Hyperscalers: Buyers & Rivals
What happened
- Hyperscalers are playing a dual role as both huge buyers of chips and as builders of alternative AI silicon. - Examples include Google’s TPU 8 series and AWS expanding Trainium/Inferentia support while signing major deals. - Startups face a tradeoff between hyperscaler silicon and traditional GPUs, so they’ll seek benchmark comparisons and migration cost analyses. (blog.google) (asiae.co.kr)
Why it matters
The biggest cloud companies are now buying enormous numbers of Nvidia chips while also pushing customers toward their own AI processors. (blog.google.com) Google used Cloud Next on April 22 to launch two eighth-generation Tensor Processing Units: TPU 8t for training models and TPU 8i for running them. Google said TPU 8t scales to 9,600 chips in one superpod and sits inside its AI Hypercomputer stack. (blog.google.com) (cloud.google.com) Amazon Web Services expanded the same play this week with Anthropic. Amazon said on April 20 that Anthropic will spend more than $100 billion on AWS over 10 years and secure up to 5 gigawatts of current and future Trainium capacity, including Trainium3 capacity expected this year. (aboutamazon.com) (anthropic.com) A training chip is the factory machine that builds an AI model; an inference chip is the checkout counter that serves answers after the model is built. Google is now splitting those jobs between TPU 8t and TPU 8i, while AWS sells Trainium for training and Inferentia for inference. (cloud.google.com) (aws.amazon.com 1) (aws.amazon.com 2) The shift comes while the same clouds keep expanding Nvidia fleets instead of replacing them outright. AWS said at Nvidia GTC in March that it plans to deploy more than 1 million Nvidia GPUs across AWS Regions starting in 2026, and Google Cloud used the same event to promote new Nvidia-based virtual machines. (virtualizationreview.com) That leaves startups with a practical choice, not a slogan. Nvidia systems usually offer the broadest software support, while AWS says Trainium3 works with the Neuron software kit and native PyTorch integration, and Google says its new TPUs are tuned for different stages of the AI lifecycle. (aws.amazon.com) (cloud.google.com) AWS is pitching cost and migration friction directly. Its Trainium page says Trn1 instances have up to 50% lower training costs than comparable Amazon Elastic Compute Cloud instances, Trn2 offers 30% to 40% better price performance than GPU-based P5e and P5en instances, and Trainium3 supports training and deployment through the same Neuron stack. (aws.amazon.com) AWS makes a similar case on inference. Its Inferentia page says Inf1 delivers up to 70% lower cost per inference than comparable Amazon Elastic Compute Cloud instances, while Inf2 is built for large language models and distributed inference with the same Neuron software layer. (aws.amazon.com) Google’s pitch is specialization at the system level. The company said TPU 8i is built for real-time inference and reinforcement learning, while TPU 8t is built for large-scale pre-training and uses a larger 3D torus network to keep giant training runs on schedule. (blog.google.com) (cloud.google.com) Nvidia still sits in the middle of the market those clouds are trying to reshape. In its February earnings call, Nvidia said quarterly data center revenue reached $62 billion and that nearly 9 gigawatts of Blackwell infrastructure are already deployed and consumed by hyperscalers, model makers, and enterprises. (fool.com) The next question for buyers is less about who has a chip and more about what it costs to switch. That is why benchmark tests, framework support, and migration work will decide whether hyperscalers stay giant customers of Nvidia, become bigger silicon rivals, or keep doing both at once. (aws.amazon.com) (cloud.google.com)
Key numbers
- Examples include Google’s TPU 8 series and AWS expanding Trainium/Inferentia support while signing major deals.
- (blog.google.com) Google used Cloud Next on April 22 to launch two eighth-generation Tensor Processing Units: TPU 8t for training models and TPU 8i for running them.
- Google said TPU 8t scales to 9,600 chips in one superpod and sits inside its AI Hypercomputer stack.
- Amazon said on April 20 that Anthropic will spend more than $100 billion on AWS over 10 years and secure up to 5 gigawatts of current and future Trainium capacity, including Trainium3 capacity expected this year.
What happens next
- (blog.google.com) Google used Cloud Next on April 22 to launch two eighth-generation Tensor Processing Units: TPU 8t for training models and TPU 8i for running them.
- Amazon said on April 20 that Anthropic will spend more than $100 billion on AWS over 10 years and secure up to 5 gigawatts of current and future Trainium capacity, including Trainium3 capacity expected this year.
- AWS said at Nvidia GTC in March that it plans to deploy more than 1 million Nvidia GPUs across AWS Regions starting in 2026, and Google Cloud used the same event to promote new Nvidia-based virtual machines.
Quick answers
What happened in Hyperscalers: Buyers & Rivals?
Hyperscalers are playing a dual role as both huge buyers of chips and as builders of alternative AI silicon. Examples include Google’s TPU 8 series and AWS expanding Trainium/Inferentia support while signing major deals. Startups face a tradeoff between hyperscaler silicon and traditional GPUs, so they’ll seek benchmark comparisons and migration cost analyses. (blog.google) (asiae.co.kr)
Why does Hyperscalers: Buyers & Rivals matter?
The biggest cloud companies are now buying enormous numbers of Nvidia chips while also pushing customers toward their own AI processors. (blog.google.com) Google used Cloud Next on April 22 to launch two eighth-generation Tensor Processing Units: TPU 8t for training models and TPU 8i for running them. Google said TPU 8t scales to 9,600 chips in one superpod and sits inside its AI Hypercomputer stack. (blog.google.com) (cloud.google.com) Amazon Web Services expanded the same play this week with Anthropic. Amazon said on April 20 that Anthropic will spend more than $100 billion on AWS over 10 years and secure up to 5 gigawatts of current and future Trainium capacity, including Trainium3 capacity expected this year. (aboutamazon.com) (anthropic.com) A training chip is the factory machine that builds an AI model; an inference chip is the checkout counter that serves answers after the model is built. Google is now splitting those jobs between TPU 8t and TPU 8i, while AWS sells Trainium for training and Inferentia for inference. (cloud.google.com) (aws.amazon.com 1) (aws.amazon.com 2) The shift comes while the same clouds keep expanding Nvidia fleets instead of replacing them outright. AWS said at Nvidia GTC in March that it plans to deploy more than 1 million Nvidia GPUs across AWS Regions starting in 2026, and Google Cloud used the same event to promote new Nvidia-based virtual machines. (virtualizationreview.com) That leaves startups with a practical choice, not a slogan. Nvidia systems usually offer the broadest software support, while AWS says Trainium3 works with the Neuron software kit and native PyTorch integration, and Google says its new TPUs are tuned for different stages of the AI lifecycle. (aws.amazon.com) (cloud.google.com) AWS is pitching cost and migration friction directly. Its Trainium page says Trn1 instances have up to 50% lower training costs than comparable Amazon Elastic Compute Cloud instances, Trn2 offers 30% to 40% better price performance than GPU-based P5e and P5en instances, and Trainium3 supports training and deployment through the same Neuron stack. (aws.amazon.com) AWS makes a similar case on inference. Its Inferentia page says Inf1 delivers up to 70% lower cost per inference than comparable Amazon Elastic Compute Cloud instances, while Inf2 is built for large language models and distributed inference with the same Neuron software layer. (aws.amazon.com) Google’s pitch is specialization at the system level. The company said TPU 8i is built for real-time inference and reinforcement learning, while TPU 8t is built for large-scale pre-training and uses a larger 3D torus network to keep giant training runs on schedule. (blog.google.com) (cloud.google.com) Nvidia still sits in the middle of the market those clouds are trying to reshape. In its February earnings call, Nvidia said quarterly data center revenue reached $62 billion and that nearly 9 gigawatts of Blackwell infrastructure are already deployed and consumed by hyperscalers, model makers, and enterprises. (fool.com) The next question for buyers is less about who has a chip and more about what it costs to switch. That is why benchmark tests, framework support, and migration work will decide whether hyperscalers stay giant customers of Nvidia, become bigger silicon rivals, or keep doing both at once. (aws.amazon.com) (cloud.google.com)
