On August 19, 2026, Marvell Technology disclosed to the SEC a commercial agreement to develop custom semiconductor products for Google, along with a warrant agreement granting Google stock purchase rights. The commercial agreement was signed on July 29, and the warrant was issued on August 18. This disclosure revealed both companies' names, the scope of the target products, and economic terms tied to actual purchase volumes.
However, the headline figures alone can be misleading. The price if Google exercises all warrant shares comes to roughly $12.2 billion—but this is neither an amount Google has already invested nor revenue Marvell has already received. Purchases of the target products remain at Google's discretion, and most of the rights are tied to cumulative revenue from those purchases. This agreement underscores the need to separate semiconductor functionality from contractual terms when asking who holds the core role in the TPU supply chain.
The $120 Billion Revenue Condition Matters More Than the "$12.2 Billion" Figure
Google received a warrant to acquire 58,970,907 shares of Marvell stock at $206.58 per share. If all shares are exercised, the total payment would be $12,182,209,968.06—approximately $12.18 billion. This figure therefore cannot be described as a "$12.2 billion investment." It represents the aggregate exercise price of the stock purchase rights, and exercising them at all remains at Google's discretion.
The breakdown of shares further clarifies the nature of the agreement. Of the total, 1,360,867 shares—about 2.3%—vest on a time basis, quarterly over one year from issuance. The remaining 57,610,040 shares—about 97.7%—vest in increments of 1/240th for every $500 million increase in Marvell's "Target Revenue."
Meeting all 240 vesting conditions would require $120 billion in cumulative Target Revenue. The measurement period runs from August 1, 2026, to January 29, 2033, with the standard expiration set at 5:00 p.m. California time on August 18, 2033. This $120 billion figure represents neither a Google purchase commitment nor a Marvell revenue forecast—it is simply the cumulative amount needed for maximum vesting.
Additionally, Google's stake in Marvell was initially capped at 4.999% or below. While Google can change this cap with 61 days' written notice, this does not mean Google immediately holds all 58,970,907 shares at the time of signing. The warrant is structured to align the interests of both companies in relation to future Target Revenue.
The Target Revenue referenced here does not encompass all of Marvell's sales to Google. Target Products are limited to semiconductors custom-developed under individual statements of work, for which Google holds the right to decide whether the complete product can be sold to third parties. The list of target products at the time of contracting has been omitted from the public filing, and future products will be added by written agreement between the two companies. How quickly the total approaches $120 billion will depend not only on Google's procurement volume but also on which products fall under this definition.
Contract Scope Doesn't Suggest a Full Transfer of Core TPU Development
Marvell's 8-K identifies the target products as custom silicon "associated with the TPU ecosystem." This specifically includes AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute. While this list is broad, it does not disclose that the design of Google's core TPU itself is being transferred to Marvell.
The warrant agreement defines "Product Launch" as the date on which Google qualifies the final production version of "Kestrel" as a product. However, whether Kestrel refers to the TPU itself, an inference accelerator, or a peripheral controller has not been disclosed. The product name alone cannot tell us its role or production timeline.
The latest generation Google officially announced in April consists of the 8th-generation TPUs, designated 8t and 8i. The 8i, targeting inference and reinforcement learning, features 288GB of HBM and 384MB of SRAM, doubles the number of physical CPU hosts per server, and transitions to Axion. Both products are slated for general availability in the second half of 2026. Comparing Google's official specifications with Marvell's disclosure suggests a contract that isn't limited to the TPU by name, but rather distributes responsibility across an entire system—spanning inference, networking, and memory peripherals. That said, no confirmed correspondence exists between Kestrel and the 8t/8i products.
Broadcom Remains Google's Partner for Future TPU Generations Through 2031
The Marvell agreement does not signify a replacement of Broadcom. In an 8-K filed April 6, Broadcom disclosed a long-term agreement to develop and supply Google's future-generation TPUs, along with agreements to supply networking for next-generation AI racks and other components through as late as 2031. While Marvell's scope and Broadcom's contract scope could potentially overlap, the disclosures do not necessarily indicate that the two companies split identical products under identical terms.
Broadcom's disclosure also references Anthropic's plan to access roughly 3.5GW of next-generation TPU compute capacity through Broadcom starting in 2027. However, this access depends on Anthropic's commercial success. Here too, planned access should not be conflated with actual usage volume.
Rather, it makes more sense to read Google's supply chain as one structured around dividing contract partners by function. The core TPU and inference accelerators can be separated, with distinct partners handling networking and memory peripherals as well. Marvell's new disclosure adds a revenue-linked warrant as an economic term to that division of labor. Given that Broadcom's long-term agreement remains in place, there is no basis for concluding that Broadcom has lost business or that an exclusivity arrangement has changed.
MediaTek's Production Plans and AMD Speculation Carry Different Levels of Certainty
At its July 31 earnings briefing, MediaTek stated that it plans to begin production in Q4 2026 on its first AI accelerator ASIC for a major U.S. cloud service provider. The company did not disclose the customer's name. MediaTek projects data center revenue exceeding $2 billion in 2026, and raised its estimated 2027 serviceable addressable market (SAM) to $80 billion, targeting a 15-20% share. A second AI accelerator ASIC is also on track for 2028 production, with yield and reliability reportedly proceeding as planned.
While these are official statements from MediaTek itself, they do not indicate that the customer is Google or that the chips will be adopted for any specific TPU generation. Even when placed side by side, Marvell's agreement, Broadcom's long-term contract, and MediaTek's production outlook each carry different levels of specificity regarding target products and customer disclosure.
Meanwhile, an excerpt from a SemiAnalysis client note offers market speculation that Google is working with AMD on a 10th-generation TPU project. Some analyses cite AMD's CPU IP and advanced packaging as possible reasons for such a partnership, but none of Google, AMD, or Broadcom has officially confirmed this. Google's most recently announced generation remains the 8t and 8i, and reports regarding AMD cannot be treated with the same level of certainty as the contract disclosures.
Stock price reactions are not the metric that matters here. What matters is which product Google ultimately qualifies as Kestrel, when and over what scope Marvell's Target Revenue accumulates, whether MediaTek discloses its customer's name, and whether Google or AMD officially confirms any collaboration. Only once these points are confirmed can we assess, product by product, which roles have actually shifted within Google's AI semiconductor supply chain.