OpenAI's viral AI agent tool OpenClaw took the tech world by storm this year—until Anthropic pulled the plug. Millions of users woke up to discover that using Claude to power their beloved agents would now cost serious money. "Our subscriptions weren't built for the usage patterns of these third-party tools," wrote Boris Cherny, head of Claude Code, on X. The message was clear: after years of subsidized access, AI labs are finally demanding returns.

Token Economics and the Math Problem

Here's what nobody in Silicon Valley wants to talk about openly: the numbers don't add up yet. Gartner estimates capital investment in AI data centers will hit $6.3 trillion between 2024 and 2029. To avoid write-downs—"an unmitigated disaster for all investors," according to senior director analyst Will Sommer—major labs need a return on invested capital of about 25 percent. Below 7 percent? You're in zombie company territory, surviving only because venture capital keeps flowing. To hit even that floor, Gartner forecasts AI companies would need to generate nearly $8.2 trillion in cumulative revenue through 2029. That means processing around 10 sextillion tokens per year—current industry estimates are somewhere between 100 and 200 quadrillion. Even assuming a generous 10 percent profit margin per token, consumption needs to grow by 50,000x to 100,000x. Let that sink in.

The Agent Tax

Reasoning models are beautiful things—until you look at the bill. These agents "think out loud" through thousands and thousands of tokens: exploring paths, launching sub-agents, verifying accuracy. A one-sentence prompt can spiral into tens of thousands of words the user never sees. Georgia Tech professor Mark Riedl puts it bluntly: "If you have thousands or millions of people using these things every single day, the inference costs really outweigh the training side of things." Companies like Box are feeling this acutely. CEO Aaron Levie notes that agent use cases have exploded while infrastructure capacity hasn't kept pace. "The use cases have exploded, and we're out of capacity," he said. OpenAI has already committed $600 billion in spending through 2030—down from a planned $1.4 trillion—which Sommer calls a "massive step down." Even the best-case revenue forecasts fall short of what's needed.

Developers Are Already Adapting

The smart money is already moving. Anaconda CEO David DeSanto spent five weeks traveling to customers and found everyone facing the same calculus: token usage up, billing costs up, caps tighter. Many are migrating to self-hosted deployments via Amazon Bedrock or Google Vertex AI for more supply chain control. Others are shifting workloads to open-source models that have dramatically improved on benchmarks recently. Eve, a legal AI startup, has seen token usage jump 100x year-over-year. The company now splits work between expensive reasoning models (25-30 percent), its own open-source variants, and cheaper alternatives from leading labs. "What open-source is really doing is putting pressure on these companies to make their cheaper models cheaper because their profit margins there are much better," said cofounder Jay Madheswaran.

Key Takeaways

  • Anthropic's OpenClaw restriction signals the beginning of aggressive third-party tool monetization across AI labs
  • Token consumption would need to grow 50,000x-100,000x by 2030 to hit even minimum ROIC targets—math that doesn't work today
  • Reasoning models are hemorrhaging money on inference costs that dwarf training expenses
  • Market consolidation is "virtually inevitable" with no more than two LLM providers per regional market surviving, per Gartner

The Bottom Line

The free era was a land grab—classic startup economics designed to capture market share before the billing meters started running. Now those meters are running, and developers building on top of these platforms need to factor in rising token costs as a permanent line item, not a temporary subsidy. Open-source alternatives aren't perfect yet, especially for coding tasks, but they're getting better fast—and that's exactly the pressure needed to keep frontier labs honest.