H200 against L40S · 141 GB and 48 GB
H200 vs L40S: what the price gap buys
At the lowest solid price an H200 is $3.49 per GPU-hour and an L40S is $0.55: 535% more for the H200. The H200 carries 141 GB of memory against 48 GB, so per GB of VRAM it is 116% dearer. Read 12 November 2026.
- H200 lowest
- $3.49
- L40S lowest
- $0.55
- Median gap
- +252%
Civo
Runcrate
$5.40 vs $1.54
H200
15 providers
L40S
14 providers
Price
The two price ranges, side by side
Solid on-demand rows only. Marketplace figures and announced prices are left out of both columns.
| Per GPU-hour | H200 | L40S | H200 against L40S |
|---|---|---|---|
| Providers with a price | 15 | 14 | +1 |
| Lowest | $3.49 | $0.55 | +535% |
| Median of provider lows | $5.40 | $1.54 | +252% |
| Highest | $10.60 | $3.50 | +203% |
| Highest over lowest | 3.0× | 6.4× | |
| Lowest at | Civo | Runcrate | |
| Highest at | Google Cloud | Oracle Cloud |
The lowest H200 price is $2.94 above the lowest L40S price. 0 of the 15 H200 providers list at or under the L40S median of $1.54.
Memory
Memory, bandwidth and price per GB
VRAM is the catalogued size. Bandwidth and the maker's memory figure come from the maker's spec page, linked on each row. Nothing stated as missing is filled in.
| GPU | Variant | Memory, maker table | Type | Bandwidth | Lowest price per TB/s-hour | Source |
|---|---|---|---|---|---|---|
| H200 | SXM | 141 GB | HBM3e | 4.8 TB/s | $0.73 | page |
| H200 | NVL | 141 GB | HBM3e | 4.8 TB/s | $0.73 | page |
| L40S | not stated | 48 GB | GDDR6 | 0.864 TB/s | $0.64 | page |
H200 · 141 GB catalogued
$0.0248
per GB-hour at the lowest price: $3.49 ÷ 141 GB
At the median: $0.0383 ($5.40 ÷ 141 GB)
L40S · 48 GB catalogued
$0.0115
per GB-hour at the lowest price: $0.55 ÷ 48 GB
At the median: $0.0320 ($1.54 ÷ 48 GB)
Price per GB-hour is a derivation: the price per GPU-hour divided by catalogued memory. It says nothing about speed, and memory is not the only reason to pick a card. The H200 carries 2.94 times the memory of the L40S.
Fit check
Three model sizes against each card
Weights only: parameters times bytes per parameter. No KV cache, activations or framework overhead. The full table is on the memory page.
| Scenario | Weights | H200s needed | L40Ss needed | H200 per hour | L40S per hour |
|---|---|---|---|---|---|
| 70B model at FP16 | 140 GB | 1 | 3 | $3.49 | $1.65 |
| 70B model at FP8 | 70 GB | 1 | 2 | $3.49 | $1.10 |
| 405B model at FP8 | 405 GB | 3 | 9 | $10.47 | $4.95 |
GPUs needed = weights ÷ catalogued VRAM, rounded up. Cost = GPUs needed × the lowest solid price per GPU-hour ($3.49 for the H200, $0.55 for the L40S). Real deployments need headroom and servers come in fixed sizes.
Same provider
Providers that list both
One row per provider: its lowest solid H200 price and its lowest solid L40S price. Sizes can differ, so the gap is a guide.
| Provider | H200 | L40S | Gap per GPU-hour | H200 against L40S |
|---|---|---|---|---|
| Civo | $3.49 | $1.29 | +$2.20 | +171% |
| CoreWeave | $6.30 | $2.25 | +$4.05 | +180% |
| DigitalOcean (Paperspace) | $4.47 | $1.57 | +$2.90 | +185% |
| Oracle Cloud | $10.00 | $3.50 | +$6.50 | +186% |
| Crusoe | $4.29 | $1.50 | +$2.79 | +186% |
| Verda | $6.50 | $2.00 | +$4.51 | +226% |
| OVHcloud | $6.20 | $1.80 | +$4.40 | +244% |
| Massed Compute | $3.62 | $0.97 | +$2.65 | +273% |
| Amazon Web Services | $7.91 | $1.86 | +$6.05 | +325% |
| RunPod | $5.29 | $1.09 | +$4.20 | +385% |
| Runcrateours, live | $4.67 | $0.55 | +$4.12 | +749% |
Listed for the H200 only: Hyperstack, Nebius, Together AI, and Google Cloud. Listed for the L40S only: CloudRift, Koyeb, and Vultr.
Questions
H200 and L40S, answered
- Is an H200 more expensive than an L40S?
- By the lowest solid price, the H200 is 535% above the L40S ($3.49 at Civo against $0.55 at Runcrate). By the median of each provider's lowest, $5.40 against $1.54, the H200 is 252% above.
- Which costs less per GB of memory?
- At the lowest listed price the H200 is $0.0248 per GB-hour ($3.49 ÷ 141 GB) and the L40S is $0.0115 ($0.55 ÷ 48 GB). At the medians, $0.0383 against $0.0320. VRAM is the catalogued size of each card.
- Which providers list both?
- 11 of them: Civo, CoreWeave, DigitalOcean (Paperspace), Oracle Cloud, Crusoe, Verda, OVHcloud, Massed Compute, Amazon Web Services, RunPod, and Runcrate. At every one the H200 costs more than the L40S; the gap runs from +171% to +749%.
- Does the H200 need fewer GPUs for the same model?
- For weights only, a 70B model at FP16 is 140 GB: 1 H200 against 3 L40Ss. At the lowest solid prices that is $3.49 against $1.65 per hour. This ignores the KV cache, activations and framework overhead.
- Where do the memory and bandwidth figures come from?
- Memory size is the catalogued VRAM. Bandwidth is from the maker's own spec pages, linked in the table above (www.nvidia.com). A figure the maker does not state stays "not stated" and gets no derived price.
Next
Other entries
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.