I don’t agree that it’s being sold below cost, at least as far as inference is concerned. One piece of data on which we could base our assessment, for example, is the DeepInfer release that DeepSeek published last week, in which it also stated that a cluster generates around $38 per GPU per hour, even at the lowest rate. In 2025, he had estimated the cost of renting a GPU at around two dollars per hour, with energy costs amounting to around 0.10 per hour per GPU. Obviously, this cannot be applied directly to Anthropic and OpenAI, as they publish figures for their own models and are likely larger organisations. However, even with costs that are many, many times higher, the margin on a single token remains substantial. It must be said that the price covers not only inference, but also training, research, free users and investments that need to be recouped. So ‘the individual token is profitable’ and ‘the business might not be sustainable’ could both be true. The separation between memory and processor isn’t going to disappear any time soon, but unified memory is already available in laptops, with an RTX Spark costing between $3,000 and $4,000 for a usable configuration. With competition – for example, from Intel, which has been working on power efficiency for years – prices will fall and on-premises models will become the norm, particularly in businesses.