Refactoring Orgs

When the intelligence itself commoditizes, the vendor stops selling the intelligence. This is not a hypothetical. It is what the frontier lab pitch has already become, if you read the last three quarters of product announcements carefully.

Here's why: The gap between best and good-enough just closed. An open-weights model at five percent of the price now clears the bar for most workflows a buyer is running. We actually ran an experiment for a week, with more than 50 coders agreeing: For us, GLM 5.2 in the pi harness is just good enough for most agentic coding work - We're finally ready to tackle that enormous AI token bill we racked up.

What is left on the shelf of frontier labs is not the model. It is the vision layer, the agent harness, the browser tool, the enterprise console, the SOC2 packet, the region-specific hosting, the safety tuning, the audit log. In other words, the plumbing. That is the actual product now. The tokens are the loss-leader.

This is a different business than the one investors in frontier providers funded. A pure inference business is a compute business, with compute margins. A plumbing business is closer to enterprise software, with the customer acquisition costs, integration cycles and sales complexity that come with it. Anthropic, which sold you a model in 2025, is now trying to sell you a platform in 2026. The people it needs to hire, the contracts it needs to sign, the ways it earns revenue all shift.