LLM.co · Private AI economicsNVIDIA B300 LLM Hosting Cost Calculator
Blackwell Ultra puts 2.3 TB of HBM in a single NVLink domain. See what that costs per hour across clouds, which models fit and when it beats per-token pricing.
Open weights: compared against API pricing and self-hosting on B300 (Blackwell Ultra).
B300 hosting costs and what fits on a node
The B300 (Blackwell Ultra) carries 288 GB of HBM3e at 8 TB/s per GPU, 13,500 dense FP4 TFLOPS, and fifth-generation NVLink across an 8-GPU HGX board: 2,304 GB in one node. That much memory in a single NVLink domain is why it has become the default target for trillion-parameter MoE models. Verified on-demand prices across 10 offerings run from $6.94 (RunPod) to $17.80 (AWS) per GPU-hour, a 2.6× spread for the same silicon. On an always-on 8-GPU node that is $40,530 versus $103,964 a month.
Price per GB of GPU memory
For memory-bound inference, the price per GB of HBM is a fairer comparison than the price per GPU:
| GPU | Memory | Cheapest verified on-demand $/GPU-hr | $/GB-hr |
|---|---|---|---|
| B300 (Blackwell Ultra) | 288 GB | $6.94 · RunPod | $0.0241 |
| B200 | 180 GB | $5.98 · RunPod | $0.0332 |
| H200 141GB | 141 GB | $3.59 · RunPod | $0.0255 |
Models that fit on one 8× B300 node
18 of the 19 open-weight models in our registry fit on a single node at their smallest official precision. The ones that don't need two nodes and multi-node serving.
| 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 | 1 |
| gpt-oss-120b | FP4 (MXFP4/NVFP4) | 75 GB | 1 | 1 |
| Mistral Small 4 119B A6B | FP4 (MXFP4/NVFP4) | 81 GB | 1 | 1 |
| Mistral Medium 3.5 128B | FP8 | 154 GB | 1 | 1 |
| Qwen3.8-Flash-Next (API: Qwen3.8-Flash) | FP8 | 207 GB | 1 | 1 |
| GLM-5.3-Flash | FP8 | 378 GB | 2 | 2 |
| Llama 4 Maverick 17B-128E Instruct | FP8 | 479 GB | 2 | 4 |
| DeepSeek-V4.1-Flash-1017 | FP8 | 587 GB | 4 | 4 |
| Mistral Large 4 (Mistral-Large-4-1T-A52B) | FP4 (MXFP4/NVFP4) | 702 GB | 4 | 4 |
| GLM-5.3 | FP8 | 869 GB | 4 | 4 |
| DeepSeek-V4-Pro-0813 | FP8 | 1,027 GB | 4 | 8 |
| Kimi K3 | FP4 (MXFP4/NVFP4) | 1,795 GB | 8 | 16 |
| Kimi K3.1 Pro | FP4 (MXFP4/NVFP4) | 1,795 GB | 8 | 16 |
| Qwen3.8-2.4T-A95B (API: Qwen3.8-Max) | FP8 | 2,813 GB | 16multi-node | 16 |
| GPU | Provider · offering | $/GPU-hr | As listed | 730 h / GPU | Source |
|---|---|---|---|---|---|
| B300 (Blackwell Ultra)B300 | RunPodCommunity Cloud | $6.94 | $6.94per GPU | $5,066 | Verified runpod.io |
| B300 (Blackwell Ultra)B300 | Hyperstackon-demand · Hyperstack VM (B300) | $7.40 | $7.40per GPU | $5,402 | Verified hyperstack.cloud |
| B300 (Blackwell Ultra)B300 SXM (B300-SXM-8-288G) | ScalewayOn-demand 8-GPU instance · B300-SXM-8-288G | $8.29 | $66.32per 8-GPU instance | $6,052 | Unverified computeprices.com |
| B300 (Blackwell Ultra)B300 (Blackwell Ultra) | Crusoeon-demand (current) · Crusoe GPU instance | $8.96 | $8.96per GPU | $6,541 | Verified crusoe.ai |
| B300 (Blackwell Ultra)B300 | RunPodSecure Cloud on-demand | $8.99 | $8.99per GPU | $6,563 | Verified runpod.io |
| B300 (Blackwell Ultra)HGX B300 | NebiusOn-demand | $9.50 | $9.50per GPU, 8-GPU nodes | $6,935 | Verified nebius.com |
| B300 (Blackwell Ultra)B300 | RunPodInstant Cluster (Clusters product) | $10.79 | $10.79per GPU, 8-GPU nodes | $7,877 | Verified runpod.io |
| B300 (Blackwell Ultra)B300 SXM6 268GB | VerdaGPU instance on-demand | $12.41 | $12.41per GPU | $9,059 | Verified verda.com |
| B300 (Blackwell Ultra)B300 SXM6 268GB | VerdaInstant cluster (16-144 GPUs) | $12.41 | $12.41per GPU, 16-GPU nodes | $9,059 | Verified verda.com |
| B300 (Blackwell Ultra)B300 (Blackwell Ultra) | Oracle Cloud (OCI)on-demand (pay as you go) · B300 bare metal (shape name unconfirmed) | $15.00 | $15.00per GPU | $10,950 | Verified oracle.com |
| B300 (Blackwell Ultra)B300 (Blackwell Ultra) | AWSon-demand · p6-b300.48xlarge | $17.80 | $142.42per 8-GPU instance | $12,995 | Verified aws.amazon.com |
| GPU | Memory | Bandwidth | Dense FP8 / FP4 TFLOPS | Power | Scale-up |
|---|---|---|---|---|---|
| NVIDIA H200 141GBHopper · SXM | 141 GBHBM3e | 4.8 TB/s | 1,979 / no FP4 | 700 W | NVLink 4, 900 GB/s per GPU |
| NVIDIA B200Blackwell · SXM | 180 GBHBM3e | 8 TB/s | 4,500 / 9,000 | 1,000 W | NVLink 5, 1.8 TB/s per GPU (NVSwitch, 14.4 TB/s per 8-GPU HGX) |
| NVIDIA B300 (Blackwell Ultra)Blackwell Ultra · SXM | 288 GBHBM3e | 8 TB/s | 4,500 / 13,500 | 1,100 W | NVLink 5, 1.8 TB/s per GPU (14.4 TB/s per 8-GPU HGX) |
| AMD Instinct MI355XCDNA 4 · OAM | 288 GBHBM3e | 8 TB/s | 5,000 / 10,100 | 1,400 W | Infinity Fabric: 7 scale-up links x 144 GB/s (~1,008 GB/s aggregate); 128 GB/s scale-out link |
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.
2,800B total · 104B active · ships as MXFP4 weights (quantization-aware training) · weights ≈ 1,561 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.
| View | Monthly | What it means |
|---|---|---|
| Cash flow | $849,800 | Capex paid up front, then opex each month |
| Depreciation | $15,205 | Straight-line (cost − residual) over 4 years: the P&L charge |
| Economic | $18,617 | Level capital cost including the 8% cost of money: used for comparisons |
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 | $31,000 | $372K | $2.14M | $5.95M |
| Dedicated cloud GPUs1 unit of 8 GPUs | $61,407 | $737K | $4.3M | $10.7M |
| Owned hardware1 unit of 16 GPUs | $48,946 | $587K | $2.1M | $4.94M |
| Hybrid30% API, 1 owned unit | $58,246 | $699K | $2.4M | $5.7M |
- 1 Owned hardware matches the API at 15.8B input tokens/mo ($48,946/mo)
- 2 Hybrid matches the API at 22.6B input tokens/mo ($69,923/mo)
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: 10B input tokens/mo, 50% growth, Kimi K3.1 Pro 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 ($31,000/month vs $48,946 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 an NVIDIA B300 cost per hour?
Verified on-demand prices in our registry vary by provider and offering. The table on this page lists each with its source and check date, cheapest first, both per GPU-hour and for 730 hours a month.
How much memory does a B300 have?
288 GB of HBM3e per GPU at about 8 TB/s, or about 2.3 TB across an 8-GPU HGX B300 node.
B300 or B200 for inference?
The B300 has 60% more memory per GPU (288 vs 180 GB) and more FP4 compute. For models that fit on 8× B200, the cheaper B200 can win on cost per token. For 1–2 TB models, one B300 node avoids multi-node serving.
Why are B300 prices so different between clouds?
Hyperscalers bundle networking, support and enterprise agreements, and price whole 8-GPU instances. Specialist GPU clouds price per GPU with fewer extras. Commitments, regions and spot capacity widen the gap further.