Google builds a four-partner chip supply chain to challenge Nvidia in AI inference

Google has announced the most diversified custom-chip supply chain in the AI industry, with four design partners and a roadmap running through the end of 2027. The company is betting that it can reduce the cloud market's dependence on Nvidia in AI inference.

Google's partner network includes four major players: Broadcom, MediaTek, Marvell and Intel. Each handles a separate area of chip architecture and manufacturing. The strategy was laid out in detail ahead of Google Cloud Next, the company's key annual event for enterprise customers and partners.

The first product of this ecosystem is the Ironwood TPU, a next-generation chip that is already shipping in the millions. It is the first sign that Google is ceasing to be merely a consumer of other companies' hardware and is steadily becoming a self-sufficient player in the AI infrastructure market.

The roadmap ends with TPU v8 chips, which will be manufactured on TSMC's 2nm process and arrive in the second half of 2027. Google's strategy splits the next generations of TPU into several independent streams, by design partner and by manufacturing capacity. This approach hedges against the supply disruptions typical of depending on a single vendor, and gives flexibility in scaling performance for different inference workloads.

Inference, meaning running already-trained models to generate answers for users, became the main battleground of the AI chip market in 2025–2026. This is where the main demand from large cloud providers is concentrated: a model is trained once, while inference runs continuously.

Nvidia retains its dominance thanks to its GPU architecture and the mature CUDA ecosystem, but Google's custom TPUs have traditionally shown better energy efficiency on transformer-model workloads.

For the market, this means stiffer competition in cloud AI hardware. If Google manages to scale its own supply chain to the required volumes, its customers will get a real alternative to Nvidia GPU clusters, with potentially lower inference cost per token and less dependence on a single supplier.