Big Tech companies Google and Amazon are accelerating sales of AI chips to outside companies, beyond use in their own clouds, drawing attention to how much they can shake Nvidia’s dominance.

Google has even rolled out financing support to buyers to expand the ecosystem for its AI chip, the TPU (Tensor Processing Unit). Amazon, the world’s largest cloud company, is also in talks to sell its Trainium AI chips to external data centre operators.

According to a recent Wall Street Journal (WSJ) report, Google is providing TPUs through financing support to the “Lake Mariner” project, an AI data centre cluster being built in western New York state.

Google provided a $3.2 billion financial guarantee for the project. The project developers will provide Anthropic with thousands of Google TPU-based computing power, WSJ reported, citing sources.

Selling through financing support is a tactic Nvidia often uses to expand demand for its GPUs.

Against that backdrop, Google also appears to be speeding up efforts to expand its position in the AI chip market with a similar tactic. After an executive reshuffle at its cloud business unit, internal pressure grew to move faster on the AI chip business, WSJ reported, citing sources.

Google recently announced it would also set up an AI cloud joint venture using TPUs with Blackstone. This joint venture is the biggest of Google’s efforts to sell and monetise its in-house chips externally. The joint venture aims to bring 500 megawatts of capacity online in 2027. That is similar to the amount of electricity supplied to a small-to-mid-sized city. The joint venture plans to sharply increase capacity afterwards.

Google will supply hardware, software and services to the joint venture, including its in-house TPU chips. In April, Google also unveiled a new processor specialised for AI inference. Earlier, Google signed deals to supply TPUs on a large scale to Anthropic and Meta.

Google has long developed TPU chip technology on the view that AI computing supply would be insufficient. It first used TPUs to train its own models and develop services, then sold them to outside companies through Google Cloud. It then recently declared it would also sell TPUs to companies that operate infrastructure outside Google’s cloud.

In addition, Google is in a position where it has to be relatively less mindful of Nvidia. It is unclear whether this is true, but amid a situation where AI chip supply is not keeping up with demand, there appears to be a trend in the AI industry to avoid offending Jensen Huang, Nvidia’s CEO. Adam Fisher (아담 피셔), a partner at Bessemer Venture Partners, said, “Some emerging cloud companies cannot break away from buying Nvidia full-stack hardware because they worry they could be disadvantaged in allocations of Nvidia GPUs or see quotas reduced, so-called ‘Jensen jail’.”

WSJ reported that companies previously avoided challenging Nvidia out of concern for Huang, but as the shortage of computing resources worsened, places like Blackstone are also working with Google.

Moves by Amazon Web Services (AWS) to expand its own AI chips are also taking clearer shape. According to a recent Bloomberg report, AWS is discussing ways to sell Trainium AI chips to other companies.

Earlier, Amazon CEO Andy Jassy (앤디 재시) said in April in his annual shareholder letter that demand for the company’s AI chips was very high and it was considering sales. “If the chip business, as an independent business entity, sells the chips produced this year to AWS and outside companies, annual revenue would reach about $50 billion,” he said. “Demand is so high that it is highly likely we will sell chips to outside companies going forward.”

The $50 billion in revenue Jassy mentioned is far short of Nvidia’s but is about the same as Intel’s annual revenue.