SANTA CLARA, 29 AUG 2026 — Marvell beat expectations and its shares fell about 8 per cent. The reason is not in the quarter. It is that the Google deal worth up to US$120bn runs through fiscal 2033, and the money starts arriving in fiscal 2029.
What the results actually said
The results were better than expected. The guidance that accompanied them, however, laid out the timeline of the Google relationship in specific terms for the first time.
Management expects fiscal 2027 revenue near US$11.5bn and fiscal 2028 near US$16.5bn, with custom revenue more than doubling year on year in fiscal 2028 and passing US$10bn in fiscal 2029. Chief executive Matt Murphy told investors that revenue from Google becomes substantially more meaningful in fiscal 2029, which is later than parts of the market had assumed.
Shares fell roughly 8 per cent to about US$223 in premarket trading.
What "up to US$120bn" is, and is not
The figure is a ceiling attached to a warrant structure, not a purchase order.
As we set out when the arrangement was disclosed, Marvell is paying Google in warrants to keep buying its chips. The warrant covers AI inference accelerators, storage controllers, networking and memory-interface controllers and near-memory computing tied to the TPU ecosystem, with a performance-based tranche linked to custom-product revenue through fiscal 2033.
The US$120bn figure is the cumulative revenue level at which the full performance tranche vests. It describes what Google would have to buy for Marvell's equity grant to pay out in full. It is not a commitment by Google to buy that much. Treating it as backlog is the error the share price corrected.
Cumulative-through-2033 is also doing work. Seven fiscal years of a large customer relationship producing US$120bn is a very different statement from US$120bn arriving on any near horizon, and the two get written identically in headlines.
Why a beat still produced a fall
Markets price expectations, not results, and the expectation here had drifted ahead of what the company was willing to underwrite.
An investor who had assumed meaningful Google revenue in fiscal 2027 or 2028 was holding a stock whose cash flows, management now says, will arrive about two years later than that. Discounting the same cash two years further out reduces its present value materially, and at Marvell's multiple that is a large number.
There is a second effect. A performance tranche vesting on cumulative revenue through fiscal 2033 dilutes existing holders if it vests, which means the good outcome carries a cost that the headline number does not mention.
None of this means the deal is bad, only that it is long. A market that had priced it as a short-term win has now adjusted.
Why custom silicon revenue is late by nature
The fiscal 2029 timing is being read as a disappointment. It is closer to the normal physics of the business, and understanding why makes the guidance less surprising and the stock reaction more so.
A custom accelerator is not a product a supplier ships from a catalogue. The customer specifies it, the two companies co-design it over roughly eighteen months to two years, it tapes out, silicon comes back, it is validated, then it is qualified into the customer's own systems, and only then does volume production begin. Revenue is negligible through all of that and then arrives in a step.
Marvell has been describing a deal whose engineering is largely still ahead of it. Fiscal 2029 for meaningful revenue on a relationship framed this year is not a slip; it is roughly what the development cycle implies.
Which raises the more interesting question about the market's reaction. If the timeline is what the physics of custom silicon always produces, the expectation that ran ahead of it was never grounded in how the product gets built. The correction was to that expectation, not to Marvell's execution.
The pattern this belongs to
This is the third instrument of its type this publication has examined in a fortnight, and they rhyme.
Nvidia's US$105bn for OpenAI's Ohio campus is a guarantee rather than a cheque. Broadcom guaranteed the resale value of its own chips to get a US$35bn loan priced. Now Marvell's largest customer relationship is expressed as a warrant with a US$120bn vesting threshold.
In each case a very large number is doing publicity work that its legal form does not support. Guarantees, warrants and performance tranches are contingent instruments: they describe conditions, not transfers. The headline treats them as money.
Nvidia's decision this week to pause its cloud revenue-sharing programme over internal antitrust concern is the first sign of the industry pulling back from one variety of this. The equity-linked variety is still expanding.
What to watch
Two numbers will tell you whether the fiscal 2029 timeline holds, and neither is US$120bn.
The first is custom revenue in fiscal 2028, which management says will more than double. That is a near-term, checkable claim and it is the leading indicator for everything after it. The second is whether the fiscal 2029 language moves again in either direction at the next two reports, because guidance that slips twice is a different situation from guidance that slips once.
The competitive context has not changed either. Marvell's own recent history is that interconnect, not custom silicon, carried the year, and a company whose narrative is custom chips while its revenue is data movement is worth watching for which line actually grows.
