Google Now Controls 25 Percent of Global AI Compute With 3.8 Million TPUs and 1.3 Million GPUs
The numbers in the new Financial Times report on Google's AI infrastructure are eye-watering. The search giant is now sitting on roughly 3.8 million Tensor Processing Units (TPUs) and another 1.3 million GPUs — enough to control about a quarter of all AI compute capacity on the planet. That makes Google the single largest AI compute owner in the world, bigger than Microsoft and Amazon combined on TPUs alone, and it explains a lot about why Gemini keeps getting cheaper while OpenAI's bills keep going up.
What 25% of Global AI Compute Actually Looks Like
Most people picture "compute" as a box of chips somewhere. The reality is closer to a small city. Google's TPU fleet alone is spread across more than a dozen mega-campuses, with single facilities pulling hundreds of megawatts. The 3.8 million figure includes every active TPU pod across v5p, v6e (Trillium), and the new Ironwood generation — and that's before you add the 1.3 million Nvidia GPUs Google rents out via Google Cloud.
By comparison, OpenAI's entire training-and-inference fleet is estimated at well under a million accelerators, and even that number leans heavily on borrowed Microsoft and CoreWeave capacity. Google built its stack in-house, financed by ad revenue, over more than ten years. That is the moat.
Why Google's TPUs Suddenly Look Like a Weapon
Until 2024, TPUs were treated as a Google-only curiosity. Then Apple confirmed it had trained Apple Intelligence on them, and Anthropic locked in a multi-gigawatt compute commitment with Google and Broadcom. TPUs are now an industry-grade product, not a private project.
The economic angle is brutal for Nvidia. Google's TPUs cost the company roughly half what an equivalent H100 buy would cost, because Google designs the silicon and Broadcom fabricates it on a margin much closer to a contract foundry than to Nvidia's 75 percent gross margin. Multiply that across 3.8 million units and you understand why Google can price Gemini 2.5 Pro at a fraction of GPT-5 and still be profitable.
The TurboQuant Bombshell That Made It Worse for Rivals
Google did not just build more compute — it figured out how to make existing compute do six times more work. Earlier this month, the company published research on a memory compression technique called TurboQuant that cut LLM memory consumption by roughly 6x with no measurable accuracy hit. That landed badly for memory-chip stocks and gave Google an additional efficiency multiplier on top of its raw chip lead.
This is part of a larger pattern. Google bought DeepMind in 2014, started TPU development the same year, signed the Broadcom partnership in 2016, and quietly accumulated capacity throughout the 2020s while Microsoft and Meta were still buying Nvidia. The hardware compounding is finally showing up in the spreadsheets.
My Take
Honestly, this should not surprise anyone who has been paying attention. Google has spent more than a decade building vertically integrated AI infrastructure while OpenAI has been renting chips by the quarter. The "OpenAI is winning" narrative was always a brand story, not a compute story. Now the compute story is catching up.
The real risk for Microsoft and Anthropic is not that Google wins on models — it is that Google decides to stop selling TPU access to outsiders the moment Gemini's enterprise share starts to scale. That would be a cold, smart move, and most people will miss the warning until the door has already closed.
Frequently Asked Questions
How many TPUs and GPUs does Google have?
Google operates approximately 3.8 million Tensor Processing Units (TPUs) and 1.3 million Nvidia GPUs as of April 2026, according to a Financial Times analysis. That gives the company roughly 25 percent of global AI compute capacity — more than any single rival.
Are Google's TPUs better than Nvidia GPUs?
For Google's own workloads, TPUs are typically cheaper per useful FLOP because Google co-designs the chip with the model. For external developers, Nvidia GPUs still offer a wider software ecosystem via CUDA, but TPUs are catching up fast through JAX and PyTorch/XLA support.
Why does Google need so much AI compute?
Google trains and serves Gemini, runs AI features in Search and Workspace, powers Google Cloud customers including Anthropic and (at times) Apple, and runs DeepMind research. With roughly 2 billion AI users across its products, the inference load alone is massive.
Who is the second-largest AI compute owner?
Microsoft is generally considered second, with most of its capacity inside Azure serving OpenAI workloads, followed by Amazon's AWS Trainium fleet and Meta's internal Llama training infrastructure.
The Bottom Line
If AI is the defining technology of the next two decades, then compute is the new oil — and Google is sitting on Saudi Arabia. Twenty-five percent global share is a structural advantage that no competitor can replicate without spending five years and tens of billions of dollars building from scratch. The AI race is not over, but the underlying balance of power just became a lot more visible.