The Bill Is Incomplete

Every company deploying AI today is looking at the same invoice and trying hard to cut it. Earlier this summer, the Wall Street Journal reported on efforts to bring that bill down. The new word describing it, tokenomics, is all about the shift from tokenmaxxing to thriftmaxxing. Companies are using cheaper models where they are good enough, saving more expensive models for targeted uses where they are necessary. Reaching for the best frontier model for every task is out, and reducing costs without losing performance drastically is the new game.

The market is finally treating this visible cost of intelligence as a real one. EY wrote a few weeks later about two failure modes: tokenmaxxing optimizes for how much gets built while budget panic for how little. But who is asking if the right things are getting built?

I have always said, for years, that industry needs to take the cost of this seriously. Cost-sustainability was a guiding principle for me from the start when I cofounded Concentric AI. I had watched too many AI startups ship dazzling demos only to fold when the bill came due at real volume.

The cost moved

Here is what has changed now. Managing token costs is absolutely the right thing to do. But it is not the bill that will eventually sink you.

In the entire history of software, the cost of building itself acted as a filter. Engineering resources were scarce, and weak ideas never made it to the top to get built. When that filter goes away, so does that quiet discipline of prioritization and far more will get built. Every new deployment will bring with it new dependencies, handoffs and coordination costs that multiply. AI drops the friction and cost of doing work faster than reducing the cost of holding all of the pieces together. And that is the gap where real costs can hide.

Why the second cost hides

The coordination cost shows up on no dashboard since it doesn’t belong to anyone. It compounds quietly as each individual system looks fine.

While it is invisible on the invoice, the effects can be visible if you know where to look. It shows up as ROI that was promised months ago but failed to materialize, as rework that was not planned, as oversight that lags, and as interacting systems with no named owners.

EY states that the cost of an agent is often invisible until too late, and that tokens are only part of the true cost. I agree. But their fix is to price every agent, to benchmark it, meter it, and assign a value metric to each one from the start. The problem: no agent carries the cost of coordination among them, it is a cost that lives in the gap in between. You can meter every agent perfectly and still miss the entire bill.

Same discipline, different line-item

Cost-sustainability was never really just about compute. It was about refusing to allow a cost to sink you at scale. When I was thinking about building a product, that was the cost of compute infrastructure and it folded companies at volume. The same thing now is repeating at the enterprise level and the cost has moved from compute to coordination.

The discipline holds but the target is new. Enterprises need to budget for coordination the same way they started budgeting for compute. Put it on the books as a real line-item. Design for it before it compounds instead of finding out when it’s too late.

The price of intelligence is falling, but there is a hidden cost that is moving. Companies who will be running at full speed in the future will be the ones who start managing that cost now. The full argument runs through my book, Coherence, arriving this Fall. If you want to follow the thinking as it develops, join the list at coherise.com. The one-page decision tool from the book is the first thing I send.