Up 13%, Then Down 6% — Same News, Two Days Apart

Here's the deal: on August 19, Marvell Technology jumped as much as 13% intraday. The filing said Marvell had expanded its custom AI chip agreement with Google and issued Google a warrant to purchase 58,970,907 Marvell shares. Multiply by the $206.58 strike and you get roughly $12.18 billion. Headlines wrote it as "Google secures a $12.2 billion stake option," and the market read it as a vote of confidence.

Two days later, on August 21, the same stock fell 6%. No new news broke. What changed is that more people had read the vesting terms in the 8-K.

Those terms go like this. Of the 58.97 million shares, only 1,360,867 vest on a clock — in equal quarterly installments over the first year following the agreement. That's 2.3% of the total. The other 97.7% vests only when Google actually buys Marvell custom products, and it unlocks in 240 equal tranches, one per $500 million of revenue. Full vesting requires Google to purchase $120 billion of Marvell silicon.

Flip the framing and the deal's real shape appears. This is not Google investing $12.2 billion in Marvell. It's Marvell offering Google roughly a 10% rebate — paid in Marvell equity instead of cash — on $120 billion of future purchases.

Marvell, Google, and the TPU Ecosystem

Marvell builds data infrastructure semiconductors. Like Broadcom, it designs custom ASICs, but its center of gravity sits somewhere different. Marvell's deep assets are in storage controllers, network interface controllers, memory interfaces, and SerDes — the chips that live where data moves. The accelerator core itself is less its home turf than everything that feeds, stores, and connects it.

Google is the one hyperscaler that has designed its own TPU since 2015. "Designed its own" invites a misreading, though. Google owns the architecture and core design; physical implementation, verification, and surrounding silicon go to partners. Broadcom has held that partner seat for years. This agreement widens the seat next to it for Marvell.

The scope is what makes it significant. The 8-K covers custom silicon programs attaching to the TPU ecosystem across five named categories: AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute. This isn't a fight over one accelerator SKU. It's Marvell claiming multiple positions inside the rack that TPUs live in.

The sequence matters too. The commercial agreement itself was signed on July 29. The warrant was issued three weeks later, on August 18, and disclosed on August 19. Contract first, equity second — which tells you this warrant is an incentive bolted onto a commercial deal, not an investment thesis.

The Vesting Terms, Laid Out

Item Detail
Total warrant shares 58,970,907
Strike price $206.58 per share
Value at strike ~$12.18B
Time-based vesting 1,360,867 shares (2.3%) — equal quarterly installments over year one
Performance vesting Remaining 57,610,040 shares — 240 tranches, one per $500M revenue
Performance window Marvell FQ3 2027 through end of FY2033
Full vesting requires $120B in cumulative Custom Products purchases by Google
If fully exercised Google becomes Marvell's fifth-largest shareholder

The timeline is the number to sit with. FQ3 2027 through the end of FY2033 is more than six years. Spread $120 billion across that and you get roughly $20 billion a year — a figure that would move Marvell into an entirely different weight class. Which is also exactly why skepticism about achievability is warranted.

Note also the wording: discretionary purchases. There is no minimum commitment. Google buys what it wants, and the warrant unlocks in proportion. Marvell has a ceiling with no floor.

The $206.58 strike is worth reading too. It sits close to where Marvell traded around issuance — not deep in the money. That means Google is co-betting on Marvell's share price rising. If Marvell trades below $206 for six years, the warrant could fully vest and still be worthless to exercise. The incentive is self-reinforcing by design: Google's orders grow Marvell, and Google captures upside from the growth it caused.

That design cuts the other way for Marvell's management, though. For the stock to rise enough for Google to exercise, custom revenue has to convert into actual profit. Custom ASIC work carries big top-line numbers and thinner margins than branded product. $120 billion of revenue is not $120 billion of anything else, and that distinction should stay attached to every reading of this deal.

What Each Side Gets

Google gets supply chain leverage. As TPU volume grows, single-partner dependency becomes a genuine risk — pricing and schedules bend to the other party's circumstances. Standing Marvell up as a second axis diversifies that dependency and lets Google play two suppliers against each other on price. The warrant is the bonus on top: Google captures Marvell's share appreciation on chips it was going to buy anyway, which functionally reduces its cost of acquisition.

Marvell gets visibility. The chronic weakness of custom silicon is that you cannot forecast when or whether contracts land. You pour years into a design and the customer shelves the program, and none of it comes back. An anchor customer committed across five product categories stabilizes the development pipeline — and the stock reaction on day one didn't hurt either.

A partially-vested outcome isn't a bad ending for Marvell either. Unvested warrants mean no dilution. The genuinely bad scenario is different: heavy investment in design headcount and mask costs, followed by Google trimming the program so the revenue never arrives. Then the development spend is stranded regardless of what the warrant does. That's where custom ASIC P&Ls have always been won or lost.

Existing Marvell shareholders pay for it. 58.97 million shares is not trivial dilution, and the 6% drop on August 21 looks like the market pricing that in late. The defense is that dilution is coupled to revenue — shares only release when money actually arrives, so the worst case (dilution without revenue) is structurally blocked.

Broadcom loses exclusivity. Broadcom fell 3% on August 19 while Marvell rose 13%, a straightforward read on its share of Google's TPU program. Broadcom answered the same week with a different axis entirely: the reported $60 billion debt financing to supply Anthropic. Widening the customer list beats defending a share of one customer.

TSMC and advanced packaging houses win either way. More custom silicon programs means more wafers and more advanced packaging regardless of which designer books the deal. Near-memory compute designs in particular push memory and logic closer together, which raises packaging difficulty — and the associated bottlenecks — along with demand.

There's one more cost embedded in this deal that doesn't appear in the filing: precedent. Microsoft and Amazon run their own silicon programs, and their procurement teams read 8-Ks. If equity incentives to anchor customers become standard practice in custom silicon, Marvell will face the same demand in every future negotiation.

Precedents — With Diverging Outcomes

Paying a customer in equity to lock in volume has become a recognizable pattern in AI infrastructure. The most cited case is OpenAI and AMD, where AMD issued warrants for up to 160 million shares tied to a 6-gigawatt GPU deployment, with vesting gated on both volume milestones and share price. AMD's stock jumped on announcement.

An older and more encouraging precedent is Amazon's warrants to logistics partners like ATSG and Air Transport Services Group, used to anchor long-term air freight contracts. Those largely worked — because Amazon's shipping volume kept growing, so vesting was a sign the relationship was performing as intended.

The failure mode exists too. In the early 2000s several semiconductor firms tied volume incentives to large customers, then watched those customers' product cycles roll over. Volume came in under half of plan. No warrants vested, so there was no dilution — but the design headcount and mask costs were gone. That's where the real money in custom silicon is at risk, not in the share count.

Which way the Marvell case runs comes down to one variable: does Google TPU volume keep compounding for six years? The trend says yes, but that trend rests on Google continuing to win in its own AI products.

How Competitors Respond

Broadcom's answer is already visible: rather than defend one customer, broaden the roster toward Anthropic and OpenAI, and bundle financing with silicon. The $60 billion debt deal reported August 20 is that strategy made concrete. Marvell binds customers with equity; Broadcom binds them with capital. The two approaches target different buyers — warrants work on a cash-rich Google, leases work on labs that need to conserve cash.

Nvidia isn't directly hit, but the direction is uncomfortable. As hyperscalers expand in-house silicon programs and fill the surrounding ecosystem with custom parts too, the space for Nvidia's complete-system pitch narrows. Its counter is already in motion via NVLink ecosystem opening and deeper networking integration.

AMD pioneered this warrant playbook, so nothing here surprises it. The side effect is that as the practice normalizes, customers start expecting it. Once "AMD gave one, Marvell gave one, why not you?" enters procurement conversations, margin structures compress across every supplier.

Memory makers should read the scope carefully. Memory interface controllers and near-memory compute being inside this contract means how memory attaches and feeds the TPU rack is now a primary battleground of custom design. Selling HBM and designing the controller that attaches HBM are different businesses, and if the latter consolidates around companies like Marvell, memory vendors end up following specs rather than setting them.

Fabless companies outside the US face a structural barrier here. To play this game, your equity has to be an asset a hyperscaler actually wants. Below a certain market cap, offering warrants doesn't move anyone. Scale itself is the moat.

So What Actually Changes

If you're an investor, filter out headlines summarizing this as "Google invests $12.2B in Marvell." Google has to decide to spend $120 billion for that $12.2 billion to complete. The metric to track isn't warrant size — it's the Custom Products revenue line in Marvell's quarterly results. How many of the 240 tranches have released is the actual progress bar on this deal.

If you work in semiconductors, the scope is more instructive than the money. Five categories bundled into one agreement signals that hyperscalers are now buying custom at the rack level. A portfolio that fills multiple sockets in a rack beats excellence at a single accelerator when contracts get awarded.

If you build on cloud, treat this as a long-run signal on TPU pricing. Dual-sourcing and price competition create room for TPU-based instance costs to fall. Whether that reaches your invoice is a separate question — Google may keep the savings as margin.

If you run an AI startup, the transferable lesson is about payment instruments. Using equity or warrants to anchor supply agreements when cash is tight isn't only a mega-cap technique. The scale differs; the mechanics don't. The design principle is to couple dilution to realized revenue — get that wrong and you own the worst combination.

If you're buying AI infrastructure for an enterprise, more deals of this shape mean better supply stability. Dual-sourced silicon lowers the odds that one supplier's incident cascades into an outage. Knowing that context lets you write sharper supply chain risk terms into cloud SLAs.

🥄 Three Things You're Probably Wondering

— Will Google actually buy $120 billion of this? Too early to say. It works out to about $20 billion a year for six years, which requires Google's TPU program to grow substantially beyond its current size. And the filing says discretionary purchases — there's no obligation. The most plausible outcome is that a meaningful fraction of the 240 tranches release, which would still be a very good result for Marvell.

— So was the market wrong when it dropped 6%? Less wrong than sequential. Day one priced the headline number and the deal scope. The following days priced the 8-K's vesting mechanics and the dilution math. Filings routinely get read after the news cycle that reported them.

— Is Broadcom being pushed out of Google? That's stronger than the evidence supports. Google's TPU program keeps growing, so a second partner doesn't necessarily shrink Broadcom's volume. What's certain is that exclusivity is gone, and that shifts pricing power toward Google.

Sources

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!