Anthropic plans to train Claude using Google Cloud TPUs and services
Plans to use up to 1 million TPUs for training and deploying Claude models
Anthropic announced an expansion of its use of Google Cloud's Tensor Processing Unit (TPU) chips. The AI company will gain access to the capacity and computing resources needed to train and deploy the next generation of Claude models. In total, Anthropic is expected to gain access to well over one gigawatt of capacity starting in 2026.
With the largest expansion of TPU usage to date, Anthropic will gain access to up to one million TPU chips and additional Google Cloud services. This will equip research and development teams with a leading AI-optimised infrastructure for years to come. The decisive factors were the price-performance ratio and efficiency, as well as the company's existing experience with training and deploying its models using Google TPUs.
“Anthropic’s choice to significantly expand its usage of TPUs reflects the strong price-performance and efficiency its teams have seen with TPUs for several years,” said Thomas Kurian, CEO at Google Cloud. “We are continuing to innovate and drive further efficiencies and increased capacity of our TPUs, building on our already mature AI accelerator portfolio, including our seventh generation TPU, Ironwood.”
Anthropic and Google Cloud announced a strategic partnership back in 2023: Anthropic uses Google Cloud's AI infrastructure to train its models and makes them available to businesses via Google Cloud's Vertex AI platform and Google Cloud Marketplace. Today, thousands of businesses use Anthropic's Claude models on Google Cloud, including Figma, Palo Alto Networks, Cursor and others.
“Anthropic and Google have a longstanding partnership and this latest expansion will help us continue to grow the compute we need to define the frontier of AI,” said Anthropic CFO Krishna Rao.
“Our customers - from Fortune 500 companies to AI-native startups - depend on Claude for their most important work, and this expanded capacity ensures we can meet our exponentially growing demand while keeping our models at the cutting edge of the industry.”
Anthropic's Claude ranks among the top three AI models https://lmarena.ai/leaderboard, particularly in areas such as web development, word processing and coding assistants.
More diverse use and faster acceptance of AI
The Anthropic Economic Index examines patterns of early AI adoption. The latest report has been expanded to include new aspects – a geographical analysis of Claude.ai conversations and an initial study of API usage in companies. It found that AI usage is geographically concentrated and strongly correlated with income, with countries such as Singapore and Canada showing higher per capita usage, while emerging economies such as India and Nigeria are underrepresented.
The United States (score 3.62) leads the way in per capita use. Canada (2.91) and the United Kingdom (2.67) also have high usage rates relative to their populations. Other major economies show lower usage, including France at 1.94, Japan at 1.86 and Germany at 1.84. There is concern that this uneven distribution of early AI adoption could exacerbate global economic inequality.
Leading countries show more diverse usage (education, science, business) and tend to focus on augmenting human capabilities rather than full automation. Usage patterns on Claude.ai show that usage in education and science has increased, with users assigning increasingly complex tasks to Claude. Directive conversations rose from 27% to 39%.
The rapid spread of AI compared to previous technologies is remarkable. While it took decades for electricity and PCs to become established, AI achieved a similar acceptance rate in just two years as the internet did in five. Everyday usefulness, easy access through integration into existing digital infrastructures, and ease of use are accelerating user acceptance.
TPUs, an alternative to GPUs?
GPUs were originally developed for processing and displaying computer graphics. A GPU consists of parallel computing units that can simultaneously perform simple, identical operations (e.g. vector calculations) on large amounts of data.
However, when training an AI and especially when making inferences, information often has to be processed in layers. Google Cloud TPUs are accelerators specifically designed for training and inferencing AI models.
Google TPUs offer higher throughput and better energy and cost efficiency for AI workloads than traditional GPU clusters. The disadvantage is that seamless scaling is only possible within the Google Cloud. Trillium TPUs are part of the Google Cloud AI Hypercomputer – an integrated system consisting of performance-optimised hardware, open software and ML frameworks, among other things.
This article was first published in German on Computing’s sister site Computing Deutschland.