← Back to Signal
Market SignalsTokenomicsPosition on PrizmalRun

The new unit of enterprise productivity is the token

Jensen pays engineers in tokens. Chamath caps them. By 2027 every serious enterprise will burn 10 billion tokens per employee per year. The question is no longer whether you use AI. It is how much of that compute is wasted.

Audry Larocque · May 5, 2026 · 6 min

Key takeaways
  1. 01Tokens are the new unit of enterprise productivity. Jensen Huang is paying engineers in them; Chamath Palihapitiya is rationing them.
  2. 02A 1,000-person enterprise consuming 10B tokens per employee per year produces a $25M annual inference bill at current mid-tier API pricing.
  3. 03Every LLM today runs at full depth on every token. Roughly 90% of that compute is wasted on tasks that do not need it.
  4. 04PrizmalRun reads token difficulty in real time via SpeCo and routes simple tokens through fewer layers. 60% FLOPs saved, 99% reasoning retained.
  5. 05The math is $15M per year recovered per 1,000 employees, with no retraining, no fine-tuning, and no model modification.

01 · The frame

The token is the unit. Almost all of it is unoptimized.

Jensen Huang announced it at GTC: engineers will earn tokens as compensation, on top of salary. Chamath Palihapitiya is already enforcing token budgets to avoid running out of money. Every serious enterprise will consume 10 billion tokens per employee per year by 2027. The question is no longer whether you use AI. It is how much of that compute is wasted.

Tokens / employee / yr
25B

Enterprise consumption in 2026 as agentic workflows scale.

Compute wasted
90%

Every LLM runs at full depth on every token, trivial or complex.

Cost / developer / day
$300

Reported by Jason Calacanis for agents at partial capacity.

Compute saved
60%

FLOPs reduction at inference time. 99% reasoning retained.

02 · The problem

Every token runs at full cost. Most of them do not need to.

Jensen Huang told GTC in March 2026 that he would give engineers half their salary again in AI tokens. That statement codifies what enterprise finance teams are already discovering. Tokens are a commodity with a cost, and the volume is accelerating faster than anyone modeled. By mid-2025, OpenRouter alone was routing 100 trillion tokens per year, with over 1 trillion tokens processed in a single day.

The structural problem is not the volume. It is the waste built into how transformers execute. Every LLM today runs at full capacity on every single token, whether the task requires deep reasoning or a trivial lookup. Agents that route correctly outperform brute-force frontier calls at scale, because the cost per token is not the issue. The issue is that most tokens cost too much for what they actually deliver.

"I'm going to give engineers probably half their base pay on top of their salary as tokens. Tokens are becoming one of the recruiting tools in Silicon Valley."
Jensen Huang · NVIDIA GTC · March 2026
"I've been forced to institute token budgets for my top developers. Without them, I'll run out of money. The usage far exceeded every projection."
Chamath Palihapitiya · All-In Podcast · February 2026
"Enterprise spending on generative AI hit $37 billion in 2025, a 3.2x increase from 2024. 98% of organizations now actively manage AI spend. The era of all-you-can-eat AI pricing is structurally unsustainable."
Enterprise Token Economics · The AI Enterprise · March 2026

03 · The math

10 billion tokens per employee. Here is what that costs.

Model this at current mid-tier API pricing of $2.50 per million input tokens. A 1,000-person enterprise consuming 10 billion tokens per employee per year generates 10 trillion tokens annually. That is a $25 million per year inference bill at full compute. With PrizmalRun cutting 60% of FLOPs while retaining 99% of reasoning quality, the same intelligence output costs $10 million. $15 million saved, without retraining a single model or modifying a single weight.

LineWithout PrizmalRunWith PrizmalRun
Employees1,0001,000
Tokens per employee per year10B10B
Total tokens per year10T10T
Average price per M tokens$2.50$2.50
Annual inference cost$25M$10M
Reasoning quality100%99%
Compute wasted~90%Governed per token
Annual savingsNone$15M

Of 10 billion tokens consumed per employee per year, SpeCo classifies and routes each one in real time. Roughly 45% are low complexity, lookups, formatting, and short completions, which skip 60 to 80% of layers. Another 35% are mid complexity, summarization, analysis, and structured generation, which skip 20 to 40% of layers. Only 20% are high complexity, multi-step reasoning, code generation, and chain of thought, which run at full depth.

04 · Measured performance

The numbers are not projected. They are measured.

All results measured under controlled A/B conditions using NVML energy instrumentation on Llama-2 7B and Mistral 7B. Full methodology in the technical appendix.

MetricWithout PrizmalRunWith PrizmalRunBusiness impact
Energy per inference95.4 J71.5 J25.1% reduction in power and cooling cost
Reasoning per joule (R/E)0.275 RPJ0.965 RPJ+251% more intelligence per energy dollar
GSM8K reasoning score28.0%55.0%+96.4% on complex multi-step tasks
CoQA coherence93.0%92.6%Quality preserved across conversational tasks
Tokens per second38.2 t/s46.9 t/s+22.8% throughput on the same hardware
Compute cost (FLOPs)Baseline-60%$15M saved annually at 1,000 employees
Tokens per watt measures volume. R/E measures intelligence.

A system that produces 5,000 trivial tokens per second scores identically to one that produces 5,000 tokens of substantive reasoning. The metric optimizes for throughput. In agentic deployments where agents run for hours, coordinate across tools, and reason dynamically, throughput is not the constraint. Intelligence per joule is.

05 · How it works

Token-level compute modulation.

Difficulty scored per token

SpeCo scores each token for difficulty in real time and adjusts layer depth, attention head count, and numeric precision before that token resolves. Simple lookups skip layers. Complex reasoning chains get the full model.

3x runtime memory reduction

KV cache depth control, sparse KV retention, and reduced context fetch width shrink the active memory footprint per inference. Models that previously exceeded device DRAM now fit and run stably.

116% reasoning per joule

More correct answers delivered per unit of energy than any static approach. GRIFFIN reaches 72%. CATS reaches 104% on its training distribution and collapses outside it.

Inference-engine native

Integrates directly into ONNX Runtime, ExecuTorch, and vendor stacks from Qualcomm, Intel, and AMD. No changes to training pipelines. Runtime control only.

Prizmal is not a model

It is the control system for AI compute. Simple tokens get less compute. Complex reasoning tokens get full depth. Every token, every inference. No retraining, no fine-tuning, no model modification.

FAQ

Is this just model routing?
No. Routing picks between models. PrizmalRun governs compute inside a single inference, layer by layer, token by token. Routing and PrizmalRun compose.
Does it require fine-tuning or retraining?
No. PrizmalRun is a runtime control system. It runs on top of any existing transformer with no weight modification.
Where does the 60% FLOPs reduction come from?
From SpeCo, Speculative Complexity Estimation. Most generated tokens are mechanically simple and do not require all transformer layers. Adaptive layer skipping reclaims that compute without changing the output distribution materially.
What about reasoning quality?
Measured at 99% retention on standard benchmarks. The control system is deliberately conservative on reasoning-heavy tokens.

Run the numbers on your own traffic

The R/E Audit measures reasoning per unit of energy on your real workload, read only and reversible.