Fireworks AI

Fireworks AI is a San Mateo inference platform founded in 2022 by a group of former Meta PyTorch engineers led by CEO Lin Qiao, who ran the PyTorch team before spinning out. Rather than building a proprietary foundation model or standing up its own datacenter campuses, Fireworks positions itself as a serving and fine-tuning layer for open-weight models: developers point at models like Llama, DeepSeek, Qwen, and Mixtral and get low-latency inference, LoRA fine-tuning, and reserved GPU capacity behind a single API. Jensen Huang has publicly called the company "the TSMC of AI factories," a framing Fireworks itself now uses in marketing.

The company is inference-focused rather than a training-hyperscaler like CoreWeave or Crusoe, and it sits closer to the software layer than to the metal. Its published on-demand GPU rates run $7 to $8 per hour for H100 and H200 (rising in September 2026), $10 to $13 per hour for B200, and up to $20 per hour for GB300, with region-restricted deployments carrying a 1.5x premium. Named customers include Cursor, Vercel, Notion, Sourcegraph, and UiPath, most of them agentic-coding or productivity products where token latency is the binding constraint. Fireworks does not publicly disclose GPU counts or MW footprint; capacity appears to sit on Nvidia silicon rented and reserved from colocation partners rather than owned campuses.

Investor demand has run well ahead of disclosed operating detail. Sequoia led a $52 million Series B in July 2024; Lightspeed and Index led a $250 million Series C in October 2025 at a $4 billion valuation; and in July 2026 Atreides Management and Index led a $1.51 billion Series D at a $17.5 billion valuation, roughly a 4x mark-up in nine months and one of the largest inference-layer rounds on record.