The AI chip market is heating up, and Google is playing a bold game. Instead of just making chips, it’s borrowing a tactic from Nvidia’s own playbook. Google is using financial guarantees and “circular financing” to lock in customers who rent its Tensor Processing Units (TPUs).

This strategy helps data centers secure cheaper loans by guaranteeing they will buy or lease Google’s chips. For example, near Niagara Falls, a $3.2 billion guarantee supports a data center cluster where Google’s TPUs power AI workloads for Anthropic, a rising AI company. This financial backing keeps demand steady and helps Google compete against Nvidia’s dominance.

Google recently announced it will sell TPUs directly to customers. It also made a $5 billion deal with Blackstone to create a cloud company aiming to rival Nvidia-backed providers. These moves show Google wants a bigger slice of the AI infrastructure market, not just through cloud services but by owning the hardware layer.

Nvidia still controls over 90% of AI chips worldwide, thanks to its CUDA software and hardware ease of use. Nvidia’s CEO is confident but acknowledges that Google’s TPU push is a real threat. The chip war is no longer just about technology but about controlling the financing and leasing networks that fuel demand.

The Rise of “Neoclouds” and Renters of AI Compute

The AI compute market has evolved into a world where few companies own their own machines. Instead, most rent GPU capacity from specialized providers called “neoclouds.” These firms lease out AI hardware to labs and companies, often backed by the chip makers themselves.

One striking example is xAI, Elon Musk’s AI company, which leases part of its supercomputer to Anthropic and Google. This creates a circular pattern where companies rent compute from each other, and chipmakers invest in the renters. Nvidia, in particular, captures billions from every AI infrastructure buildout while holding stakes in many buyers.

These financial loops help fund the rapid expansion of AI but also concentrate risk. If demand slows or orders get canceled, many companies feel the impact quickly. Rental prices for Nvidia’s top GPUs have dropped by up to 75% from their peak, signaling potential overcapacity or weakening demand.

Anthropic’s Gigawatt Bet Changes AI Infrastructure

Anthropic recently signed letters of intent to lease more than 1 gigawatt of data center power across the U.S. This commitment covers over 200,000 GPUs and marks a big shift from renting cloud capacity to owning dedicated infrastructure.

Google backs Anthropic with up to $40 billion in financial guarantees, including $10 billion immediately and $30 billion tied to performance goals. This support helps Anthropic secure cheaper leases and ensures data center operators trust the payments will come through.

Owning infrastructure gives Anthropic control over power contracts, cooling, and hardware choices. This is a big deal for enterprise clients who want data kept in specific locations for compliance. Dedicated data centers also build customer loyalty because switching costs get higher.

Anthropic’s move forces competitors like OpenAI and Microsoft to rethink their strategies. OpenAI relies heavily on Microsoft Azure but may need to lease dedicated data centers to keep up. Meanwhile, Google benefits twice—first from backing Anthropic’s leases, and second by supplying TPUs for their AI workloads.

Amazon’s Chip Ambitions Add Another Twist

Amazon is also stepping up. It’s working on selling its custom AI chips to other data centers. These chips power Amazon’s own cloud but could soon be licensed to competitors. This could shake up the market by offering a third major chip option alongside Nvidia and Google.

Amazon’s AI chief says demand for its chips is stronger than expected. However, Amazon faces engineering challenges to adapt its hardware for outside users. Still, if successful, this move could diversify the AI chip market and force Nvidia and Google to innovate faster.

For businesses and AI developers, these changes mean more choices but also more complexity. Different chips require software adjustments, and switching systems can be costly. But competition can also drive down prices and improve performance.

The AI chip and infrastructure race is no longer just about technology. It’s about who controls the money, the leases, and the access to compute power. Companies that own or guarantee infrastructure will likely shape enterprise AI for years to come.