Reading the memory specifications.

A GPU’s memory capacity tells us how much data it can hold. Bandwidth describes how quickly that data can move. Neither number, taken alone, tells us how many useful tasks a system will finish.

NVIDIA’s 2022 Hopper architecture overview describes the H100 SXM5 configuration with 80 GB of HBM3 memory. That is a useful historical example of a capacity specification; it is not evidence of current HBM4 availability or a forecast of supply.

Start with the workload

A memory estimate needs a model, numerical precision, batch size, and context length. Model weights are only part of the footprint: intermediate results, caches, and training state can also occupy memory. Changing these assumptions changes how many accelerators a deployment needs.

Three measurements to keep separate

• Capacity: can the workload fit, and what has to be split across devices?
• Bandwidth: how much time is spent moving data relative to computing?
• System output: what throughput is achieved at the required latency and quality?

This distinction matters for financial models. Doubling installed memory does not by itself imply twice the billable output. A customer may use the additional capacity for longer contexts, larger models, or spare capacity; each has different revenue implications.

The evidence needed for a supply thesis

To argue that memory constrains industry growth, we would need evidence on qualified shipments, yields, packaging, and customer allocations. Product specifications answer a technical question. They do not establish supplier pricing power or how long a shortage will last.

Research commentary. Examples explain a method; they are not investment recommendations. Sources were reviewed on 2 October 2026.