Marvell Technology

Marvell Technology (NASDAQ: MRVL) makes two things the AI buildout needs and neither is a GPU. One is custom silicon — application-specific chips designed to a single customer’s spec, the ASIC work that lets a hyperscaler run its own accelerator instead of buying Nvidia’s merchant part. The other is the wiring between chips: retimers, switches, and — after a 2025 acquisition — optical interconnect, the market Astera Labs occupies from the interconnect-only side. Marvell does both jobs, which is why its stock reacts to two stories at once: whether hyperscalers keep designing their own silicon, and whether Marvell wins the contract when they do.

The custom-silicon thesis is the standing bear case against merchant GPU pricing power: if Amazon, Google, and Microsoft can each have an accelerator built to their own workload, Nvidia’s margin has somewhere to go. Marvell is one of two companies — Broadcom is the other — that does that design work at scale. TEXXR’s corpus holds the coverage built on top of that thesis: a stock that moved on an Nvidia CEO’s endorsement, a Google chip-talk leak, a Microsoft chip-talk leak that named a rival instead, and quarterly numbers that eventually caught up. What the record does not yet hold, as of this writing, is a single confirmed edge tying Marvell to a named hyperscaler’s shipped, revenue-generating custom chip. Every relationship in the graph below is a completed acquisition, an investment, or a talk.

The stock’s own arithmetic makes the same point. Marvell closed at $80.37 on 2025-07-31 and $187.56 on 2026-07-31 — up 133.4% over that year — but the peak came and went in between: a close of $316.43 on 2026-06-04, down 40.7% to the July close a year-end holder would have actually realized. The gain is real. So is the round trip.

Marvell’s fiscal year runs a quarter ahead of the calendar: a “Q4” report lands in March, a “Q1” in May. The AI-era story starts with its December 2024 report, when Marvell beat Q3 estimates on custom-silicon and interconnect demand and, five days later, the Wall Street Journal noted its market cap had passed Intel’s — with CEO Matt Murphy floated, that same week, as a potential Intel CEO candidate. It didn’t hold: two of the next three quarterly reports closed down double digits the session after — -19.8% on 2025-03-06 despite a Q4 revenue beat, because guidance trailed the highest estimates, and -18.6% on 2025-08-29 after a Q2 report beat on revenue but missed on the Q3 data-center forecast — the SiliconANGLE headline named the fear directly: concern for custom silicon’s long-term prospects, even as the same piece argued the business looked strong. Between the two selloffs, Marvell sold its automotive Ethernet unit for $2.5 billion to concentrate on the AI bet — betting its own balance sheet on the thesis the stock kept punishing it for.

Two acquisition stories ran in parallel through late 2025, and only one happened. In November, Bloomberg reported SoftBank had explored taking Marvell over to merge it with Arm, the chip industry’s largest deal had it happened — reported, never confirmed, and the stock barely moved (92.90 the day before to 93.33 the day of, +0.5%). In December, The Information reported Marvell was in advanced talks to buy Celestial, a photonics startup, for up to $5 billion; the same day Marvell confirmed the deal at $3.25 billion, targeted to close in fiscal Q1 2027. Marvell’s 10-K says it closed on 2 February 2026; TEXXR’s corpus has no article on the close.

The confirmed hardware partnership arrived in spring 2026. On 2026-03-31, Nvidia announced a $2 billion investment alongside a plan to collaborate on silicon photonics — the two companies’ interconnect roadmaps, not a custom-chip deal — and the stock closed up 12.8%. Two months later, at Computex, Jensen Huang called Marvell the next trillion-dollar company, and the stock closed up 32.52% on 2026-06-02 ($219.43 to $290.79) — its largest single-session gain in TEXXR’s price history for the name. Three days later Marvell and Flex were set to join the S&P 500, the stock jumping more than 6% after hours — after closing down 16.74% that same session ($316.43 to $263.47) amid a broader tech selloff. The whipsaw continued: -9.4% on 2026-06-23 ($307.86 to $279.04) in a second selloff, then a slower bleed to $187.56 by 2026-07-31 — the 40.7% drawdown from the Computex-week peak cited above.

The record’s most important entry for the custom-silicon thesis is its smallest one. On 2026-04-20, The Information reported Google was in talks with Marvell to co-develop a memory processing unit and a new TPU for AI inference — Marvell helping build the chips that let a hyperscaler need Nvidia less. It is, in TEXXR’s corpus, still a talk: no follow-up article confirms a signed agreement, a shipping product, or a dollar of revenue. The corpus holds the opposite signal too. On 2025-12-06, The Information reported Microsoft was in talks with Broadcom to design future custom chips — talks that would mean Microsoft switching its custom-silicon business away from Marvell. A design win at one hyperscaler and a defection risk at another, both unconfirmed, inside the same six months: the thesis is being negotiated, not executed.

The most recent 25 hyperedges TEXXR’s graph has recorded under “Marvell”, running March through August 2026, split 20 financial, 4 partnership, 1 personnel — no launch, competitor or regulatory edges. Financial edges are the graph’s record of the market reacting to a catalyst — a stock move, an earnings print, an index addition — not of the business doing something new. Four-fifths of the recent record on Marvell, by this count, is reaction.

The two acquisition edges are instructive. One, first recorded 2025-12-02, carries status rumored: Marvell in advanced talks to buy Celestial AI. The other, recorded a day later, carries status confirmed: Marvell announcing the deal. That one-day flip is the graph doing exactly what it should — but it has not recorded the deal’s close on 2 February 2026, which Marvell’s 10-K reports — the harder confirmation that would tell a reader the two companies’ interconnect roadmaps actually merged.

The Google talks sit differently. Two edges from the April report, filed under “Marvell Technology” — Google and Marvell, partnership, memory processing unit (edge 297140), and TPU (edge 297141) — still carry status rumored. The successor arrived in August under “Marvell”: two confirmed edges on an expanded chip-development agreement and the Google warrant. The graph did not link the two, and neither records a shipped chip — the design-win-versus-shipment distinction rendered as data.

Positioning: how much of MRVL is sold short, on the record brokers file twice a month. It lags by design — read it as where the crowd stood, not where it stands.

• Whether the Google deal produces shipments. The talks reported in April became, by August, an expanded development agreement. A confirmed shipment or revenue edge is the most direct test of whether the custom-silicon thesis is shipping or still being negotiated.
• The Microsoft-Broadcom talks. Recorded rumored in December 2025: Microsoft potentially moving its custom-chip business from Marvell to a rival. A confirmed edge would be a customer-concentration loss, not a competitor gaining share in the abstract.
• Whether the record catches up to the Celestial AI close. Announced 2025-12-02 at $3.25 billion or more, it closed on 2 February 2026, by Marvell’s 10-K; TEXXR has no article on the close.
• Whether the Nvidia silicon-photonics collaboration produces a launch edge. The $2 billion investment, announced 2026-03-31, is an interconnect deal, not a custom-chip one — worth tracking separately so the two don’t get conflated.
• Whether the graph’s edge mix moves off financial. Twenty of the last twenty-five edges recorded a market reaction; four recorded a partnership. All 25 sit inside TEXXR’s 2026 extraction, which only then began labelling edges “competitor”; a shift toward launch or competitor edges within that regime would mark the record catching up to product, the way Astera Labs’ has not yet caught up to its own customer concentration.
• Whether the Google warrant produces volume. The August warrant lets Google buy up to $12.2 billion of shares, vesting one tranche per $500 million of custom-product purchases, by Marvell’s 8-K — a customer taking equity rather than signing a supply contract, which is the same pattern the record shows elsewhere in the buildout. Custom silicon is the buyer’s route around Nvidia’s share of the split; the chokepoint rotation is where that fight over the split gets tracked.

Marvell Technology lives or dies by the same constraint as these names — small equal-weighted groups, one constraint each. Not investment products: see all categories.