LLM.co · Private AI economicsNVIDIA H100 LLM Hosting Cost Calculator
The most available inference GPU, priced across specialist clouds and hyperscalers. Size a model to it and see when it undercuts an API.
Open weights: compared against API pricing and self-hosting on H100 80GB.
H100 hosting costs in a Blackwell world
The H100 SXM (80 GB HBM3, 3.35 TB/s, 1,979 dense FP8 TFLOPS) is no longer the flagship, but it is the most widely available inference GPU and usually the cheapest per hour. Verified on-demand prices span $1.99 (RunPod) to $12.29 (Microsoft Azure): a 6.2× gap between specialist GPU clouds and hyperscalers. It has no native FP4 math, so 4-bit models such as gpt-oss load through weight-only kernels. They save memory but not compute.
Commitments
Reserved capacity changes the math for steady workloads:
- Hyperstack reserved ('starting from'): $2.72/GPU-hr (term not stated)
- AWS 3-yr reserved (standard, no upfront): $2.97/GPU-hr (3-year)
- AWS 1-yr reserved (convertible, no upfront): $5.42/GPU-hr (1-year)
- Lambda 1-Click Cluster 256+ GPUs (2 weeks to 1 year term): $5.54/GPU-hr (2 weeks to 1 year)
- Lambda 1-Click Cluster 64+ GPUs (2 weeks to 1 year term): $5.85/GPU-hr (2 weeks to 1 year)
- Lambda 1-Click Cluster 16+ GPUs (2 weeks to 1 year term): $6.16/GPU-hr (2 weeks to 1 year)
- Google Cloud 1-yr Compute Resource CUD: $6.59/GPU-hr (1-year)
What a single H100 can serve
Gemma 4 12B Unified (instruction-tuned), gpt-oss-20b, Gemma 4 26B A4B (instruction-tuned), Gemma 4 31B (instruction-tuned), Qwen3.8-27B, Llama 4 Scout 17B-16E Instruct, gpt-oss-120b fit on one 80 GB H100 at the precision shown below. A single-GPU replica avoids tensor parallelism entirely, which makes it the simplest private deployment to run: one GPU, one model, no interconnect. Larger models need 2, 4 or 8 GPUs with NVLink, and the largest MoE models need two or more nodes.
| Model | Precision | Min memory | Min GPUs | Production GPUs |
|---|---|---|---|---|
| Gemma 4 12B Unified (instruction-tuned) | INT4 | 8 GB | 1 | 1 |
| gpt-oss-20b | FP4 (MXFP4/NVFP4) | 16 GB | 1 | 1 |
| Gemma 4 26B A4B (instruction-tuned) | INT4 | 16 GB | 1 | 1 |
| Gemma 4 31B (instruction-tuned) | INT4 | 27 GB | 1 | 1 |
| Qwen3.8-27B | FP8 | 32 GB | 1 | 1 |
| Llama 4 Scout 17B-16E Instruct | INT4 | 69 GB | 1 | 2 |
| gpt-oss-120b | FP4 (MXFP4/NVFP4) | 75 GB | 1 | 2 |
| Mistral Small 4 119B A6B | FP4 (MXFP4/NVFP4) | 81 GB | 2 | 2 |
| Mistral Medium 3.5 128B | FP8 | 154 GB | 2 | 4 |
| Qwen3.8-Flash-Next (API: Qwen3.8-Flash) | FP8 | 207 GB | 4 | 4 |
| GLM-5.3-Flash | FP8 | 378 GB | 8 | 8 |
| Llama 4 Maverick 17B-128E Instruct | FP8 | 479 GB | 8 | 8 |
| DeepSeek-V4.1-Flash-1017 | FP8 | 587 GB | 8 | 16 |
| Mistral Large 4 (Mistral-Large-4-1T-A52B) | FP4 (MXFP4/NVFP4) | 702 GB | 16multi-node | 16 |
| GLM-5.3 | FP8 | 869 GB | 16multi-node | 16 |
| DeepSeek-V4-Pro-0813 | FP8 | 1,027 GB | 16multi-node | 24 |
| Kimi K3 | FP4 (MXFP4/NVFP4) | 1,795 GB | 24multi-node | 32 |
| Kimi K3.1 Pro | FP4 (MXFP4/NVFP4) | 1,795 GB | 24multi-node | 32 |
| Qwen3.8-2.4T-A95B (API: Qwen3.8-Max) | FP8 | 2,813 GB | 40multi-node | 48 |
The calculators below open on gpt-oss-120b on one H100, the smallest real private-LLM deployment.
| GPU | Provider · offering | $/GPU-hr | As listed | 730 h / GPU | Source |
|---|---|---|---|---|---|
| H100 80GBH100 PCIe 80GB | RunPodCommunity Cloud | $1.99 | $1.99per GPU | $1,453 | Verified runpod.io |
| H100 80GBH100 SXM5 | Fluidstackon-demand · Fluidstack H100 SXM5 | $2.10 | $2.10per GPU | $1,533 | Unverified gpuperhour.com |
| H100 80GBH100 NVLink | Hyperstackon-demand · Hyperstack VM (H100 NVLink) | $2.60 | $2.60per GPU | $1,898 | Verified hyperstack.cloud |
| H100 80GBH100 SXM 80GB | RunPodCommunity Cloud | $2.69 | $2.69per GPU | $1,964 | Verified runpod.io |
| H100 80GBH100 SXM | Hyperstackreserved ('starting from') · Hyperstack VM (H100 SXM) · term not stated | $2.72 | $2.72per GPU | $1,986 | Verified hyperstack.cloud |
| H100 80GBH100 PCIe 80GB | RunPodSecure Cloud on-demand | $2.89 | $2.89per GPU | $2,110 | Verified runpod.io |
| H100 80GBH100 80GB | AWS3-yr reserved (standard, no upfront) · p5.48xlarge · 3-year | $2.97 | $23.78per 8-GPU instance | $2,170 | Verified aws.amazon.com |
| H100 80GBH100 SXM | Hyperstackon-demand · Hyperstack VM (H100 SXM) | $3.20 | $3.20per GPU | $2,336 | Verified hyperstack.cloud |
| H100 80GBH100 PCIe 80GB | LambdaOn-demand 1x instance · H100 PCIe 1x | $3.79 | $3.79per GPU | $2,767 | Verified lambda.com |
| H100 80GBH100 80GB | Crusoeon-demand (current) · Crusoe GPU instance | $3.90 | $3.90per GPU | $2,847 | Verified crusoe.ai |
| H100 80GBH100 SXM 80GB | RunPodSecure Cloud on-demand | $4.39 | $4.39per GPU | $3,205 | Verified runpod.io |
| H100 80GBHGX H100 | NebiusOn-demand | $4.50 | $4.50per GPU, 8-GPU nodes | $3,285 | Verified nebius.com |
| H100 80GBH100 SXM 80GB | LambdaOn-demand 8x instance · HGX H100 8x | $4.59 | $4.59per GPU, 8-GPU nodes | $3,351 | Verified lambda.com |
| H100 80GBH100 SXM 80GB | LambdaOn-demand 4x instance · H100 SXM 4x | $4.69 | $4.69per GPU, 4-GPU nodes | $3,424 | Verified lambda.com |
| H100 80GBH100 SXM 80GB | RunPodInstant Cluster (Clusters product) | $4.79 | $4.79per GPU | $3,497 | Verified runpod.io |
| H100 80GBH100 SXM 80GB | LambdaOn-demand 2x instance · H100 SXM 2x | $4.79 | $4.79per GPU, 2-GPU nodes | $3,497 | Verified lambda.com |
| H100 80GBH100 SXM 80GB | LambdaOn-demand 1x instance · H100 SXM 1x | $4.89 | $4.89per GPU | $3,570 | Verified lambda.com |
| H100 80GBH100 SXM5 80GB | VerdaGPU instance on-demand | $5.24 | $5.24per GPU | $3,826 | Verified verda.com |
| H100 80GBH100 80GB | AWS1-yr reserved (convertible, no upfront) · p5.48xlarge · 1-year | $5.42 | $43.34per 8-GPU instance | $3,955 | Verified aws.amazon.com |
| H100 80GBH100 | Lambda1-Click Cluster 256+ GPUs (2 weeks to 1 year term) · 1cc-h100-256 · 2 weeks to 1 year | $5.54 | $5.54per GPU, 8-GPU nodes | $4,044 | Verified lambda.com |
| H100 80GBH100 | Lambda1-Click Cluster 64+ GPUs (2 weeks to 1 year term) · 1cc-h100-64 · 2 weeks to 1 year | $5.85 | $5.85per GPU, 8-GPU nodes | $4,271 | Verified lambda.com |
| H100 80GBHGX H100 | CoreWeaveOn-demand 8-GPU node · HGX H100 | $6.16 | $49.24per 8-GPU instance | $4,493 | Verified coreweave.com |
| H100 80GBH100 | Lambda1-Click Cluster 16+ GPUs (2 weeks to 1 year term) · 1cc-h100-16 · 2 weeks to 1 year | $6.16 | $6.16per GPU, 8-GPU nodes | $4,497 | Verified lambda.com |
| H100 80GBH100 80GB | Google Cloud1-yr Compute Resource CUD · a3-highgpu-8g · 1-year | $6.59 | $52.70per 8-GPU instance | $4,809 | Verified cloud.google.com |
| H100 80GBH100 80GB | AWSon-demand · p5.48xlarge | $6.88 | $55.04per 8-GPU instance | $5,022 | Verified aws.amazon.com |
| H100 80GBH100 NVL 94GB | Microsoft Azureon-demand (single GPU) · Standard_NC40ads_H100_v5 | $6.98 | $6.98per 1-GPU instance | $5,095 | Verified azure.microsoft.com |
| H100 80GBH100 80GB | Oracle Cloud (OCI)on-demand (pay as you go) · BM.GPU.H100.8 | $10.00 | $10.00per GPU | $7,300 | Verified oracle.com |
| H100 80GBH100 80GB | Google Cloudon-demand · a3-highgpu-8g | $11.06 | $88.49per 8-GPU instance | $8,075 | Verified cloud.google.com |
| H100 80GBH100 80GB | Microsoft Azureon-demand · Standard_ND96isr_H100_v5 | $12.29 | $98.32per 8-GPU instance | $8,972 | Verified azure.microsoft.com |
Registry last checked 2026-11-14. Prices older than 45 days are flagged stale. Verified means read from the provider's own page on that date; unverified means the official page couldn't be read and the figure came from a third-party listing (named in the link); estimatemeans a reseller or analyst figure. A missing price is shown as unavailable, never filled in. Dense TFLOPS are half the sparse figures vendors headline. “n/p” means not published.
API token cost calculator
Estimate pay-as-you-go inference spend from users and requests, or from monthly token totals. Cached input, cache writes and reasoning tokens are each priced separately and never double-counted.
Reasoning tokens bill at the output rate and are added to output below, once.
220K tasks → 660K model calls a month. Context overhead is 60% of input tokens.
Dedicated GPU hosting cost calculator
Rent GPUs by the hour and serve an open-weight model yourself. Compare providers for the same configuration, and see whether the model fits, how fast it might run and what each token costs at your utilization.
117B total · 5.1B active · ships as MXFP4 MoE weights · weights ≈ 65 GB at FP4 (MXFP4/NVFP4)
Sold as whole nodes means you pay for every GPU in the node, even if your replica needs fewer. Same storage, network and management assumptions applied to every row. Throughput is a planning estimate (memory-bandwidth roofline, 60% efficiency). Load-test before you commit.
Private LLM hardware ownership calculator
Buy the servers and run them in your facility or a colocation cage. Cash flow, accounting depreciation and economic cost are shown separately, because they answer different questions.
No public price for H100 80GB servers.NVIDIA and AMD don't publish datacenter list prices; enter a vendor quote to see ownership costs.
Enter a server price (and power draw, if the GPU has no published TDP) to model ownership.
We size, procure, install and operate private LLM hardware, and hand you the keys.
API vs. self-hosting break-even calculator
The four strategies on one chart. Slide the volume and watch where fixed infrastructure overtakes per-token pricing. Infrastructure adds replicas as volume grows, and a line breaks where a strategy runs out of capacity.
| Strategy | Monthly | Annual | 3-yr | 5-yr |
|---|---|---|---|---|
| API inferencepay per token | $270 | $3,240 | $18.7K | $51.8K |
| Dedicated cloud GPUs1 unit of 1 GPUs | $10,353 | $124K | $373K | $621K |
| Owned hardwareno price: complete its calculator | — | — | — | — |
| Hybridno price: complete its calculator | — | — | — | — |
Financial break-even is not proof of capacity. It's the volume where a fixed monthly bill equals the API bill. Whether one replica can serve that volume depends on throughput, so the chart adds units as volume grows and breaks the line where a strategy runs out of room.
Agentic AI infrastructure cost calculator
Agents call models many times per task, resend context at every step and retry when tools fail. Start from a preset, then route each share of traffic to a premium API, a low-cost API or a private model.
Long repository context, many tool calls per change, heavy reasoning. Illustrative starting values, not benchmarks.
Agent workloads resend context on every step, so input tokens dominate: 9.11B input vs 524M output here. Caching and routing usually save more than haggling over output prices.
Enterprise deployment recommendation
A deterministic recommendation from your calculator inputs plus the factors cost can't capture: data sensitivity, regulation, team capability and licensing.
Also uses: 1B input tokens/mo, 50% growth, gpt-oss-120b licensing, and the infrastructure from calculators 2–3.
API-first deployment
Use pay-as-you-go APIs now; revisit when volume or data rules change.
- At this volume the API is the lowest-cost option ($270/month vs $10,353 self-hosted).
- Pay-per-token has no idle capacity, no capex and nothing to operate.
The rules are fixed and run in your browser, so the same inputs always give the same answer. Cost is one input among several: the cheapest option loses if it fails on data handling, team capability, licensing or capacity.
Stop renting intelligence. Start owning your AI infrastructure.
An engineer reviews your scenario against real quotes, your traffic logs and your data rules, then returns a costed architecture: API, dedicated, owned or hybrid.
Get a custom enterprise LLM infrastructure assessment.
Ask for an infrastructure assessment, a private LLM deployment quote or a consultation. You decide whether your calculator scenario is attached. We use it only to answer you.
LLM hosting cost questions.
How much does it cost to rent an H100?
It depends heavily on the provider. Our registry's verified on-demand rates are listed on this page, cheapest first. Reserved one- and three-year terms cost less for steady workloads.
What LLMs can run on a single H100?
Models whose weights plus runtime headroom fit in 80 GB, such as gpt-oss-120b in its native MXFP4 or 30B-class dense models at FP8. The fit table on this page lists every model in the registry.
Is an H100 still worth it for inference?
For models that fit one GPU or one 8-GPU node, often yes. It is cheaper per hour than Blackwell and widely available. For very large MoE models, B200/B300 or MI355X nodes need fewer GPUs and avoid multi-node serving.
Can I buy H100s instead of renting?
Yes, but NVIDIA publishes no list price; get vendor quotes and enter them in the ownership calculator to see payback against renting.