☀️ AI Morning Minute: TPU
NVIDIA makes the chip everyone talks about. Google makes the one that runs everything Google does.
When people talk about AI hardware, NVIDIA dominates the conversation. But Google has been building its own AI chips since 2013, long before most people had heard the word “GPU” in the context of AI. The Tensor Processing Unit, or TPU, is Google’s custom silicon for AI workloads, and it powers Gemini, Google Search, Google Photos, Google Maps, and most of what Google AI does at scale. Anthropic runs the majority of Claude on them too.
What it means
A TPU is an application-specific integrated circuit, or ASIC, designed from the ground up for the matrix math that AI models rely on. Unlike a GPU, which is a general-purpose graphics chip repurposed for AI, a TPU does exactly one thing: the specific mathematical operations that neural networks need. That specialization makes it faster and more energy-efficient for AI workloads than hardware designed to do many things.
Google built the first TPU in 2015 after projecting that demand from AI inference would require doubling its data center capacity every few years. Rather than pay that cost, they built a chip that could handle AI workloads more efficiently. The current generation, TPU 8, released in April 2026, is the first time Google has split the architecture into two specialized designs: the TPU 8t for training and the TPU 8i for inference. The 8i delivers an 80% performance-per-dollar improvement over its predecessor for low-latency inference on large models.
Anthropic announced in October 2025 that it would access up to one million TPU chips worth tens of billions of dollars, one of the largest single AI infrastructure deals ever announced.
Why it matters
The chip that powers the competition runs on Google’s hardware. Anthropic, one of OpenAI’s main rivals, trains and serves Claude predominantly on TPUs. That means Google’s infrastructure is underneath a significant portion of the AI industry’s output, not just Google’s own products.
TPUs are purpose-built for what AI actually needs. GPUs handle graphics, video, scientific computing, and AI. TPUs handle AI and nothing else. That focus translates into better performance per watt for AI workloads, which matters enormously when you’re running hundreds of millions of queries a day. Google stated that TPU v7 delivers double the performance per watt of its predecessor.
Google is now selling TPU access externally and exploring putting them in space. The company has begun signing agreements to supply TPU hardware to customers who need on-premises infrastructure. Separately, Google announced Project Suncatcher, an initiative to launch solar-powered satellites carrying TPUs by early 2027, exploring orbital compute at scale.
Simple example
A restaurant that cooks everything on a general-purpose stove can make any dish. A pizzeria that installs a dedicated wood-fired oven makes better pizza faster, at lower cost per pie, than the general-purpose kitchen ever could. It can’t do anything else with that oven, but it doesn’t need to.
A GPU is the general-purpose stove. A TPU is the pizza oven, built specifically for the one thing Google needs it to do billions of times a day.