AI demand cannot be estimated by counting GPUs alone. HBM per accelerator, CPU memory, rack volume and the ASIC mix all matter.
Memory per rack
A GB300 NVL72 rack contains 20.736 TB of HBM and 17.280 TB of CPU memory, or 38.016 TB in total. The Vera Rubin model uses 20.736 TB of HBM and 54 TB of CPU memory, or 74.736 TB.
The 36.720 TB increase is mainly on the CPU side. Higher final HBM configurations would make the current model conservative.
NVIDIA annual demand
In 2026, 65,000 GB300 racks plus 10,000 Rubin racks consume 3.2184 EB. In 2027, 85,000 Rubin racks consume 6.3526 EB, an increase of 3.1342 EB.
The conversion is capacity per rack in TB multiplied by rack volume, divided by 1,000,000. Annual totals and year-on-year increments remain separate.
ASICs and total increment
The base model for Google, AWS, Meta, Microsoft, OpenAI and other ASICs adds 2.705 EB. Combined with NVIDIA, the strictly additive increment is 5.839 EB.
This is incremental AI accelerator demand, not total global DRAM demand in 2027.
For industry research only. This is not investment advice. Forecasts depend on assumptions and may differ materially from actual results.
Methodology