Google secures $12.2 billion Marvell stock warrant in expanded AI chip deal

Google has agreed to pay one of its newest chip suppliers largely in stock rather than cash, according to a securities filing made public this week that lays out one of the largest custom-silicon commitments in the artificial intelligence industry to date. Marvell Technology disclosed in a Form 8-K filed with the Securities and Exchange Commission that it issued Google a warrant on August 18 to purchase up to 58,970,907 shares of Marvell common stock, a stake worth roughly $12.2 billion if fully exercised. The warrant is Google's reward, in effect, for agreeing to keep buying Marvell-made chips for the better part of the next decade.

The underlying commercial agreement, signed July 29, expands a relationship that already made Marvell one of a handful of companies helping Google build the specialized processors that power its search engine, Gemini models and cloud business. Under the new terms, Marvell will design a broader family of custom semiconductors that attach to Google's Tensor Processing Unit ecosystem, including AI inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute hardware, according to the text of the warrant agreement filed as an exhibit to Marvell's disclosure.

The mechanics of an unusual payment

Rather than negotiating a conventional supply contract, Google structured its side of the deal as an equity incentive tied directly to how much it ultimately spends. The warrant carries an exercise price of $206.58 per share and remains exercisable until August 18, 2033. Of the total shares, 1,360,867 are time-based and vest in equal quarterly installments over the agreement's first year regardless of how much Google buys. The remaining 57.6 million shares are performance-based: they vest in 240 equal tranches, with one tranche unlocking for every $500 million in qualifying custom-product revenue that Marvell books from Google's purchases, running from Marvell's third fiscal quarter of 2027 through the end of fiscal 2033.

The arithmetic behind that structure is what has drawn the most attention from Wall Street. For Google's holding company, Alphabet, to collect the full warrant, Marvell would need to record about $120 billion in cumulative revenue from Google's custom-chip purchases by early 2033. That figure dwarfs Marvell's entire annual revenue today and would represent one of the largest single-customer commitments in semiconductor history.

• Total warrant size: up to 58,970,907 Marvell shares, worth roughly $12.2 billion if fully exercised
• Exercise price: $206.58 per share, expiring August 18, 2033
• Vesting: one tranche per $500 million in Google purchases, 240 tranches in total
• Implied ceiling: about $120 billion in custom-silicon revenue for Marvell through fiscal 2033

Because the shares vest against actual purchases rather than a fixed order, Google faces no penalty if its AI infrastructure spending slows, while Marvell absorbs the dilution risk entirely. A commentary published by the Motley Fool laid out the trade Marvell's existing shareholders are effectively making, noting that full vesting would dilute existing holders by roughly 7 percent.

"Every $500 million of purchases hands Google another block of Marvell stock at $206.58 per share," Motley Fool analyst Daniel Sparks wrote, describing how the arrangement converts what would ordinarily be routine supplier spending into a growing equity position for the search giant.

A hedge against reliance on a single supplier

Google has relied on Broadcom for years to help design and manufacture its Tensor Processing Units, the custom chips it uses instead of, or alongside, Nvidia's graphics processors to train and run its AI models. That relationship has made Broadcom one of the best-performing chip stocks of the past two years, as investors bet that Google's TPU buildout would keep expanding. The Marvell agreement does not appear to replace that arrangement, but it does something Google's cloud rivals have already been doing: spreading critical chip work across more than one vendor.

Trade publication The Register reported that Broadcom shares slipped roughly 3 to 4 percent on the news, while Marvell's stock jumped, reflecting investors' read that Google is deliberately building leverage over its most important chip supplier rather than ending the relationship outright. The publication noted that Broadcom raised licensing and support costs sharply after its 2023 acquisition of VMware, an experience that has made hyperscale cloud customers wary of concentrating too much purchasing power with any single vendor.

Marvell brings its own distinct capabilities to the table beyond conventional chip design. The company's 2025 acquisition of Celestial AI added silicon-photonics interconnect technology, including a Photonic Fabric network interface card that can let servers share pooled memory across racks up to 50 meters apart — a capability that could let Google treat memory as a shared resource across a data center rather than a fixed allocation tied to each server. Combined with Marvell's existing SerDes networking components and compute-express-link memory controllers, the expanded partnership positions Marvell to compete for a wider slice of the infrastructure that surrounds, rather than replaces, the TPU itself.

The immediate winners are clear enough. Marvell's shareholders benefit from a long-term revenue commitment even if it comes at the cost of dilution, and Google locks in supply and pricing leverage over a supplier it needs for years of AI buildout. Less clear is what the deal means for Broadcom, whose custom-chip unit has been one of the primary drivers of its stock price rally over the past two years on the strength of its TPU work with Google. A prolonged shift of any meaningful share of Google's custom-silicon budget toward Marvell would complicate the growth story Broadcom has told investors, even if Google's total AI infrastructure spending keeps climbing enough to leave both suppliers with more business than they had before.

Part of a broader race for custom silicon

Google began developing its own AI accelerators internally in 2015, first disclosing the existence of the Tensor Processing Unit publicly in 2016 and later making TPUs available to outside cloud customers. What started as a way to run Google's internal search and translation workloads more efficiently has since become the backbone of the infrastructure used to train and serve the company's Gemini family of models, and a product Google now rents to external customers as an alternative to Nvidia GPUs. Google is far from alone in trying to reduce its dependence on Nvidia's general-purpose graphics processors and on any one custom-silicon partner. Amazon has built its own Trainium chips with help from Marvell and Alchip, Microsoft has developed its Maia accelerators, and Meta has pursued a similar strategy with its MTIA chips. Each of the major cloud providers has concluded that owning more of its own silicon roadmap, even at enormous up-front cost, gives it more control over AI infrastructure economics than buying exclusively from merchant chip vendors.

What distinguishes the Google-Marvell agreement is its scope. Rather than covering a single accelerator chip, the deal spans an entire family of supporting silicon — storage, networking, memory and near-memory compute — that together determine how efficiently a data center full of TPUs can actually be used. Industry analysts covering the announcement described the shift as evidence that the TPU is becoming the center of a broader custom-silicon platform rather than a standalone product, with data movement, storage access and memory management increasingly treated as separate semiconductor opportunities in their own right.

Market reaction and the road to 2033

Marvell's stock rose as much as 13 percent in the days following the disclosure, with shares trading near $237 to $243, comfortably above the warrant's $206.58 strike price and putting the instrument already worth more than $2 billion above its exercise cost even before further vesting. Analysts at Wedbush and elsewhere framed the deal as validation of the broader argument that custom silicon, not merchant GPUs alone, will define the next phase of AI infrastructure spending. A dispatch carried by Reuters noted that the warrant disclosure came without an accompanying press release from either company, surfacing instead through routine securities filings — an indication of how central these financial instruments have become to chip-supply negotiations even when neither party wants to publicize the terms.

Marvell has not detailed how quickly it expects Google's purchases to accumulate toward the $500 million thresholds that trigger each vesting tranche, and the company's own investor communications, including material posted to its newsroom, have so far said little beyond confirming the expanded scope of the partnership. Marvell is scheduled to report second-quarter fiscal 2027 results and hold an investor day on October 6, occasions likely to bring the first detailed guidance on how the Google relationship will show up in its revenue forecasts.

For now, the arrangement leaves Google with an unusual dual position: a customer locking in years of chip supply, and, if it keeps spending, a shareholder in the company selling it those chips. Whether that structure becomes a template other hyperscalers copy with their own suppliers, or remains a one-off born of Google's specific leverage as one of the largest buyers of advanced semiconductors in the world, is likely to become clearer as the AI infrastructure buildout continues through the rest of the decade.