Tokenmaxxing: giving employees frontier AI is not AI transformation

Somewhere inside Meta, an engineer was probably just trying to centre a div.

In mid-2026, reports described Meta employees consuming a staggering volume of AI tokens: roughly 73.7 trillion in a single month. At the estimated rates going around, that works out to something like 221 million dollars a month, or about 2.65 billion dollars a year, spent on internal AI prompts and queries. One widely shared analogy pointed out that the annual figure is in the range of 9,000 engineering salaries. The numbers are estimates from reporting, not audited Meta disclosures, so treat them as directional. The direction is the point.

It is the habit of running as many AI calls as possible, often in parallel, to maximize output or to climb a usage metric, without regard to whether the spend produces anything useful. At Meta it reportedly had a competitive edge: an internal leaderboard ranking the top users by tokens consumed. Consumption became the scoreboard. Meta's reported response was to rein it in with a centralized AI gateway that tracks spend by team, sets budgets, and fires alerts on spikes.

Here is the uncomfortable part for everyone else. Meta is not an outlier because it uses too much AI. It is an early, visible example of a mistake most enterprises are quietly making: confusing access to AI with transformation by AI.

Buy every employee a frontier-model subscription and you have bought access. You have not changed how the business runs. Access produces a thousand private, unmonitored, unrepeatable interactions: a marketer rewriting an email, an analyst pasting a spreadsheet, an engineer retrying a hallucinated function five times. Some of it helps. Most of it evaporates. None of it is a system.

Transformation is the opposite of scattered. Enterprise AI transformation is the work of turning the knowledge and processes already inside the organization into systems that run in production: agents that carry a workflow end to end, automation wired into the tools people actually use, and models that operate a defined part of the business day after day. The measure is not tokens consumed. It is work done, reliably, with a number attached.

The reason tokenmaxxing gets out of control is that nobody is watching the return. We used to judge engineering by clean architecture and uptime. It is strange to now judge it by how fast a team can burn through a compute budget. Frontier access with no instrumentation is a credit card with no statement.

Managed AI operations is the unglamorous discipline that prevents this. It means every AI system in production has an owner, a budget, monitored usage, and a clear line from cost to value. It asks the question tokenmaxxing never does: is this spend shipping better work, or are we subsidizing the world's most expensive autocomplete? Meta's own fix, a gateway with budgets and alerts, is a late, centralized version of exactly this. Better to build the discipline in from the start than to bolt it on after the first billion.

There is a second lesson hiding in the token bill. Metered, per-token frontier APIs turn usage into an open-ended variable cost. The more your organization leans on AI, the less predictable the number at the bottom of the invoice.

For workloads that are heavy, repetitive, and sensitive, owning the model changes the economics. Open-weight models sized and quantized for your own hardware move AI from a metered utility to a fixed asset. You pay for the infrastructure once and run inference against it, with predictable cost and no per-call data exposure. That is not the right answer for every workload, and the largest model is rarely the right one for a private deployment. But for the high-volume core of an operation, it is the difference between a habit you cannot see and an asset you control.

If your AI strategy is a pile of individual subscriptions, you do not have a strategy. You have Meta's problem at a smaller scale, minus the leaderboard.

The move is not to use less AI. It is to convert access into systems: pick the workflows where AI does real work, build them to run in production, instrument the cost and the return, and own the models where the volume and the sensitivity justify it. That is enterprise AI transformation. Everything else is tokenmaxxing with a nicer dashboard.