NVIDIA H100 GPU: Specs, VRAM, Price & Benchmarks (2026)
The NVIDIA H100 (80GB HBM3) launched in 2022. This guide covers its full specs, VRAM, SXM/PCIe differences where they apply, AI performance, and live per-GPU rental pricing on Aquanode.
Short answer on cost: the H100 rents from $2.19 per GPU per hour on Aquanode. Out of capacity right now.
How much VRAM does the H100 have?
The H100 has 80GB HBM3, with 3.35 TB/s of peak memory bandwidth.
- VRAM: H100 80GB HBM3 vs A100 80GB HBM2e
- VRAM: H100 80GB HBM3 vs H200 141GB HBM3e
What fits in 80GB HBM3 of VRAM
| Model | Precision | Fits? |
|---|---|---|
| Llama 3 8B / similar 7-9B models | FP16 | ~16GB. Fits with plenty of headroom for a large KV cache. |
| Llama 3.1 70B | FP16 | ~140GB weights alone. Does not fit on one 80GB card; needs 2 GPUs. |
| Llama 3.1 70B | INT4 (AWQ/GPTQ) | ~35-40GB. Fits comfortably on a single card, room for context. |
| Mixtral 8x7B | FP16 | ~94GB weights. Just over the 80GB limit; INT4 (~24GB) fits easily. |
Approximate, based on published parameter counts and standard bytes-per-parameter rules of thumb (FP16 ≈ 2 bytes/param, INT4 ≈ 0.5-0.6 bytes/param). Real footprint also depends on KV-cache size and framework overhead.
What can the H100 run?
Popular open models from small to frontier scale, with the memory each needs and how many H100 cards (80GB HBM3 each) that takes.
| Model | As published | FP8 | INT4 |
|---|---|---|---|
| Qwen/Qwen3-8B 8.2B | BF16: ~18.3 GB, 1 GPU | FP8: ~9.2 GB, 1 GPU | INT4: ~4.6 GB, 1 GPU |
| Qwen/Qwen2.5-14B-Instruct 14.8B | BF16: ~33 GB, 1 GPU | FP8: ~16.5 GB, 1 GPU | INT4: ~8.3 GB, 1 GPU |
| Qwen/Qwen3-32B 32.8B | BF16: ~73.2 GB, 1 GPU | FP8: ~36.6 GB, 1 GPU | INT4: ~18.3 GB, 1 GPU |
| Qwen/Qwen-72B 72.3B | BF16: ~162 GB, 3 GPUs | FP8: ~80.8 GB, 2 GPUs | INT4: ~40.4 GB, 1 GPU |
| MiniMaxAI/MiniMax-M2.7 228.7B | FP8: ~256 GB, 4 GPUs | – | INT4: ~128 GB, 2 GPUs |
| deepseek-ai/DeepSeek-R1 684.5B | FP8: ~765 GB, 10 GPUs | – | INT4: ~383 GB, 5 GPUs |
Estimates: weights at the stated precision plus a flat 20% for KV cache and overhead, at a moderate context length. A dash means the precision is not offered for that model (it is already published at that size). INT4 needs a published quantized checkpoint. Open any model for a per-GPU breakdown, or use the H100 VRAM calculator.
H100 VRAM calculator: check which models fit in its memory at each precision.
All models that fit in 80 GB: the open models whose weights and overhead fit, at native, FP8 and INT4 precision.
H100 specs
H100 SXM5
| VRAM | 80GB HBM3 |
| Memory bandwidth | 3.35 TB/s |
| TDP | Up to 700W (configurable) |
| Form factor | SXM (DGX H100 / HGX H100 platform only) |
| Interconnect | NVLink 4, 900 GB/s bidirectional |
H100 PCIe
| VRAM | 80GB HBM2e |
| Memory bandwidth | ~2 TB/s |
| TDP | 350W |
| Form factor | Standard PCIe, dual-slot |
Specs sourced from the vendor's public datasheet/product page. See the source.
Related reading: H100 vs H200, A100 vs H100, H100 and H200: SXM vs NVL vs PCIe, and The best GPUs for AI, ranked.
GPU Glossary: What is VRAM?, HBM, Tensor Cores, CUDA Cores, TFLOPS, NVLink vs PCIe
H100 SXM5 vs PCIe: which should you use?
- Memory bandwidth: SXM5 delivers 3.35 TB/s vs ~2 TB/s on PCIe.
- Interconnect: SXM5 has NVLink 4, 900 GB/s bidirectional. PCIe has no dedicated GPU-to-GPU interconnect.
- Power and form factor: SXM5 is Up to 700W (configurable) in a SXM (DGX H100 / HGX H100 platform only) form factor. PCIe is 350W in a Standard PCIe, dual-slot form factor.
For large-scale distributed training, the variant with NVLink and the highest memory bandwidth is usually the right choice, since those advantages compound across a multi-GPU cluster. For single-GPU inference or fine-tuning, the lower-power variant often gives the same usable VRAM at a lower hourly rate.
H100 AI performance
The default choice for training and fine-tuning models up to the ~30-40B range on a single card, and for high-throughput FP8 inference. Its 1,979 TFLOPS dense FP8 figure and 900 GB/s NVLink make 8-GPU pretraining and serving runs practical. It's also the card with the deepest cloud/on-prem availability, so it's usually the cheapest per-FLOP hour to actually get.
- Dense FP16/BF16 tensor throughput: 989 TFLOPS
- Dense FP8 tensor throughput: 1,979 TFLOPS
- Memory bandwidth: 3.35 TB/s
80GB isn't enough to hold a 70B-class model in FP16 on one card, and it has no native FP4 support, so it can't run the newest FP4-native model formats at full precision the way Blackwell parts can.
See how it stacks up against other cards in the GPU benchmarks and specs table.
The cheapest H100 offer right now ($2.19/GPU/hr) is about 50% below the market median of $4.39/GPU/hr. Out of capacity right now.
H100 price: what does it cost?
Buying. $25,000-$30,000 (PCIe card, standalone) (commonly quoted street price (no fixed retail; OEM channel)). A complete 8x SXM DGX H100 system runs roughly $250,000-$400,000. Secondary-market H100 pricing peaked around $80,000-$120,000 in 2023 and has moderated since, but standalone PCIe cards remain well above the figure above through most resellers.
Renting. The cheapest current on-demand rate for the H100 on Aquanode is $2.78/GPU/hr. Live rates range from $2.19 to $4.39 per GPU per hour, with a median of $4.39/GPU/hr. Billed by the offer's own terms; the table below shows every live rate. Out of capacity right now.
| Region | $/GPU/hr | Available | VRAM | vCPU | RAM |
|---|---|---|---|---|---|
| – | $2.19 | 0 | 80 GB | – | – |
| South Korea | $2.77 | 1 | 94 GB | 64 | 336 GB |
| Canada | $2.78 | 1 | 80 GB | 28 | 180 GB |
| Bkk, Th | $2.98 | 2 | 80 GB | 49 | 862.5 GB |
| Des Moines, Us | $3.00 | 1 | 80 GB | 20 | 128 GB |
| Noida, In | $3.76 | 8 | 80 GB | 24 | 200 GB |
| Finland | $4.95 | 32 | 80 GB | 16 | 200 GB |
| Virginia, Us | $7.57 | 0 | 80 GB | 16 | 256 GB |
H100 price history
Aquanode stores one snapshot of its GPU price index per UTC day. For the H100 that is 5 days so far, 2026-10-06 to 2026-10-10, so this is a short history, not a long-run trend. The lowest per-GPU rate was $2.19 on both 2026-10-06 and 2026-10-10.
| Day (UTC) | Lowest per GPU hour | Median per GPU hour | Data-center lowest per GPU hour | Offers |
|---|---|---|---|---|
| 2026-10-10 | $2.19 | $3.51 | $2.78 | 74 |
| 2026-10-09 | $2.19 | $3.65 | $3.00 | 74 |
| 2026-10-08 | $2.19 | $3.51 | $2.78 | 71 |
| 2026-10-07 | $2.19 | $3.64 | $2.78 | 72 |
| 2026-10-06 | $2.19 | $3.51 | $2.78 | 73 |
Each row is the stored daily snapshot of the live index, copied as recorded. A dash means that day stored no figure. The same series is available as JSON and summarised in the monthly GPU price report.
How this price is calculated
All prices on this page are normalized to a per-GPU hourly rate using each offer's authoritative GPU count, so that raw price is divided by the number of GPUs it actually covers; some offers report price as already per-GPU, so those are used as-listed. An offer with a missing, zero, or invalid GPU count is excluded entirely rather than published at a guessed rate.
No offers were excluded from this snapshot for a missing or invalid price. No offers were dropped as price outliers in this snapshot.
Only the cheapest qualifying offer per provider is shown in the table above. This page regenerates at most once per hour.
H100 vs A100: how do they compare?
- VRAM: H100 80GB HBM3 vs A100 80GB HBM2e
- Memory bandwidth: 3.35 TB/s vs 2,039 GB/s
- Dense FP16 tensor throughput: 989 TFLOPS vs 312 TFLOPS
On Aquanode right now, A100 starts at $0.936/GPU/hr against H100's $2.19/GPU/hr, about 57% less.
H100 vs H200: how do they compare?
- VRAM: H100 80GB HBM3 vs H200 141GB HBM3e
- Memory bandwidth: 3.35 TB/s vs 4.8 TB/s
- Dense FP8 tensor throughput: 1,979 TFLOPS vs 1,979 TFLOPS
On Aquanode right now, H100 starts at $2.19/GPU/hr against H200's $3.95/GPU/hr, about 45% less.
Compare the H100 with other GPUs
Compare H100 with
- B200 vs H100
- H100 vs H200
- A100 vs H100
- DGX A100 vs H100
- H100 vs V100
- AMD MI300X vs H100
- H100 vs L40S
- H100 vs L40
- A40 vs H100
- H100 vs RTX A6000
- H100 vs RTX A5000
- H100 vs RTX A4000
- H100 vs L4
- H100 vs T4
- H100 vs RTX 6000 Ada
- H100 vs RTX PRO 6000
- H100 vs RTX PRO 6000 WS
- H100 vs RTX PRO 6000 SE
- H100 vs RTX PRO 5000
- H100 vs RTX 5090
- H100 vs RTX 5080
- H100 vs RTX 5070 Ti
- H100 vs RTX 5070
- H100 vs RTX 5060 Ti
- H100 vs RTX 4090
- H100 vs RTX 4080 Super
- H100 vs RTX 4080
- H100 vs RTX 4070 Ti
- H100 vs RTX 4070 Super
- H100 vs RTX 4070
- H100 vs RTX 4060 Ti
- H100 vs RTX 3090
- H100 vs RTX 3080
- H100 vs RTX 3070
- H100 vs RTX 3060
- B300 vs H100
- AMD MI355X vs H100
- AMD MI325X vs H100
- H100 vs H100 NVL
- H100 vs RTX PRO 4500
- H100 vs RTX PRO 4500 SE
- H100 vs RTX PRO 4000
- H100 vs RTX 5880 Ada
- H100 vs RTX 5000 Ada
- H100 vs RTX 4000 Ada
- H100 vs RTX 4000 SFF Ada
- H100 vs RTX 2000 Ada
- H100 vs RTX 4070 Ti Super
- H100 vs RTX A4500
- H100 vs RTX 2080 Ti
- H100 vs RTX 3080 Ti
- H100 vs RTX 3070 Ti
- H100 vs RTX 3060 Ti
- GTX 1080 Ti vs H100
- H100 vs RTX 4060
- H100 vs RTX 5060
- GTX 1660 Super vs H100
- H100 vs RTX 2060
- H100 vs Radeon RX 9070 XT
- H100 vs Radeon RX 6700 XT
- H100 vs RTX 6000
- A16 vs H100
- H100 vs P40
- H100 vs P4
- H100 vs Quadro P2000
- H100 vs Quadro M4000
Get notified when the price drops
GPU supply moves hourly. Tell us what you're waiting for and we'll email you when a matching offer appears across any provider we track.
Rent a H100 on Aquanode
- On-demand instances from $2.19/GPU/hr, billed by the provider's own terms, with no hardware procurement or long-term commitment.
- 100 live offers across 8 regions today.
- Set a price/availability alert above to hear the moment a cheaper or newly-available H100 offer appears.
- Compare every H100 offer side by side, or browse the full multi-provider GPU marketplace.
Good for
The default choice for training and fine-tuning models up to the ~30-40B range on a single card, and for high-throughput FP8 inference. Its 1,979 TFLOPS dense FP8 figure and 900 GB/s NVLink make 8-GPU pretraining and serving runs practical. It's also the card with the deepest cloud/on-prem availability, so it's usually the cheapest per-FLOP hour to actually get.
Not good for
80GB isn't enough to hold a 70B-class model in FP16 on one card, and it has no native FP4 support, so it can't run the newest FP4-native model formats at full precision the way Blackwell parts can.
H100 FAQs
How much VRAM does the H100 have?
The H100 has 80GB HBM3, with 3.35 TB/s of peak memory bandwidth.
What is the NVIDIA H100?
The NVIDIA H100 is a GPU released in 2022, with 80GB HBM3 of memory and a Up to 700W (configurable) power envelope. See the full spec table above for interconnect, form factor and tensor-throughput details.
How much does it cost to rent a H100?
Live H100 rental prices currently range from $2.19 to $4.39 per GPU per hour, with a median of $4.39 per GPU per hour. Out of capacity right now.
What's the cheapest H100 rate?
The lowest current H100 rate on Aquanode is $2.19 per GPU per hour in . Out of capacity right now.
What is the difference between H100 SXM5 and PCIe?
SXM5 has 3.35 TB/s of memory bandwidth and NVLink 4, 900 GB/s bidirectional. PCIe has ~2 TB/s, no dedicated GPU-to-GPU interconnect. Both carry 80GB HBM3 / 80GB HBM2e respectively.
How does the H100 compare to the A100?
VRAM: H100 80GB HBM3 vs A100 80GB HBM2e Memory bandwidth: 3.35 TB/s vs 2,039 GB/s Dense FP16 tensor throughput: 989 TFLOPS vs 312 TFLOPS On Aquanode right now, A100 starts at $0.936/GPU/hr against H100's $2.19/GPU/hr, about 57% less. See the full H100 vs A100 comparison for a shared-provider price breakdown.
How does the H100 compare to the H200?
VRAM: H100 80GB HBM3 vs H200 141GB HBM3e Memory bandwidth: 3.35 TB/s vs 4.8 TB/s Dense FP8 tensor throughput: 1,979 TFLOPS vs 1,979 TFLOPS On Aquanode right now, H100 starts at $2.19/GPU/hr against H200's $3.95/GPU/hr, about 45% less. See the full H100 vs H200 comparison for a shared-provider price breakdown.
What is the H100 good for?
The default choice for training and fine-tuning models up to the ~30-40B range on a single card, and for high-throughput FP8 inference. Its 1,979 TFLOPS dense FP8 figure and 900 GB/s NVLink make 8-GPU pretraining and serving runs practical. It's also the card with the deepest cloud/on-prem availability, so it's usually the cheapest per-FLOP hour to actually get.
What are the H100's limitations?
80GB isn't enough to hold a 70B-class model in FP16 on one card, and it has no native FP4 support, so it can't run the newest FP4-native model formats at full precision the way Blackwell parts can.
Is renting cheaper than buying?
Renting avoids the upfront hardware cost and lets you match spend to actual usage. A rented H100 at $2.19/hr only costs money while it's running, whereas buying ties up capital in hardware that keeps depreciating whether it's in use or not. Buying outright runs $25,000-$30,000 (PCIe card, standalone) (commonly quoted street price (no fixed retail; OEM channel)). Which is cheaper depends on how continuously you'd run it; short or bursty workloads usually favor renting. Out of capacity right now.
How is the H100 price calculated?
All prices are normalized to a per-GPU hourly rate using each offer's authoritative GPU count, which the raw price is divided by; some offers report price as already per-GPU. Offers whose price can't be safely normalized, or whose rate is an extreme outlier against the rest of the market, are excluded.
H100 price by region
Related guides
Other models in the same generation, then the rest of the GPU index.
