OpenAI expects to spend some $100 billion over the next five years, just on renting backup servers from cloud providers for periods of peak demand.

This, The Information reports, is on top of a projected $350bn spend on cloud services to support its generative AI training and inference workloads.

This additional compute is expected to help the company deal with spikes in demand, such as when a surge of Studio Ghibli-inspired images on ChatGPT crashed the company's servers in March.

Future feature launches, and even expected hardware releases, could cause more sudden demand spikes. Not being able to handle such moments, the company fears, could lead to users turning to Google, Meta, or other competitors.

The servers are expected to be used for research workloads during periods of lower demand. The company could potentially give back the excess capacity to cloud providers, but only if they agree to the terms. Alternatively, OpenAI could resell the capacity, again something that depends on the cloud providers' terms.

In total, OpenAI expects to spend around $85bn a year on cloud services, as it ramps up its own data center infrastructure through Stargate.

Stargate includes contracts with Oracle and CoreWeave, but the company also plans to build its own data centers. OpenAI has a long-term relationship with Microsoft and has begun to use Google Cloud.

Oracle is set to become the largest provider of OpenAI services, with the two companies reportedly signing a $300 billion deal starting in 2027.

OpenAI CEO Sam Altman has called compute access the key to maintaining an advantage in the AI race, and has told colleagues that the company will eventually need more power than the entire US grid for its data centers.

The company currently runs at a loss, and hopes to turn a profit in 2030, The Information reports. OpenAI projects its revenue will rise to roughly $200bn in 2030 from $13bn this year, but the business is expected to lose $115bn before it reaches a profit.