Memory bottlenecks¶

The semiconductor industry fixates on the glamorous components while the unglamorous dependencies quietly become the constraint. Everyone discusses GPUs, process nodes and computational performance; memory receives attention only when it becomes the choke point, which it increasingly is. This is rather like obsessing over the quality of a carriage’s horses while the wheels work loose, which is fine until you are stranded halfway to your destination wondering why nobody mentioned the wheel situation.

High Bandwidth Memory has graduated from obscure specification to supply chain bottleneck with remarkable speed. AI accelerators need enormous memory bandwidth to keep their computational units fed, HBM provides it through 3D stacking, and only three manufacturers produce it at scale. A state-of-the-art GPU without adequate memory is a kitchen full of talented chefs with no ingredients: technically impressive, not actually useful.

Why HBM is special and difficult¶

HBM achieves its bandwidth through literal vertical integration: memory dies thinned to around 50 micrometres, stacked, and connected by thousands of through-silicon vias, the whole assembly bonded with tolerances that make conventional manufacturing look forgiving. A defect in any layer can ruin the stack, testing can only happen after assembly, and the middle of a stack is hard to cool, all of which keeps yields tense and prices at several times conventional memory. The premium is paid anyway, because an expensive processor idling while it waits for data is money evaporating continuously; for AI workloads, bandwidth is the limiting factor, and the larger the models grow, the more it binds.

Three suppliers and a queue¶

Meaningful HBM production belongs to SK Hynix, Samsung and Micron, in that approximate order of comfort, which means the AI hardware supply chain depends not only on one dominant foundry but also on a three-firm memory oligopoly, with the leader supplying the bulk of the dominant GPU maker’s needs. Capacity has not scaled with demand, because HBM was a niche product when the fabs were planned and because a production line costs billions that memory makers commit cautiously, having been burned by cycles before. Packaging adds a further gate: integrating stacks and processor on an interposer is its own scarce capability with its own queue. Each new generation, HBM3 to HBM3e to HBM4, resets yields and prices while the previous generation is still ramping, so the constraint rolls forward rather than resolving.

What it costs¶

Memory can account for a third to a half of an AI accelerator’s bill of materials, which makes it a strategic input negotiated in volume agreements rather than a component bought from a catalogue. The costs propagate: training gets dearer because the clusters do, inference gets dearer because serving models needs the bandwidth, and, since late 2025, consumer electronics get dearer because the AI buildout has begun outbidding phones and laptops for the same memory supply. The bottleneck eases only at the speed of new fabs and matured yields, which is to say in years.

The clerk’s brief¶

January 2026: Undersupply, as scheduled¶

Sourceability’s 2026 outlook, published January 2026, described DRAM undersupply persisting through the first quarter, HBM4 pricing dynamics arriving with the new generation, and price increases on leading foundry nodes flowing through to buyers. None of this was a surprise to anyone reading the file; all of it was priced accordingly. The clerks note that a shortage which arrives on schedule is not a shortage so much as a business model with good calendar discipline.

December 2025: A customer the size of a country¶

CNBC’s reporting in December 2025 described the dominant AI chipmaker’s shift to LPDDR as a seismic event for the memory supply chain, adding a buyer on the scale of a major smartphone maker to a market where an imbalance of one or two per cent normally moves prices sharply and the imbalance was running near three. Memory prices were expected to rise 30 per cent in the quarter, with consumer brands warning customers openly. The wheels, the clerks observe, are now setting the price of the horses.

The stack gauge¶

The gauge has read tight since the file opened, and the reasons have not changed: three suppliers, billion-euro lines committed cautiously, packaging as a second gate, and each new HBM generation resetting yields while demand compounds. What changed in late 2025 is who feels it: with the December 2025 memory squeeze the constraint stopped being a data centre procurement problem and became a consumer price, which tends to be the moment such files acquire political readers. The clerks’ standing assessment: the memory oligopoly is the quietest chokepoint in the section, the one least discussed in speeches about sovereignty, and, on current readings, the one most reliably collecting its rent.