Anthropic to use Google's AI chips to train Claude chatbot in multi-billion deal

Anthropic is expanding its deal with Google to use as many as one million of the tech giant's artificial intelligence (AI) chips, worth tens of billions of dollars, as the startup races to advance its AI systems in the competitive market.

Under the deal announced on Thursday, Anthropic will have access to more than one gigawatt of computing capacity, coming online in 2026, to train the next generations of its large language model (LLM) Claude on Google's in-house tensor processing units (TPUs), which were traditionally reserved for internal use.

American AI company Anthropic said it chose the TPUs due to their price-performance ratio and efficiency, as well as its existing experience in training and serving its Claude models with the processors.

The deal is the latest sign of insatiable chip demand in the AI industry, where companies are rushing to develop technology that can match or surpass human intelligence.

Alphabet-owned Google, whose TPUs are available for rent on Google Cloud and serve an alternative to supply-constrained Nvidia chips, will also provide additional cloud computing services to Anthropic.

Rival OpenAI recently signed multiple deals that may cost over $1 trillion to secure about 26 gigawatts of computing capacity, enough to power roughly 20 million U.S. homes. One gigawatt of compute can cost roughly $50 billion, industry executives have said.

ChatGPT-maker OpenAI is actively using Nvidia's graphics processing units and AMD's AI chips to power its growing demand.

Anthropic emphasizes AI safety and building models for enterprise use cases. Its models have helped power a boom in vibe coding startups such as Cursor.