Quick Answer
The Nvidia H100 price runs about $31,000 for a new 80GB card, $25,000 to $35,000 on the secondary market, and $250,000 to $320,000 for a new 8-GPU HGX system. Renting runs $2.89 to $6.98 per GPU-hour on demand across the market, with the September 2026 market median at $3.25 and rates volatile rather than steadily falling.
Last updated: October 5, 2026
The H100 GPU was supposed to be winding down. Blackwell shipped, Rubin is in production, and the four-year-old chip should be depreciating in the corner. Instead, H100 rental prices spiked 22% in the month to September 7, 2026, and the Ornn OCPI-H100 index still settled at $2.92 per GPU-hour on October 1, up 3.2% over thirty days against a three-month range of $2.46 to $3.17.
One-year contract rates are up roughly 38% since October 2025, and on-demand capacity is sold out across providers. The market wrote the H100 off and then bid it up. Four years into its life, the H100 has the rental curve of a chip that just launched.
That reversal is the story of GPU budgeting in late 2026, and it changes the buy-rent-skip math this page exists to answer.
It also raises the stakes on knowing your effective rate: CloudZero’s State of AI Costs report found GPU spend sitting as the largest raw line in most AI budgets, and a line that big moving 22% in a month is not a procurement detail.
How much does an Nvidia H100 cost to buy?
The numbers that matter, current market:
| What you’re buying | Price (Q4 2026) |
|---|---|
| Single H100 80GB card, new | ~$31,000 (live retail listings) |
| Single H100, secondary market | $25,000 to $35,000 per GPU |
| 8x H100 HGX system, new | $250,000 to $320,000 (~$285,000 typical) |
| 8x H100 system, refurbished | Substantially less: new runs about 70% more than equivalent used capacity |
| DGX H100 (Nvidia’s own 8-GPU system) | Six figures, quote-based through partners; the DGX H100 price tracks integrator HGX builds |
Purchase prices have held steadier than rental rates this year, which improved the case for owning.
The price picture splits into list, street, and used, and there is still no official H100 MSRP: Nvidia has never published a list price, it sells through OEMs and integrators, and every published number, this page included, is a market observation. The difference is whether the observations carry a date.
The used market went from underrated story to seller’s market. Azure ran V100 instances for about 7.5 years, in August 2026 CoreWeave CEO Mike Intrator said the company had contracted 2020-vintage A100s out to 2029 at full freight. Five-year-old silicon with a waiting list is not a depreciation curve anyone modeled.
2026 State of AI Spend Report
Why Finance Caps What It Can’t See
Learn how finance leaders are managing AI spend, and how they’re approaching FY2027
How much does it cost to rent an H100?
The H100 GPU price per hour, on demand, October 2026:
| Provider | Per GPU-hour (on-demand) | Notes |
|---|---|---|
| Market median (all providers) | $3.25 to $3.28 | Per the aiMultiple and Ornn indexes, September 2026 |
| Spot / marketplace floor | $2.20 to $2.89 | Spheron listed H100 at $2.98 on-demand and $2.21 spot on October 2; Bank of America pegged broad spot at ~$2.80 in August, near all-time highs for older silicon |
| Thunder Compute | $3.20 | Stable on-demand, per-minute billing |
| Lambda | $3.29 to $3.99 | SXM at the top of the range; no egress fees |
| CoreWeave | $4.25 (PCIe) to $6.16 (per GPU on 8x HGX) | InfiniBand, enterprise SLAs |
| GCP (A3 instances) | ~$2.25 spot; on-demand well above | The GCP A3 H100 SKU; spot is interruptible |
| AWS (p5.48xlarge, 8x H100) | $55.04/hour per instance, ~$6.88 per GPU | Effective rates lower with capacity blocks |
| Azure (ND H100 v5) | Up to $6.98 | The Azure ND H100 v5 series |
The table’s fine print changed this year, because the market flipped underneath it. A year ago the footnote was “spot occasionally dips below $2.” In September 2026 the footnote is “on-demand capacity is sold out across GPU types,” per SemiAnalysis, which also reports contract renewals landing at the exact rates signed in 2023 and some H100 contracts renewing through 2028.
Renters are subletting cluster capacity like apartments during a Grand Prix; SemiAnalysis has floated the term Neocloud slumlords for the emerging middlemen.
The dedicated-cloud versus hyperscaler spread still exists, roughly $3.20 to $4 against $6.88 to $6.98, and it still buys the same silicon wrapped in SLAs, compliance, and ecosystem. Whether that wrapper is worth 2x is a workload question, not a loyalty question. What changed is the floor under both tiers: it rose.
What did Blackwell do to H100 prices?
Repriced everything three times, in opposite directions:
Move one, the 2025 cuts. AWS cut H100 on-demand pricing 44% in June 2025 and the market followed. That was B-series inventory arriving and demand curves bending, and it produced the cheapest H100s in the chip’s history by October 2025, with one-year contracts touching $1.70 per hour.
Move two, the value repositioning. With B200 spot running around $5.66 per hour per Bank of America’s August 2026 tracking, and B300 (Blackwell Ultra, 288GB) taking the memory crown, the H100 settled in as the price-performance workhorse: mature software stack, every tutorial on the internet, no driver archaeology. Ecosystem maturity is a discount that never appears on a rate card, because engineer-hours cost more than GPU-hours.
Move three, the math flip. B200 delivers roughly 2.5x H100 training performance, so the newer chip is often cheaper per result despite the higher sticker. Hourly rate is the wrong metric; cost per training run is the right one, and it sometimes points at the more expensive GPU. That remains the most common error in GPU budgeting.
That’s the H100 news cycle through mid-2026. What Blackwell did not do is kill H100 demand, and that surprise gets its own section.
Why are H100 prices going back up in 2026?
The price reversed because supply froze while demand compounded. Nvidia moved production to Blackwell and Rubin, so the H100 fleet stopped growing and only shrinks from here. Meanwhile every deployed model is a permanent inference load, and inference does not care that the silicon is four years old.
The receipts, all dated: Ornn reported H100 rentals up 22% in the month to September 7, 2026, with its OCPI-H100 index settling at $2.92 on October 1. SemiAnalysis tracked one-year contract midpoints from $1.70 (October 2025) to $2.35 (March 2026), a 38% climb, with on-demand sold out and Blackwell lead times stretching months.
Bank of America’s early-August note put spot near all-time highs for older silicon. AMD warned partners of double-digit price increases on its graphics cards into 2026 on memory inflation, a signal of the same DRAM squeeze pressing on data-center supply, and the January 2026 export-control shift permitting conditional H200 sales to China added a demand source mid-cycle.
The market is now big enough to trade. Mordor Intelligence sizes GPU rental at about $52 billion in 2026, up from $34.6 billion in 2025 and projected toward $199 billion by 2031.
The first US-regulated compute futures list on October 5, 2026: CME Group’s Silicon Data H100 and B200 Rental Index Futures (tickers GPU1 and GPU2 on NYMEX, pending regulatory review). Each contract is one month of GPU time, 730 GPU-hours, cash-settled against indexes of neocloud on-demand rates, with monthly expiries listed as far as 36 months out.
ICE announced a competing suite with Ornn, and the CFTC is taking public comment on compute derivatives.
The H100 successor story now includes a 36-month forward curve. GPU-hours become a traded commodity with a hedgeable price on October 5, which tells finance teams something rate cards never could: the market itself is unsure where this goes.
For a finance leader, the strategic read is short. You can’t control the market rate, and after 2026 nobody should claim they can predict it. You control exactly one variable in the GPU equation, and it’s utilization. The rest of this page is about that variable.
H100 vs H200 vs B200: which should you pay for?
| GPU | Buy (approx.) | Rent (September 2026) | The honest positioning |
|---|---|---|---|
| A100 80GB | Secondary market only | $1.09 to $5.07/hr, median $1.76 | The budget workhorse; fine for 7B-class work, contracted to 2029 at some clouds |
| H100 80GB | ~$31K/card; $25-35K used | $2.89 to $6.98/hr, median $3.25 | The value pick with the deepest software maturity per dollar |
| H200 | Quote-based premium | Median $4.40/hr | 141GB on the same architecture; the inference-heavy pick |
| B200 (Blackwell) | Reservation-heavy | Spot ~$5.66/hr; rate cards still scarce | ~2.5x H100 training performance; often cheapest per result |
The H100 vs A100 call is the easiest in the table: the A100 at $1.76 median is roughly half the H100’s rate for meaningfully less performance, which makes it the right rental for smaller models and the wrong one the moment jobs start queueing.
The H100 vs H200 call is a memory question: pay the ~35% rental premium when models are memory-bound, skip it when they’re not. Spec context for both comparisons: the Nvidia H100 carries 80GB of HBM3 at up to 700W in SXM form, and that 80GB ceiling is exactly what the H200’s 141GB exists to relieve.
Decision rule in one sentence: memory-bound inference favors H200, training throughput favors B200 per result, budget inference favors A100, and everything cost-sensitive with a mature toolchain favors the H100 at 2026 rates, rebound included.
Buy, rent, or skip GPUs entirely?
The three-way decision, with the math that decides it:
Rent when utilization is uncertain. At $3.99/hour, a rented H100 costs about $35,000 per GPU-year at full utilization, more than buying the card. But renting converts a six-figure guess into an hourly experiment you can end on Friday, and in a rising-rate market it also preserves the option to jump generations when Rubin capacity reaches the open market in late 2026 (first-wave Rubin allocation is dominated by hyperscaler commitments, with Microsoft and Oracle taking the earliest NVL72 systems).
Buy (probably used) when utilization is proven. An 8x used system pays back against rentals in about seven months at full utilization, roughly $168,000 for used capacity against $23,300 a month renting eight GPUs around the clock, and the 2026 rebound shortened that payback: rising rental rates make owned, utilized hardware more valuable, and the resale tail got stronger, not weaker. At 30% utilization, payback still stretches to about two years. Measure from three months of actual GPU-hours on rental invoices, not a roadmap deck’s optimism.
Skip GPUs when an API does the job. Many inference workloads on rented H100s cost a fraction on a per-token API, and the gap widened again in 2026: OpenAI cut GPT-5.6 Luna 80% in July while H100 rates rose 22% in a month. Token prices fall as GPU rents climb. Price your monthly volume against the LLM API pricing comparison and OpenAI’s rates, check the per-unit view in the inference cost guide, and keep the GPUs only if the infrastructure bill wins or the workload can’t leave your walls.
What is the real cost of an idle H100?
The H100 cost that matters is the effective one: an H100 at $3.99/hour costs the same whether it is training a model or sitting idle behind a misconfigured queue.
At 40% utilization your effective H100 GPU cost isn’t $3.99; it’s nearly $10 an hour, which puts your discount GPU cloud above hyperscaler list prices. And the 2026 rebound made idle time strictly more expensive: the market spent 2025 negotiating rates down 44%, took the win back in 2026, and most teams’ idle time out-ate both moves without appearing on any invoice.
GPU spend also has a visibility problem the rates hide. A team renting from Lambda for training, RunPod for experiments, and AWS for production inference gets three invoices in three units: per GPU-hour, per instance-hour with eight GPUs bundled, per GPU-second. None can say what a single training run cost end to end.
Put numbers on it: a fine-tuning run on 8 H100s for 36 hours at $3.99 costs $1,149; the same run on an idle-prone cluster stretching to 60 wall-clock hours costs $1,915, a 67% premium with no line item explaining it. H100 waste never arrives as one big number. It arrives as a thousand reasonable-looking ones.
That’s the problem CloudZero works on, and the mechanics are specific to GPUs. Native AWS, Azure, and GCP integrations pull hyperscaler GPU spend, and AnyCost ingests specialized GPU cloud bills into one normalized view in the AI Hub.
Dimensions allocate it to cost per training run, per model, per team, no tags required. Budgets and forecasting put forward numbers on the biggest checks in AI, and anomaly detection catches the job that left 64 GPUs running over a weekend before the invoice announces it.
It works at AI scale in production: Upstart, an AI lending platform, used this operating model to drive $20 million in cloud cost reduction while its AI workloads kept growing.
The industry got very good at negotiating GPU hourly rates and stayed bad at knowing what the hours produced, and only one of those skills shows up in AI ROI. The fastest way to see the difference is the self-guided tour, which walks the GPU allocation view with no sales call attached.
The full spend pattern across the market is in CloudZero’s AI spend management guide, and the adoption numbers driving all of this demand live in the AI statistics roundup.
To see your own GPU spend allocated to cost per training run and cost per model, schedule a demo.
Frequently asked questions about H100 pricing
