For customers
AI Price Guide: what AI infrastructure costs in November 2026
Rough ballpark prices for buying AI hardware, renting GPU cloud and calling AI models, so you can set a budget before you ask for quotes. Prices checked on 18 November 2026.
Read this first: These are indicative street and list prices from public trackers, in US dollars, before UK VAT, import duty and delivery. They are not offers. Always get a firm quote before you commit.
1. Buy: AI hardware
| Item | Typical price (USD) | What it is for | Notes |
|---|---|---|---|
| NVIDIA H100 80GB GPU | $25,000 to $40,000 | Training and inference. The established workhorse. | Used cards sell for roughly $8,000 to $28,000. |
| NVIDIA H200 141GB GPU | About $31,000 to $32,000 | Bigger models that need more memory. | An 8-GPU H200 server is about $315,000. |
| NVIDIA B200 GPU | $30,000 to $50,000 | Current-generation training and inference. | Waitlists reported. Most customers buy complete systems. |
| 8-GPU H100 server (DGX class) | $250,000 to $400,000 | One-node training or serving a large model. | Often the entry point for serious in-house AI. |
| 8-GPU B200 server (DGX class) | $300,000 to $520,000 | Highest single-node performance. | Estimates vary by source and configuration. |
| 8-GPU B300 server (DGX class) | $300,000 to $400,000 | Largest memory per GPU. | Newest. Limited supply. |
| GB200 NVL72 rack (72 GPUs) | $2 million to $3 million | Large-scale training and inference, liquid cooled. | GB300 NVL72 is estimated at $3 million to $4 million. |
| AMD MI300X GPU | $10,000 to $15,000 | Memory-heavy inference at lower cost. | One source only. Check availability. |
Purchase prices have held or risen during 2026, partly because memory costs are rising. Used H100 prices have fallen as newer chips arrive. The figures cover the GPU or server only. Networking, storage, power, cooling, software and installation are extra, and we price the whole system for you.
2. Rent: GPU cloud
| GPU | Big cloud providers (AWS, Azure, Google) | GPU cloud specialists | Market median |
|---|---|---|---|
| H100 | $6.88 to $12.29 per GPU-hour | $4.39 to $6.16 | About $4.30 |
| H200 | $7.91 to $10.60 | $4.47 to $6.31 | About $5.30 |
| B200 | About $14.24 (AWS) | $7.69 to $8.60 | About $10.30 |
| AMD MI300X | About $6.00 (Azure) | About $2.59 | About $3.40 (one provider) |
| A100 80GB | Not tracked | $1.00 to $1.50 | About $1.40 |
| RTX 4090 and 5090 | Not offered | $0.40 to $1.20 on marketplaces | About $0.90 and $1.15 |
On-demand rates per GPU per hour, in US dollars. Committing for one to three years usually cuts the rate by roughly 25 to 56 percent. Spot capacity is cheaper again but can be interrupted. GPU rental prices rose through 2026: trackers show H100 medians up by about 13 to 31 percent on a year ago.
3. Call: AI models and agents
| Tier | Typical price per million tokens (input / output) | Examples as listed in November 2026 |
|---|---|---|
| Flagship | $2 to $10 in, $6 to $50 out | GPT-6 Astra $10 / $50. Claude Fable 5.2 $8 / $40. Gemini 4 Argon $2 / $10 until 30 November, then $4 / $20. Grok 4.7 $2 / $6. |
| Mid-tier | $1 to $2 in, $5 to $10 out | GPT-6.1 Sol $2 / $10. Claude Sonnet 5.5 $2 / $10. Gemini 3.5 Flash $1.50 / $9. Mistral Medium 3.5 $1.50 / $7.50. |
| Small and fast | $0.10 to $0.30 in, $0.40 to $1.50 out | GPT-6 Luna $0.10 / $0.50. Claude Haiku 5.5 $0.10 / $0.50. Gemini 3.1 Flash-Lite $0.25 / $1.50. DeepSeek V4.1 Flash $0.30 / $1.20 at peak. |
| Open-weight models, hosted | About $0.15 in, $0.60 out | GPT-OSS 120B on Groq, Fireworks and Together. |
Prices in US dollars, taken from public price trackers. Check each supplier before you commit, because model prices change often and some are promotional. Batch jobs are usually half price, and repeated input that the model has already seen can cost 90 percent less. Our LLM models and token costs page explains how to estimate your own bill.
What moves the price
- Memory. HBM and DRAM costs are pushing up GPU and server prices through 2026.
- Volume and timing. Larger orders and flexible delivery dates get better prices and shorter waits.
- Export controls. Some destinations and end users need licences, which can add time and cost.
- Support and service level. Four-hour on-site support costs more than next-day parts replacement.
- UK costs. VAT, import duty, delivery and power at your site are not in these figures.
Tell us what you need once. Up to four suppliers that fit will send you a quote. Not sure yet? Start with a fixed-fee assessment.
Request for QuoteBook an assessment
New here? Create a free account to ask questions and keep every reply in one place.
Reach customers who are ready to talk. We check your company and product facts, and label every paid relationship.
Become a supplier partnerSee the supplier directory
Want to apply online? Create a supplier account and we will vet and approve it.
Common questions
Are these prices exact?
No. They are rough ballpark figures from public price trackers and reseller listings in November 2026. NVIDIA does not publish list prices for data-centre GPUs, so street prices move with supply, volume and delivery date. Use them to set a budget, then request firm quotes.
Do the prices include VAT, import duty and delivery?
No. Figures are in US dollars before UK VAT, import duty, shipping, installation and support. Quotes we collect for you are in pounds where the supplier can price that way.
Should we buy hardware or rent GPU cloud?
It depends on how busy the hardware will be and for how long. As a guide, renting eight H100 GPUs at a typical GPU cloud rate for a month costs roughly 9 percent of the price of an 8-GPU H100 server, so rental is cheaper for short or uncertain projects and buying can win for steady, long-running use. We model both for you.
Why do GPU cloud prices differ so much?
The big cloud providers charge more for on-demand capacity than GPU cloud specialists, and committing for one to three years cuts the rate sharply. Region, networking and support level also change the price.
How do I get a firm price?
Send a Request for Quote. We screen it, pass it to suitable suppliers and bring back comparable quotes.
Where the numbers come from
Hardware: IntuitionLabs (updated October 2026), GPUSmith (October 2026) and Amnic (June 2026). GPU cloud: GPUPerHour (18 November 2026), IoT Digital Twin PLM (14 November 2026) and Ornn (17 November 2026). Models: Morph (verified 10 November 2026), IoT Digital Twin PLM (15 November 2026) and Maximem (30 October 2026). These are independent trackers, not supplier price lists, and they sometimes disagree. Where they did, we show the range.
Next steps
Related: LLM models and token costs · Neoclouds and GPU cloud
Know roughly what you need?
Send one brief and get comparable quotes.
