memujo
Technology4 min read

Google's $120 Billion Marvell Deal: How Chip Suppliers Are Paying Big Tech for the Privilege of Supplying Them

Marvell issued Google a warrant for 58.97 million shares, roughly 7% of the chipmaker, at $206.58 per share. Most of the shares vest as Google's discretionary chip purchases accumulate, with one block for every $500 million spent. Google controls the pace entirely, and nothing obligates it to spend a dollar.

By Memujo Editorial

Google's custom silicon strategy took a major step forward this week as Marvell Technology disclosed an SEC filing that revealed the scale and unusual structure of its partnership with Alphabet's search division. The deal, signed on July 29 but made public via an August 19 filing, represents one of the most significant custom chip agreements to come out of the AI infrastructure build-out.

At the core of the deal is a warrant that gives Google the right to purchase up to 58.97 million Marvell shares at a fixed price of $206.58 per share. In total, exercising every warrant would cost approximately $12.18 billion, and Marvell shares jumped nearly 10% on the news, meaning the package is already worth more than $2 billion more than it would cost to exercise.

What Google is buying

Marvell will develop custom products across a broad range of silicon attached to Google's Tensor Processing Unit ecosystem:

  • AI inference accelerators for serving large language models and other AI workloads at scale
  • Storage controllers for the massive data pipelines feeding AI training and inference
  • Network interface controllers for connecting thousands of TPUs in a single cluster
  • Memory interface controllers for keeping data flowing to accelerators
  • Near-memory compute for processing data where it sits rather than shuttling it across the system

The breadth of the agreement is notable. Google is not just buying a single accelerator, it is outsourcing the design of multiple semiconductor functions that surround the TPU itself. The filing even identifies a custom product code-named "Kestrel," though it does not disclose its architecture or function.

How the warrant works

The warrant structure is the part that most analysts are paying attention to, because it inverts the usual dynamic between a buyer and a supplier.

Most of the warrant's shares, 57.61 million of them, vest in performance-based tranches as Google's discretionary chip purchases accumulate. For each $500 million of qualifying revenue that Marvell recognizes from Google-related products, approximately 240,000 shares vest. There are up to 240 such revenue thresholds, implying that full vesting corresponds to as much as $120 billion in cumulative qualifying revenue during the measuring period running from August 1, 2026 through January 29, 2033.

A smaller slice of 1.36 million shares vests on a time-based schedule during the first year regardless of purchases.

Critically, the filing makes clear that Google's purchases are entirely discretionary. Nothing obligates the company to spend a dollar on these products. Google controls the pace, and the warrant converts its spending decisions into equity upside.

Why this matters

There are three things to watch for in the market reaction.

First, the Broadcom implication. Google has relied on Broadcom to design its TPUs since 2015. The two companies expanded a long-term agreement in April 2026 covering future TPU generations with component supply running through 2031. The Marvell deal does not give Marvell exclusivity, but it does mean Google is adding a second supplier and negotiating equity from both. Broadcom's shares fell roughly 5% on the news, reflecting investor concern about competitive pressure on a supplier that has enjoyed a decade-long monopoly on Google's TPU design.

Second, the OpenAI-AMD precedent. A similar structure emerged last October when AMD announced a partnership to supply OpenAI with GPUs and issued the AI company a warrant for up to 160 million AMD shares, with tranches vesting as deployments scaled. The Google-Marvell version is more conventional with a real exercise price and not the one-cent strike that AMD offered, but the direction is the same: the biggest buyers of AI silicon have started collecting equity from their suppliers for the privilege of supplying them.

Third, the Google advantage. Google's Q2 capital expenditure was $44.9 billion, roughly double the year-ago figure, and the company spent $132 billion over the trailing twelve months. At this pace, the $120 billion qualifying revenue threshold is achievable within a few years. Meanwhile, Google's Q2 earnings already included a $98 billion net gain primarily from unrealized gains on equity securities, meaning the warrant shares, once vested, would land in exactly that same bucket.

The broader picture

The Marvell-Google deal is one data point in a much larger trend. Every major cloud provider, from Amazon and Meta through Microsoft and Google, is investing heavily in custom silicon to reduce dependence on Nvidia. Google's TPUs have been in production since 2015 and now power a significant share of Alphabet's AI workloads. The question is no longer whether hyperscalers should build custom chips, but how many suppliers they need to keep prices competitive and supply chains resilient.

Marvell's entry into Google's TPU supply chain suggests the answer is "more, not fewer." And the warrant structure tells you something else: in the current AI build-out, buyers with enough spend to move needle are the ones with leverage.

Marvell shares are currently trading near $243, well above the warrant exercise price of $206.58. If Google exercises the full warrant, it will own roughly 7% of the company, a material stake that gives the search giant a permanent equity position in one of the most important infrastructure plays in the AI supply chain.

The filing was made by Marvell Technology on August 19, 2026. Google declined to comment on the warrant terms.

  • #google
  • #marvell
  • #ai-chips
  • #tpu
  • #custom-silicon
  • #securities-filing