GPU Infrastructure Market Timeline

OpenAI's ChatGPT release (November 2022) triggered unprecedented GPU demand. Training large language models and serving inference at scale requires thousands of GPUs. Demand exceeded NVIDIA forecasts by 3-5x.

Public cloud pricing remained stable ($1-2/hour for H100), but scarcity created value. Being able to buy any GPU at all became a competitive advantage. Small operators with supplier relationships charged premiums ($2.50-3.00/hour for scarce capacity).

This phase created opportunity for neocloud founders (many started in 2022-2023 solving their own bottleneck). It also drove irrational FOMO. Everyone raised money to build capacity despite uncertain demand durability.

NVIDIA couldn't ship enough H100s and A100s to meet demand. Lead times stretched to 6+ months. Operators with allocation agreements paid spot prices 40-60% above list.

This created extraordinary margins: buying H100s at cost ($10-12k wholesale) and selling 4-5 month leases for $4-5k monthly revenue. Neocloud startups exploded in funding: CoreWeave, Lambda, Crusoe, Modal, Replicate all raised Series A/B at $1B+ valuations. Hyperscalers faced inventory pressure and started raising prices and tightening SLAs.

The market felt infinite. Operators built capacity with 2-year payback assumptions. Irrational exuberance peaked in Q4 2023.

CoreWeave filed for IPO (eventually withdrawn, relisted 2025) signalling neocloud maturity and venture confidence. However, 2024 exposed the reality beneath hype. Demand growth slowed.

New model launches consumed existing capacity rather than demanding new clusters. Hyperscaler supply normalised (AWS/Azure/GCP shipped planned capacity). Spot market pricing fell 30-40% as supply exceeded peak demand.

Operators who built capacity in 2023 found utilisation falling to 40-60% in H1 2024. Gross margins compressed: $4k/month per H100 fell to $2.5-3k as pricing normalised. This killed dozens of marginal operators. Winners (CoreWeave, Lambda) survived because they had diversified customer bases, strong unit economics, and venture runway.

NVIDIA's B200 (current-gen, up to 1,000W, massive yield improvements) shipments began scaling in Q1 2025. B200 is 2-4x performance/watt versus H100.

Operators face upgrade pressure: deploy B200 or risk margin compression as customers migrate. This requires $800M-1.2B capex for scaled operators; only achievable with revenue stability or capital access.

This is consolidating the market. Well-capitalised operators (hyperscalers, large neoclouds) deploy B200. Marginal operators cannot afford the upgrade and gradually fade. Pricing for B200 capacity is 15-25% lower than H100 despite superior performance, reflecting structural margin compression.

By late 2025/early 2026, the market consolidates into clear winners and losers. Hyperscalers maintain 70% capacity share. CoreWeave and Lambda are likely to be the dominant neoclouds.

Crusoe survives through carbon/location advantage. Smaller operators either become regional specialists, get acquired, or fail. Pricing pressure will continue: B200 pricing could fall to $0.40-0.60/hour for basic inference within 24 months (from current $0.75-1.00).

This compresses margins to 15-25% for undifferentiated capacity. Winners will be those with operational scale (low opex), customer lock-in (software, integration, SLAs), capital access for next-gen upgrades, or true differentiation (carbon, specialisation, location). The 'easy money' phase (2022-2024) is over. The next 24 months reward operational excellence and customer intimacy, not hype and capital.

2022-2023 saw infinite demand and supply constraints; every operator could raise money and deploy capacity profitably

2024 normalized supply and slowed demand growth, killing marginal operators and exposing bad capital allocation

Blackwell deployment (2025) triggered consolidation: only well-capitalized operators can afford next-gen upgrades

CoreWeave's IPO attempt and Lambda's maturation mark transition from startup phase to infrastructure maturity

Next 24 months: pricing compression, operator consolidation, competitive advantage shifting from capital to operations and customer lock-in