GPU Cloud Market Structure

AWS, Azure, and GCP own roughly 70% of public GPU cloud capacity. They offer breadth (every GPU generation, every region, ML service integration), scale (massive inventory, global redundancy), and operational excellence. Their margins are opaque but likely 20-30% gross after amortising capex and opex.

Customers choose hyperscalers for reliability, SLA credibility, and ecosystem lock-in (native Sagemaker, Vertex AI integration). Their moat is capital, relationships, and operational maturity, not sustainable technological advantage. Neocloud competitors cannot beat them on price and reliability simultaneously.

Scaled neoclouds (>$100M funding, >2,000 GPUs deployed) occupy the second tier. They succeed through specialisation, geographic density, or price advantage.

CoreWeave dominates bare-metal enterprise training (strong enterprise sales, long contracts, regional capacity). Lambda dominates developer-focused managed inference (free tier, deep PyTorch/TensorFlow integration, low-friction billing).

Crusoe Energy differentiates on stranded power and carbon reduction (cheaper capacity in geographies with excess power). Each has 10-20% estimated market share by revenue. Their margins are higher than hyperscalers (35-50% gross) because they're lean, but reach is narrower. These operators win by choosing a customer segment and serving it better than hyperscalers, not by competing head-to-head.

Mid-market operators (1scale.com, Paperspace, Replicate) own 10-15% of market by revenue. They succeed through deep vertical integration (Paperspace's Gradient ML platform, Replicate's inference hosting) or regional advantage (cheaper operations in emerging markets, local customer support). Aggregators (Vast.ai, RunPod) operate peer-to-peer or spot markets, listing spare capacity from larger operators and independent miners.

Aggregators own 5-10% of revenue but are growing fast because they're capital-light and charge 20-30% commission. They don't build capacity; they arbitrage utilisation across others' infrastructure. This creates a 'tail of long-tail' dynamic where thousands of small operators farm out capacity through aggregators.

NVIDIA controls allocation, pricing, and upgrade cycles via hardware scarcity and driver support. During supply constraints (2023-2024), NVIDIA effectively dictated which operators could scale.

As supply normalises, NVIDIA's influence shifts to ecosystem (CUDA dominance, software stack optionality) and customer relationships. Consolidation is accelerating.

Large operators acquire or fold smaller competitors because opex scale matters more than differentiation as the market matures. CoreWeave is likely to acquire 2-3 regional players by 2026. The market will likely settle into 4-5 major operators (hyperscalers plus 1-2 scaled neoclouds) plus a long tail of specialists and aggregators.

Market divides into hyperscalers (70% capacity, 20-30% margins), scaled neoclouds (15-20% capacity, 35-50% margins), and tail operators

Scaled neoclouds succeed by specialization (bare-metal enterprise, developer tools, carbon reduction) not by competing head-to-head with hyperscalers

Aggregators are growing fast because they're capital-light; they arbitrage spare capacity and charge 20-30% commission

NVIDIA's role shifts from allocation (supply-constrained) to ecosystem (CUDA dominance) and upstream relationships

Consolidation will likely settle into 4-5 major operators plus a long tail; competitive advantage flows to operational scale and customer lock-in