Marvell’s $12.2B Google Warrant Is a $120B AI Revenue Test
Google’s $12.2 billion Marvell warrant is not an investment. It is a $120 billion performance review, unlocked roughly $500 million at a time.
Google has put a $12.2 billion carrot in front of Marvell Technology. The catch is brutal: Marvell has to earn it, roughly $500 million at a time.
That is the real story behind the custom-chip agreement disclosed this week between Google and Marvell. Forget the usual AI press-release fog. This is a commercial relationship dressed up as an equity deal, and the structure matters more than the headline.
Marvell issued Google a warrant to buy up to 58,970,907 shares at an exercise price of $206.58 a share. At full exercise, that is about $12.18 billion. But nearly all of those shares only vest as Google and its affiliates generate revenue for Marvell from a broad basket of custom semiconductor products through the end of Marvell’s fiscal 2033.
The number worth paying attention to is not $12.2 billion. It is $120 billion.
Google Has Turned Equity Into a Procurement Weapon
Most founders see a giant customer asking for warrants and think, “Fantastic — validation.” Sometimes it is. Sometimes it is a customer using your cap table as a crowbar.
In this case, Google gets the right to own a meaningful slice of Marvell only if it sends substantial business its way. The performance component of the warrant vests in 240 tranches, with each tranche tied to $500 million in qualifying custom-product revenue. Do the maths: that is roughly $120 billion of cumulative revenue required to fully unlock the performance-based shares.
That is not a soft handshake at a conference. It is a seven-year commercial scoreboard.
Marvell’s agreement with Google covers custom AI inference accelerators, storage, networking, memory-interface controllers and near-memory computing technology. In plain English: Google is not merely shopping for one chip. It is looking at the guts of the AI factory — compute, memory and the plumbing connecting it all.
The first small portion of the warrant vests over time. The overwhelming majority depends on revenue. That distinction matters because it changes the incentives.
Google is saying: we want you motivated, funded and pointed at our problems for years.
Marvell is saying: we will accept dilution if the customer becomes enormous enough to make the dilution worthwhile.
That is sensible on both sides. It is also a reminder that in AI infrastructure, the customer with the data centres has more leverage than the supplier with the clever silicon.
This Is About Escaping the Nvidia Tax — Without Pretending Nvidia Is Dead
Everyone wants to frame every chip deal as “the Nvidia killer.” That is lazy nonsense.
Nvidia remains the benchmark for general-purpose AI training and increasingly for inference. Its hardware, software and developer ecosystem remain staggeringly hard to dislodge. Google signing Marvell does not mean Nvidia has been knocked off the perch. It means Google is doing what every serious hyperscaler should do: refusing to build a trillion-dollar computing business around one supplier’s economics.
Custom silicon is not about winning a beauty contest against Nvidia. It is about control.
If Google can tailor chips and networking to its own workloads, data-centre architecture and models, it may be able to lower operating costs, improve performance per watt and reduce supply-chain dependence. Those gains compound when you operate at Google scale. Saving a modest amount per unit is meaningless to a startup. Saving it across a global AI fleet is the difference between a clever product and a licence to print money.
Google already has long experience designing its own Tensor Processing Units. What is changing is the breadth of the battle. AI infrastructure is no longer just an accelerator sitting in a rack. The bottlenecks are memory, networking, power, heat and how quickly data moves between machines.
That is why Marvell matters.
The company sits in the less glamorous bit of the AI boom: custom silicon and connectivity. It is not selling the public a chatbot. It is selling the picks, shovels and high-speed pipes needed when thousands of chips need to act like one machine.
In March, Nvidia itself invested $2 billion in Marvell and agreed to collaborate around Nvidia’s NVLink Fusion ecosystem, including silicon photonics and AI networking. Then Marvell’s acquisition of Celestial AI brought optical-interconnect technology into the fold for an upfront $3.25 billion, with further contingent consideration available if revenue milestones are met.
That tells you where the serious money thinks the next bottleneck sits: not just in chips, but in moving data between them fast enough without setting the power bill on fire.
The Overlooked Angle: Google Is Buying Optionality, Not Just Chips
The superficial take is that Google has made a huge investment in Marvell. It has not — at least not yet.
A warrant is an option. Google has the right, not the obligation, to buy shares at $206.58. And it gets more of that right as Marvell delivers qualifying revenue.
That makes this an unusually disciplined deal for Google.
If Marvell executes, Google gets a supplier deeply aligned to its roadmap and a potentially valuable equity position. If Marvell fails to deliver, Google has not written a $12.2 billion cheque upfront and crossed its fingers. The equity upside is tied to commercial success.
This is what smart strategic capital looks like. It is not charity. It is not a vanity investment. It is a mechanism for forcing incentives into the same lane.
There is another reason this should make founders sit up: Google has made revenue itself the currency of trust.
Too many startup deals are built around slides, forecasts and strategic adjectives. “Partnership” is one of the most abused words in business. It can mean anything from a signed procurement commitment to two blokes drinking bad coffee after a panel discussion.
This arrangement is different. The rewards are attached to a measurable output: actual revenue from actual products.
That is how you should structure important commercial relationships wherever possible. Not necessarily with warrants — that tool is expensive and specialised — but with clear milestones, reciprocal commitments and consequences for failure.
Why Broadcom Should Pay Attention — But Not Panic
Google’s custom-chip work has long been associated with Broadcom, particularly around TPUs. Reports earlier this year indicated Google was exploring new work with Marvell on inference-oriented chips and memory technology while also looking to diversify elements of its supply chain.
That has investors reaching for the dramatic headline: Google dumps Broadcom.
Slow down.
Google itself has said it remains engaged with Broadcom for the long term. And it would be reckless to assume a company running one of the world’s biggest computing estates would replace a major supplier overnight because a new partner has appeared.
But diversification is not nothing. It is the beginning of negotiation power.
When a hyperscaler builds credible alternatives, incumbent suppliers have to compete harder on price, delivery, technology and flexibility. That is the strategic threat. Not that Broadcom wakes up tomorrow with no Google revenue, but that Google has more choices by the time the next enormous contract is negotiated.
For Marvell, the prize is obvious: a path into a deeper share of Google’s AI stack. The danger is equally obvious: when your customer is Google, the revenue can be gigantic, but the customer concentration can become terrifying.
A company can become more valuable and more fragile at the same time. That is one of the oldest tricks in business.
The Contrarian Verdict: This Is Better News for Operators Than Speculators
Marvell’s share price will do what share prices do: sprint, wobble and give armchair analysts something to shout about online.
But if you are treating this as a quick stock-market punt, you are looking at the least interesting part of the deal.
The meaningful lesson is that AI is moving from experimentation to industrial procurement.
The first phase was about model capability: bigger models, better benchmarks, more demos. The next phase is about whether AI can be delivered cheaply, reliably and at enormous scale. That is an operations game.
It rewards companies that can do three things:
1. Build a product tied to a painful, expensive bottleneck. 2. Prove they can deliver at enterprise scale. 3. Structure contracts so growth is earned, not merely promised.
Marvell’s deal does not guarantee $120 billion in revenue. It establishes the ceiling implied by the warrant structure, not a cheque that has already cleared. Google still has to buy, Marvell still has to execute, and AI demand still has to justify the infrastructure build-out.
That is a lot of “ifs.” But they are the right ifs. They are operational ifs, not marketing ifs.
What This Means for You
If you are a founder, stop chasing strategic partnerships that do not change your bank account or product roadmap.
Ask three questions before celebrating any big-name deal:
- What exactly must each side deliver? - What gets measured, and how often? - What does the other side lose if it does not perform?
If the answers are vague, you do not have a partnership. You have a press release.
If you are an operator, look for the bottleneck behind the obvious trend. AI may be the headline, but the money is often hiding in the tedious bit: data movement, workflow redesign, compliance, integration, procurement or unit economics. The flashy layer attracts attention. The constraint captures value.
And if you are an investor, separate confirmed revenue from theoretical upside. A $12.2 billion warrant sounds like money changing hands. It is not. The more useful question is whether Marvell can win, manufacture and support enough qualifying business to earn it.
That is the whole game.
Google has not handed Marvell a trophy. It has handed it a very expensive performance review. The companies that get rich in this AI cycle will be the ones that can pass those reviews for years, not the ones that make the loudest noise this quarter.