Marvell Gives Google a $12.2 Billion Option in Chip Deal

A warrant for 58.97 million shares at $206.58 apiece binds Marvell’s next seven years to Google’s chip orders, turning corporate strategy into a long-dated bet on AI infrastructure demand.

Key Takeaways

• Marvell issued Google a warrant for 58.97 million shares at $206.58 apiece, worth up to $12.2 billion, tied to a custom AI chip agreement signed July 29, 2026.
• Only 1.36 million shares vest automatically, while the remaining 57.6 million vest solely as Google buys chips, requiring $120 billion in cumulative orders to fully vest.
• The agreement gives Marvell design relationships with all three major American hyperscalers for the first time, marking Google’s move toward a second custom silicon supplier alongside Broadcom.

The Terms of Alignment

On July 29, in a filing that would not surface publicly for another three weeks, Marvell Technology and Google signed a commercial agreement that reshapes the geography of custom AI silicon. Marvell would design a family of chips built to plug directly into Google’s Tensor Processing Unit ecosystem: inference accelerators, storage controllers, network interface controllers, memory interface controllers, and the near memory compute components that keep data moving at the speed modern AI training demands. The mechanism binding the two companies together surfaced only when Marvell filed its Form 8-K on August 19, disclosing that it had issued Google a warrant for up to 58,970,907 shares at $206.58 apiece, a position that could reach $12.2 billion if every share vests and every option is exercised.

Wall Street did not wait for nuance. Marvell’s stock, which closed the prior session at $234.33, surged as much as 14 percent through the trading day, trading in a range from roughly $228 to $245 before settling with double-digit gains. Broadcom, the incumbent supplier that has built Google’s TPUs for most of the past decade, fell as much as 5.8 percent, the market’s blunt verdict on what a second major vendor might mean for a relationship long treated as exclusive. Alphabet’s own shares barely moved, a detail worth sitting with: investors were not repricing Google’s cost structure. They were repricing Marvell’s position within it.

The Architecture of the Warrant

Warrants of this size invite a certain credulity, as though $12.2 billion were simply promised. It was not. Of the 58.97 million shares, only 1,360,867 carry a fixed vesting schedule, releasing in equal quarterly installments across the year following execution. Everything else, the remaining 57.61 million shares, vests exclusively against performance: one tranche of roughly 240,000 shares for every $500 million in qualifying Custom Products revenue Google generates for Marvell, across 240 tranches, from Marvell’s fiscal third quarter of 2027 through the close of fiscal 2033. Should Google’s purchases fail to scale, the warrant simply lapses, unexercised and largely forgotten.

The remaining terms reward close reading. Google may exercise vested shares for cash or through a net exercise structure that requires no cash outlay at all, meaning the $12.2 billion headline describes a ceiling, not a receivable Marvell can plan around. The instrument cannot move beyond Google’s controlled affiliates without Marvell’s consent, carries trading volume restrictions once shares vest, and was issued under the private placement exemption in Section 4(a)(2) of the Securities Act, sparing both companies a public registration process. None of this is unusual for an instrument of its kind. What is unusual is the discipline built into it: Google earns equity only by buying, and Marvell dilutes only when revenue justifies it.

A Third Seat at the Table

Google’s decision to bring in a second chip designer did not emerge from dissatisfaction with Broadcom, which extended its own TPU and AI rack partnership with Google in April through an agreement running to 2031, one that also opened roughly 3.5 gigawatts of TPU based compute to Anthropic beginning in 2027. It emerged from a different kind of pressure: the sheer scale of inference demand now moving through Google’s infrastructure, and a strategic instinct, familiar to any procurement executive, that a single supplier for mission critical silicon is a fragility disguised as efficiency. As inference workloads increasingly favor chips built for a narrow purpose over general purpose graphics processors, Google’s appetite for custom silicon has begun to outpace what any one vendor can comfortably absorb.

Marvell entered that opening with credentials already in hand. The company has built custom accelerators for Amazon’s Trainium program, developed the electro optics and interconnect technology that AI clusters require to move information between chips without bottlenecking, and spent years cultivating exactly the kind of hyperscale relationships this deal now completes. With Google added to its roster, Marvell holds design relationships with all three major American hyperscalers, a milestone analysts have taken to calling the completion of the hyperscaler sweep. Morningstar’s William Kerwin, speaking to Reuters, offered the clearest single line on what this means: a genuine win for Marvell, though one that reflects an expanding pool of demand at Google rather than a fracture in its relationship with Broadcom.

The Market’s Verdict

Numbers alone rarely explain a stock move this violent, but in this case they come close. Marvell entered Wednesday’s session already carrying constructive analyst sentiment: Stifel had reiterated a Buy rating and a $350 price target on strength in the data center segment, TD Cowen had lifted its target to $225 from $200 just two days earlier on Marvell’s share of the optical signal processing market, and UBS had held its own Buy rating steady. Into that setup landed a filing that gave the bulls a concrete reason for conviction rather than a narrative one, and the stock behaved accordingly.

The timing carries its own irony. Marvell’s own guidance, issued May 27 alongside record first quarter revenue of $2.418 billion, called for the current quarter to reach $2.7 billion at the midpoint, up 35 percent year over year, a forecast set roughly two months before the Google agreement was even signed. None of that guidance reflects this deal. The real test arrives on August 27, when Marvell reports second quarter results and management faces the questions the filing left open: what margin the Google program carries, how quickly qualifying revenue might accumulate, and whether one relationship can meaningfully ease the customer concentration that has shadowed Marvell’s story for years. Options markets are already pricing a swing of roughly 14 percent around that report, a figure that says as much about unresolved uncertainty as it does about excitement.

The Distance Still to Travel

Strip away the enthusiasm and the arithmetic imposes a useful discipline. Two hundred forty tranches at $500 million apiece total $120 billion in qualifying revenue, spread across a measurement window of roughly six and a half years, an average run rate near $18.5 billion annually. Marvell’s entire company generated $8.195 billion in revenue for all of fiscal 2026, up 42 percent year over year. For the warrant to vest in full, Google’s purchases from Marvell alone would need to exceed, by a wide margin, everything Marvell currently sells across every product line and every customer combined. That is not a bar out of reach. It is a bar that reframes what success would actually require.

This is why the warrant deserves to be read less as a windfall and more as an instrument of discipline, one that ties Google’s financial upside precisely to Marvell’s ability to deliver at a scale it has never yet reached. It gives Marvell a credible claim on a third hyperscale relationship and gives investors, provided they read past the headline figure, a genuinely useful signal about where the company believes its next several years of growth will come from. Dilution estimates land in a narrow band, roughly 7 percent on a raw share basis and closer to 6.3 percent fully diluted according to Reuters, either way a real cost borne by existing shareholders in exchange for an option whose value depends entirely on execution still to come. Wednesday’s filing did not answer whether that execution will materialize. It simply set the terms on which the answer will eventually be judged.