Company file
A cloud and custom-silicon builder pairing TPU architecture with data centers, networking, energy, and frontier-model demand.
Alphabet reported $91.4 billion of 2025 CapEx, mostly for servers, data centers, and networking, while Google described Ironwood TPU superpods with 9,216 chips. It shows the physical scale of custom compute without claiming a supplier commitment that was not disclosed.
Operating profile
How Google builds access
- Operating model
- Cloud platform and custom-silicon builder
- Controls through
- TPU systems, data centers, networking, software, and long-duration cloud demand
- Physical stack
- TPUs, servers, memory, data-center networking, power, facilities, and cloud operations
Custom silicon is a full-stack choice
Google's TPU program joins accelerator design with the servers, data centers, networking, and software that make a cloud platform useful. Alphabet reported $91.4 billion of 2025 CapEx, about 60% for servers and 40% for data centers and networking. That spending is not a supplier commitment balance, but it makes the physical context clear: custom silicon has to be supported by an entire fleet before customers can use it.
The TPU is designed as a system, not an isolated device
Google described Ironwood TPU superpods with 9,216 chips. The useful point is not the chip count by itself. A superpod represents an integrated design problem involving compute, memory, network fabric, thermal design, software scheduling, and data-center operations. The system becomes valuable when all of those layers can be brought into service together for a customer workload.
Anthropic turns TPU design into a demand signal
Anthropic said it planned to expand its use of Google Cloud technologies to up to one million TPUs, an expansion worth tens of billions of dollars that it expected to bring well over one gigawatt online in 2026. In April 2026, Anthropic added a Google and Broadcom agreement for multiple gigawatts of next-generation TPU capacity expected to begin coming online in 2027. These are cloud and capacity commitments, not Alphabet CapEx or a disclosed wafer allocation.
Demand has to meet a real physical fleet
A frontier-model customer can provide a powerful demand signal, but Google still has to translate that signal into powered sites, servers, networking, custom silicon, and operating capacity. The strongest infrastructure programs preserve the distinctions between these layers. A planned gigawatt, a cloud-services commitment, a new TPU generation, and annual CapEx all describe different aspects of the same buildout rather than one interchangeable total.
Cloud customers create a learning loop for the platform
Large workloads reveal where a fleet needs more performance, bandwidth, capacity, or operational refinement. That feedback can inform the next TPU generation, the next superpod design, and the next data-center deployment. Google's distinctive advantage is the opportunity to connect a customer's real demand with its own silicon, software, and operating teams, then use that learning to improve the next physical buildout.
What to watch
Watch whether TPU design, data-center construction, power availability, networking, and customer adoption can expand with a compatible rhythm. Google's compelling role is to show how a cloud platform can make custom silicon useful at scale. Its advantage is not factory ownership. It is the ability to design, deploy, operate, and continuously improve the full environment in which a custom accelerator works.