Partnering with Sail: Agents Without Limits
Sequoia Lauren Reeder and Konstantine Buhler
Sequoia says Sail is building the infra for AI agents that can run for hours or days. The pitch: make long-running agents cheap enough to be normal.
Based on reporting by Sequoia, Lauren Reeder and Konstantine Buhler — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
For years, “using AI” mostly meant a person typing a prompt and waiting. Sail is betting the next phase looks very different: software that keeps working on a problem for hours, or days, without tapping out.
That’s the gap it wants to fill. Sequoia says the demand for agents is already here, but the infrastructure to run them at serious scale is not. Sail’s answer is an inference platform built to handle long-running tasks across a wide range of compute.
The company’s core trick is a harder, thriftier inference stack. Its line is simple: there are no bad chips, only bad prices. Instead of relying on only the newest or most desirable hardware, Sail spreads work across providers, uses underused data-center capacity, and tries to find a job for new accelerators as soon as they appear. The point is to squeeze more out of every chip, and to do it cheaply enough that an agent can chew through billions of tokens on one task without the bill going wild.
Then there’s Sailboxes, the piece meant for agents that don’t want to clock out. They scale CPU and memory as needed, and when an agent is idle because it’s waiting on an event or a tool call, the bill drops close to nothing. Customers can leave them running for months. Sequoia says they’re basically immortal.
The people behind it fit the pitch. Neil Movva worked on GPUs at NVIDIA, then on AI accelerators at Apple, then returned to GPUs at Together AI. Samir Menon also came from Apple, where he built highly reliable systems at billion-user scale. Sequoia says the team is expanding quickly, and that the whole operation is aimed at one thing: making intelligence abundant by removing the usual limits on tokens, scale, and runtime.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI that actually matters: not another chatbot demo, but the plumbing that lets agents run long enough to do real work. The industry loves shiny model releases; the boring infrastructure layer is where the useful money usually hides. Also, “there are no bad chips, only bad prices” is exactly the kind of line that sounds ridiculous until it starts making procurement people nervous.
Read more about this at: Sequoia