Google secures a $12.2B stake in Marvell, the only chipmaker designing for all three hyperscalers

Marvell Technology stock rallied nearly 10 percent Wednesday after the company disclosed a deal giving Google the right to buy up to $12.2 billion in Marvell shares.

The agreement is tied to a custom chip partnership. Marvell will design semiconductor products that attach to Google's tensor processing unit (TPU) ecosystem, the company said in a securities filing.

The warrant covers 58.97 million shares at $206.58 per share. That is roughly 7 percent of the company.

The deal was signed on July 29. The warrant was issued on August 18, according to a US Securities and Exchange Commission (SEC) Form 8-K filing.

The structure links ownership to purchasing volume. Only 1.36 million shares vest in the first year. The remaining 57.6 million shares vest in 240 equal tranches.

One tranche vests for every $500 million in custom products revenue Marvell records from Google. The schedule runs through fiscal year 2033.

If Google exercises the entire warrant, it would have generated approximately $120 billion in cumulative chip purchases from Marvell. The structure mirrors a warrant Marvell issued to Amazon Web Services in December 2024 tied to Trainium and Inferentia chip procurement. The Google deal is 14 times larger by share count.

Three hyperscalers, zero precedent

The deal makes Marvell the only custom chip design firm with active partnerships at all three major US hyperscalers: Amazon Web Services' Trainium processors, Microsoft's Maia 200/300 accelerators, and now Google's TPU ecosystem.

Broadcom, which holds approximately 70 percent of the custom AI chip co-design market, has not replicated that sweep. Broadcom shares fell roughly 5 percent on Wednesday.

The products Marvell will build for Google span the full AI memory infrastructure stack: artificial intelligence (AI) inference accelerators, storage controllers, network interface controllers, and near-memory compute solutions. Near-memory compute places processing logic next to data storage. It cuts the delay of moving data between compute units and memory.

That architecture matters because the binding constraint in large language model inference is memory access, not processing speed. A model generating output tokens must read its weights from memory on every forward pass. At frontier scale, the memory-access cost multiplied across millions of simultaneous requests is the limiting factor for both speed and cost per token.

Marvell's Structera X, a CXL (Compute Express Link)-based memory expansion platform built on technology from its $540 million acquisition of XConn Technologies in February 2026, demonstrated 4.8 times higher inference throughput and an 82.7 percent reduction in time-to-first-token for an eight-graphics processing unit (GPU) server cluster using 16 terabytes of pooled memory.

The valuation leaves no room for error

Marvell shares traded above $230 on Wednesday, a 40 percent recovery from a July trough of $163. The stock remains roughly 35 percent below its 52-week high of $329.88, set on June 4.

The company reports second-quarter fiscal 2027 results on August 27. Analysts expect non-GAAP (non-Generally Accepted Accounting Principles) earnings per share of $0.93. That would be a 39 percent increase from a year ago.

Revenue is expected at approximately $2.71 billion, up 35 percent year over year. Management guided data center revenue to rise in the mid-to-high teens quarter over quarter. That is approximately 45 percent growth from a year ago.

At roughly 54 times forward earnings, Marvell trades at a 124 percent premium to the semiconductor sector median. The spread between the lowest analyst price target ($195 at Goldman Sachs) and the highest ($400) spans $205, reflecting genuinely divergent views on how much of the custom silicon revenue ramp will materialize.

One loose thread: the Google revenue timeline is undefined. Marvell signaled that CXL and PCIe switching products will begin revenue contribution in the second half of fiscal 2027. Those products are ramping toward a $50 million annualized run rate by Q4 of that year.

But custom silicon revenue from the Google partnership is not expected to be material in the near term. The earnings call on August 27 will be the first chance to hear management frame the timeline.

Source