GPU capacity in 2026: contracted, scarce and repricing

For two years the market expected GPU prices to fall as supply caught up. In 2026 the opposite happened. Demand for training and, increasingly, inference has absorbed new supply, hyperscalers are buying factory output years ahead, and the price of capacity has risen. For investors in neoclouds, GPU-backed lending or data centres hosting AI tenants, this changes the shape of the opportunity.

Prices are rising, not falling

Runpod reported that H100 rental contract pricing rose from around $1.70 per GPU-hour in October 2025 to $2.35 in March 2026. That is a striking move for a GPU generation that is no longer the newest, and it reflects how much of the newer Blackwell supply is already committed under long-term contracts.

Hyperscalers are reported to be buying out factory commitments three years or more ahead. Uncommitted capacity, particularly of current-generation systems, is becoming harder to find.

Contracts have lengthened

The clearest signal is in the contracts. In April 2026, Axe Compute signed a $260m, 36-month agreement for a dedicated cluster of 2,304 NVIDIA B300 GPUs in a US Tier III facility. Working through the arithmetic:

1. 1.$260m ÷ 36 months = about $7.2m a month.
2. 2.$7.2m ÷ 2,304 GPUs = about $3,135 per GPU per month.
3. 3.At about 730 hours a month, that is roughly $4.30 per GPU-hour, contracted for three years.

At the scale end, CoreWeave reports about 1.5 GW of capacity live and 4.2 GW contracted. Multi-year, take-or-pay structures are what make GPU fleets financeable.

Costs are rising too

Hardware costs are moving up alongside prices. Server DRAM rose by around 90% in the first quarter of 2026, according to TrendForce, as AI servers absorb memory supply. High-bandwidth memory remains tightly allocated. Power is the other cost: a 1,000-GPU GB300 cluster needs about 2.6 MW of facility power, and the next generation is expected to need closer to 5 MW per 1,000 GPUs.

Where the risk sits

GPU-backed assets carry risks that conventional data-centre investors are less used to:

• Utilisation: revenue depends on the fleet being sold, not only installed.
• Technology cycle: each new generation reprices the previous one; contract tenor should match a realistic economic life.
• Counterparty: a strong three-year contract is worth far more than higher spot pricing.
• Delivery and power: GPUs that arrive before the power and cooling are ready earn nothing while financing costs run.
• Export controls: end-user and destination checks are mandatory, and rules can change. The US moved the UAE into its unrestricted tier in July 2026, opening Gulf deployments that were previously constrained.

What this means for investors

• Value contracted revenue over spot pricing: tenor, take-or-pay terms and counterparty credit drive financeability.
• Match GPU delivery to power and cooling readiness; idle hardware is the most expensive failure in the model.
• Stress-test residual value at the end of the contract against the next hardware generation.
• Confirm export-control compliance and end-user verification for every deployment.
• Look at the host facility: density, liquid cooling and power headroom determine which hardware it can run.
• Nodus secures GPU allocation for neoclouds and enterprise AI programmes and screens compute-backed opportunities for investors.