Anthropic, Google expand cloud deal worth tens of billions of dollars

Anthropic plans to expand its use of Google Cloud technologies, including up to 1 million Tensor Processing Units (TPUs), to increase its AI compute resources by 2026.

Anthropic, a US-based AI startup, said the deal is valued at tens of billions of dollars and will deliver over a gigawatt of capacity.

The company now has more than 300,000 business customers and has seen a near sevenfold rise in large accounts—customers generating over US$100,000 in annualized revenue—over the past year.

Anthropic uses a combination of Google TPUs, Amazon Trainium, and Nvidia GPUs for its AI training.

The company said it continues to work with Amazon as its main training partner, including on Project Rainier, a large-scale compute cluster across several US data centers.

🔗 Source: Anthropic

🧠 Food for thought

Implications, context, and why it matters.

TPU plans unclear; Ironwood details add context

Anthropic plans to use up to one million TPUs (Tensor Processing Units) and to bring well over a gigawatt of power in 2026, but it has not said which TPU generation it will deploy. Ironwood is Google’s seventh-generation TPU for inference that scales to 9,216 liquid-cooled chips per pod (a large cluster of TPU chips) 1.
Each Ironwood chip delivers 4,614 teraflops with 192GB of High Bandwidth Memory (HBM). That is 6x more than the Trillium generation, a prior TPU family 2. Anthropic has not detailed whether it will use Ironwood or older lines, the terms of any binding deal, or the delivery schedule through 2026.
Google cites 42.5 exaflops per 9,216-chip pod at FP8 (8-bit floating point) 2. Supercomputers like El Capitan rate FP64 (64-bit, double-precision), which makes direct speed comparisons misleading 2.

Multi-chip AI plans open room for cross-platform tools and tuning

Anthropic runs on Google TPUs, Amazon Trainium (AWS-designed AI training chips) and NVIDIA GPUs (graphics processing units used as AI accelerators). This mix fuels demand for tools that move and tune models across accelerators. The OpenXLA compiler framework (an open-source compiler stack for machine learning) lets PyTorch (a popular machine learning framework) run on XLA (Accelerated Linear Algebra) devices like TPUs with small code changes 3, while the Portable JIT Runtime (PJRT) API offers standard interfaces across hardware 4.
Enterprise AI teams and AI-native startups on multi-cloud need compiler bridges and job schedulers. They also need orchestration and cost to performance benchmarking. OpenXLA uses a modular setup with backing from Google and Amazon Web Services (AWS) 4. AMD and Intel also back it, as do NVIDIA and others 4. That backing opens room for optimization layers and profilers. It also paves a path for workload managers across TPU, Trainium, and GPU platforms.