NVIDIA H100 price guide: what buyers actually pay
AI & GPU infrastructure buying · Published 2026-07-25
As of Q3 2026 a single NVIDIA H100 80GB accelerator typically trades in the region of 25,000 to 40,000 US dollars per card on the new market, with the SXM module form commanding more than the PCIe card, and secondary-market units clearing well below that band depending on condition and provenance. A complete 8-GPU server built around them runs to several hundred thousand dollars once platform, memory, storage and support are included.
Those are indicative market ranges, not a quotation: real pricing moves with supply, configuration and destination, and this page is refreshed quarterly. For a formal number on a specific configuration, the honest answer is always a proforma invoice. This guide sits under our GPU server buying guide, which covers specification and platform choice.
At a glance
- H100 80GB: roughly 25,000 to 40,000 USD per card, new, as of Q3 2026
- SXM modules price above PCIe cards and are sold as part of a baseboard
- Secondary market clears materially lower, with provenance driving the spread
- An 8-GPU node is a several-hundred-thousand-dollar system, not eight card prices
- Export controls restrict where top-tier accelerators can lawfully ship
- Cloud rental prices per GPU-hour, and can beat ownership below a utilization threshold
What you are actually buying
| Variant | Memory | Form | Relative price |
|---|---|---|---|
| H100 PCIe | 80GB HBM2e | Standard double-width card | Lowest of the H100 family |
| H100 SXM5 | 80GB HBM3 | Module on an HGX baseboard | Above PCIe, higher bandwidth and power |
| H100 NVL | 94GB per card, paired | Dual-card bridged configuration | Premium, aimed at large-model inference |
| H200 | 141GB HBM3e | SXM and PCIe | Above H100, successor generation |
Buyers searching for an H100 price often mean one of these four and get quoted another. Fix the variant before comparing offers, because the memory capacity and interconnect differences are exactly what the price gap pays for.
Why the price is what it is
Four forces set the number.
Manufacturing scarcity
Advanced packaging and high-bandwidth memory supply have been the binding constraints on datacenter GPU output through this cycle. Capacity has expanded, but allocation to the largest customers still shapes what reaches everyone else.
Demand concentration
Hyperscalers and large AI labs buy in fleet quantities on contracts. Smaller buyers compete for the remainder, which is why list-style pricing and channel pricing can diverge so widely for the same part.
Export controls
United States export rules restrict shipment of top-tier AI accelerators to certain destinations. That fragments the market: supply concentrates in permitted markets and thins in restricted ones, and grey-market premiums in restricted regions are a symptom of that, not an opportunity. We decline transactions that cannot be lawfully fulfilled, and any seller who does not ask about end use is a warning sign rather than a bargain.
Generational rotation
As newer accelerator generations ship in volume, previous-generation silicon rotates out of hyperscale fleets and into the secondary market. That is the single biggest deflationary force on H100 pricing and the reason a refurbished strategy is now genuinely mainstream for teams that were priced out at launch.
New versus secondary market
Secondary-market H100s clear below new pricing, sometimes substantially, but the spread is wide because you are pricing risk as much as silicon. What separates a good buy from a bad one:
- Per-card burn-in and memory test results, not a batch certificate
- Documented provenance: what fleet, what environment, why decommissioned
- Written condition grade and a stated seller warranty window
- Firmware and vendor-lock status disclosed up front
- An accountable counterparty who will take a failed card back
Datacenter accelerators are built for continuous duty, so a well-documented used card is a reasonable industrial purchase. An undocumented one is a lottery ticket. Our GPU and accelerator listings state condition and grading per line for that reason.
The system price, not the card price
Almost nobody buys a bare H100. Budget the whole node:
| Line | Share of an 8-GPU build |
|---|---|
| Accelerators | The dominant line by far |
| HGX platform, CPUs, chassis | Significant, tens of thousands |
| System memory and NVMe | Meaningful on training builds |
| Networking and fabric | Rises sharply for multi-node clusters |
| Power, cooling, rack | Facility cost, often overlooked |
An 8-GPU node also draws power on the order of 10 kW under load, which is a facility decision before it is a purchasing one.
Buy or rent
Cloud providers rent H100 capacity by the GPU-hour. Renting wins for bursty, exploratory or short-project work and for teams without facilities. Owning wins when utilization is sustained, when data residency or privacy constrains where workloads run, or when the amortized hardware cost per hour beats the rental rate at your duty cycle. Run that arithmetic with your real expected utilization rather than a hoped-for one: the break-even is usually about how many hours per week the cards are actually busy.
How to get a real number
Send the exact variant, quantity, condition acceptance (new or graded refurbished), destination country and whether you need the platform as well. That is enough for a proforma invoice with lead time and Incoterms. Prices for this part move week to week, so quotations carry validity windows, and anyone quoting a firm price for an indefinite period is not pricing this market seriously.
FAQ
How much does one NVIDIA H100 cost?
Roughly 25,000 to 40,000 US dollars per card on the new market as of Q3 2026, varying by variant, quantity and channel, with secondary-market units clearing lower. Treat any single figure as indicative until it is on a proforma invoice.
Why is the NVIDIA H100 so expensive?
Constrained advanced packaging and high-bandwidth memory supply, demand concentrated among very large buyers, export controls fragmenting where supply can go, and the fact that each card carries substantial HBM content. It is a supply-and-allocation story more than a manufacturing-cost story.
Is an H100 better than a consumer card like a 5090?
For datacenter AI work, yes, and they are not really substitutes. The H100 carries far more high-bandwidth memory, ECC protection, datacenter driver support, continuous-duty engineering and multi-GPU interconnect. Consumer cards win on price per unit of raw compute for single-card workloads that fit in their memory.
How much does H100 cost per hour in the cloud?
Cloud pricing is quoted per GPU-hour and varies widely between hyperscalers and specialist GPU clouds, and with commitment length. Compare the rate against your amortized cost of ownership at your genuine utilization before deciding.
Can I buy H100 GPUs anywhere in the world?
No. Export controls restrict shipment of top-tier accelerators to certain destinations, and lawful sellers will ask for end-user and end-use information on controlled models. We supply only where the transaction can be lawfully fulfilled.
Sourcing this hardware?
AXISLINK supplies business buyers worldwide from Hong Kong and mainland China, with export documentation and freight arranged. Send your requirements and receive a formal quotation, typically within 24 hours.
Request a quote