The world can't keep up with the pace of AI companies

Code agents have become the first AI product to create massive paid demand. Revenue at OpenAI and Anthropic is growing rapidly, and Anthropic's revenue has tripled compared to the start of the year. Claude Code's share of commits, based on public GitHub repositories, also jumped from 2%→4% in January, with projections suggesting it could reach 20%+ by year-end.

If a $100-a-month subscription can deliver a 10-30x ROI relative to a developer's salary, then even very small amounts of automation are enough to justify the purchase. This trend isn't simply hype—it's a signal of real demand that actually reduces work hours.

Still, the limits are clear. Some figures are year-end projections, and the ways they're counted differ. AI companies are still operating at a loss, agents remain unstable, and enterprise work is blocked by requirements, architecture, review, testing, and compliance, so they can't immediately replace humans.

Even so, the rising tide matters more than the 'bubble' framing. Just as reasoning, image generation, voice, and Go eventually became possible, agents are following the same path, and users are increasingly learning to work around the tools.

The industry structure is tied together in a long value chain of AI labs, large cloud providers, and chip makers. Anthropic has been in 10x annual growth mode for three years running, and while Dario expects things to slow down around 2026, growth is still accelerating for now. Since more than half of revenue goes into research, the cash flow gap is widening—to generate $30B in annual revenue, someone has to commit $80B to infrastructure. Meanwhile, Amazon is pouring about $200B, Google $180B, Meta $125B, and Microsoft $105B into capital expenditures.

The bottleneck has shifted to a different place each year.

• 2023: TSMC's CoWoS packaging
• 2024: HBM memory shortage
• 2025: Data center power supply
• 2026: Transmission limits of the US power grid

Memory accounts for a large portion of GPU costs, and with SK Hynix and Samsung controlling 90% of the market, pricing pressure could continue. xAI has shown that data centers can be built up quickly, but the power grid and transmission infrastructure can't keep up with that pace. Workarounds like industrial gas turbines may hold things together for now, but ultimately, the conclusion is that the world is failing to keep up with the growth speed of AI companies.