The Lake Mariner Bet: How Google Is Bankrolling Its Own Chip Ecosystem
There's a playbook in the AI chip world, and for years, Nvidia owned every page of it. Now Google is borrowing the whole thing.
The story starts in western New York, at a data center campus called Lake Mariner, operated by a company called TeraWulf. Google has committed $3.2 billion in financial backstops there — not to buy chips outright, not to train its own models, but to underwrite the infrastructure that houses its Tensor Processing Units. Computing power flows from that facility to AI cloud platform Fluidstack, and ultimately to Anthropic for training its Claude models.
That's the structure. But here's what it actually means: Google isn't just designing chips anymore. It's financing the entire world around them.
Why Google Is Copying the Nvidia Playbook
Nvidia didn't win the AI hardware market purely by building fast chips. It won by making those chips affordable to use — backing leases, helping customers finance massive compute clusters, lowering the barrier to say yes. That financial muscle turned CUDA dependency into something that felt less like a vendor relationship and more like essential infrastructure.
Google watched that strategy work and decided to replicate it. The Lake Mariner backstop is part of a deliberate pattern: use financial guarantees to make TPU adoption easier for customers, then watch demand for those chips follow the money. Alphabet is, in other words, borrowing from Nvidia's own strategic playbook — and deploying it at serious scale.
It's a smart move. And honestly, it might be the only kind of move that actually works at this level.
The Numbers Behind Google's TPU Push
The numbers here are worth sitting with.
Google's TPU business has already generated tens of billions in revenue, according to disclosures from Broadcom CEO Hock Tan. The company expects to ship 4.3 million TPU units in 2026, scaling to 35 million by 2028. And according to TrendForce, custom AI chip sales are projected to grow 45% in 2026 — nearly triple the 16% growth rate expected for standard GPUs.
That gap tells you something important about where the market is heading. When custom silicon is growing at three times the pace of commodity GPUs, the direction of travel is clear. The question is whether Google can capture enough of that momentum to put real pressure on Nvidia's dominance.
Beyond Lake Mariner: Intel, Marvell, and the Software Problem
Lake Mariner is just one front in a wider campaign.
Google has placed an order with Intel to manufacture more than three million TPUs in 2028. It's also in active discussions with Marvell Technology to co-design new custom chips. And it has introduced TorchTPU — a software layer built specifically to help developers migrate away from Nvidia's deeply entrenched CUDA ecosystem.
That last piece is arguably the hardest problem on the list. Nvidia's moat isn't just performance numbers or supply chain scale. It's the years of developer tooling, research libraries, and workflow integrations baked into CUDA that have made it feel indispensable. Getting developers to move is a fundamentally different challenge than getting data centers to install different hardware. TorchTPU is Google's attempt to make that migration feel less like starting over from scratch.
The $35 Billion Deal and the Firepower Behind It
Zoom out and the full scale of the ambition becomes clearer.
A $35 billion deal linking Google, Broadcom, and Anthropic across five U.S. data centers is being financed by Apollo Global Management and Blackstone. This isn't a chip company doing chip company things. This is a tech giant reshaping how AI infrastructure gets built and financed — at a level that requires the biggest private equity firms in the world to make it happen.
And here's a telling detail: both Google and Amazon have reportedly given Nvidia CEO Jensen Huang advance notice of their custom chip plans. That's not a hostile move — it's a careful one. These companies still depend on Nvidia as a primary GPU supplier. They're navigating a delicate balance: building alternatives while trying not to burn the relationship they still need.
Can Google Actually Challenge Nvidia's Dominance?
Here's where it's worth being honest.
Nvidia's lead isn't just financial or logistical — it's cultural and technical. The AI developer community has built careers, workflows, and entire research practices on top of CUDA. That kind of embedded knowledge doesn't evaporate because Google offers a more attractive financing structure on TPU infrastructure.
What Google is building is real, and the numbers are genuinely significant. TPU shipments scaling from 4.3 million to 35 million units in two years is not a minor play. But overcoming Nvidia's software moat and performance edge at the same time — those are different problems than arranging multi-billion-dollar data center deals, no matter how sophisticated the structure.
The AI chip race is far from over. What Google has made clear is that it's willing to spend whatever it takes to stay in it.