MI300X vs H100
Specifications, today's prices from 90 providers, price per unit of memory and compute, and who rents each cheapest. Better values are highlighted; lower is better for prices.
Specs, prices, value
Highlighted cells are the better of the two: more memory and compute, lower prices, lower cost per unit.
| MI300X | H100 | Difference | |
|---|---|---|---|
| Memory & compute | |||
| VRAM | 192 GB | 80 GB | +140% |
| Memory bandwidth | 5.30 TB/s | 3.35 TB/s | +58% |
| FP16 tensor | 1,307.4 TFLOPS | 989 TFLOPS | +32% |
| FP32 | 163.4 TFLOPS | 67 TFLOPS | +144% |
| Power | 750 W | 700 W | +7% |
| Prices today | |||
| Cheapest on-demand | $1.65 | $1.35 | +22% |
| Median on-demand | $2.71 | $3.17 | -14% |
| Highest on-demand | $6.00 | $8.98 | -33% |
| Cheapest spot | $0.99 | $0.67 | +48% |
| Cheapest reserved | $1.32 | $1.08 | +22% |
| Current offers (n) | 23 | 197 | -88% |
| Providers with a fresh offer | 14 | 76 | -82% |
| Value | |||
| Per GB of VRAM · hour (cheapest) | $0.009 | $0.017 | -49% |
| Per GB of VRAM · hour (median) | $0.014 | $0.040 | -64% |
| Per PFLOP FP16 · hour (median) | $2.074 | $3.201 | -35% |
| Per TB/s · hour (median) | $0.512 | $0.945 | -46% |
| Trend | |||
| 30-day median change | +1.5% | +3.7% | — |
| 90-day median change | +1.5% | +12.0% | — |
How the two have moved
Both medians indexed to 100 ninety days ago, so the slopes compare even though the price levels differ.
Cheapest providers today · on-demand · current offersWho rents each one cheapest
- 1
GPUBankTrust 931× $1.65 - 2
TensorWaveTrust 648× $1.71 - 3
OpenRelay1× $1.99
- 1
GPUBankTrust 931× $1.35 - 2
GcoreTrust 801× $1.71 - 3
LiumTrust 401× $1.75
Choose one, choose the other
Choose the MI300X if…
- your model needs more than 80 GB per GPU: 192 GB against 80 GB.
- the hourly price matters most: median $2.71 against $3.17, 14% less.
- you serve LLMs: memory bandwidth per dollar is better ($0.512 per TB/s-hour against $0.945).
Choose the H100 if…
- you want choice: 76 providers have a fresh offer today against 14.
Questions about MI300X vs H100
Which is cheaper to rent, the MI300X or the H100?
Today the cheapest current on-demand MI300X is $1.65 per GPU-hour and the cheapest H100 is $1.35; medians are $2.71 and $3.17. Cheapest excludes offers marked unavailable and counts one floor per marketplace.
How much memory do the MI300X and H100 have?
192 GB of HBM3 on the MI300X, 80 GB of HBM3 on the H100, at 5.30 and 3.35 TB/s.
Where do these numbers come from?
From 220 current offers read on the providers' own pages and APIs over the last 45 days (reference 00:32 UTC), or in our October 2026 market study for the few providers our collector does not read. Specifications come from manufacturer datasheets. Every price on the GPU pages links to its receipt and shows its read date.
