Many of these largest customers “are not subject to long-term contracts with us,” Anthropic writes. That means they could lower their spending on Anthropic’s models at any time, either because of a shift to competing models, a reduced reliance on AI altogether, or if they “develop their own solutions that compete with our models.” That last risk is particularly acute, Anthropic notes, for “software-as-a-service and other software companies, whose offerings may incorporate AI functionality similar to, or dependent on, the capabilities we provide.”
What’s more, Anthropic notes that many of its large customers “may depend on venture capital or other external financing to fund their operations.” That kind of spending on AI services could drop quickly if and when the venture capital spigot starts running dry.
Collaborators or competitors?
Anthropic finds itself in an awkward position: Companies like Amazon, Google, and Microsoft are both its biggest cloud-computing providers and some of its biggest competitors in AI research. That situation “creates complex dynamics that could give rise to conflicts of interest and adversely affect our access to compute,” as Anthropic puts it.
These companies’ position as Anthropic’s cloud providers gives them the unique opportunity to raise Anthropic’s prices, terminate contracts with Anthropic, or establish more favorable relationships with competitors, “for competitive reasons or otherwise,” the company writes. Their position also gives these cloud providers a competitive advantage through “visibility into [Anthropic’s] pricing and commercial terms across the platforms they operate.”
Anthropic is working to secure direct data center leases to avoid buying on-demand cloud compute resources from its own competition. But the $125 billion in data center commitments Anthropic has made come with their own problems. As Anthropic writes, “if our compute requirements or technology needs change, we may find ourselves with leased or dedicated infrastructure that is poorly suited to our actual needs, and technological changes in AI hardware or training methodologies could render such infrastructure less efficient than newer alternatives.”