Google builds chip supply chain with four partners to challenge Nvidia in AI inference

Google has announced the construction of the most diversified custom chip supply chain in the AI industry—with four design partners and a roadmap extending through the end of 2027. The company is betting on reducing the cloud market's dependence on Nvidia in the AI inference segment. Google's partner network includes four major players: Broadcom, MediaTek, Marvell, and Intel. Each is responsible for a separate direction in chip architecture and manufacturing. The strategy was detailed 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 being shipped in the millions. This is the first signal that Google is ceasing to be merely a consumer of third-party hardware and is systematically becoming a self-sufficient player in the AI infrastructure market. The roadmap's final milestone is marked by TPU v8 chips, which will be manufactured using TSMC's 2nm process and will arrive in the second half of 2027.

Google's strategy divides the next generations of TPU into several independent streams—by design partners and manufacturing capacity. This approach insures against supply disruptions characteristic of single-vendor dependency and provides flexibility in scaling performance for different inference tasks.

Inference—that is, running already-trained models to generate answers for users—has become the main battlefield of the AI chip market in 2025–2026. This is where the primary demand of major cloud providers is concentrated: model training happens once, while inference runs continuously. Nvidia maintains dominance thanks to its GPU architecture and mature CUDA ecosystem; however, Google's custom TPUs traditionally demonstrate superior energy efficiency on transformer model tasks.

For the market, this means intensifying competition in the cloud AI hardware segment. If Google succeeds in scaling its own supply chain to the required volumes, its customers will have a real alternative to Nvidia's GPU clusters—with potentially lower cost per inference token and reduced dependence on a single supplier.