The AI Compute Market

AI compute is the infrastructure that runs machine learning training and inference. In 2026, it is the fastest-growing segment of the global technology infrastructure market, with the GPU rental market alone valued at approximately $52 billion and projected to reach $198 billion by 2031. Global AI capex crossed $765 billion in 2026, surpassing oil and gas investment for the first time.

Market Size and Growth

Market Structure

The AI compute market has three layers of participants:

GPU Manufacturers

NVIDIA dominates AI training GPU supply with the H100 (Hopper), H200, B200, and GB200 (Blackwell generation). AMD competes with the MI300X and MI350. Intel offers Gaudi accelerators for some inference workloads. Google manufactures TPUs for internal and limited external use.

GPU supply constraints were the defining story of 2023 and 2024. By 2026, Blackwell supply has expanded substantially, creating downward pressure on H100 spot prices while B200 pricing settles in the $4.00 to $6.50 per GPU per hour range.

Cloud Providers

Two types of cloud provider serve the GPU compute market:

• Hyperscalers. AWS, Azure, Google Cloud, and Oracle OCI offer GPU compute as part of broad cloud platforms. They charge a premium for bundled services: networking, storage, compliance, enterprise support, and integration with existing enterprise tools. Hyperscalers are the preferred option for enterprises that need GPU compute tightly integrated with existing cloud architecture.
• Neoclouds. CoreWeave, Lambda Labs, Crusoe, Nebius, and Voltage Park focus exclusively or primarily on GPU compute. Lower overhead translates to lower per-GPU prices. CoreWeave is the largest neocloud by capacity, with billions of dollars of private credit financing backing its H100 and B200 buildout.
• GPU Marketplaces. RunPod, Vast.ai, TensorDock, and FluidStack aggregate supply from independent GPU owners, creating a spot-market layer with the lowest prices and most variable reliability.

AI Buyers

Demand comes from AI labs (OpenAI, Anthropic, xAI, DeepMind), hyperscalers building internal AI infrastructure, enterprises running AI applications, and startups. The mix of buyer types affects market dynamics: lab demand is lumpy and project-driven; enterprise demand is more continuous and price-sensitive.

Pricing Dynamics

GPU rental prices move based on several factors. Supply is the dominant driver: when NVIDIA ships more GPUs, prices fall. The Blackwell ramp drove H100 on-demand prices from above $4.00 per GPU per hour in 2023 to below $3.00 by mid-2026.

The market is bifurcated by commitment structure. Spot and on-demand prices are observable and volatile. Reserved and committed capacity prices are negotiated bilaterally, rarely disclosed, and consistently 20 to 40 percent below spot equivalents.

Regional variation is significant. Data center power costs, grid connectivity, and local demand patterns create price differences across geographies. Texas and Virginia data centers compete on price differently than Singapore or Europe.

The Financialization of Compute

The defining development of 2026 is the formal financialization of GPU compute: CME Group and ICE are launching standardized futures contracts tied to GPU rental price indexes. This transition, from infrastructure expense to tradable financial instrument, mirrors the path followed by oil in the 1980s and natural gas in the 1990s.

The GPU rental market now has the preconditions that typically produce a derivatives market: sufficient scale, observable spot prices, standardized products (GPU models), and multiple buyers and sellers with different risk exposures. Futures allow AI companies to hedge compute budget risk, providers to hedge revenue, and financial participants to take directional positions on GPU price trends.

What Comes Next

The AI compute market is still in early innings. Several developments will shape the next three to five years:

• Rubin architecture GPU launch by NVIDIA, likely 2027, creating the next generational displacement cycle
• Power infrastructure as the binding constraint: grid interconnect delays and transformer lead times are limiting new data center energization
• Private credit market for GPU-backed financing maturing, with more standardized loan structures and secondary market for GPU assets
• AI inference demand growth partially offsetting training demand as models become production-ready
• ETF market expansion as Roundhill, Defiance, VanEck, and others launch compute-exposed funds