AI model training costs rise, and customers help pay over time

In short: Building the most advanced AI models is getting very expensive, and while AI companies pay upfront, customers often cover part of the cost later through pricing.

What's going on

Training a “frontier” AI model, meaning one of the most capable models available, now commonly costs tens of millions of dollars and can exceed $100 million for a single training run. Training is the long, resource-heavy step where the model learns from large amounts of data, like a student going through an intensive multi-year course all at once.

Some estimates in the sources suggest the biggest training runs could pass $1 billion by 2027 if current trends continue. These large runs are not funded like a typical software project. Instead, a small group of well-funded companies and their cloud partners are paying most of the upfront bill.

The sources point to OpenAI, Anthropic, Google DeepMind, Meta AI, and xAI as key organizations covering these training costs, often with support from cloud providers like Microsoft, Amazon, and Google. Customers usually do not see a separate “training fee.” They pay indirectly through subscriptions, usage-based pricing, or API fees (an API is a paid doorway that lets other apps send requests to the AI).

Many businesses avoid paying for full training themselves. Instead, they fine-tune existing models, which is like giving a trained employee a short course for a specific job, or they use hosted AI through paid APIs.

What to watch

As training costs rise, fewer companies may be able to afford building top-tier models, which can concentrate power in the hands of the biggest players. For everyone else, the practical impact may show up in AI pricing, limits on usage, and which companies can offer the most capable tools at sustainable costs.

Source: NYTimes