Anthropic disclosed this week that over 1,000 enterprise customers now spend more than $1 million a year on Claude, double from 500 in February. Revenue run-rate hit $30 billion. Alongside it: a 3.5-gigawatt compute deal with Google and Broadcom. The narrative is familiar. Bigger model, bigger deal, race to the frontier. But the companies getting real value this year aren’t winning on model choice. They’re winning on what surrounds the model. The model is rented. The harness around it is yours.
Why It Matters
Most organizations treat AI like a calculator. You ask, it answers, you move on. Every conversation starts from zero. That is why AI feels disposable: no matter how good the model gets, nothing it learns sticks.
The teams pulling ahead do the opposite. They codify what their AI learns into shared documents and feedback logs the next session inherits. One practitioner reported rejection rates dropping from 40% to 15% in four weeks, not because the model improved, but because the feedback loop did. Memory compounds. Models don’t.
The Decision
You are renting intelligence. The question worth asking this week: what part of what your AI does is actually yours to keep? If you switched vendors tomorrow, what would survive the move? If the answer is “nothing,” you don’t have an AI strategy. You have a subscription.
What To Do This Week
- Ask one team: “Where do we write down what the AI gets wrong, so it doesn’t get it wrong again next week?”
- Pick one high-volume AI workflow and add a single feedback log, one line per correction, shared across the team.
- Audit your top three AI tools: how much of your context lives inside the vendor versus inside files you own?
What Not To Do
Don’t pay for amnesia. Don’t let a vendor pitch you “smarter prompts” as a substitute for memory your team controls. Don’t measure AI progress by how big the model got this quarter. And don’t confuse buying access with building capability. Access expires. Capability compounds. Only one of those shows up on next year’s P&L.
Signal Boost
Ethan Mollick, “A Guide to Which AI to Use in the Agentic Era” - a crisp argument that the real competition in AI is harness infrastructure, not model quality. Read it this week to see why your memory layer is the only durable edge.