Google Ties Up With Marvell on Custom AI Silicon, Takes Warrants for 59 Million Shares to Cut Broadcom Reliance

Google secured warrants to purchase 59 million Marvell Technology shares as part of a multi-year custom AI silicon agreement disclosed on August 20, in a deal explicitly framed by the parties as a way to cut Google's reliance on Broadcom for AI accelerator interconnect and packaging. The agreement, which pairs Marvell's custom-silicon and high-speed-serdes capabilities with Google's TPU roadmap, is the third major non-NVIDIA AI silicon announcement in two weeks and accelerates the structural disaggregation of the AI accelerator stack that has been underway since 2024.

What the Warrants Actually Lock In

The 59 million Marvell warrants vest on a multi-year schedule tied to Marvell's delivery of custom AI silicon, advanced packaging and high-speed interconnect IP into Google's TPU and Gemini training infrastructure. While the strike price was not disclosed, the warrant count represents the equity component of a deal that, according to two analysts familiar with the agreement, includes multi-billion-dollar direct-purchase commitments over the next three years. The deal is structured so that Google's effective cost-of-silicon decreases as Marvell hits delivery milestones; Marvell's market capitalization, which closed at roughly $78 billion before the announcement, rose 9 percent on the news.

Why Google Is Reducing Broadcom Exposure

Broadcom has historically been Google's primary partner for custom AI ASICs — Google's first-generation TPU and several subsequent generations were co-developed with Broadcom, and Broadcom remains a key supplier of high-speed networking ASICs for Google's data centers. The Marvell deal does not terminate that relationship; rather, it gives Google a second qualified supplier for the most strategically sensitive parts of the AI compute stack. Reducing single-supplier dependency on advanced packaging and on custom-serdes IP is the kind of structural hedge that AI hyperscalers have been quietly building across 2025 and 2026, partly in response to NVIDIA's vertical integration and partly in response to geopolitical supply-chain risk in Taiwan.

The Broader Silicon Disaggregation

The Google–Marvell agreement is the third major non-NVIDIA AI silicon announcement in two weeks. On August 15 at Cloud Next, Google announced TPU v7 Ironwood commitments with Anthropic, Apple and CoreWeave. On August 18, AMD's data-center segment disclosed a $4.5 billion backlog through the second half of 2026, driven largely by hyperscaler demand for its MI400-series accelerators. The three announcements together tell a consistent story: the AI compute stack is moving from a single-vendor-dominated GPU path to a multi-vendor mix that includes custom ASICs, TPU-class accelerators and high-end AMD GPUs. The market implication is that NVIDIA's pricing power, while still dominant, faces structural headwinds that compound with each new hyperscaler announcement.

The Marvell Custom-Silicon Pipeline

For Marvell, the deal extends a transformation that began with the 2024 Inphi acquisition. Marvell's data-center segment, which reported $1.5 billion in revenue in its most recent quarter, has been repositioning around AI custom silicon and advanced interconnect. The Google warrants are a multi-year validation of that strategy and effectively give Marvell a strategic customer counterpart to Amazon's Trainium and Microsoft's Maia programs. For Google's TPU roadmap specifically, the Marvell partnership allows Google to integrate Marvell's advanced 3-nanometer and 2-nanometer packaging and high-speed interconnect IP directly into the next two TPU generations, with first silicon expected in 2027.

What to Watch Through Year-End

Three near-term checkpoints follow. NVIDIA's Q2 FY2027 earnings on August 26 will show whether the multi-vendor hedging by hyperscalers has begun to compress data-center revenue growth at the top end of the range. AMD's MI400 ramp through the second half of 2026 will show whether non-NVIDIA GPU demand is reaching the scale needed to support a true three-vendor AI compute market. And Marvell's next earnings call, scheduled for late September, will provide the first quantitative disclosure of how much of the Google commitment is recognized in backlog and how the warrant vesting schedule maps to revenue.