Renting an NVIDIA H200 costs roughly 160% more than it did on January 1, 2026. That is not a projection or a broker's pitch; it is the year-to-date move tracked across rental data through mid-October. If your team reserved Hopper capacity last winter and let the contract lapse, you are now shopping in a market that repriced under you.
The good news: this is a procurement problem with known failure modes, and a buyer who does five things in the right order usually pays meaningfully less than one who clicks deploy on the first dashboard they see.
What the Rental Numbers Actually Say
Two datasets describe the same climb. Ornn Data's index, republished by a16z, puts the H200 up 143% since the start of the year, the B200 up 111%, the RTX PRO 6000 up 73% and the H100 up 54%, with the six-year-old A100 still managing a 9% gain. A separate October 24 tally has the H200 at roughly 160%, the B200 near 120%, the RTX PRO 6000 around 65% and the H100 near 35%.
Different windows, same shape. The H200 leads by a wide margin because of what it carries: 141 GB of HBM3e at 4.8 TB/s, or 1.76x the H100's memory. When people talk about AI token consumption pushing GPU prices up, the chip that gets bought first is the one that holds the most context per card.
The chronology matters as much as the totals. Most GPUs bottomed in the first quarter, with the H100 and B200 dropping about 23% below their January levels before rebounding in April and May. The H200 broke later, spiking in late May, giving back roughly a quarter of that move within weeks, then grinding upward into October.
What a GPU-hour costs right now
Median on-demand H200 rentals land between $4.27 and $4.94 per GPU-hour across tracked providers. Spot capacity can touch $1.47 and runs up to about $3.88. Hyperscalers sit at the other end, averaging $7.91 to more than $10.60 per hour. For context, at $5.40 an hour, one GPU bills about $3,942 a month, and an eight-GPU node about $31,536. Buying instead of renting is no shortcut either: an H200 runs $32,000 to $47,000 per GPU, or roughly $300,000 for an eight-GPU HGX node.
- Spot: $1.47 to $3.88 per GPU-hour, interruptible, best for training runs you can checkpoint
- Median on-demand: $4.27 to $4.94, the honest benchmark for 24/7 inference
- Hyperscaler on-demand: $7.91 to $10.60 and up, the price of compliance, regions and support
The Five-Step Buying Sequence
1. Price the workload, not the card
Before you ask a broker for a discount, compute your tokens per GPU-hour and your monthly GPU-hours. An inference service pinned at 60% utilization is really paying 1.67x the sticker rate for its useful compute, and that number, not the list rate, is what determines whether a reserved contract wins. Teams that skip this step negotiate on the wrong axis and end up with a multi-year commitment sized for a workload that has already moved to a cheaper model.
2. Benchmark against spot before you sign anything
Spot is the price of your indifference to interruption. If your pipeline checkpoints every few minutes and can tolerate eviction, a spot H200 at a fraction of the on-demand rate buys you more total compute than a marginally discounted reserved card. If it cannot, stop pretending: write the checkpointing requirement into the contract review, and price the engineering time it would take to get there.
3. Cap the escalation, not just the discount
Providers have already moved. Nebius raised its on-demand H200 from $4.50 to $5.40 per GPU-hour on October 1, a 20% step, and lifted H100 rates 17% in the same adjustment. A 15% discount off list is worthless if list moves 20% three weeks into the term. Ask for a fixed rate or a hard annual escalator ceiling, and get the renewal formula in writing.
4. Mix generations deliberately
The H100 rose roughly 35% year to date against the H200's 160%, and it was still the cheaper card per unit of memory bandwidth at the January baseline. For fine-tuning and medium-context serving, a mixed fleet of H100s for throughput plus a small H200 pool for long-context work often beats an all-H200 allocation. The A100 is not a joke either; at a 9% gain it is the most stable price in the lineup and still runs plenty of 70B-class workloads.
5. Decide whether to hedge, then stop
GPU-hours now have a derivatives market. CME listed GPU rental futures on October 5, and on October 26 Injective launched the SICI perpetual, an on-chain index tracking rental prices for five NVIDIA GPUs with up to 10x leverage. For a team spending six figures a year on compute, a modest futures position can blunt a rate spike. For everyone else, hedging is a distraction from step one, and leveraged exposure to a volatile index is a new risk rather than a solution to the old one.
The Short Version
- Know your real cost per useful GPU-hour at your actual utilization before any call
- Use spot for interruptible work and pay for reserved capacity only where you need 24/7 uptime
- Negotiate a fixed rate or an escalation cap, not a percentage discount off a moving list
- Run H100s where throughput is enough and reserve H200s for long context
- Re-price the whole fleet the week a provider announces a rate change
Buying at these prices is not about finding the cheapest number on a dashboard. It is about knowing which number you are actually paying, and making sure the one that moves next year is capped by a contract you can point to. An H200 at $10.60 an hour and the same card at $4.94 are running identical silicon; the difference is how much of the surge you agreed to absorb, and when you decided to notice it.
Comments