Google and Marvell just signed an AI chip deal that headlines are calling worth “$120 billion” – but that’s not a check Google is writing today. The real financial commitment behind the Marvell-Google AI chip deal is a $12.2 billion stock warrant that only pays out as Google actually buys more custom AI chips over the next several years. The agreement covers new AI inference chips and the supporting hardware built around Google’s own TPUs, and it’s one of the clearest signs yet that Big Tech is racing to build its own AI silicon instead of relying only on Nvidia GPUs.
Key Takeaways
- The “$120 billion” figure is Marvell’s projected revenue from Google through 2033, not the price Google is paying right now.
- Google’s actual commitment is a warrant to buy about 59 million Marvell shares, worth up to $12.2 billion if it’s fully exercised.
- The warrant vests in stages, unlocking more shares only as Google hits specific chip-purchasing milestones with Marvell.
- The deal covers AI inference chips plus the storage, network, and memory hardware that supports Google’s TPUs, not a replacement for TPUs.
- It’s part of a bigger trend: major AI companies are increasingly building or backing custom chips instead of depending only on Nvidia GPUs.
The Marvell-Google AI Chip Deal: What Was Actually Announced
On August 19, 2026, Marvell disclosed an expanded, multi-year commercial agreement with Google covering custom semiconductor products tied to Google’s Tensor Processing Unit (TPU) ecosystem. The underlying deal was actually signed back on July 29, 2026, but the financial details came out later. As part of it, Marvell issued Google a warrant to purchase up to roughly 58.97 million Marvell shares at $206.58 each, exercisable anytime through August 2033, as Reuters reported.
The agreement covers several categories of custom silicon: AI inference accelerators (chips built to run already-trained AI models quickly and cheaply), plus storage controllers, network interface controllers, memory interface controllers, and near-memory compute hardware. None of this replaces Google’s TPUs, which the company already designs in-house with help from Broadcom. Instead, it’s the supporting cast: the specialized chips that move data in and out of a TPU cluster fast enough to keep it running efficiently. If you want the bigger picture of what an “AI factory” actually looks like end to end, we broke that down in our explainer on how Nvidia’s Vera Rubin platform works.
Wait, Is It $120 Billion or $12.2 Billion?
Both numbers are real, they just measure completely different things. The $12.2 billion is the maximum value of the stock warrant itself: 58.97 million Marvell shares at the $206.58 strike price. The $120 billion is a projection of Marvell’s total revenue from Google over the life of the agreement, through 2033, if every single vesting condition is met.
Here’s the mechanism, according to a detailed breakdown from The Next Platform: only 1.36 million of the warrant shares vest automatically over the first four quarters, essentially a modest starting bet. The remaining roughly 57.6 million shares are split into around 240 separate tranches, each one unlocking only after Google buys another $500 million worth of qualifying chips from Marvell. In other words, Google earns the right to buy more Marvell stock at a fixed price, but only in proportion to how much hardware it actually orders. If Google never scales up its purchases, most of that warrant never vests, and the $120 billion revenue figure never materializes.
Investors clearly liked what they saw anyway. Marvell’s stock jumped enough on the news to add about $18.6 billion in market capitalization in a single day, which is actually more than the warrant itself is worth. That reaction was less about the specific dollar figures and more about what the deal signals: Marvell landing a second major hyperscaler customer for custom AI silicon, alongside its existing work for Amazon.
Why Are Tech Giants Building Their Own AI Chips Instead of Just Buying GPUs?
This deal is one data point in a much bigger shift happening across the AI industry: the biggest AI companies are increasingly designing their own chips rather than buying everything off the shelf from Nvidia.
The Nvidia bottleneck
Nvidia’s GPUs (graphics processing units) are the default choice for AI training and inference because they’re powerful and flexible enough to run almost any kind of AI workload. But that flexibility comes at a cost, literally. Demand for Nvidia’s top chips has outstripped supply for years, prices stay high, and every company competing for AI capacity is bidding for the same limited pool of hardware. For a company running AI at Google’s scale, that’s an expensive dependency to have on a single supplier.
Why custom silicon wins at scale
A general-purpose GPU has to be good at a wide range of tasks for a wide range of customers. A custom chip, by contrast, can be designed to do exactly one job extremely well, like running a specific kind of AI model’s inference step, and nothing else. When you’re running that one job billions of times a day, as Google does with search and Gemini, a purpose-built chip can end up cheaper and more power-efficient than a general-purpose GPU doing the same work. That’s the logic behind Google’s TPUs, Amazon’s Trainium chips, and now this expanded Marvell partnership covering everything around the TPU itself. We go deeper into how these massive AI data centers are built and powered in our guide to AI supercomputing platforms, and in our look at what August 2026’s AI price war actually means for the market, since chip costs are a big part of why AI pricing keeps shifting.
What This Means Going Forward
Nothing about this deal changes what you’ll notice using Google Search, Gemini, or any other Google AI product tomorrow. The timelines are long: qualification for parts of the program is targeted around late 2027, with broader deployment expected in new data centers by 2028, and the full agreement running through 2033.
What it does signal is a continued, structural shift in how AI infrastructure gets built. Google now has two major custom-silicon partners (Broadcom for TPUs themselves, and an expanded role for Marvell around them), which reduces its dependence on any single supplier, including Nvidia. Expect more deals shaped like this one: chipmakers trading warrants or equity stakes for guaranteed, revenue-linked purchase commitments from the handful of companies that can actually afford to build AI at planetary scale.
Is Google buying Marvell?
No. Google isn’t acquiring Marvell or taking direct ownership today. It received a stock warrant, which is an option to buy Marvell shares later at a fixed price, as part of a multi-year custom chip supply agreement.
What is a stock warrant, in plain English?
A warrant is a contract that gives its holder the right, but not the obligation, to buy a company’s stock at a fixed price before a set expiration date. Google can buy Marvell shares at $206.58 each anytime through August 2033, but only the portion of the warrant that has vested based on how many chips it has purchased.
What’s the difference between a TPU and a GPU?
A TPU (Tensor Processing Unit) is a chip Google designs specifically for its own AI workloads, like training and running models such as Gemini. A GPU (Graphics Processing Unit), Nvidia’s specialty, was originally built for graphics rendering and later adapted into a flexible, general-purpose AI chip. TPUs can be more efficient for Google’s specific needs, while GPUs remain more flexible for a wider range of customers and tasks.
Why did Marvell’s stock jump on the news?
Investors read the deal as proof that Marvell can win large, multi-year custom-chip contracts with major hyperscalers beyond its existing customers. Marvell’s market value rose by about $18.6 billion in a single day, which is actually more than the total value of the warrant itself.
When will these new chips actually be used?
Google and Marvell have set a qualification target around late 2027 for parts of the program, with broader deployment expected around 2028 inside new data centers built for large-scale AI inference.
The headline number will keep getting rounded up to “$120 billion” in tweets and TV chyrons, but the real story is smaller and more interesting: Google is paying for AI chip capacity with future stock upside instead of cash up front, and it’s betting that its own bill for AI hardware is only going to grow.
