3 FY27 non-GAAP gross margin 57.5–58.5%; Q3 FY27 non-GAAP EPS $1.05–1.15 against GAAP EPS $0.48–0.58; FY28 non-GAAP tax rate 13%.

Bear / base / bull, on the FY28 line: bear — data center grows 25% not 60%, margin holds, EPS ~$3.90, and a 25x multiple on a decelerating custom-silicon franchise gives ~$98; base — the table above, 30x on $6.76 gives ~$203; bull — 60% holds and mix lifts margin to 60%, EPS ~$7.40, 35x gives ~$259. The base case is below today's price. That is the finding, and it says the market is already paying for the FY28 guide plus more.

Two adjustments the owner should make before trusting any of it. First, the GAAP/non-GAAP gap is not noise: Q3 FY27 is guided at $0.48-0.58 GAAP against $1.05-1.15 non-GAAP, a ~$0.55 quarterly wedge, ~$2bn a year of stock compensation and acquired-intangible amortisation — and the share count above does not absorb the continuing issuance that wedge represents. Second, the FY28 revenue line is management's 60%, compounded; it is a guide, not a forecast, and the kill criteria are written to catch it breaking.

Proposed buy price and proposed size are owner-only and are not stated here. Liquidity does not constrain sizing: ~$5.8bn traded on 2026-09-18.

S9 — Monitoring

• Next earnings: Q3 FY27 results, expected early December 2026 — derived from filing cadence (Q3 FY26 results 8-K filed 2025-12-02, 10-Q 2025-12-03), not from a calendar feed.
• Filing feed: the Tier 0 EDGAR sweep already covers cik 1835632 (watchlist.yaml carries it), so 8-K, 10-Q, S-3, 13D/G and Form 4 arrive without further setup.
• Leading indicator between quarters: the Distributor A percentage in each 10-Q's significant-customer table. It is disclosed quarterly, it moved 34% → 44% in a year, and it is the single number that decides whether this is a diversified custom-silicon franchise or one channel.
• Trigger that converts this to a buy: data-center revenue on track to the guided 60% for FY28 and non-GAAP gross margin holding at or above 58% and the price at or below ~30x the model FY28 EPS — i.e. a FY28 multiple that does not already assume the guide. On today's model that is roughly $203, about 17% below the 2026-09-18 close.

Evidence that was thin, and why

• No earnings-call transcript was read. Guidance came from Equibles' structured GetGuidance rows (one call), which carry the 8-K source links but not the Q&A. The gate-minimum spec asks for two transcripts and the Q&A; this run traded them for budget across three names.
• The efts.sec.gov full-text sweep for "changes in accountants" / "restatement" / NT 10-K was not run. The auditor and restatement reads above come from the EDGAR form census and the proxy instead — weaker evidence, and an item-4.01 8-K would not be visible in a form census.
• No analyst coverage count. Not available from the rungs used without spending an Equibles call; the skill lists it under S2.
• Proxy compensation structure not extracted — only the ownership table was read. What the bonus is paid on is unread.
• Next earnings date is derived from filing cadence, not a calendar feed; ROIC.ai's earnings calendar needs a paid plan on this account (established in slot names-ai-capex-1).
• The 2026-06-30 13F quarter is incompletely ingested — see S5. Every "exit" reading is suppressed accordingly.