Google enters talks with Marvell Technology to develop efficient AI inference chips

Alphabet is reportedly in advanced negotiations with Marvell Technology to develop two specialised AI chips, marking a significant step in Google’s quest to challenge Nvidia’s market dominance.

According to reports surfaced on Sunday, the partnership focuses on a bespoke memory processing unit and a next-generation Tensor Processing Unit (TPU) specifically optimised for AI inference.

The companies aim to finalise these designs as early as next year, with the memory unit scheduled for test production shortly thereafter to alleviate critical data bottlenecks in large-scale models.

TPU sales have emerged as a primary engine for Google Cloud’s revenue growth, particularly following a landmark deal in early 2026 where Meta Platforms agreed to lease Google's hardware.

By diversifying its chip architecture with Marvell alongside existing partners like Broadcom, Google intends to drive down the immense costs associated with serving AI models to hundreds of millions of users.

This strategic pivot toward high-efficiency inference silicon is viewed by industry analysts as a vital move to secure long-term commercial returns on Alphabet's multi-billion-pound AI investments.

Simultaneously, Alphabet is expanding its influence within the public sector through a high-stakes negotiation with the US Department of Defence.

The proposed agreement would allow the Pentagon to deploy Google’s Gemini AI models within classified military environments to streamline administrative tasks and reduce operational overheads.

To address ethical concerns, Google has reportedly insisted on strict contractual safeguards, including prohibitions against domestic mass surveillance and the deployment of autonomous weaponry without human oversight.

This double-pronged approach—advancing hardware independence while securing elite government contracts—signals Google’s determination to lead the global AI infrastructure race.