Up until now, Alphabet's (GOOG +0.14%)(GOOGL +0.14%) Google has kept its artificial intelligence hardware efforts to itself. The company has been working on its tensor processing units (TPUs) for nearly a decade, unveiling the first iteration back in 2017. This was well ahead of ChatGPT and the subsequent AI boom that catapulted Nvidia (NASDAQ: NVDA) to the forefront of the AI chip market.
Originally, Google's TPUs were designed to accelerate computations used by the company's various services. TPUs were later made available to Google Cloud customers for running AI workloads. Google's TPUs are application-specific integrated circuits (ASICs), which are designed at the hardware level to perform specific tasks efficiently. This contrasts with Nvidia's GPUs, which are more general-purpose processors.
Image source: Getty Images.
Google is ramping up its AI hardware ambitions
Nvidia has had to contend with tech giants like Alphabet, Amazon, and Microsoft designing their own AI chips and installing them in their own data centers, but that hasn't stopped the company from dominating the market for AI accelerators. There's competition from AMD, but Nvidia and its proprietary CUDA software have been impossible to beat.
Google is now reportedly eyeing Nvidia's massive market share. According to The Information, Google is talking with potential customers about deploying its TPUs. Meta Platforms is a potential TPU customer, with a multi-billion-dollar deal reportedly being discussed.
Google's ambitions appear to be ramping up, with Google Cloud executives reportedly seeing an opportunity to capture 10% of Nvidia's annual revenue. That would translate into many billions of dollars in new revenue and make a significant dent in Nvidia's dominance.
Because Google's TPUs are ASICs, they can achieve a level of efficiency that is likely to be appealing to customers who are building massive AI data centers with enormous energy requirements. Google's latest Ironwood TPUs are twice as power-efficient as their predecessors and 30 times more power-efficient than the first TPUs made available through Google Cloud in 2018.
The downside is that Google's TPUs are a different architecture from Nvidia's industry-standard GPUs. Customers who are already heavily invested in Nvidia's GPUs and the CUDA ecosystem will face an uphill battle to adopt TPUs. A company like Meta has the resources to make it happen if the benefits are significant enough, so Google is likely to be limited to the largest tech giants as potential TPU customers.
NASDAQ: GOOG
Key Data Points
A new threat to Nvidia
The biggest threat facing Nvidia is the potential for the AI boom to turn into an AI bust. Outside of that, competition is likely to gradually erode the company's dominance. AMD has been making some inroads, and now Google is directly targeting Nvidia's largest customers.
It will take quite a while for any of these threats to show up in Nvidia's results. The company's cloud GPUs are currently sold out, with customers making multi-year plans to build massive AI data centers using its chips. For the time being, Nvidia's dominance will remain intact.
In the longer term, once the frantic rush to build out capacity has run its course, efficiency will reign supreme. Google's TPUs could be an attractive alternative to NVIDIA's GPUs when energy efficiency is more important than raw performance, especially considering that power production is a limiting factor for tech giants building out AI data centers.
Google's apparent shift in AI hardware strategy represents a meaningful long-term risk for Nvidia. With nearly a decade of iteration under its belt, Google and its TPUs could eventually steal away a meaningful chunk of market share from Nvidia.
