Google is copying Nvidia’s chequebook tactics to muscle into AI chips, while customers worry about upsetting Jensen Huang.
According to the Wall Street Journal, the search giant has been using financial guarantees to win data-centre customers for its tensor processing units, or TPUs.
On Lake Ontario’s southern shore, near Niagara Falls, Google has backed an AI data-centre cluster called Lake Mariner with a $3.2 billion guarantee. The site’s developers will rent computing power from thousands of Google microprocessors to Anthropic, according to people familiar with the arrangement.
It is the same wheeze Nvidia has used to make demand for its AI chips even hotter, and Google now fancies a bigger chunk of the market. Until recently, Nvidia had the market mostly to itself, with GPUs prized for training and running AI models. Now the AI race is increasingly about grabbing compute wherever it can be found.
TeraWulf co-founder and chief technology officer Nazar Khan said: “You have all these very well-capitalised companies who are big believers that this market around compute is going to have tremendous value. They want to be in the game; they don’t want to be left behind.”
Google has copied Nvidia’s habit of helping data centres raise cheaper debt and using circular financing, where investment money comes back as chip sales.
Alphabet chief executive Sundar Pichai’s outfit has become more urgent after a shake-up inside Google Cloud, while Nvidia remains tightly linked to OpenAI.
Nvidia chief executive Jensen Huang has talked down Google’s chances, saying in April that Nvidia had much wider reach than any TPU or custom ASIC.
“Our market reach is far greater than any TPU or ASIC can possibly have. I would love to hear them demonstrate the cost advantage of TPUs. It makes no sense in my mind,” Huang said.
Google has since struck a $5 billion deal with Blackstone to create a cloud-services company to compete with Nvidia-backed outfits CoreWeave and Nebius.
Google DeepMind chief scientist Jeff Dean said the company started thinking about specialised hardware in 2013, when speech recognition at mass scale looked like it could double Google’s computer needs.
Google first kept TPUs for itself, using them for AI models, search and other products, before offering them through Google Cloud as demand exploded. In May, Google said it would sell chips directly to customers and unveiled its first TPU built for inference, the query-serving side of AI computing.
Google Cloud vice president of AI and computing infrastructure Mark Lohmeyer said: “We’re seeing a set of customers that might not have considered it in the past.”
Citadel Securities chief technology officer Josh Woods said his firm can run key workloads at a 30 per cent lower cost and up to four times faster with TPUs.
Nvidia still has a monster moat, including CUDA, networking gear and a market share estimated at more than 90 per cent, which gives buyers plenty to fret about.
