Google Cloud's Chip Strategy Explained by CEO

Matthew Berman just sat down with Google Cloud CEO Thomas Kurian, and the interview surfaces something genuinely interesting about the AI infrastructure race: while everyone else is screaming about compute constraints, Google's quietly playing a different game entirely.

The setup seems paradoxical at first. Anthropic and OpenAI can't get enough GPUs. They're compute-starved, publicly begging for more chips. Meanwhile, Google Cloud is training its own models, selling inference to customers, letting competitors build on their infrastructure, and selling TPUs directly. How?

The Long Game Nobody Saw Coming

Kurian's answer isn't some brilliant last-minute pivot. It's boring, actually—they started planning years ago. When everyone else was waiting to see how AI would shake out, Google was locking in real estate for data centers, diversifying energy sources, and shifting from construction to manufacturing for faster deployment.

"We don't build them in construction. We shifted a lot more to manufacturing because manufacturing can always do faster than you can do with construction," Kurian explains. It's the kind of operational detail that sounds trivial until you realize it's compounding across dozens of facilities.

But the real differentiator? They own the silicon. Twelfth year building TPUs now, with the eighth generation about to drop. And here's where it gets interesting: demand isn't just from AI labs anymore. Citadel's using TPUs for algorithmic trading. Department of Energy for high-performance computing. The chips are becoming general-purpose infrastructure.

The Economics of Owning vs. Reselling

When Berman pushes on the obvious question—if AGI is the prize and compute is the bottleneck, why not hoard everything for your own models?—Kurian's response cuts through the hype: "You have to make money to fund all of this."

Even Google, with all its cash cows, needs cash flow. "No matter which lab you're in, venture capital cannot fund you indefinitely," he notes. And as training costs explode, if you're losing money on inference while burning capital on training, "the number of sources you can go to gets smaller."

The unit economics argument is where Google's vertical integration really shows. They're not reselling someone else's IP at razor-thin margins. They control the chip, so in a capacity-constrained market, their costs stay predictable while competitors scramble. "We make great margins no matter which way we're selling it because we own our own IP," Kurian says.

This creates a fascinating feedback loop. Because they're selling to multiple markets—their own models, other AI labs, financial firms, government—they get better terms from supply chain vendors. More volume, more leverage. Diversification compounds.

The TPU vs. Nvidia Question

Jensen Huang recently claimed Nvidia offers the best total cost of ownership per token. Kurian's response? "We have a lot of customers who say we are the best total cost of ownership." Then the kicker: "We have more demand than we can possibly meet from all the other AI labs."

Not exactly a spec-sheet debate. If other frontier labs are choosing TPUs despite having Nvidia relationships, that's a different kind of signal.

The technical explanation gets into system architecture: TPU v8 connects 9,600 chips on a single optical torus network with "incredibly high bandwidth, super predictable latency." Two petabytes of memory in a single system—"like a hundred times the size of all the Library of Congress digitized." But Kurian keeps circling back to the stack above the chip: Jax, PyTorch optimization, XLA, Pathways. It's not just silicon, it's the whole system.

The Job Displacement Conversation Nobody's Having Honestly

The interview takes an unexpectedly grounded turn when discussing AI and employment. Public sentiment on data centers sits around 20% favorability. People are worried about energy costs, jobs, whether this technology actually benefits them or just enriches tech companies.

Kurian walks through Google's approach: behind-the-meter energy that can feed back to the grid during shortages, investing in alternative energy sources, industry-leading power usage effectiveness (PUE), distributing data centers to avoid burdening individual states. The operational stuff.

But the more interesting thread is how they're framing AI deployment. Signal, Germany's largest health insurer, deployed Gemini agents to help teams work. The anxiety was immediate—this means layoffs, right? Except they haven't cut anyone. Instead, questions that took 23 minutes to research now take seconds. Better service, same headcount.

Or the American Society for Clinical Oncology building an AI system to help doctors navigate treatment guidelines. The patient has breast cancer but they're also diabetic and certain chemo is contraindicated and the overlapping rules are impossibly complex. The AI helps, but it has to be 100% accurate because hallucinations in oncology aren't acceptable trade-offs.

Google Cloud itself is hiring—sales, deployed engineers, new product teams. Even as their existing engineers get more productive with AI, demand is growing faster. Kurian acknowledges other companies are making different choices (looking at you, Block, cutting half your workforce), but frames it as a demand equation: "We're seeing plenty of demand and so we're investing."

What This Actually Means

The interview reveals a strategic divergence that's easy to miss in the AGI race discourse. While venture-backed labs optimize for the headline—biggest model, most parameters, AGI timeline—Google's playing the infrastructure game. Owning the stack from chip to compiler gives them optionality everyone else lacks.

It also surfaces a tension about what AI deployment actually looks like beyond the demos. The wealth advisor Citigroup is building on Gemini—making financial advice accessible to people who can't afford private banking—that's a different narrative than "AI will replace your job." Whether it plays out that way at scale remains genuinely open.

The compute constraint story might be more nuanced than the standard telling suggests. Maybe it's not just about who can buy the most GPUs. Maybe it's about who planned for this moment years ago, who owns their economics, and who can actually sustain the capital requirements without hitting a venture funding wall.

For now, Google's betting that owning the picks and shovels—while also mining yourself—is the better play than going all-in on a single mine.

— Tyler Nakamura