Google’s $12 billion Marvell deal bets on custom silicon for future TPUs

The agreement valued at $12 billion brings Marvell into a deeper role supporting Google’s upcoming tensor processing units, shifting the effort from a mere research venture to a full vertical integration push. Existing ties already see Marvell providing ASICs for Google’s networking silicon and earlier TPU models; the new pact expands that relationship across multiple generations and may encompass inference‑focused TPUs aimed at cloud clients. Initial coverage by the Wall Street Journal disclosed the transaction, yet details such as a public roadmap, launch dates, or technical specifications remain absent beyond the “strategic partnership” label.

Questions arise concerning Google’s parallel work with Broadcom, another key custom silicon collaborator in networking. Marvell’s recent advances in high‑speed SerDes and coherent DSPs for data‑center interconnects hint at a possible consolidation of compute and networking IP under a single supplier. An alternative motive could involve alleviating capacity bottlenecks, given the intense demand for TSMC’s advanced packaging services; securing Marvell’s design talent might guarantee priority for Google’s projects.

Should the capital flow extend to TPU v6 and TPU v7, production timelines could stretch to 2026–2027, fitting Google’s long‑range AI hardware outlook. This horizon coincides with speculation about architectures beyond transformers, though it remains unclear whether the funding will simply scale existing designs or venture into entirely new concepts. Market reaction saw Marvell’s shares climb 20% following the announcement, reflecting confidence in its custom ASIC expertise. The ultimate effect will depend on whether future TPUs stay confined to internal workloads or become available on Google Cloud Platform at rates that challenge NVIDIA’s H200 and undefined families.

Unresolved technical aspects include whether the contract encompasses optical interconnect intellectual property or is restricted to compute dies, with reports offering conflicting interpretations. As the broader cloud AI accelerator arena grows—featuring AWS Trainium2, Azure Maia, and merchant silicon from AMD and Intel—Google’s forthcoming strategy will influence its standing amid increasing competition.

The version displayed here has only undergone formatting tweaks and the removal of internal web links. The perspective conveyed originates from a single Google employee and does not represent the entire corporation. Our platform serves merely as a conduit for this document, which surfaces several compelling observations. SemiAnalysis operates as an ad‑free, reader‑supported outlet. An uncomfortable reality acknowledges that neither our organization nor OpenAI seems positioned to dominate the ongoing arms race. While internal disputes have persisted, a third party has been quietly gaining advantage. Challenges once labeled as “major open problems” have already been resolved and are in active use; for example, large language models now run on a Pixel 6, achieving about 5 tokens per second.