Google infrastructure lead: electricity is the fundamental bottleneck to AI expansion

In a recent interview with Sequoia Capital partner Sonia Huang, Google's head of AI infrastructure, Amin Vahdat, laid out the company's approach to AI infrastructure, covering compute trends, technology roadmaps and key constraints. He said Google's capital expenditure is expected to exceed $200 billion in 2026, with most of it going toward data center construction.

Vahdat said the rise of long-horizon agents is fundamentally changing data center design, pushing up demand not only for accelerators but also for CPUs, networking and storage. To keep pace, Google needs to roughly double its service-side token generation capacity every six months, with gains coming mainly from ongoing software and model optimization rather than hardware alone.

He described electricity as the most fundamental bottleneck to AI expansion, ahead of chips or manufacturing capacity. Google's preferred approach is to interconnect with utilities, using statistical multiplexing to reduce redundancy costs, while taking on the cost of new transmission lines and substations. Where shortfalls remain, the company may temporarily build local solar and storage capacity.

Vahdat also said TPU inference and training chips were split into separate product lines for the first time in 2026, with both retaining the ability to handle the other's workloads. He discussed optical circuit switching for fault tolerance and network reconfiguration, goodput as a more meaningful performance metric, and longer-term compute forms including highly integrated racks and orbital data centers, which Google is pursuing as a moonshot project, though cooling, servicing and inter-component communication remain key challenges.