How much does it cost to rent an H100 in 2026?

On-demand H100 rates span $1.38-1.49/hr on marketplace and boutique clouds, roughly $2-3/hr on major neoclouds, and $6-12+/hr on hyperscalers (AWS/Azure/GCP list up to $11.68-12.29/hr). The Silicon Data H100 Rental Index sat near $2.53/GPU-hour in mid-2026, and 1-year reserved contracts priced around $2.35/hr in March 2026.

Why did H100 prices fall from $8 per hour?

2023's ~$8/hr reflected extreme scarcity during the first LLM training rush. Massive neocloud capacity additions through 2024-2025, H200/Blackwell availability, and competition drove 1-year contract rates to a low near $1.70/hr in October 2025 before inference demand pushed prices back up.

Are H100 prices rising again?

Yes. One-year contract rates rose almost 40% from the $1.70/hr October 2025 low to $2.35/hr by March 2026, and the spot index jumped 10% in four weeks over the 2025-26 year-end ($2.00 to $2.20). Drivers: inference workload growth, Blackwell supply absorbed by hyperscalers, and enterprises extending H100 fleet life.

What is the difference between spot and reserved GPU pricing?

Spot/on-demand is pay-as-you-go with no commitment — highest unit price, instant elasticity. Reserved (1-36 months) discounts 30-60% versus on-demand in exchange for committed spend. On Compute Exchange auctions, reserved H100 listings ranged $1.07-10.14/hr with a $5.52 median, showing how wide the same-hardware spread remains.

How much does a B200 cost to rent compared to an H100?

B200 on-demand spans $3.70-6.00/hr on neoclouds (Spheron from $3.70, Lambda $4.99) and up to $14.24/hr on AWS p6 — published rates span $3.75 to $27 across all providers. That is roughly 2-3x H100 pricing for about 2.5-3x training throughput, so B200 economics per token are competitive when available.

Is it cheaper to buy or rent H100s?

New H100s cost $25,000-40,000; used units traded at $6,000-22,000 on secondary markets in mid-2026. At a $2.35/hr reserved rate, a rented GPU costs ~$20,600/year — near the hardware price — so ownership wins beyond ~18-24 months of high utilization if you can handle power, colocation, and resale risk. Below that, renting wins.

Where are H100 prices headed?

Futures markets now price this question: CME (with Silicon Data) and ICE (with Ornn) both announced GPU compute futures in 2026. Consensus mechanics point to firm near-term pricing — inference demand growing, H100 supply no longer expanding — followed by erosion as Blackwell/Rubin volume lands and expiring reservations flood the secondary market in late 2026.

H100 Rental Prices in 2026: History, Benchmarks, and Where They Go Next

H100 rental costs $1.40-2.60/hr on neoclouds and $6-12/hr on hyperscalers in 2026. Prices fell from $8/hr in 2023, bottomed in 2025, then rose 40%.

Renting an NVIDIA H100 in mid-2026 costs anywhere from $1.38/hr on GPU marketplaces to $12+/hr on hyperscaler on-demand — an 8x spread for identical silicon. The reference benchmark, the Silicon Data H100 Rental Index, sat near $2.53/GPU-hour, while 1-year reserved contracts priced around $2.35/hr in March 2026, up almost 40% from the October 2025 low of $1.70. The H100 has traveled from $8/hr scarcity pricing in 2023 to commodity status — and now back into a tightening cycle.

Key takeaways

Current benchmark: ~$2.53/GPU-hour (Silicon Data index, mid-2026); 1-year contracts ~$2.35/hr (March 2026).
Price history: ~$8/hr peak (2023) → steady decline through 2024 → $1-2/hr trough (2025, low of $1.70 in October) → +40% rebound into 2026.
The spread is the story: marketplace median $1.70/hr, neocloud median $2.31, hyperscaler median $6.11 — hyperscalers charge ~3.3x the non-hyperscaler median for the same GPU, per Compute Exchange auction data.
B200 comparison: $3.70-6.00/hr neocloud on-demand, up to $14.24/hr on AWS; full published range $3.75-27/hr.
Buy vs. rent crossover: used H100s at $6,000-22,000 vs. ~$20,600/year at reserved rental rates puts the ownership breakeven near 18-24 months of high utilization.
Futures have arrived: CME (Silicon Data index) and ICE (Ornn) GPU compute futures make the forward curve public for the first time.

GPU pricing pairs with facility economics: AI-ready colocation rates are in our colocation index and SEA data center catalog; live GPU comparisons in /gpu/.

H100 price history: 2023-2026

Period	1-yr contract / benchmark rate	Market condition
Mid-2023	~$8.00/hr	Peak scarcity; allocation by relationship
2024	$3-5/hr, declining	Neocloud capacity wave lands
H1 2025	$2-3/hr	H200/Blackwell arrive; competition compresses margins
Oct 2025	$1.70/hr (low)	Supply overshoot; fears of H100 obsolescence
Dec 2025 → Jan 2026	$2.00 → $2.20 spot (+10% in 4 weeks)	Inference demand inflection
Mar 2026	$2.35/hr 1-yr contracts (+40% off low)	Tightening; Blackwell absorbed upstream
Mid-2026	~$2.53/hr index	Firm, volatile

History per Silicon Data, SemiAnalysis’s 1-year rental index, and the January 2026 spike analysis.

The 2025 trough taught the market two things. First, GPU rental is brutally competitive at the commodity end — anyone can rack H100s, and margins collapsed accordingly. Second, demand is not only training: inference growth re-priced “obsolete” hardware. An H100 that lost its training-frontier status remained the workhorse for serving models, and at $1.70/hr it was suddenly the cheapest FLOPS in the market. Buyers noticed; the rebound followed.

On-demand pricing by provider type, mid-2026

Provider type	H100 on-demand ($/GPU-hr)	Median (Compute Exchange data)
Marketplaces / boutique clouds	$1.38-1.49 low end	$1.70
Neoclouds (Lambda, Nebius, RunPod, Spheron, etc.)	$1.90-3.50	$2.31
Hyperscalers (AWS, Azure, GCP) on-demand	$6-12+ (up to $11.68-12.29)	$6.11
Reserved listings, all types	$1.07-10.14	$5.52

Sources: IntuitionLabs GPU pricing index, Compute Exchange, AIMultiple.

Why the hyperscaler premium survives: enterprise procurement inertia, egress and data-gravity lock-in, compliance certifications, and bundled credits. For workloads without those constraints, the same GPU costs 60-75% less one tier down. That arbitrage is the neocloud business model — and its compression is why neocloud margins depend on reserved commitments rather than spot.

Spot vs. reserved: how to buy

Spot/on-demand suits bursty experimentation and short fine-tunes. You pay the top of the curve for the right to walk away.
Reserved (1-12 months) discounts 30-60% off on-demand. The March 2026 auction data shows the catch: reserved listings spanned $1.07-10.14/hr — provider selection matters more than contract type.
Multi-year (24-36 months) locks today’s rate against the forward curve. With futures now trading, treasurers can separate the capacity decision from the price decision (see below).
Secondary/used hardware became a real channel in July 2026 when Compute Exchange opened its used-GPU marketplace: used H100s at $6,000-22,000 versus $25,000-40,000 new, with buyer requests ranging from hundreds to tens of thousands of units.

Rule of thumb: at $2.35/hr reserved, a year of one GPU is ~$20,600. If your utilization is durably above ~60% and your horizon exceeds 18-24 months, priced-down ownership (especially used) plus colocation — get a facility benchmark via /quote/ — undercuts renting. Below that, rent.

B200 comparison

Metric	H100 (SXM 80GB)	B200 (192GB)
Neocloud on-demand	$1.90-3.50/hr	$3.70-6.00/hr
Hyperscaler on-demand	up to $11.68-12.29/hr	up to $14.24/hr (AWS p6); published range to $27
8-GPU node, neocloud	~$15-25/hr	~$29.60/hr (Spheron)
8-GPU node, AWS	—	~~$113.93/hr (~~$82,030/month)
Training throughput vs H100	1x	~2.5-3x

Per Spheron’s B200 pricing survey and Silicon Data’s B200 index, which recorded the B200 moving from ~$4.40/hr to spikes above $6.00/hr within a single month in early 2026 — Blackwell remains supply-constrained and correspondingly volatile (spot up 48% between February and April 2026). On per-token economics, a $4.50/hr B200 at ~2.7x throughput beats a $2.35/hr H100; the constraint is getting allocation at all.

Cluster math: what an H100 deployment actually costs

Unit rates obscure deployment-scale economics. Three ways to run 64 H100s (8 nodes) for a year at mid-2026 prices:

Sourcing model	Unit assumption	Annual cost, 64 GPUs
Hyperscaler on-demand	$10/GPU-hr average	~$5.6M
Neocloud 1-yr reserved	$2.35/GPU-hr	~$1.32M
Own used hardware + colocation	$14k/GPU used + ~$196/kW/mo colo + power	~$0.9M capex + ~$0.35M/yr opex

The ownership line assumes used H100s at the mid of the $6,000-22,000 secondary range, a ~45 kW/node draw with facility overhead, and wholesale colocation at index rates (current $/kW here). The spread between the first and third rows — roughly 6x on an annualized basis — is why sophisticated buyers treat GPU sourcing as a portfolio decision: hyperscaler on-demand for burst, reserved neocloud for base load, owned hardware for permanent inference fleets. It is also why hyperscaler list prices keep losing share of benchmark indices even as absolute hyperscaler GPU revenue grows.

Two frictions keep the arbitrage from closing. Enterprise data gravity: moving training data out of a hyperscaler incurs egress fees that can erase a year of rental savings on large corpora. And operational capability: owned fleets need failure management, InfiniBand/NVLink fabric expertise, and a facility partner — the practical hurdle our quote service exists to price.

Where prices go next

Four forces, two in each direction:

Upward: (1) inference demand compounding — every deployed model is a permanent load; (2) H100 supply is now fixed — production has shifted to Blackwell/Rubin, so the H100 fleet only shrinks from here.

Downward: (3) Blackwell volume ramps through 2026-27, and each B200 displaces ~2.5-3 H100-equivalents of demand; (4) expiring 2023-24 reservations and fleet upgrades are pushing A100s/H100s onto the secondary market in size in late 2026.

The honest answer is that the forward curve is no longer a matter of opinion: CME Group (referencing the Silicon Data H100 index) and ICE with Ornn are both launching cash-settled GPU compute futures, covering H100, H200, B200 and beyond. Once liquid, those contracts — not analyst forecasts — will be the H100 price outlook. Our working base case: H100 rates hold the $2-3/hr band through 2026, then grind down toward power-plus-margin economics ($1-1.50/hr) as Blackwell saturates in 2027-28. Track the live numbers on our GPU price page and stats dashboard.