Sail Research, a San Francisco startup building inference infrastructure purpose-built for long-horizon AI agents, emerged from stealth on Wednesday with $80 million in combined seed and Series A funding at a $450 million valuation, The SaaS News reports.
The Series A was led by Kleiner Perkins, and the seed by Sequoia, with participation from Redpoint Ventures, Theory Ventures, CRV, A\*, Abstract Ventures, and angels including former Stanford president John Hennessy, Cadence CEO Lip-Bu Tan, and FlashAttention creator Tri Dao. Fortune first reported the company's launch, noting that Sail is betting the explosion of AI agents will force enterprises to rethink the economics of computing.
Founded in 2026 by Neil Movva and Samir Menon, Sail provides a high-throughput inference stack and a sandbox environment called Sailboxes that allows autonomous agents to run for extended durations. Unlike chat-focused inference — where latency is measured in milliseconds — long-horizon agents can consume millions of tokens over hours or days, creating fundamentally different infrastructure requirements that existing cloud inference stacks were never designed to handle.
Analysis
The second-order effect of long-horizon agents is a restructuring of the cloud compute market itself. If agents become the dominant AI workload — running continuously, spawning sub-tasks, consuming tokens around the clock — the pricing model shifts from per-request to per-agent-hour, and the infrastructure requirements flip from low-latency chat serving to stateful, high-throughput batch processing. That's not a configuration tweak; it's a different architecture. Sail's bet is that this shift creates a wedge large enough for a new infrastructure layer between the hyperscalers and the application layer. The risk is that AWS, Azure, and GCP see the same shift and build agent-optimized inference into their existing stacks before Sail achieves escape velocity. But the investor roster — Sequoia, Kleiner Perkins, and the architect of FlashAttention on the cap table — suggests this isn't a naive bet. The presence of Tri Dao alone signals that Sail is thinking about the problem at the kernel level, not just as a packaging exercise on top of existing GPU infrastructure.
