There is a tax in your AI invoice you didn’t sign up for.
It is not labelled. It does not appear on a line item. It is paid in tokens, billed by the foundation model provider, and absorbed into your cloud expense category before anyone in finance has the chance to ask what it is for.
It is the largest single cost category in enterprise AI today. And nobody is calling it what it is.
What the tax is
The token tax is the cost of reconstructing intelligence the system already produced.
Every probabilistic AI generation event is stateless. The system does not remember what it produced last time. It does not remember the user. It does not remember the company. It does not remember the policy. It does not remember the strategy. It does not remember anything.
So every query starts cold. Context reconstruction tokens. Reasoning reconstruction tokens. Identity reconstruction tokens. All paid for. All discarded the moment the response is delivered. All paid for again on the next query.
That is the tax. It runs in the background of every enterprise AI deployment in the world right now. And nobody has the line item.
How much it costs
The math here is conservative because we are not building for vendor argument. We are building for board-level clarity.
A mid-sized enterprise deployment runs roughly 50,000 commercially meaningful queries per month. The average context reconstruction overhead per query — based on standard system prompt patterns, RAG retrieval payloads, persona scaffolding, and instruction repetition — sits between 4,000 and 9,000 input tokens before any output is generated.
Take the midpoint. 6,500 input tokens per query, paid for, doing nothing except reminding the system of things it already had access to last time.
50,000 queries multiplied by 6,500 tokens equals 325 million tokens per month.
At standard enterprise API pricing that is a five to six figure monthly cost — paid every single month — for the system to remember what it did yesterday.
Run that for a year. Run it across multiple enterprise use cases. Run it across multiple departments. The numbers stop being trivial very quickly.
The three layers the tax runs on
The token tax is not a single line. It runs three deep.
Reconstruction tax: the cost of rebuilding context that should have been persistent.
Redundancy tax: the cost of repeating reasoning the system has already performed elsewhere or earlier.
Drift tax: the cost of human review and correction caused by probabilistic outputs that vary across identical queries.
Most enterprises see only the first layer because that is the visible API line item. The other two are absorbed into operational cost — productivity overhead, quality assurance time, manual moderation, the invisible tax on trust that accumulates every time the same question yields a different answer.
Add all three layers together and the real cost of probabilistic AI infrastructure is two to four times the published API spend.
Why the vendor won’t help you here
Foundation model providers do not have a financial incentive to help you reduce token consumption.
They sell tokens. That is the business model. The economics of their entire enterprise are built around customers consuming as many tokens as possible. Helping you architect a system that reduces token consumption by seventy percent would shrink their revenue by seventy percent. They are not going to do that.
This is not a moral failing. It is an alignment of interests problem. You are paying for tokens. They are paid by the token. The system between you and them is engineered to maximise token throughput per query, because higher throughput means higher revenue.
The architecture that breaks the tax has to come from somewhere else. Not from the vendor selling the tax.
What actually removes the tax
A deterministic cognitive layer. Sitting in front of the foundation model. Scoring intent. Retrieving pre-compiled artefacts. Routing structured reasoning through verb-based execution rather than probabilistic generation. Sending the foundation model only the lightweight rendering tasks that genuinely benefit from generative output.
Token consumption per query drops sixty to eighty percent. Drift drops to zero on structured tasks. Audit trail becomes deterministic. Cost compresses with usage rather than scaling against you.
This is not theoretical. This is in production. SDCI is the architecture. MatrixOS is the deployment shell. Nine platforms are running on it today.
The bottom line
You can keep paying a tax you did not sign up for, on infrastructure that gets more expensive the more you use it, billed by a vendor whose business model depends on you not asking why the bill is what it is.
Or you can put deterministic cognition in front of the model and stop paying for the same reasoning to be rebuilt thousands of times a day.
That is the choice. Most CFOs have not yet been told there is a choice.
Martin Lucas is founder and CEO of TMX Group, inventor of SDCI™ — Synthetic Deterministic Cognitive Intelligence — and author of the Human Architecture Series. The MatrixOS portfolio runs nine SaaS platforms on a single deterministic cognitive engine.
