The race for artificial intelligence computing capacity has reached an unprecedented financial scale. According to an investigation by The Information, AI developer Anthropic has locked in binding infrastructure and compute agreements worth up to $517 billion over the past eleven months. The contracts secure an immense capacity allocation totaling approximately 14.8 gigawatts of power over multi-year periods. This extraordinary commitment highlights the massive capital required to train future model generations and sustain rapidly scaling commercial inference workloads.
The secured infrastructure relies on a diversified procurement strategy designed to mitigate supply bottlenecks and hardware dependencies. Anthropic is distributing its commitments across several major cloud providers and specialized hyperscalers. Key arrangements include Amazon Web Services utilizing its custom Trainium processors, as well as Google and Broadcom for access to custom Tensor Processing Units. Additionally, the company has contracted compute resources from dedicated cloud operators such as CoreWeave and Fluidstack, alongside Microsoft Azure.
By diversifying across multiple architectures, Anthropic aims to shield its operations from single-vendor constraints. The contractual commitment of 14.8 gigawatts matches the energy consumption of entire metropolitan regions and marks a watershed moment in data center scaling. These capacities are slated not only for training advanced frontier models, but increasingly to support enterprise inference through commercial application programming interfaces. As autonomous agents and multimodal models handle continuous operations, the inference demand of live deployments is expanding far faster than raw pretraining requirements.
This infrastructure push coincides with escalating technological competition among chipmakers seeking to challenge Nvidia. Google is aggressively positioning its latest TPUv7 architecture, code-named Ironwood, for external inference workloads. Recent benchmarks and economic data indicate that Ironwood delivers up to 50 percent more inference performance per dollar compared to Nvidia systems based on the B200 and B300 series. To help engineering teams migrate away from established Nvidia stacks, Google has introduced open-source accelerator agents based on Gemini, which automatically convert PyTorch pipelines to JAX and optimize Pallas kernels for Cloud TPUs.
These unprecedented capital allocations illustrate how leading AI research is evolving from agile software development into energy-intensive heavy industry. While smaller teams face steep barriers to accessing large clusters, frontier labs are locking in decade-long power purchase agreements and silicon production lines. For Anthropic, these commitments provide the foundation needed to keep pace with rivals like OpenAI, ensuring that sufficient compute remains available as automated workflows and enterprise tooling enter widespread commercial use.