Tokenomics: where the dollars actually go on a nine-figure AI spend

A large enterprise signs a $100M AI transformation. Eighteen months later they have slide decks, a few pilots in production, a renewal quote, and a bill nobody itemized honestly. Let's itemize it. Same scope, two ways to buy it.

The bill nobody breaks down

Here is where a nine-figure program goes when a systems integrator and a token vendor run it. The line items are the ones that never make it onto the one-pager, scaled to a round $100M so the proportions are the point, not the exact figures.

Where the money is really going

Add it up by what you're buying, not what the invoice calls it: roughly $0.60 of every dollar buys human hours to wire the same plumbing every enterprise wires; another quarter re-reasons, at token prices, outcomes that have exactly one correct shape. You are renting the answer to a question with a fixed answer, over and over, and renting it back the next year.

The same scope, on your own node

Now buy the outcome. Capture the domain as a validated schema once, generate the boring 80% deterministically, and run inference on a committed GPU node instead of a metered endpoint. The two biggest evaporating line items above, inference and retraining, both change shape. Inference stops being a meter and becomes a fixed cost. Retraining stops being a multi-day event and becomes a signed Mempack™ delta you distill in an afternoon.

Start with the meter. The traditional program spent $14M renting inference at premium per-token rates, a bill that grows with every call. A committed 8×H100 node, reserved for a year, is a flat line whatever you push through it.

How the experts stay current: teacher-student distillation

You don't retrain a 70B model from scratch to teach it a new rule. You distill. A frontier teacher reasons over the new domain traces and labels them; a small student, the model already living on your node, learns them in a short QLoRA pass. The output isn't a new brain. It's a Mempack™: a low-rank adapter signed against the base checkpoint, plus the knowledge-graph edges that ground it. Hours, not days. (The process that produces and governs these packs is patent pending.)

Here is one update cycle for a single domain expert, itemized. The only metered spend is the teacher pass, and it's bounded before you start.

Full retrain vs. distilled patch

The difference isn't only price; it's the shape of the artifact and how it moves.

That last row is the point. A branch office or edge node doesn't pull a new model; it pulls a 300 MB signed delta, smaller than a phone OS update, verifies the signature against the base checkpoint, and swaps the adapter in without a redeploy. That's what prescriptive means: you ship the correction, not the whole brain, and you ship it the same day the rule changed.

The one number that matters

It isn't the raw savings, though those are real: metered inference becomes a fixed node, and multi-day retraining becomes a $1.5K afternoon. The deeper point is what happens to the dollars you do spend.

In the traditional program spend evaporates: it buys hours and tokens that leave nothing behind. Here spend converts into an asset: every teacher pass distills into a signed Mempack™ you own, every validated rule widens the deterministic core, and the knowledge graph you build this quarter grounds next quarter's distillation for free. The bill goes down while what you own goes up. That's tokenomics: not "how cheap are the tokens," but who owns the compounding asset when the engagement ends. Rent the answer to a fixed question forever, or distill it once into a patch you keep. See the paths →

Figures are illustrative: a committed 8×H100 node and a QLoRA distillation cycle, scaled against the $100M reference program to show proportion, not a quote.