✦ The Google AgreementNarrow moat

In August 2026 Marvell filed an 8-K disclosing a commercial agreement with Google to design custom semiconductor products that attach to Google's TPU ecosystem: AI inference accelerators, storage controllers and network interface controllers¹. With Amazon, Microsoft and Meta already engaged, this gives Marvell a relationship with every major American hyperscaler running a custom-silicon program.

The symbolism is larger than the initial revenue. Google's TPU is the oldest and most successful custom accelerator program in existence and has been co-designed with Broadcom across seven generations. A customer that sophisticated deciding it wants a second design partner — even for adjacent silicon rather than the accelerator core — is the strongest available evidence that hyperscalers will deliberately fund an alternative to their incumbent supplier.

The realistic framing is that Marvell has been invited to the periphery of a program someone else owns. Inference accelerators, storage controllers and NICs are valuable, but they are not the TPU. Watch whether the engagement expands toward the accelerator itself in later generations, and watch for disclosed revenue. An adjacent-silicon relationship that never becomes a core one is a decent business; becoming Google's second accelerator source would be a different company entirely.