Published August 21, 2026 · Updated August 21, 2026
Compute Prices
Compute prices - what it costs to run GPU workloads in the cloud - are not published on a central exchange, do not follow a standard format, and change without notice. This page explains how AI compute is priced, what the price ranges look like in 2026, and what drives price differences across providers and contract types.
How AI Compute Is Priced
GPU cloud providers offer compute under several contract structures, each with different price levels and risk characteristics:
On-Demand (Spot/Hourly)
Pay per GPU per hour with no commitment. Highest per-hour price. No cancellation risk. Standard for development, experimentation, and short-run inference. On-demand prices are the basis for most public GPU price indexes and compute futures settlement.
Reserved Capacity (1–3 Year Commitments)
Pre-purchase GPU capacity for a fixed term at a discounted rate. Typical discount: 20–40% below on-demand. Common for hyperscalers (AWS Reserved Instances, Azure Reserved VMs). Risk: you pay whether or not you use the capacity.
Committed Capacity / Take-or-Pay
Common in neocloud contracts. The AI company commits to a minimum monthly GPU spend in exchange for capacity guarantees and volume pricing. CoreWeave, Lambda, and Crusoe all offer committed capacity structures. Prices are negotiated bilaterally and rarely published.
Spot / Preemptible
Available from some providers at significant discounts (30–70% off on-demand) in exchange for preemption risk - the provider can reclaim capacity with short notice. RunPod, Vast.ai, and some hyperscaler spot markets offer this. Not suitable for long training runs without checkpoint infrastructure.
2026 GPU Price Ranges (On-Demand)
The following price ranges reflect on-demand GPU rental costs as of mid-2026, sourced from public provider pricing pages. Prices vary by region, availability zone, and interconnect configuration. ComputeWatcher does not publish real-time prices on this page - subscribe to the weekly digest for price movement summaries.
| GPU Model | Architecture | Est. On-Demand Range ($/GPU/hr) | Notes |
|---|---|---|---|
| NVIDIA H100 SXM | Hopper | $2.00–$3.50 | Wide range; hyperscalers higher, neoclouds lower |
| NVIDIA B200 | Blackwell | $4.00–$6.50 | Ramping supply; pricing settling |
| NVIDIA H200 | Hopper+ | $3.00–$4.50 | Higher memory bandwidth premium |
| NVIDIA A100 SXM | Ampere | $1.50–$2.50 | Declining as H100 displaces |
| AMD MI300X | CDNA3 | $2.00–$3.50 | CoreWeave, Oracle OCI availability |
Note: Prices are illustrative ranges based on publicly available data. Actual prices depend on provider, region, commitment, and availability. Unknown negotiated or committed rates are excluded.
Hyperscaler vs. Neocloud Pricing
A consistent pattern in compute pricing is the hyperscaler premium: AWS, Azure, and Google Cloud charge materially more for GPU compute than neoclouds like CoreWeave, Lambda, Crusoe, and Nebius. The premium reflects the hyperscaler bundle - integrated networking, storage, security, compliance, and support - that enterprises pay for.
The Silicon Data index tracks this split explicitly. The neocloud index typically runs 15–30% below the hyperscaler equivalent for the same GPU model in equivalent configurations.
What Drives Price Changes
- GPU supply. When NVIDIA ships more H100s or B200s, spot prices fall. When supply is constrained, prices spike. The Blackwell ramp in 2025–2026 drove significant H100 price softening.
- AI training demand cycles. Prices spike before major model release deadlines and fall during consolidation periods.
- Competition among providers. New entrants with fresh hardware, like Nebius and Crusoe, price aggressively to acquire customers.
- Power and infrastructure costs. Data center energy costs are rising, putting a floor under compute prices in constrained power markets.
- Generational transitions. H100 prices fell when B200 arrived. A100 prices fell when H100 arrived. The pattern repeats.
Live Compute Price Tracker - Coming Soon
ComputeWatcher is building a daily compute price tracker covering H100, B200, H200, A100, and MI300X across 15+ providers. Subscribe to the weekly digest to be notified at launch.
Get the weekly compute market digest
GPU prices, futures updates, ETF filings, financing deals - every Monday.
No sponsors. Unsubscribe anytime.