H100 and H200 Prices in 2026: Buy New, Refurbished or Rent?

Hopper GPUs did not get cheap when Blackwell arrived. In 2026, H100 and H200 prices have held up, used Hopper hardware resells close to new, and cloud rental rates rose in the first half of the year. Price trackers put a new H100 at roughly $25,000 to $40,000 and rental at about $3 to $7 per GPU-hour outside the most expensive hyperscalers. Whether to buy or rent comes down to utilization and time horizon. This guide gathers the reported figures as of September 2026, shows the break-even arithmetic, and explains where refurbished hardware fits. All prices below are reported street or list figures, not Nexus Compute prices.

How much does an H100 cost in 2026?

Reported prices vary by source, form factor and channel, so treat these as ranges rather than quotes:

• New H100 80GB: $25,000 to $40,000, with PCIe at the lower end and SXM5 at a premium, according to Compute Exchange in April 2026. CloudZero puts a single new H100 at about $31,000 as of August 2026.
• Refurbished H100: $21,000 to $34,000, roughly 80 to 85% of new, per Compute Exchange.
• Used H100: $15,000 to $28,000, roughly 60 to 70% of new, per Compute Exchange.
• 8-GPU HGX H100 server: CloudZero reports $250,000 to $320,000 new, while Compute Exchange says complete systems routinely exceed $350,000. The gap reflects differences in CPU, memory, storage and networking, and rising DDR5 and NVMe prices.

How much does an H200 cost?

The H200 has the same Hopper compute as the H100 but 141GB of HBM3E at 4.8 TB/s (NVIDIA), which matters most for memory-bound inference. Mercatus reported in May 2026 that H200 SXM GPUs sell for $32,000 to $40,000 and the PCIe H200 NVL for $28,000 to $34,000, with 8-GPU HGX H200 servers at $320,000 to $420,000. It also noted that the H200 secondary market barely exists yet, so refurbished H200s are rare. The H200 NVL is a dual-slot, air-cooled PCIe card rated up to 600W, which makes it easier to host than SXM systems. Our H200 server guide and H100 vs H200 comparison explain which workloads justify the extra memory.

How much does it cost to rent an H100 or H200?

A comparison updated October 1, 2026 by Thunder Compute, itself a GPU cloud, lists on-demand H100 prices from $2.99 per GPU-hour at Vultr and $3.29 at Lambda to $6.88 at AWS, $6.98 at Azure and $10.98 at Google Cloud. It also notes that H100 availability is extremely low across providers. AWS's rate reflects price cuts of up to 45% on P4 and P5 instances made in June 2025 (AWS).

The trend since then has run the other way. SemiAnalysis reported that H100 one-year rental contracts rose almost 40%, from a low of $1.70 per GPU-hour in October 2025 to $2.35 by March 2026, and that on-demand capacity was sold out across GPU types (SemiAnalysis). For H200, Mercatus reports on-demand rates of about $2.50 to $3.50 at specialist providers and $4.50 to $7.00 at hyperscalers.

Why have Hopper prices stayed high?

Three forces have kept H100 and H200 prices firm well into Blackwell's life. First, inference demand grew faster than supply. SemiAnalysis attributes the 2026 rental squeeze to rapid adoption of coding assistants, multi-agent workflows running at high concurrency, and image and video generation, and concludes that rental pricing is more likely to keep rising than to fall (SemiAnalysis). Second, Blackwell supply is itself allocated, so buyers who cannot get B200 or B300 in time fall back on Hopper. Third, the rest of the server has become more expensive: DDR5 memory and NVMe drives have risen sharply in 2026, which lifts the total for any 8-GPU system regardless of GPU price. Our posts on server RAM prices and why server prices went up cover those components.

When does buying an H100 server beat renting?

Here is illustrative arithmetic using the reported figures above. Treat it as a method, not a forecast, and substitute your own quotes.

• Renting eight H100s on demand at about $3.00 per GPU-hour costs $24 an hour, or about $210,000 a year if they run around the clock.
• On a one-year contract at SemiAnalysis's March 2026 figure of $2.35, the same eight GPUs cost about $165,000 a year at full use.
• A new 8-GPU HGX H100 server at around $300,000 therefore pays back in under two years against on-demand specialist rates at full utilization, before running costs.
• At 50% utilization, rented cost halves to about $105,000 a year, and payback stretches toward three years.
• Add running costs. The DGX H100 is rated at 10.2kW maximum (NVIDIA datasheet); at an assumed $0.12 per kWh that is up to about $10,700 a year in electricity before cooling, plus colocation or rack space, support contracts and staff time.

The pattern is consistent: ownership wins when GPUs stay busy for three years or more and you have somewhere suitable to put them. Renting wins for bursty work, short projects, experiments, and teams without power or cooling for 10kW servers. Our build vs buy vs cloud guide and 3-year TCO guide set out the full line items.

Is a refurbished or used H100 worth it?

The discount is smaller than many buyers expect. Compute Exchange estimates that once warranty gaps and failure rates are counted, refurbished H100s deliver only about a 9% effective cost advantage over 24 months, compared with a 15 to 20% sticker discount. Used GPUs also raise questions about provenance, prior workload, firmware and manufacturer warranty transfer. For production systems, that narrow saving rarely covers the risk. Our guides to new vs refurbished servers and grey market GPU risks cover what to check if you do consider it.

Should you buy Hopper at all in late 2026?

Sometimes. H100 and H200 remain sound choices for inference, fine-tuning and FP64-heavy work, and they run in air-cooled servers at lower power than Blackwell. But at these prices, compare against B200 on cost per unit of work, since Blackwell often delivers more throughput per dollar on new projects. For smaller inference loads, PCIe cards such as the H200 NVL or RTX PRO 6000 Blackwell Server Edition can avoid an HGX purchase altogether.

How do you get a reliable H100 or H200 quote?

• Specify the exact form factor: SXM in HGX, PCIe, or NVL with NVLink bridges.
• Ask for the full system bill of materials, because memory and storage now move the total significantly.
• Confirm the warranty is the manufacturer's and is registered to your organization.
• Check quote validity; many suppliers now hold prices for days rather than weeks.
• Get rental and reserved quotes for the same capacity so the comparison is like for like.

Nexus Compute supplies new H100 and H200 GPUs and servers through authorized distribution, with the full manufacturer warranty, and checks compatibility against your platform before dispatch. Send us your requirements and we will quote within 48 business hours, with purchase orders and net terms for qualified organizations. If renting is the better answer for your workload, we will tell you.