不只卖 Token：Modal 如何把混乱算力、沙箱与代理开发连成平台

Bernhardsson describes compute as a heterogeneous and only partly organized market, where reliability, location, networking, and contract terms matter enough that aggregating capacity is more than matching a GPU name to a price.

Modal's custom file system, container runtime, and scheduling work are presented as foundations for fast cold starts and elastic applications, while open-source components remain useful where they complement rather than erase that infrastructure advantage.

The sandbox product began as an asymmetric bet made cheap by existing infrastructure, and Bernhardsson emphasizes that its later surge required repeated rewrites rather than vindicating a fully formed prediction of the eventual market size.

Bernhardsson expects token serving to face commoditization as optimization techniques spread, so he sees a more durable position in an adjacent product portfolio that supports inference, sandboxes, training, and the surrounding workflow.

Although he remains bullish on larger language models and NVIDIA demand, Bernhardsson argues that many workloads can run economically on older GPUs, with memory fit offering a useful heuristic rather than eliminating the need to test performance.

A broad platform strategy still requires ruthless prioritization and deliberate sequencing, and Bernhardsson's later reflection is that some dependent projects should proceed in parallel when waiting would squander a fast-moving opportunity.

Bernhardsson reports especially strong agent adoption among senior engineers who can specify extensive tests and supervise results, but he keeps humans responsible for reliability tail risks and sees delivery automation as increasingly important.

Agents benefit from the same fast feedback, clear errors, and documentation as developers, yet stale training data makes API deprecation harder, while Modal tracks whether agents recommend it when its name is absent from the prompt and Bernhardsson describes a possible incumbency flywheel rather than a directly measured causal mechanism.

His next infrastructure bets target low-latency inference, training, and richer execution environments, while he treats agent architectures and RL services as unsettled areas where productization and lasting primitives matter more than following every fashionable pattern.

Bernhardsson distinguishes capital-intensive GPU lessors from software-like platforms and warns that partially prepaid compute commitments can leave model labs dependent on future funding, so strong AI demand does not guarantee that every supplier or lab will succeed.