Google Did Not Just Buy $12B of Marvell — It Got Warrants

Coverage of the Marvell-Google warrant agreement, disclosed in an 8-K dated August 18, 2026, led with the headline number as if Google had taken a $12 billion stake in Marvell. The filing says something more interesting and much less dramatic: Google received a warrant — a conditional right to buy up to 58,970,907 Marvell shares at $206.58 apiece — and 97.7% of those shares vest only if Google actually buys custom chips from Marvell, in $500 million increments, through fiscal 2033.

No money changed hands at signing. No shares moved. If Google never hits the purchasing targets, the bulk of the warrant never vests, and anything unexercised expires on August 18, 2033. What the document actually records is a supply commitment wearing an equity wrapper — and the wrapper is what made the headlines.

This post reads the 8-K and the warrant instrument itself: what the filing says versus what circulated, how the 240-tranche vesting mechanic works, the five product categories the agreement covers, what it means for Broadcom's decade-long TPU relationship with Google, and how to read the next equity-linked supply deal before the headline reads it for you.

Start with the sequence, because it undercuts the "sudden $12B deal" framing on its own. Per Marvell's Form 8-K, Marvell and Google entered into the underlying commercial agreement on July 29, 2026 — covering Marvell's development of custom semiconductor products for Google. The warrant itself was issued August 18, 2026, roughly three weeks later, and the 8-K carrying both disclosures is dated the same day. The share-price move and the coverage came on August 19.

The instrument is a warrant "to purchase up to an aggregate of 58,970,907 shares" at an exercise price of $206.58 per share, subject to customary adjustments. Multiply the two and you get the headline: 58,970,907 × $206.58 ≈ $12.18 billion, which press coverage rounded to $12.2B. That number is a ceiling on a hypothetical — the maximum Google could spend if every share vested and Google exercised all of them — not a transaction that occurred.

The load-bearing sentence in the whole filing is the vesting mechanic, and it deserves to be read verbatim rather than paraphrased through a headline.

Two more structural details round out the instrument. First, the warrant was issued under the Section 4(a)(2) private-placement exemption — a negotiated instrument between two companies, not a public offering. It cannot be transferred beyond controlled affiliates without Marvell's consent, and Google holds customary registration rights for eventual resale of shares it actually vests and exercises. Second, per the warrant instrument filed as Exhibit 4.1, the whole thing expires at 5:00 p.m. California time on August 18, 2033 — the seventh anniversary of issuance. Anything unvested or unexercised by then is gone.

Notably, there was no accompanying investor-relations press release. Marvell's newsroom shows nothing between August 4 and the filing date for this transaction — the 8-K itself is the disclosure. Google issued no separate public statement that this research pass could locate either. Every deal term in circulation traces back to one SEC filing and its exhibit, which is exactly why reading the filing beats reading the coverage.

The gap between the circulating framing and the document is wide enough to change what kind of story this is. Laid side by side, the corrections stack up:

None of this makes the deal small. A supplier does not hand a customer an equity instrument of this size for a routine purchase order, and Google does not negotiate a seven-year vesting schedule for a relationship it considers experimental. But the size lives in the commitment structure, not in any cash that moved. The honest one-line summary is: Marvell agreed to give Google equity upside proportional to how much custom silicon Google actually buys, and the maximum version of that upside prices out around $12.2B.

The vesting schedule is where the filing gets genuinely interesting, and it is the part of the deal a single headline number cannot carry. The 58,970,907 shares split into two very unequal pools. A sliver — 1,360,867 shares, about 2.3% — is time-based, vesting in equal quarterly installments over the first year regardless of purchasing. Read that as consideration for signing: the only part of the warrant Google earns just by showing up.

The other 57,610,040 shares — 97.7% — vest purely on performance: 240 equal tranches, one for each $500 million in Custom Products revenue from Google and its affiliates, counted from Marvell's third quarter of fiscal 2027 through the end of its fiscal 2033. Run the arithmetic and the scale of the commitment comes into focus: 240 tranches × $500M means a full vest requires $120 billion in qualifying purchases — roughly six years of sustained buying. Each tranche vests about 240,042 shares (57,610,040 ÷ 240), which at the $206.58 strike is roughly $49.6M of stock per $500M of revenue — about ten dollars of qualifying purchasing for every dollar of nominal warrant value earned.

One more definition from the warrant instrument sharpens the supply-commitment reading further. "Custom Products revenue" is not any dollar Google spends with Marvell. Under Exhibit 4.1, a qualifying product needs a written Statement of Work or Product Exhibit under the commercial arrangements and a "Sale Control Right" — a contractual right for Google to determine whether Marvell may sell that product to anyone else. In other words, the vesting counter only ticks on custom silicon Google effectively controls the distribution of. This is co-designed, customer-controlled product revenue — the deepest form of supplier lock-in short of an acquisition — not general merchant sales that happen to have Google's name on the invoice.

The deal's scope compresses easily into a three-item phrase — AI inference accelerators plus storage and network interface controllers. The 8-K's own language is broader, and worth quoting in full: "The expanded partnership spans a comprehensive range of custom silicon programs that attach to the TPU ecosystem, including AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute."

That is five categories, and the two the shorthand loses — memory interface controllers and near-memory compute — are the tell for what Google is actually buying. This is not a second source for TPUs. It is the connective tissue around the TPU: the silicon that moves data in, out, and near the accelerator.

The pattern across the five categories mirrors what other hyperscalers have been doing with their own custom-silicon programs — we traced the same accelerator-plus-plumbing logic in Amazon's custom AI chip strategy. The compute headline gets the attention; the interconnect, memory, and storage silicon around it is where a second design partner earns a decade of revenue.

The market read the deal through Broadcom, and the same-day tape shows it. Marvell shares rose about 10% on August 19 — reported intraday snapshots ranged from roughly 7% to 13% depending on the hour, so treat any single percentage as a snapshot, not the move. Broadcom went the other way: CNBC reported the stock down 5% that afternoon, on pace for its lowest close since early July, and framed the tie-up as supporting "the bear case against Broadcom: that Google will increasingly partner with other suppliers to develop and implement its custom AI compute systems" — CNBC's editorial read, worth labeling as such.

The facts underneath are less dramatic than the price action. Broadcom has co-designed Google's TPUs for about a decade, with the TPU now in its eighth generation per CNBC's reporting, and Broadcom disclosed in April 2026 an agreement to develop and supply future TPU generations — plus networking and other components — through 2031. Nothing in Marvell's filing touches that agreement. Google has added a second design partner for the silicon around the TPU; it has not swapped out the partner designing the TPU itself.

Broadcom's own CEO had already priced this in publicly, months before the filing. Asked on the company's early-June earnings call about the possibility of Google adding suppliers, Hock Tan gave an answer that reads like a preview of this exact 8-K:

Scale matters here too. Per the same CNBC reporting, Broadcom expects roughly $56B in AI chip revenue in fiscal 2026 and more than $100B in fiscal 2027, with a custom-silicon customer roster that now spans OpenAI, Anthropic, Meta, and Apple alongside Google. Against that base, a second Google design partner is a diversification signal — a real one — rather than an existential event. The Google-infrastructure orbit has room for many simultaneous bets; the Blackstone-Google TPU neocloud joint venture is another deal in the same ecosystem that reads very differently from its headline.

Earlier this week, we unpacked NVIDIA's financing guarantee for OpenAI's Ohio campus. The warrant is that post's equity-side cousin, and it deserves the same discipline extended one step further. A guarantee overstates committed money by reporting a ceiling; a warrant overstates it twice over, because the headline value requires both the vesting conditions to be met over years and the holder to then pay real cash to exercise. Neither instrument moves a dollar on signing day. Both produce a ten-figure number for the front page.

The warrant variant of this pattern has precedent. In October 2025, AMD issued OpenAI a performance-based warrant for up to 160 million AMD shares tied to gigawatt-scale deployment milestones. But the structural differences between that deal and this one are exactly where the reading skill lives:

The comparison shows why "warrant deal" is not one thing. AMD's structure hands the customer most of the economics through vesting itself; Marvell's hands the customer an option that is only worth exercising if the relationship — and Marvell's execution — makes the shares worth more than $206.58 by then. For Marvell, that is the point: it lands a formalized anchor customer for its custom silicon roadmap without cutting a cash rebate that would hit margins today. For Google, the equity upside is a rebate that only pays if the purchasing actually happens. Incentives point the same direction on both sides of the table, which is what a well-constructed supply agreement looks like.

What we deliberately will not do is estimate what the deal is "really worth" to either party. The filing gives the maximum share count, the strike, the tranche mechanic, and the window; any "expected value" figure on top of that would be an invented number wearing a spreadsheet costume. The honest position is that the deal's value is unknowable today because it depends on six years of purchasing decisions the filing explicitly labels discretionary.

AI infrastructure is now financed through instruments — guarantees, warrants, prepayments, vehicles — that all share one property: they generate a headline number far larger than any money that moved. We built the general version of this skill in how to read an AI valuation leak; the equity-linked supply agreement deserves its own four questions. Ask them in order before repeating any headline figure:

The same four questions belong on the buy side of any AI contract, not just the analyst side — they sit alongside the checklist in our AI procurement questions to ask before you sign. Looking forward, the direction of travel seems clear enough to state: as custom-silicon commitments stretch toward decade-length windows, expect more supply relationships to be papered as performance-vested equity rather than simple purchase agreements, because the structure lets both sides announce commitment without booking obligation. That makes filings-first reading a compounding skill — every one of these deals will produce a headline number, and the filing will keep being the only place the conditions live. If your team is negotiating AI infrastructure or vendor commitments and needs this discipline applied before signature, our AI transformation practice starts engagements with exactly this kind of document-level diligence.