Memory Analyst
2026 to 2030 HBM sensitivity model

AI HBM Demand Calculator

Additive memory model: resident weights plus KV cache plus scratch, slack, and redundancy.

Assumptions

Demand case
Supply case
2030 HBM demand
0.0 EB
bucket total
Supply minus demand
0.0 EB
2030 balance
Weight bucket
0.0 EB
resident weights
KV bucket
0.0 EB
resident cache
Demand / supply
0.00x
fixed utilization

Demand path by grossed memory bucket

Bars stack to final required HBM after redundancy and fixed utilization. Hover a year to inspect the exact weight, KV, scratch, supply, and gap numbers.

Formula audit

v2.2 buckets + supply split

Yearly model

Year Tokens/mo Replica index Context bucket Weights KV cache Scratch Total demand RoW supply China eq. Supply Balance
Sources: NVIDIA H100 architecture, NVIDIA Vera Rubin platform, NVIDIA Dynamo KV cache, vLLM PagedAttention, LMCache KV reuse research, DeepSeek-R1 model card, Goldman Sachs token-growth frame, TrendForce HBM supply and wafer allocation context, TrendForce HBM4 supplier landscape, TrendForce China domestic HBM reporting.

Disclaimer

Memory Analyst is an independent research site for discussion and education. Nothing on this site is investment, legal, tax, accounting, or procurement advice, and nothing should be read as a recommendation to buy or sell any security, private investment, memory product, GPU, contract, or related asset.

The calculator, charts, and model outputs are scenario analysis built from public information, estimates, simplifications, and user-selected assumptions. They may be wrong, stale, incomplete, internally inconsistent, or inappropriate for any specific decision. Changing one assumption can materially change the apparent supply-demand balance.

HBM supply, AI demand, model architecture, export controls, pricing, yields, packaging capacity, and serving efficiency can change quickly. Do your own work, consult appropriate advisers, and check primary sources before making financial, technical, operational, or strategic decisions.