A new H100 GPU usually lands in the five-figure range, while full eight-GPU systems often climb into the mid six figures.
The H100 is one of those parts that makes people stop and stare at the quote. You’ll see numbers that look close together at first, then realize they refer to totally different things: a single PCIe card, a dual-GPU NVL setup, an eight-GPU HGX tray, or a whole server packed with CPUs, RAM, networking, and support.
That’s why there isn’t one neat answer with one neat price tag. The real answer depends on what kind of H100 you mean, where you’re buying it, whether it’s new or used, and whether you’re buying hardware at all or renting it by the hour in the cloud.
If you just want the fastest clean read, a single H100 card usually sells somewhere from the high twenty-thousands to the high sixties in current channel listings, with packaged enterprise SKUs often sitting on the high end. H100 NVL listings can land in the low thirty-thousands, while complete eight-GPU servers can run from roughly a quarter-million dollars up to well past half a million once the full platform is included.
What most buyers are paying right now
How Much Does A H100 Cost? The plain answer
Most people asking this question are trying to budget for one of three paths. Path one is a single accelerator card for an existing server. Path two is a ready-built server with several H100 GPUs already inside. Path three is cloud rental, where you skip the capital spend and pay per hour.
For a single card, the market range is broad. Enterprise resellers have shown H100 PCIe listings in the tens of thousands, and the spread can be huge from one vendor bundle to the next. Some listings sit around the low thirty-thousands for H100 NVL class hardware. Others push past sixty thousand for OEM-tied H100 80 GB PCIe parts that are sold with business support, channel margin, and platform compatibility baked into the price.
That sounds wild until you look at what buyers are paying for the server around the card. Once you move into 4-GPU and 8-GPU systems, the sticker isn’t just about the GPUs anymore. You’re paying for host CPUs, huge memory pools, high-speed storage, power delivery, thermal design, chassis layout, fabric, warranty terms, and sometimes software rights.
That’s why two quotes can differ by six figures even when both include “H100” in the name. One is a bare card. The other is a full production node.
H100 pricing by card, server, and cloud rental
The H100 family shows up in a few shapes, and each shape lands in a different price band.
Single PCIe cards
This is the version many buyers picture first. It drops into a supported server, draws serious power, and gives you the cleanest “one GPU” comparison. In practice, single-card pricing is still messy because channel sellers often price the same silicon through OEM part numbers, support bundles, and customer-specific contracts. Street listings still put it firmly in the five-figure bracket.
H100 NVL cards
These are aimed at dense inference and large-memory jobs in PCIe-based systems. They don’t always map neatly to the same buying path as a standard H100 PCIe card, so the price can land lower or higher than you’d expect from the label alone. In live reseller listings, this class often appears around the low thirty-thousands, though some OEM-flavored listings jump much higher.
HGX and eight-GPU servers
This is where buyers stop talking about the price of “an H100” and start talking about the price of a machine room decision. A full eight-GPU box can land deep into six figures. The GPU silicon is still the star of the quote, but the rest of the platform drives a lot of the final bill.
Cloud rental
If buying hardware feels too steep, renting H100 capacity can make the entry point far easier. Yet the math can fool you. Cloud providers usually sell you an instance, not one card. A node with eight H100 GPUs may look pricey by the hour, but the per-GPU rate can be much easier to digest when split across the full box.
A quick stop at NVIDIA’s H100 product page also shows why simple price talk breaks down fast. NVIDIA lays out multiple H100 variants and deployment styles, not one universal retail SKU with one public list price. That’s a big hint that partner quotes and channel listings drive most real-world buying.
There’s also a timing angle. H100 pricing isn’t frozen. Supply pressure has eased from the peak shortage era, but demand from AI training, fine-tuning, and heavy inference is still strong. So the market has cooled from the frenzy days without turning cheap. That leaves buyers in a zone where the H100 is no longer impossible to source, yet still expensive enough that every detail on the quote matters.
What drives the price up or down
First is form factor. A PCIe card and an SXM-based system are not priced the same way because they don’t live in the same thermal and platform world. SXM gear usually sits in denser, pricier systems built for top-end throughput.
Second is memory and interconnect. Buyers care about HBM capacity, bandwidth, and how the GPUs talk to each other. That matters a lot for large model work, and it changes what a platform is worth to the person signing the purchase order.
Third is support. Enterprise hardware isn’t sold like a consumer graphics card. The quote may include warranty terms, business support, certified compatibility, firmware handling, or install help. Those line items can make a steep quote look steeper, yet they also cut risk for teams that can’t afford downtime.
Fourth is the route to market. Distributor pricing, OEM pricing, grey-market pricing, used-market pricing, and cloud pricing all tell different stories. A used H100 pulled from a datacenter cycle is not the same buying event as a fresh OEM part on a corporate quote.
| Buying route | What you’re getting | Current market range |
|---|---|---|
| Used single H100 card | Prior-use hardware, usually card-only | About $18,000–$25,000 |
| Lower-end new channel listing | New card with lighter bundling | About $25,000–$35,000 |
| H100 NVL street listing | 94 GB NVL-class board or bundle | About $30,000–$35,000 |
| Mainstream new H100 PCIe quote | New 80 GB PCIe card through enterprise reseller | About $35,000–$50,000 |
| High enterprise reseller quote | OEM-tied H100 80 GB PCIe SKU | About $50,000–$70,000 |
| 4-GPU server build | Chassis, CPUs, RAM, storage, networking, 4 GPUs | About $180,000–$320,000 |
| 8-GPU H100 server | Full production node with 8 GPUs | About $250,000–$500,000+ |
| DGX-class branded system | Turnkey AI appliance with software and support layers | Often $350,000 and up |
Those ranges are broad on purpose. They reflect what buyers tend to run into in current public listings and enterprise quotes, not a neat MSRP sheet. With H100 hardware, the exact package matters more than the headline name.
Why one public price can mislead you
Let’s say you find a listing for an H100 at thirty-two thousand dollars and another at sixty-seven thousand. It’s tempting to think one seller is just overcharging. Sometimes that’s true. Still, just as often, you’re staring at two different channels and two different bundles.
One may be a cleaner add-in board with fewer extras. The other may be an OEM-qualified part intended for a narrow list of supported systems. That second quote can carry firmware validation, compatibility limits, channel margin, and enterprise warranty terms that don’t show up in a stripped listing.
The same trap shows up in servers. A box with eight H100 GPUs may sound like “eight times the card price,” but it rarely works that way. Host platform costs are heavy at this tier. High-core CPUs, giant DDR footprints, NVMe storage, power supply overhead, fabric, rails, and service contracts all pile on fast.
If you’re trying to budget from a search result alone, use a range first. Then pin the quote down by asking what is and isn’t included. That single step can save you from a bad comparison.
Cloud math changes the answer
For many teams, buying an H100 outright isn’t the best first move. Renting one can be easier on cash flow, easier on deployment time, and easier on risk. You lose the long-term asset, but you skip a large upfront check and the wait tied to hardware procurement.
The trick is to read cloud pricing the right way. On AWS, the P5 instance page shows H100-backed instances sold as full nodes, not single cards. Current public trackers place a p5.48xlarge on-demand instance around the mid-fifties per hour in the lowest-price regions, and that box includes eight H100 GPUs. Split that out and you’re in the ballpark of about seven dollars per GPU hour before you count the rest of the node resources.
Specialist GPU clouds can land lower or higher than that depending on whether the quote is for a full node, a bare GPU component, reserved use, or spot-style capacity. That’s why a cloud H100 can look cheap in one screenshot and expensive in the next. You may not be buying the same shape of service.
There’s a rough break-even idea that helps. If you need H100 compute all day, every day, for a long stretch, owned hardware starts to look better. If your work comes in bursts, or your model path is still changing, rental often wins because it keeps you from locking a pile of cash into one generation of silicon.
| Use case | Better fit | Why the money works that way |
|---|---|---|
| Short training run | Cloud rental | No big upfront spend and fast access |
| Daily production inference | Owned server or long reservation | High steady use can justify the purchase |
| Research team with spiky demand | Cloud rental | You pay only when jobs are running |
| Large internal AI platform | Owned cluster | Lower long-run cost if the GPUs stay busy |
| Testing before a hardware buy | Cloud first | Cheap way to validate model fit and memory needs |
Should you buy one now or wait
That depends on whether you need H100-class performance or just want it. Those are not the same thing.
If your workload is tightly tuned to H100 memory bandwidth, interconnect, or training speed, then the answer is simple: price the exact platform you need and move. Lost time can cost more than the hardware.
If your workload is lighter, the H100 may be overkill. A lower-tier accelerator or a rental-first plan can cut your spend hard without wrecking your results. That’s where buyers save the most money: not by haggling a few points off an H100 quote, but by checking whether they needed an H100 at all.
There’s also the generational angle. As newer accelerators grab attention, H100 prices tend to soften at the edges. That doesn’t mean a collapse. It means the used market gets more active, cloud competition gets sharper, and the premium attached to “must have one today” eases a bit.
What to budget before you click buy
If you’re budgeting for a single new H100 card, a sane working number is thirty-five to fifty thousand dollars, then a buffer on top if you’re buying through a large enterprise channel. If you’re shopping used, you may find deals below that, though the trade-off is more risk around condition, warranty, and provenance.
If you’re budgeting for a full server, don’t start with “GPU price times card count.” Start with the full platform target. Four-GPU systems can clear two hundred thousand dollars fast. Eight-GPU systems can push far beyond that once the machine is production-ready.
If you’re budgeting for cloud, estimate the hourly burn from the whole instance, not from the chip name alone. Then multiply by your real training or inference hours each month. That number tells you more than any forum thread ever will.
So, how much does an H100 cost in plain English? A lot. A single card still lives in premium five-figure territory, and a serious H100 server is a capital purchase, not an impulse buy. The better question is whether you should buy the card, buy the box, or rent the compute. Once you answer that, the price starts making sense.
References & Sources
- NVIDIA.“NVIDIA H100 GPU.”Shows the H100 family, deployment styles, and product variants that explain why there is no single public one-price answer.
- Amazon Web Services.“Amazon EC2 P5 Instances.”Confirms that AWS P5 instances are built around NVIDIA H100 GPUs, which supports the cloud-rental pricing section.