Marvell just got a Google route into TPU-attached custom AI silicon—tightening the “ASIC duopoly” pressure on NVIDIA’s inference economics

Marvell disclosed a commercial agreement with Google where Marvell develops custom semiconductor products intended to “attach to the TPU ecosystem,” spanning custom silicon programs for AI inference and supporting data-center functions. The equity component is explicit: Google received a warrant for up to 58,970,907 shares at an exercise price of $206.58—an implied value of about $12.2B—creating a direct linkage between Google’s custom-silicon engagement and Marvell’s delivered revenue.

In the disclosed agreement, Marvell’s custom products are not limited to a single accelerator-like block. The filing describes custom semiconductor products that attach to the TPU ecosystem, including AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and “near-memory compute.”

Hyperscalers buying custom silicon usually optimize three things at once: performance-per-watt, throughput scaling, and total system cost (including the surrounding I/O and memory hierarchy). By giving Marvell a TPU-attached, inference-oriented pathway, Google increases the credible set of silicon design partners beyond any single supplier-centered strategy. The investor takeaway is that NVIDIA faces a second-order effect: even when GPUs remain the primary training/inference engines, faster near-TPU cost-down via custom-attached components can tighten the pricing/ROI envelope.

Marvell’s financial trajectory shows a swing from losses into profitability by FY2026. FY2025 revenue rose to $5.77B from $5.51B in FY2024, then accelerated to $8.19B in FY2026. With operating income turning positive in FY2026 ($1.34B), the company is in a better position to absorb the R&D intensity typically required for custom-silicon programs—making the Google-linked revenue trigger more material to earnings power than it would be for a cash-constrained supplier.

NVIDIA’s economics are defended by ecosystem lock-in (software, toolchains, and performance consistency) and by the breadth of GPU coverage. But custom TPU-attached silicon—especially inference-targeted components—can change procurement discretion in the cluster design phase. If Google’s “Frozen” approach (private chip codenames mentioned in the broader market narrative) results in measurable inference efficiency gains using TPU-attached surrounding components, NVIDIA can face margin pressure through accelerated buyer substitution: fewer “must-have” GPU compute units for the same inference workload.