Amazon Faces Rising AI Costs with Anthropic

Amazon Confronted with Rising AI Costs from Anthropic

Amazon has invested billions of dollars in Anthropic and hosts its models. The reward? A bill that is set to be hefty, calculated on a token basis starting next year, which is already prompting the giant to seek a Plan B. Amazon's infrastructure has long been one of the pillars for Anthropic to deploy its models, but the end of this partnership seems to be approaching.

According to The Information, the contract between Amazon and Anthropic was renegotiated earlier this year. It now includes billing on a token basis, these text fragments that serve as a unit of measurement for AI model consumption, a shift that could significantly increase Amazon's costs, although the company denies any additional expenses.

Kiro, Quick, Alexa: Amazon's Tools Dependent on Claude

The new pricing structure will only take effect next year, but its mere prospect is already causing concern in Seattle. Amazon has indeed built part of its internal tools on Anthropic's models. Its programming agent Kiro, its professional assistant Quick, and even Alexa for Shopping, the shopping version of its voice assistant, all rely on Claude. Each request from these services will soon translate into a line on the bill, and the meter runs quickly when millions of users are engaging with the AI throughout the day.

Just a few months ago, Amazon internally displayed a ranking encouraging its employees to consume as many tokens as possible. Now, it is scrutinizing every token spent.

Facing "AIflation"

In the face of this threat of "AIflation," the giant is exploring other avenues, leading it straight to Anthropic's major rival, OpenAI. Earlier this year, Amazon committed to investing $50 billion in Sam Altman's company, opening its infrastructure in exchange for access to its models. This rapprochement increasingly resembles an exit strategy, as the relationship with Anthropic has noticeably cooled.

After promising up to $25 billion in additional funding to its protégé, Amazon has seen Anthropic rise in stature and sign a massive cloud agreement with Google. The startup has since reached a size that attracts all eyes, including the less favorable ones.

Amazon's Model

Amazon's model is not just about investing in Anthropic but also about funding AI labs whose models it then hosts on AWS, even competing with them using its own tools. This triple play mechanism—funding, hosting, then competing—has been dissected by Clubic. In France, Bouygues is already using Claude via Amazon Bedrock to analyze contracts at its subsidiary Equans.

From the startup's perspective, it is hard not to see retaliatory measures in light of recent news surrounding Mythos. Anthropic's flagship model has been subject to an export ban by the Trump administration for several weeks. Amazon's president, Andy Jassy, is said to be behind these drastic measures as he reportedly alerted senior officials (including Treasury Secretary Scott Bessent, already heated after the initial announcement of Mythos) after researchers managed to circumvent the safeguards of the model, which are intended to prevent its use for cyberattacks. Amazon has neither confirmed nor denied the rumors, a nuance that Anthropic has likely not overlooked in its push for its shareholder to pay up.