GPU Cloud & NeoCloud — Market

Verified claims and product-axis read for GPU Cloud & NeoCloud. Every fact below is sourced; every product judgment traces back to underlying signals.

Verified facts

• CoreWeave's customer count as of S-1 was disclosed at ~30 customers with top 2 accounting for >75% revenue. ↗ (financial)
• Crusoe's ~$3.4B Stargate-related debt facility was led by Blue Owl and JPMorgan in 2025. ↗ (financial)
• Lambda's 'Stack' OSS toolkit for ML environments has >50K GitHub stars driving developer-led pipeline. ↗ (other)
• Together AI is the maintainer of the OSS RedPajama dataset and FlashAttention research releases. ↗ (other)
• Groq Cloud offers Llama 3.1 70B inference at ~$0.59 input / $0.79 output per million tokens. ↗ (financial)
• Cloudflare disclosed AI inference revenue is not yet broken out as a segment in 10-K filings. ↗ (financial)
• Nebius reported it deployed NVIDIA HGX H200 platforms in production before peer European neoclouds. ↗ (other)
• Applied Digital signed an additional ~$5B 15-year lease extension with CoreWeave during 2024. ↗ (financial)
• CoreWeave's reported gross margin in early 2025 was ~70-75% on cloud-services revenue per S-1. ↗ (financial)
• Fireworks AI offers fine-tuning at ~$0.50/M training tokens on Llama 3 8B, undercutting OpenAI fine-tune pricing. ↗ (financial)

Top products (engine read)

Distributed AI Training Reference Architecture — multi-node GPU training stack / reference architecture

Opportunity: Enterprises buying H100/B200 fleets discover that fault-tolerance, checkpointing, and topology-aware scheduling — not raw FLOPS — gate time-to-train; CoreWeave's content + Anyscale/Ray hiring signal a productized 'training reliability' layer.

Mid-tier — clearly hot as a content/POV play but the actual productization is fragmented across CoreWeave clusters, Anyscale Ray, and in-house customer code. Opportunity for a turnkey 'training control plane' SKU sits open.

Hyperscaler-Dedicated GPU Capacity Contracts — multi-year reserved GPU compute contract / take-or-pay capacity

Opportunity: Hyperscaler Azure/OpenAI capacity shortfalls (verified: Microsoft → CoreWeave ~$10B backfill, OpenAI → CoreWeave $11.9B) created take-or-pay demand that neoclouds with siting+power can monetize on debt-financed builds. Nebius's $17.4B Microsoft deal is the structural template.

This is the product line that defines the topic. Nebius and CoreWeave are the only public proof points; market is asking who's third (1ecdd2c9). Whoever signs the next-named contract gets re-rated.

Data Center Networking (800VDC / 800G fabric for AI racks) — AI-rack power & networking infrastructure

Opportunity: $4.55B per GW DC-networking TAM emerging as Blackwell-era 600kW racks force 800VDC retrofit; Vultr's broad multi-region DC technician push is the staffing footprint of a network/power buildout, not a SaaS launch.

Adjacent product — not a neocloud SKU directly but the substrate the neocloud SKUs ride on. Vultr is the only neocloud in evidence with broad DC-tech hiring across 4 metros; signals continued physical expansion.

Inference-as-a-Service Platform — managed/dedicated inference endpoint

Opportunity: Production AI workloads are shifting from training spend to inference spend; verified facts show CoreWeave claims fastest inference on Kimi K2.6 and Together AI shipped 'Together Dedicated Endpoints' for reserved-throughput pricing — neoclouds are racing to own the reliability/latency layer that hyperscalers under-serve.

The single most concentrated product theme in the evidence — CoreWeave has six inference-themed launches and Together AI hires three inference engineers in this snapshot. Whoever wins per-token economics on open models (Kimi/Llama/Qwen) takes the next leg of neocloud margin.

See the Products and Strategy modules for the full product list and forward-looking judgment.