Neil Movva's Sail is betting that background agents will shift inference demand from roughly fifty-fifty with real-time work toward ninety-ten in favor of background workloads, making cost per token the metric that matters.
Movva says Sail will buy cheap, interruptible capacity like small one-megawatt data centers at 95 percent uptime, because background agents tolerate interruptions and Sail does not sell low latency.
Movva is contrarian on NVIDIA, arguing performance per watt has barely improved from Hopper to Blackwell to Rubin, and that losing TSMC would hurt but not be catastrophic.
Sail bets background AI agents make cheap tokens everything
Sail is betting that AI agents working in the background for hours or days will dominate inference, so cost per token matters more than speed. That lets it buy cheap, scattered, less reliable compute and use any chip with good flops per dollar. Movva sees memory, power and chip supply as the real constraints.
Takeaways
Agentic inference is shifting to background work. Movva says long-running agents matter more than low-latency chat, so cost per token beats speed. He predicts background workloads move from roughly fifty-fifty with real-time toward ninety-ten. Sail's goal is the lowest cost per token in the industry.
Sail's edge is scavenging cheap compute. Sail buys any chip for any duration and uses cheap power in small one-megawatt data centers with minimal redundancy. Background agents tolerate interruptions, so uptime as low as 95 percent works. Sail avoids bidding against Anthropic or OpenAI for capacity.
Sail optimizes flops per dollar, not latency. NVLink is mandatory for low latency inference, which Sail does not need. It will use AMD and challengers such as Etched and SambaNova, and plans to support TPUs and Trainium. Movva says there are no bad chips, only bad pricing.
Memory, not compute, is the hardware bottleneck. The KV cache grows with every conversation and can exceed model weights. Cerebras SRAM bandwidth is about 21 petabytes per second versus about 10 terabytes for Blackwell HBM. Movva expects a hybrid, with Cerebras-style chips for MLP layers and GPUs for attention.
Data is running out, so self-improvement follows. Movva calls the internet a one-time data subsidy, with about 30tn tokens of high-quality text. Next gains come from RL environments and expert human feedback, where labs now spend far more than they once spent on data.
Cheap intelligence pays most on verifiable problems. Examples include deep research across 10,000 or more sources and cybersecurity agents that attack software so it can be patched. Proactive personal assistants running in the background are another. Human taste is unsolved, so Movva leaves writing and art to people.
Nvidia's dominance is less secure than assumed. Movva says performance per watt has barely improved from Hopper to Rubin. Losing TSMC would hurt but not be catastrophic, with Intel at worst about twice as bad. NVIDIA allocates chips strategically and will not sell tokens itself.
Labs, investors and chip founders
Frontier labs pay a premium to stay three to six months ahead, still worth it, but Movva doubts it lasts because enterprises adopt slowly.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
He doubts distillation can be prevented, and says open source will not disappear.
He compares semiconductors at 19 to 21 percent of the S&P 500 to the dot-com era and says token demand is real because people use tokens immediately.
Chip entrepreneurs should master the bottlenecks, which are TSMC wafer capacity, HBM, advanced packaging and power.