H200 · 141 GB · weights-only arithmetic
What 141 GB holds
An H200 has 141 GB of memory against 80 GB on an H100: 1.76 times as much. For weights alone a 70B model at FP16 is 140 GB, which is 1 H200 or 2 H100s. The tables below do the sum for seven model sizes at three precisions.
- H200 memory
- 141 GB
- H100 memory
- 80 GB
- Ratio
- 1.76×
HBM3e, 4.8 TB/s
SXM 3.35 TB/s
141 ÷ 80
7B at FP1614 GB of weights
13B at FP1626 GB of weights
34B at FP1668 GB of weights
70B at FP16140 GB of weights
123B at FP16246 GB of weights
180B at FP16360 GB of weights
405B at FP16810 GB of weights
The rule
Parameters times bytes
Scenario inputs, not benchmarks. The parameter counts are round sizes we chose; the bytes per parameter are the precision.
Weights in GB = parameters in billions × bytes per parameter. FP16 is 2 bytes, FP8 is 1 byte and INT4 is 0.5 byte. GPUs needed = weights ÷ 141 GB for the H200 (or ÷ 80 GB for the H100), rounded up.
One GPU holds, before any overhead, 70.5B parameters at FP16, 141B at FP8 and 282B at INT4. The H100 holds 40B, 80B and 160B.
Cost per hour = GPUs needed × the lowest solid price per GPU-hour: $3.49 for the H200 at Civo and $1.99 for the H100 at Voltage Park. Those prices come from the ledger, read 12 November 2026.
The difference
Where 141 GB changes the count
Counts are the arithmetic minimum. Servers come in fixed sizes, and real runs need headroom.
Across the 21 model and precision pairs below, the H200 needs fewer GPUs than the H100 in 9, the same number in 12. The pairs where it matters: 70B at FP16 (1 against 2), 123B at FP16 (2 against 4), 180B at FP16 (3 against 5), 405B at FP16 (6 against 11), 123B at FP8 (1 against 2), 180B at FP8 (2 against 3), 405B at FP8 (3 against 6), 180B at INT4 (1 against 2), and 405B at INT4 (2 against 3).
Small models gain nothing: anything that fits on one H100 also fits on one H200. The gain shows where the weights pass 80 GB, and again at each multiple after that.
Fewer GPUs is not always a lower bill. At the lowest solid prices ($3.49 for the H200, $1.99 for the H100) the H200 costs less per hour in 6 of the 21 pairs and more in 15; the rest are level. The cost columns in the tables below show each pair.
The tables
Seven sizes, three precisions
FP16 2 bytes per parameter
| Model | Weights | H200s needed | H100s needed | H200 per hour | H100 per hour |
|---|---|---|---|---|---|
| 7B | 7 × 2 = 14 GB | 1 | 1 | $3.49 | $1.99 |
| 13B | 13 × 2 = 26 GB | 1 | 1 | $3.49 | $1.99 |
| 34B | 34 × 2 = 68 GB | 1 | 1 | $3.49 | $1.99 |
| 70Bfewer H200s | 70 × 2 = 140 GB | 1 | 2 | $3.49 | $3.98 |
| 123Bfewer H200s | 123 × 2 = 246 GB | 2 | 4 | $6.98 | $7.96 |
| 180Bfewer H200s | 180 × 2 = 360 GB | 3 | 5 | $10.47 | $9.95 |
| 405Bfewer H200s | 405 × 2 = 810 GB | 6 | 11 | $20.94 | $21.89 |
FP8 1 byte per parameter
| Model | Weights | H200s needed | H100s needed | H200 per hour | H100 per hour |
|---|---|---|---|---|---|
| 7B | 7 × 1 = 7 GB | 1 | 1 | $3.49 | $1.99 |
| 13B | 13 × 1 = 13 GB | 1 | 1 | $3.49 | $1.99 |
| 34B | 34 × 1 = 34 GB | 1 | 1 | $3.49 | $1.99 |
| 70B | 70 × 1 = 70 GB | 1 | 1 | $3.49 | $1.99 |
| 123Bfewer H200s | 123 × 1 = 123 GB | 1 | 2 | $3.49 | $3.98 |
| 180Bfewer H200s | 180 × 1 = 180 GB | 2 | 3 | $6.98 | $5.97 |
| 405Bfewer H200s | 405 × 1 = 405 GB | 3 | 6 | $10.47 | $11.94 |
INT4 0.5 bytes per parameter
| Model | Weights | H200s needed | H100s needed | H200 per hour | H100 per hour |
|---|---|---|---|---|---|
| 7B | 7 × 0.5 = 3.5 GB | 1 | 1 | $3.49 | $1.99 |
| 13B | 13 × 0.5 = 6.5 GB | 1 | 1 | $3.49 | $1.99 |
| 34B | 34 × 0.5 = 17 GB | 1 | 1 | $3.49 | $1.99 |
| 70B | 70 × 0.5 = 35 GB | 1 | 1 | $3.49 | $1.99 |
| 123B | 123 × 0.5 = 61.5 GB | 1 | 1 | $3.49 | $1.99 |
| 180Bfewer H200s | 180 × 0.5 = 90 GB | 1 | 2 | $3.49 | $3.98 |
| 405Bfewer H200s | 405 × 0.5 = 202.5 GB | 2 | 3 | $6.98 | $5.97 |
Questions
141 GB, answered
- What size of model fits on one H200?
- By weights alone, FP16: up to 70B in our list (140 GB of 141 GB); FP8: up to 123B in our list (123 GB of 141 GB); INT4: up to 180B in our list (90 GB of 141 GB). One GPU holds 70.5B parameters at FP16, 141B at FP8 and 282B at INT4 before any overhead.
- Does a 70B model fit on one H200?
- Its FP16 weights are 70 × 2 = 140 GB against 141 GB, so the weights fit with 1 GB to spare. That is almost no room for the KV cache and activations, so a real deployment usually needs more memory or lower precision. An H100 needs 2 GPUs for the same weights.
- Where does 141 GB change the number of GPUs?
- In 9 of the 21 model and precision pairs in the tables the H200 needs fewer GPUs than the H100; in 12 the count is the same. The largest saving in our list is 5 GPUs.
- What does it cost to hold 405B at FP8?
- 405 × 1 = 405 GB of weights: 3 H200s at $3.49 = $10.47 per hour, against 6 H100s at $1.99 = $11.94 per hour. These use each GPU's lowest solid listed price; a server has to be bought in the sizes a provider offers.
- Is this a performance estimate?
- No. It is weights-only arithmetic on stated parameter counts. It excludes KV cache, activations and framework overhead, and it says nothing about speed, batch size or context length.
Disclosure
H200 Price Ledger is published by the Runcrate team, who also run Runcrate, an AI cloud. Runcrate is listed in the table for H200, marked "ours, live". It is sorted by price like every other row, with no pinned position. Other providers' prices are read from their own public pricing pages. We have no paid placement.