A founder argues slow overnight AI agents matter more than faster chatbots

Neil Movva, founder of Sale Research, a company that sells cheap access to open source AI models, was defending an unpopular bet when he said it. Asked why anyone would want slower AI when everyone waiting on a chatbot wants the fastest answer possible, he answered that the goal is to stop waiting altogether.

"The best latency is no latency at all," he said. Movva argues that AI should be proactive and work in the background, finishing overnight so the work is done when you wake up.

What he means is that AI should run on human time. You do not manage a colleague every five minutes, he said, you hand over a high level task and check back in a day or a week later. In that world, it does not matter if the model writes 10 tokens, or small chunks of words, per second instead of 100, if slower speed brings greater efficiency.

Movva argues that existing AI providers focused on low latency inference after they were "pulled in that direction by one" big customer, the coding assistant Cursor, and he called that the right choice a year ago. He believes the future is long horizon tasks where a system runs "for hours or days at a time."

That shift fits how the hardware works, he argues. A graphics processor, the chip that runs AI, is a throughput machine that is happiest with a lot of work to chew through at full use. Forcing it to spit out instant answers fights that happy path. It is, he said, a choice between "narrow and fast or wide and slow."

To explain it, he uses traffic. Instant answers are like a private car taking a direct route downtown. High throughput AI is like a bus, slower because it serves many riders at once by grouping their work together. His plan is to build the best possible bus.

The payoff, in his view, is scale without human limits. Live coding help is bounded by attention, but background work is not, and he predicts roughly 50-50 background versus real-time workloads by the end of this year, shifting eventually so background work makes up about nine in ten jobs. His examples are deep research across more than 10,000 sources, and a customer, Parallel Web Systems, that wants to index the whole internet and watch it change in real time.

That tolerance for slowness shapes his business. If a job fails and must move to another chip, one step might take minutes longer. He argues that delay barely registers when the whole task already stretches across hours overnight. He said average throughput would stay very competitive, but the slowest requests would not be controlled, and in return he would offer "unbeatable economics."