Google is not just building better AI models anymore. It is rebuilding the hardware behind them.
Full report via Fortune
https://fortune.com/2026/04/23/google-rolls-out-latest-custom-ai-chips/
According to reports, Google has rolled out its latest custom chips, the eighth generation of its Tensor Processing Units, called TPU 8t and TPU 8i. But the real story is not the names, it is the split.
For the first time, Google is designing separate chips for different stages of AI.
The TPU 8t is built for training massive models, the heavy lifting that happens before AI systems go live. The TPU 8i, on the other hand, is designed for inference, the part where AI actually responds to users in real time.
That division is intentional.
As Google explained, these chips are “tailored to an agentic age,” where AI systems are no longer just answering questions but actively reasoning, planning, and executing tasks.
And that shift changes everything.
Training used to be the main focus in AI. Build the biggest model, feed it the most data, and that was the advantage. But now, inference is becoming just as important, because AI is being used live, inside apps, workflows, and real time systems.
So instead of one general purpose chip, Google is building specialised infrastructure for each phase.
There is also a business angle here.
These chips are not just internal tools. They power Google’s cloud, which it rents out to companies like OpenAI, Anthropic, and Meta. That means better chips are not just about performance, they are about attracting developers and owning more of the AI stack.
And there is competition.
Nvidia still dominates AI hardware, but moves like this show Google is trying to reduce its dependence and build its own vertically integrated system, from models to infrastructure to deployment.
What this really signals is a deeper shift.
AI is no longer just a software race. It is a hardware race too.
And as models become more capable and more embedded in everyday tools, the companies that control the infrastructure underneath may end up controlling the entire ecosystem.
So the question is no longer just who builds the smartest AI.
It is who owns the machines that run it.
