H100 rental price
median price per GPU-hour$2.55 / GPU-hr
Live chart · updates with every published snapshot
Most memory on an AI accelerator
The largest on-package memory capacity on any reviewed datacenter AI accelerator, held by the current specification leader.
on-package memory capacity256 GBon the AMD Instinct MI325X
Behind the number
The full field
- AMD Instinct MI325X256
- AMD Instinct MI300X192
- NVIDIA B200 SXM180
- NVIDIA H200 SXM141
- NVIDIA H100 SXM80
The board in rank order. The headline number is the top-ranked row, highlighted.
| Rank | Entry | on-package memory capacity |
|---|---|---|
| 1 | AMD Instinct MI325X | 256 GB |
| 2 | AMD Instinct MI300X | 192 GB |
| 3 | NVIDIA B200 SXM | 180 GB |
| 4 | NVIDIA H200 SXM | 141 GB |
| 5 | NVIDIA H100 SXM | 80 GB |
256 GB on every publication so far. A line appears here as soon as the number moves.
About this metric
Model size is bounded by memory before it is bounded by anything else: a model must fit in accelerator memory, in one device or across many, before a single token can be generated. This chart tracks the ceiling, the largest on-package memory capacity on any datacenter AI accelerator in the reviewed cross-vendor specification matrix.
Capacities come from first-party vendor specifications, reviewed and republished per accelerator on the underlying board alongside memory bandwidth. Only shipping datacenter parts with published specifications are included; roadmap claims and unannounced configurations are not, so the ceiling moves when hardware ships, not when it is teased.
Every jump in the memory ceiling changes inference economics: a model that previously needed two accelerators suddenly needs one, and serving costs drop by roughly half for that workload. The vendor holding the ceiling also holds a real pricing lever, which is why memory capacity, not raw compute, has become the headline number in accelerator launches.
Related charts