Marvell Technology — Bull & Bear Memo

A thinking anchor, not a thesis or a valuation. Every figure comes from Marvell SEC filings; nothing outside them was used.

company
Marvell Technology, Inc. (Nasdaq: MRVL) — fabless data-infrastructure semiconductors; one reportable segment
sources used
FY2026 Form 10-K (year ended Jan 31, 2026); Forms 10-Q for Q1 FY2027 (May 2, 2026) and Q2 FY2027 (Aug 1, 2026); Forms 8-K dated Apr 15, Jun 11 and Aug 19, 2026; FY2022–FY2025 10-Ks for history
latest period
Q2 fiscal 2027, quarter ended August 1, 2026 (10-Q filed August 28, 2026)
missing / stale
No transcripts, investor presentations or management guidance in sources — the earnings 8-Ks hold only the cover page, not the press-release exhibit, so no forward guidance appears anywhere in this memo. Insider-trading file is empty. One segment reported: no custom-silicon vs. merchant split, no customer names, nothing below 10%. The Google warrant's grant-date fair value is not yet in any filing.

01Business in one line

Marvell is a fabless designer of data-infrastructure silicon selling two things into AI data centers: custom accelerators and adjacent silicon built to one hyperscaler's specification (custom ASICs and XPUs), and merchant connectivity parts — optical DSPs, retimers, co-packaged and linear optics, switching — that attach to racks whichever accelerator wins. It manufactures nothing: TSMC is sole source for all its advanced-node wafers.

Net revenue was $8,194.6M in fiscal 2026, 74% data center. Q2 fiscal 2027 revenue was $2,739.3M, 79% data center.

Type: fast grower on cyclical bones

Revenue grew 42.1% in fiscal 2026 and 36.5% year over year in Q2 fiscal 2027, but that demand sits inside a capital-spending cycle controlled by a handful of buyers. The classification drives the rest of this memo: the bear case is not slow decay, it is a step function — a socket re-bid and lost, or a capex budget cut.

02Bull case — Peter Lynch pitch

The simple reason this stock could work is that several hyperscalers have decided they need someone other than Broadcom to build their custom AI silicon, and they are handing Marvell multi-year programs and helping fund them with their own capital.

Click a card to open it. Switch views to see each assumption beside the attack on it.

Why the market might be missing it: the reported profit line is noisy in both directions. Fiscal 2026 net income of $2,670.1M included a $1.8B pre-tax gain on the automotive ethernet sale, while first-half fiscal 2027 net income of $342.5M absorbed a $433.7M non-cash increase in the fair value of the Celestial earnout. Neither figure describes the operating business, which earned a 16.8% operating margin on $2.7B of revenue last quarter.

03Bear case — Munger invert

The most likely way I lose money is that Marvell is renting its growth from four customers who hold every lever — price, volume, second sourcing, and now Marvell's own equity — and one of them re-bids a socket at the next node.

Ordered by permanence of the damage, not to mirror the bull sequence.

04Signals to monitor

Each signal tracks a named assumption or risk above. Nothing here is a forecast — these are the observable values as last filed.

06Bottom line

Marvell could work because several hyperscalers now need a credible second source for custom AI silicon and are committing multi-year programs, and in NVIDIA's case $2.0B of capital, to build one.

For that to pay, the sockets already won have to reach volume and be re-won at 2nm while R&D falls back below the 25.3% of revenue it reached in fiscal 2026.

The thesis most likely breaks not through slow erosion but through a single re-bid lost at a customer that is currently 44% of revenue, against $8.5B of wafer commitments that do not disappear with the order.

I would change my mind if gross margin fell while data center revenue was still growing — evidence the volume is being bought with price — or, on the other side, if Marvell began disclosing custom-silicon revenue separately and it compounded across more than one named customer.

§Sources & gaps

Periodic reports 3 used
  • FY2026 Form 10-K, year ended January 31, 2026 (filed Mar 11, 2026) — business, competition, end-market revenue, customer concentration, warrants, financial statements
  • Form 10-Q, quarter ended August 1, 2026 (filed Aug 28, 2026) — latest income statement, balance sheet, cash flow, Notes 3/4/7/9/10/15, MD&A, risk factors
  • Form 10-Q, quarter ended May 2, 2026 — Series A preferred issuance, Celestial and XConn purchase accounting
Current reports (8-K) 3 used
  • August 19, 2026 — Google commercial agreement dated July 29, 2026 and warrant for up to 58,970,907 shares at $206.58, vesting one tranche per $500M of custom-product revenue through fiscal 2033
  • June 11, 2026 — CFO transition: Willem Meintjes resigned effective June 15, 2026; Daniel Durn (previously Adobe, Applied Materials, NXP, GlobalFoundries) appointed CFO from the board
  • April 15, 2026 — $1.0B of 5.300% senior notes due 2036, proceeds to repay the 1.650% notes due 2026
Historical comparison 4 filings
  • FY2022–FY2025 Forms 10-K — end-market mix history, gross-margin and R&D trend, restructuring, prior-year revenue
Known blind spots. No earnings-call transcripts, investor presentations or management guidance were available, so nothing forward-looking from management appears in this memo. The earnings 8-Ks in the folder contain only cover pages, not the press-release exhibits. The insider-trading extract is empty. Marvell reports a single segment and discloses neither custom-silicon versus merchant revenue nor any customer name; "Customer A" and "Distributor A" are the filings' own labels. The grant-date fair value of the August 2026 warrant — the number that sizes its contra-revenue effect — should first appear in the Q3 fiscal 2027 10-Q.