Google's Compute Empire and the Coming Agent Economy

welcome to Briefly, AI. A daily podcast made by AI, about AI, using a grumpy AI voice of a real human called Harry. Let's get into it.

Right. Google now controls a quarter of the world's AI compute. Not a quarter of the market — a quarter of the entire planet's training infrastructure. Let's talk about what that actually means.

Epoch AI just put numbers to something most people suspected but hadn't quantified. Google's sitting on roughly 3.8 million TPUs and 1.3 million GPUs. That's not just impressive kit — that's structural advantage. When Thomas Kurian, Google Cloud's CEO, says demand and revenue justify the spend, he's being polite. What he means is: we're printing money, and the moat's getting wider.

Here's why it matters. If you're OpenAI, Anthropic, or any lab trying to train the next frontier model, you're either renting from Google, Microsoft, or AWS — or you're building your own data centre, which takes years and costs billions. Google doesn't have that problem. They own the infrastructure. They can run experiments other labs can't afford to try. And every time a competitor writes them a cheque for compute, Google learns exactly what's being built and how much juice it takes.

This isn't hypothetical. Oracle just closed a $16 billion bond to build a Michigan campus for OpenAI. We covered that last week. Anthropic took $40 billion from Google with heavy infrastructure strings attached. The labs building the models are increasingly tenants. And the landlords are starting to look like the real power brokers.

Now, speaking of Anthropic — they've been busy. This week they published research on something called Project Deal. Sounds boring. It's not. They built a test marketplace where Claude agents represented buyers and sellers, and then let them negotiate real transactions for real goods with real money. No humans in the loop. Agent-versus-agent commerce.

And it worked. The agents struck deals, haggled over price, adapted their strategies based on what the other side was doing. Anthropic's framing it as a research paper, but let's be honest — this is a product demo with footnotes. They're showing you what's coming: a world where your AI assistant doesn't just book your holiday, it negotiates the rate with the hotel's AI, which is doing the same thing for a thousand other bookings at the same time.

If that sounds efficient, it is. If it also sounds like it could get weird very fast, you're right about that too. What happens when two agents deadlock? Who's liable when an agent agrees to terms you didn't want? What does competition law look like when the entire economy's running on algorithmic negotiation? Nobody really knows yet. But the infrastructure's being built anyway.

Right, last one — and this one's about sovereignty, which is a word we're going to hear a lot more of. Cohere, the Canadian AI lab, is merging with Germany's Aleph Alpha. Backing the deal: Schwarz Group, which owns Lidl and has extremely deep pockets and a strong opinion about not depending on American AI.

The framing here is blunt. Europe wants an alternative to OpenAI, Google, and Anthropic that doesn't route through US data centres or US export controls. Cohere's good at language models. Aleph Alpha's got European regulatory credibility and government relationships. Together, they're positioning as the sovereign option — the stack you pick when data residency isn't just compliance, it's geopolitics.

This matters even if you're not European. Because the AI supply chain is fragmenting. We've already got a US-China split. Now we're watching a transatlantic one take shape. Different models, different rules, different ideas about what's acceptable. If you're building anything that crosses borders, you're going to need to navigate that. And if you're a lab trying to scale globally, you've just lost a chunk of the map.

That's your lot. Google's tightening its grip on the physical infrastructure. Anthropic's agents are learning to haggle. And Europe's building its own stack. Three stories, one thread: the AI economy's consolidating fast, and the decisions being made right now about who owns what are going to shape the next decade.

If any of that was useful, share it. If not, well — blame the machine. See you next time.