GPU capacity is a shape problem, not a supply problem

The core problem. Modal founder Erik Bernhardsson frames the AI bottleneck as a shape mismatch. GPU capacity is sold as three-year rectangles that begin months in the future, while real demand is bursty and shifts overnight. Modal grew from about $100 in monthly revenue in November 2022 to $30,000 in January 2023, and he says revenue is at least 50x what it was two years ago. The numbers are unaudited, but they explain the market tension.

Capacity reality. Modal now rents from almost 25 cloud providers, often under three-year prepaid deals, and has a four-person procurement desk plus a ten-person bare-metal team to manage bad drives and GPUs falling off hosts. Bernhardsson is openly wary of buying hardware: prices normally fall, so anyone buying at the top risks a repeat of the Hunt brothers silver trade. Scarcity today means Nvidia and neoclouds capture the rent, but he treats that as temporary.

The platform bet. Modal's answer is aggregation. With thousands of customers, individual spikes smooth out, letting Modal offer on-demand capacity instead of rectangles. The same logic applies to Modal Sandboxes, now about a third of revenue. The median sandbox lasts roughly eight seconds, and the product only took off once LLMs became good at generating full apps, with usage from Lovable, Ramp's background agents, and Anthropic's managed agents.

What this means for agents. Good developer experience and good agent experience are nearly the same thing: docs, error messages, fast feedback loops. The less obvious cost is that agents are trained on old Modal syntax, so deprecating features now breaks their behavior. Modal runs internal benchmarks asking agents to choose a sandbox vendor and keeps public examples current. That is a new, real operational burden.

Org and culture changes. Bernhardsson says code reviews are optional at Modal, and the company compensates with strong CI/CD, decoupled systems, and agent-written tests. He also reports a swing back toward individual contributor work, with managers and ICs carrying equal prestige and everyone, including the CEO, writing code. That works because Modal has unusually strong automated controls; many teams would be naive to copy it without those safeguards.

What's missing. The interview offers production experience, not independent evidence. Revenue and usage claims come from the founder, and predictions about non-Nvidia accelerators, GPU derivatives, and agents finding vulnerabilities are speculative. The structural lesson still stands: when capacity is sold in rigid blocks, whoever can slice it into demand-shaped pieces has leverage. Engineering leaders should let that shape their procurement and platform decisions.