On April 10, 2026, CoreWeave signed a multi-year capacity deal with Anthropic — months after parallel commitments with OpenAI and Meta. The same week, Nscale crossed a $14.6B valuation after $3.1B raised across three rounds in eighteen months, with Nvidia, Lenovo, Dell, Citadel, and Jane Street on the cap table. AI infrastructure stopped being a software thesis and became a real-estate-power-and-silicon thesis, and the 68 companies on this page sit along that stack.
The compute-pact era
The top of the table is defined by who has signed multi-year offtake agreements with frontier labs. CoreWeave at $21B raised has Anthropic, OpenAI, and Meta. Lambda ($1.98B Series E, November 2025) and Crusoe ($1.38B Series E, October 2025) are racing to lock in similar contracts on the GPU-cloud tier. Cerebras ($2.45B Series G) and Groq ($2.33B Series E, September 2025) compete on inference silicon — Cerebras with wafer-scale, Groq with LPUs that claim sub-millisecond token latency. Etched is staking the same ground with Sohu, an ASIC built only for transformer inference. NVIDIA is investing across the layer it sells GPUs into.
Layers underneath the GPU
- Inference runtimes. Modal ($111M Series B), Baseten ($60M Series B), Anyscale ($260M Series C, commercial Ray), Fireworks ($285M Series C), and Together AI ($267.5M Series C) compete on cold-start latency and per-token cost.
- Vector and retrieval. Qdrant ($85M Series B, Rust-built in Germany) and LanceDB ($41.5M Series A) anchor the pure-play substrate. Pinecone ($228M Series C) and Weaviate ($168M Series C) round it out — Postgres, Elasticsearch, and Snowflake all ship native vector now.
- Registries and platforms. Hugging Face ($470M Series D) is the community layer. Databricks ($5.6B through a Series L closed December 2025) is the lakehouse incumbent that has absorbed the AI-platform mandate.
- Sovereign and frontier outliers. Ineffable Intelligence raised a $1.1B seed in April 2026 in the UK targeting reinforcement-learning-driven superintelligence. HUMAIN in Saudi Arabia, Krutrim Cloud, and Sarvam AI in India are building national stacks tied to government compute commitments.
What 2026 squeezes out
Three forces are doing the squeezing. AWS, GCP, and Azure keep collapsing inference, vector, and agent runtimes into managed services — hitting horizontal startups hardest. The inference-cost curve keeps dropping, so margin lives with whoever owns the lowest-latency, highest-utilization minutes. And vertical demand — sovereign AI in the UAE and Saudi Arabia, defense buyers in the US — is rewarding integrated stacks (G42, HUMAIN) over horizontal plays. The startups holding ground in 2026 are the ones with measurable performance advantages or vertical lock-in.