The End of Cheap Tokens

Armin Ronacher and Ben Vinegar discuss agentic coding economics. They cover infrastructure costs, declining subsidies, Pi joining Earendil, and the data and platforms that may determine control of future coding models.

Key Points Covered

• AI demand is raising costs beyond GPUs: The hosts connect more expensive RAM and storage to inference infrastructure, caching, energy, and data-center build-out, while treating further price moves as predictions rather than certainties 09:59–16:39(opens in a new tab).
• Security harnesses make existing models materially more capable: Purpose-built systems can find large numbers of real vulnerabilities, creating pressure for maintainers and companies to run comparable audits rather than dismiss AI-assisted reports as slop 16:39–22:11(opens in a new tab).
• Enterprises are starting to constrain model and token use: Once companies pay API rates at workforce scale, cache efficiency, extensions, and per-developer limits become budget concerns; the hosts see standardization beginning 22:11–25:46(opens in a new tab).
• The subsidy era is starting to unwind: Subscription restrictions and SaaS products moving toward usage pricing expose the cost of long-running agents, while downstream vendors have little room for margins when inference dominates their expenses 25:46–32:29(opens in a new tab).
• Higher prices will test whether agent usage produces durable value: Ben argues that less-subsidized pricing could force more selective use. Armin compares the transition to industrial investment cycles in which productivity rose but many firms misjudged capital needs 32:29–35:55(opens in a new tab).
• Open models need useful agent traces, not only model weights: Armin argues that open-weight providers need authentic coding trajectories for reinforcement learning, but voluntary sharing faces privacy, consent, and critical-mass problems 51:43–57:13(opens in a new tab).
• GitHub's dominance is no longer unquestioned: Reliability problems and agent-driven load are prompting projects and investors to explore alternatives, although GitHub's integrations and ecosystem still make leaving costly 57:13–1:19:16(opens in a new tab).
• Training data is becoming a commercial asset with unresolved consent: The hosts expect more companies to sell valuable usage data to model labs. They argue that clear disclosure and user control should matter even when current incentives reward quiet extraction 1:25:08–1:38:02(opens in a new tab).

Full video: https://www.youtube.com/watch?v=JM1sIVIZYRg(opens in a new tab)