The worst-hit name in a bad market wasn't the chipmaker everyone watches. It was the company that feeds it.

On Tuesday, as the AI-related stocks fell for a second day, the heaviest blow didn't land on Nvidia, the company everyone names. Nvidia fell about 4%. Micron — a maker of memory chips — fell about 13%, ahead of its results due today. In Asia overnight, South Korea's two memory giants, Samsung and SK Hynix, fell around 12% each, dragging the country's whole market down 10%. The market punished memory harder than it punished the processor. That is worth stopping on, because it points straight at the real bottleneck of the entire AI build-out.

Here is the mechanism, simplifying only slightly. An AI chip is astonishingly fast at doing sums. But it is useless if it can't be fed data quickly enough to keep those sums coming — like a chef who can chop faster than anyone alive, but only if the ingredients arrive at the same speed. The ingredient-delivery system for an AI chip is a particular kind of memory called high-bandwidth memory, or HBM: stacks of memory mounted right next to the processor, shifting data into it at enormous rates. Too little HBM, or HBM that isn't fast enough, and the expensive chip sits half-idle. The processor gets the headlines; the memory beside it decides whether the processor can actually run.

And HBM, at the cutting edge, is made by essentially three companies: SK Hynix, Samsung and Micron. That is the whole world's supply of the part that sets how fast the machines can think. Nvidia can design the fastest chip on earth and still not ship a working AI system without memory from that very short list. When you hear that AI is "compute-constrained", a large part of what that means, in physical terms, is that there isn't enough HBM and it can't be made fast enough.

This is the systems lesson under the market noise. The machines that are quietly taking over more of the work — the models drafting the email, sorting the claim, reading the scan — rest on a physical base far narrower than their reach suggests. Enormous capability; three suppliers. That is efficient when it works and brittle when it doesn't, and it is why a wobble in confidence about AI strikes the memory makers hardest: that is where the bet is most concentrated and least substitutable. You can't conjure a fourth HBM maker quickly — the knowledge and the kit take years.

Britain sits almost entirely outside this. It makes none of these chips and none of this memory. What Britain does is host the buildings that consume them — the data centres now competing for grid connections and, on a day like today, for the same electricity and cooling that a record heatwave is straining. The country has chosen to be a place where the machines run, not a place that makes the parts they run on. That is a deliberate position, and it leaves the UK's AI ambitions doubly dependent: on three foreign memory makers for the silicon, and on its own overstretched grid for the power.

None of this is a forecast about share prices, which this desk doesn't make. It is a map of where the real constraint lives. The market spent Tuesday rediscovering that the AI story runs through a very small number of factories making a part almost no one outside the industry can name.

**Why it matters here.** Britain's AI plans depend on memory it doesn't make and power it is struggling to supply. The week's sharpest market move is a reminder that the bottleneck isn't the famous chip — it's the quieter part beside it, and the UK is exposed at both ends of that chain.