$80M · AI + Sail Research

What they're building

Sail builds inference infrastructure specifically for AI agents that run for hours or days, not chatbots that answer in seconds. The core insight: existing inference platforms (including Together AI, where CEO Neil Movva worked) optimize for low latency because a human is waiting at the prompt. Agents don't wait. They care about throughput, reliability, and cost across thousands of concurrent calls. Sail deliberately sacrifices real-time responsiveness to pack far more work into the same GPUs, claiming up to 10x lower cost per token. The second product is Sailboxes: persistent sandboxed environments where agents run continuously in the cloud, billed only for time actually working. The API is OpenAI- and Anthropic-compatible and serves open-source models like DeepSeek, Kimi, and GLM, with LoRA fine-tune support and a Tinker integration. In a recent benchmark, Sail topped BrowseComp-Plus with 90.72% accuracy at up to 10x lower cost. Customer Detail.dev runs code-review agents on Sail that spend 3-4 hours combing entire codebases for bugs.

Why this matters

Getting both Sequoia (seed lead) and Kleiner Perkins (Series A lead) into the same company, plus angels like Alphabet chairman John Hennessy and Intel CEO Lip-Bu Tan, is about as strong an investor signal as exists at this stage. The $450M valuation for a company that launched its inference service in March tells you the bet size. The market timing argument: token prices have collapsed, yet enterprise AI bills have tripled because agentic workflows consume 50-500x more tokens than chat. Goldman Sachs forecasts a 24x increase in token consumption by 2030. Fortune reported Nebius recently paid $643M for the 20-person inference startup Eigen AI, so acquirers are hungry for exactly this talent. Movva's framing: 'Sail exists to make intelligence abundant.' The risk worth knowing: frontier labs building their own inference could commoditize this layer, and Together AI is a formidable (and fellow Kleiner-backed) incumbent.