Will AI get more expensive in the future? Probably not!

There currently appears to be a fairly widespread belief that AI use is heavily subsidised by the AI labs, and that using AI models will get much more expensive in the future. I am sceptical of this view.

It is true that frontier AI labs such as OpenAI (of ChatGPT fame) and Anthropic (who make Claude) are widely reported to be losing tens of billions of dollars every year. But a lab losing money overall is not necessarily losing money each time someone uses one of its models. Untangling this difference is crucial, and can teach us a lot about how we are likely to interact with AI in the future.

Indeed, a recent blog post from DeepSeek (a Chinese AI lab) contains indications that inference (running a model to answer prompts once it has been trained) is already quite profitable.

What DeepSeek published

News coverage of AI advances the recent years has mostly focused on training ever better models. This blog post, however, open-sources the inference engine DeepSeek use to run their trained models to answer prompts. It essentially boils down to several quite impressive optimisations that make handling and responding to prompts more cost-effective.

Even more interestingly, the blog post includes a claim that they are able to make approximately "$38 per card per hour (based on off-peak prices)" running the DeepSeek V4.1 Flash model on this inference engine. Since prices outside off-peak hours are higher, we can take this as a lower-bound estimate.

Based on this, we can also make a very rough guess at the profitability of running inference. While DeepSeek have not revealed which graphics cards they are using, the median price of renting even the most expensive graphics cards on vast.ai on 23 October 2026 was less than $10 per hour.

Even with these relatively conservative estimates, we end up at a very high gross margin of approximately 73%. This matches quite well with other independent estimates of the margins of running inference.

If inference is profitable, where are all the losses coming from?

To simplify slightly, an AI lab has two main costs: training a model, and running inference.

If inference is profitable, the losses must come from training. Training a model is very expensive, with typical figures in the hundreds of millions, and maybe billions of dollars reported over the past few years. Like inference, it requires a large number of expensive graphics cards, which require electricity and cooling. Unlike inference, however, training does not generate any money in itself (aside from a theoretical sale or licence of a trained model, which does not seem to be common).

The labs are not training new models for fun, of course. OpenAI and Anthropic are making billion-dollar bets that people will constantly want better models, and will be willing to pay more for them. They are in an arms race to build the next leading (“frontier”) model.

Do we actually need the next model?

This is the big question. There are already signs that companies are not automatically switching to the best and latest models, and are often staying a generation behind to save money.

I tried out DeepSeek V4.1 Flash (this is the model the $38 per card figure relates to) while researching this blog post, and it was very good. I only encountered one funny oddity, where the model answered one of my English prompts entirely in Chinese, for no apparent reason. That quirk aside, it seemed good enough for most tasks.

If the appeal of the frontier models is fading, then I see little reason to fear that using AI will become significantly more expensive. In fact, the cost of running inference has been falling dramatically over the past few years for equivalent performance, with drops of 47% per quarter being estimated.

Where does this leave us?

Overall, it seems very likely the world will spend more on AI in the future than it does now. However, because of the constant improvements in inference efficiency, it seems unlikely to me that the cost per prompt will increase.

Instead, with falling costs, it is likely that people will use AI more, and that more people will start using it. This would be a textbook example of the Jevons paradox: with efficiency improvements, total use of a resource tends to rise rather than fall.

Competition should also help. Many models, including DeepSeek V4.1 Flash, have open weights, so lots of companies are already competing to run them for you. As improvements like DeepSeek's inference engine spread, running these models should keep getting cheaper. This should also keep the cost of the closed-weight models (such as those behind ChatGPT and Claude) in check. They may cost more, at least at first. But people will only pay extra if they are clearly better, which is unlikely to stay the case for long with open-weight models consistently keeping up with the improvements in the closed-weight models.

None of this is guaranteed. Still, the engineering is moving in one direction: doing more with less. As long as that continues, I would expect AI to become cheaper, and for most of us to use it more, rather than less.