Anthropic Calls for Slowdown in Frontier AI Development, But Why Is $517 Billion Compute Investment Still Expanding?

TradingKey - On September 12, Eastern Time, Anthropic CEO Dario Amodei published "We Must Pace the Frontier," calling for a slowdown in the rate of capability advancement for frontier AI models. Meanwhile, the company continues to expand its long-term compute infrastructure investments.

[Source: X]

According to data compiled by The Information, Anthropic has signed computing capacity agreements totaling approximately $517 billion over the past 11 months, equivalent to about 14.8 gigawatts in power capacity, with related expenditures to be executed over multiple years in the future.

Amodei advocates reserving more time for model safety research, testing, and independent evaluation, while explicitly stating that this does not mean halting model training or technological progress. Anthropic continues to expand its compute investments, primarily for Claude training, risk assessment, product deployment, and inference services.

Anthropic Compute Agreements Reportedly Total $517 Billion

According to statistics from The Information, Anthropic has signed computing power agreements worth approximately $517 billion over the past 11 months, covering cloud services, chips, and data center projects, with related expenditures occurring over the next decade.

Among the announced partnerships, Anthropic has committed to spending over $100 billion with Amazon (AMZN) AWS over the next decade to secure up to 5 gigawatts of additional capacity. The company currently uses more than 1 million Trainium2 chips to train and run Claude. The agreement covers Graviton as well as Trainium2 through Trainium4, and includes options to purchase future products.

Anthropic also obtains computing resources through Google (GOOGL) Cloud and Microsoft (MSFT) Azure. The company has expanded its TPU collaboration with Google and Broadcom (AVGO), with several gigawatts of new capacity set to come online starting in 2027; Anthropic has also committed to purchasing $30 billion in Azure computing capacity, with options for up to 1 gigawatt of additional capacity.

According to The Information, Anthropic has also signed long-term computing infrastructure agreements with companies including Fluidstack and SpaceX (SPCX).

Slowing Frontier AI Does Not Mean Stopping Model Training

In "We Must Pace the Frontier," Amodei proposed that the pace of capability advancements in frontier AI models should be controlled to leave more time for training environment governance, safety research, model testing, and independent evaluation, ensuring that risk management capabilities can keep pace with model progress.

Anthropic committed to bringing in third-party evaluation teams with access levels near those of internal employees to inspect model training processes and safety measures, evaluate completed models and training pipelines, and report relevant risk incidents. Amodei also advocated for frontier AI companies in democratic nations to establish shared safety standards and urged governments to strengthen international coordination.

Amodei made it clear that controlling the pace of development does not mean halting model training or technological progress. He also suggested that the industry could discuss setting limits on training compute, training methods, and AI involvement in developing next-generation models.

Anthropic's existing compute agreements are primarily used for training and running Claude, covering frontier model R&D, enterprise deployment, and inference services. The company's agreement with AWS is explicitly designated for training and running Claude, while its partnerships with Google and Broadcom support frontier Claude models and customer demand.

Revenue Growth Faces Long-Term Compute Cost Test

According to Reuters, as of the end of July 2026, Anthropic's annualized revenue run rate exceeded $65 billion, up from approximately $9 billion at the end of 2025. The Financial Times reported that the company's second-quarter revenue was about $11.5 billion, and it expects to record a positive adjusted operating profit for the second consecutive quarter in the third quarter.

Adjusted operating profit excludes items such as stock-based compensation. The Financial Times also reported that Anthropic disclosed a gross margin exceeding 80%, a metric that does not yet account for revenue sharing paid to distribution partners such as Amazon or model training expenses. The payment schedule and actual usage of long-term compute contracts will affect the company's future capital requirements.

Anthropic completed a $65 billion Series H funding round in May, reaching a post-money valuation of $965 billion. According to Reuters, Anthropic is seeking to raise up to $100 billion in this IPO, which could correspond to a valuation of approximately $2 trillion.

Meanwhile, Nvidia (NVDA) is considering subscribing to up to $10 billion in shares as an anchor investor. The relevant proposal remains under discussion.

Anthropic advocates controlling the pace of capability enhancements in frontier models while continuing to expand the compute resources required for Claude's training, deployment, and inference services. As long-term contracts enter the execution phase, actual payment amounts, data center commissioning progress, and compute utilization rates will impact the company's cash flow.

Whether revenue growth can cover the costs of model training, inference services, and long-term compute infrastructure will be a key factor determining Anthropic's capital requirements, profitability, and IPO valuation.