Nvidia warned its biggest customers this weekend: the price of its AI servers is climbing by more than 15%, and the cause is memory, whose cost is soaring.
The same day, SoftBank places the largest retail bond in Japan's history, one trillion yen, to fund its commitments to OpenAI; Samsung announces a record $79 billion return to its shareholders "under the pressure of the AI boom"; Alibaba raises $10.2 billion in Hong Kong for its infrastructure; memory maker YMTC is preparing its stock market listing on the "memory boom".
Four signals, one direction: the physical layer of AI, memory first, is the bottleneck, and it costs more.
The dominant story says the opposite, that AI is getting cheaper to run every month. That holds for the sticker price of a token, pushed down by competition between labs; we have written it ourselves. It turns fragile when the hardware that produces those tokens jumps 15% at once and memory is scarce.
The price paid by the end user is the result of two opposing forces, competition pushing down and a silicon shortage pushing up. Nothing guarantees the first always wins.
The skill gaining value in a French team is managing the real cost. Measure today the cost per request of each use in production. Spot the cases whose economics hold only if the price keeps falling, an assistant handling millions of calls, a bulk document analysis.
Prepare for those a fallback plan before the bill moves: a smaller model at measured quality, batch processing outside peak hours, caching of repeated answers. The AI boom rests on a silicon supply chain with real limits. The advantage will go to those who know their cost in use, not to those who assume it is heading to zero.
The Masteria editorial team