Analysts flag Alphabet’s homegrown AI chips as a growing threat to Nvidia’s data-center dominance
- Alphabet’s Google Cloud used its April 22 chip event to sharpen the case that its in-house tensor processing units can handle more artificial-intelligence work now dominated by Nvidia graphics processors. - Google split its newest eighth-generation tensor processing units into separate chips for training and inference, while highlighting Ironwood’s inference focus and broader customer use of custom silicon in cloud services. - Amazon is making the same bet with Trainium, adding to analyst concerns that hyperscalers may keep more AI spending in-house instead of buying only Nvidia systems. (cnbc.com)
Alphabet used an April 22 Google Cloud event to press a simple argument: more artificial-intelligence workloads can run on Google’s own chips instead of Nvidia’s. (cnbc.com) Google said its eighth-generation tensor processing unit lineup will split training and inference into separate processors, a break from earlier designs that handled both jobs on one chip family. Both products are due later in 2026. (cnbc.com) That change follows Google Cloud Next in April 2025, when the company introduced Ironwood, its seventh-generation tensor processing unit and its first chip built specifically for inference, the step where trained models generate answers for users. (blog.google) Google said Ironwood delivers 10 times the peak performance of TPU v5p and more than four times better performance per chip than TPU v6e, which Google also calls Trillium. (cloud.google.com) Bloomberg reported on April 20 that Google was trying to build on deals with Meta and Anthropic as it pushed newer tensor processing units beyond internal use and deeper into cloud sales. (bloomberg.com) Amazon Web Services has been making the same case with Trainium. In December 2024, Amazon said its Trn2 instances with Trainium2 offered 30% to 40% better price-performance than current-generation graphics-processor-based EC2 instances. (aboutamazon.com) Amazon expanded that effort in late 2025, saying Project Rainier came online with nearly half a million Trainium2 chips, one of the world’s largest artificial-intelligence compute clusters. (aboutamazon.com) The pressure point for Nvidia is not that Google or Amazon can replace every graphics processor overnight. It is that the biggest cloud companies can shift some training and a growing share of inference onto chips they design themselves. (cnbc.com 1) (cnbc.com 2) Nvidia still dominates the market for general-purpose artificial-intelligence computing, and Google remains a major Nvidia customer even as it rents tensor processing units to cloud clients. (cnbc.com) The latest analyst framing is narrower than “Nvidia versus one rival.” It is a fight over which parts of the artificial-intelligence stack stay on merchant chips and which parts get absorbed by hyperscalers building their own. (fool.com 1) (fool.com 2)
