NVIDIA CEO Jensen Huang recently argued that top engineers should spend the equivalent of half their annual salary in tokens.
“If that $500,000 engineer did not consume at least $250,000 worth of tokens, I'm going to be deeply alarmed,” Huang said.
This idea has a name: tokenmaxxing. Tokenmaxxing is the practice of maximizing token consumption as a signal of how deeply employees are embracing AI tools. The term borrows from a pattern in internet slang where "-maxxing" means pushing something to its absolute limit. It took hold in early 2026, when stories broke about mega-corporations like Meta and Amazon ranking employees by token usage on internal leaderboards.
As more engineering teams began deliberately maximizing token usage, some finance teams started pushing back by implementing token budgets: per-user or per-team spending caps designed to put a ceiling on consumption. Tokenmaxxing has become one of the most commonly tracked AI adoption metrics, and token budgets are the most common attempt to control it. The question is whether either one is actually useful.
The short answer is that neither measures what matters. Token count captures how much AI was used, not whether that use created value, and budgets cap spend without ever connecting to outcomes.
Why Is Token Count a Terrible Metric for AI Productivity?
Tokenmaxxing is built on the assumption that token count is a reliable signal of AI productivity. The problem is that token count only tells you how much AI was used, not whether that use created value. If your dashboard only displays token count, an engineer burning 10,000 tokens to generate $1 of business value looks identical to one burning 50 tokens to generate $100: both show up as just usage.
The problem compounds when you factor in how AI actually works at scale, especially agentic AI. Token count only captures the model bill. Every agent action triggers downstream activity that never shows up in a token dashboard:
- API calls to third-party services
- Data enrichment and verification costs
- Retry loops when agents fail and restart
- Downstream tool calls triggered by a single prompt
- Infrastructure costs across multi-agent workflows
Unlike standard generative AI, where a user submits a prompt and the model returns a response, agentic systems operate through iterative reasoning loops and tool invocations. According to research by Information Matters, agentic systems consume five to nine times more tokens per workflow than standard generative AI. An organization tracking token consumption can significantly undercount its real AI costs while having no idea which of those costs are generating returns.
Token budgets are a blunt-force instrument; they keep spend in check, but they don't address the question of ROI. Capping spend per user or per team sets a ceiling on consumption but does nothing to link that consumption to outcomes. A team can stay well within budget and still have no idea whether their AI spend is working.
Token count as a metric also distorts the teams it's meant to measure. Teams that optimize by using better prompts, smarter model routing, and cleaner architecture consume fewer tokens than teams that don’t. On a token count leaderboard, the disciplined engineers often look like laggards.
What Happens When Organizations Use Token Count Anyway?
Any metric that becomes visible, ranked, or discussed inside an organization shapes behavior, even without a formal mandate. Give engineers a target, and they will try to hit it. The question is whether hitting it means anything. With token count, the answer is often unclear.
Daithi Walsh, Head of Product Management at Revenium, has seen token maxing firsthand.
Walsh recalls a conversation with a principal DevOps engineer with a strong track record in high-growth startups, someone who had spent decades optimizing infrastructure cost, improving resilience, and helping companies meet demanding scale objectives.
The engineer’s organization had started measuring token burn as a success metric. But the engineer’s work did not require heavy AI usage to deliver meaningful business outcomes. As a result, his individual and team performance optics started to look weak.
The response was predictable: optimize for the new success criteria by automating token consumption. That allowed him to keep delivering real operational value while also protecting his team’s standing and reasserting himself as a top performer.
For Walsh, this is a textbook example of Goodhart’s Law: when a measure becomes a target, it stops being a good measure. Token consumption can be a useful adoption signal when observed correctly. But once people are judged against it, the metric becomes gameable. The organization is no longer measuring productive AI use or business impact. It is just measuring the ability to generate any token activity - teaching staff how to waste tokens”
At Uber, engineers were actively encouraged to use AI coding tools and were ranked on internal leaderboards based on usage. Adoption surged, but so did costs: Uber burned through its entire planned AI budget just months into the year. Another report recently surfaced about an unnamed company that burned through $500 million on Claude in a single month after providing employees with unlimited access and no governance controls.
When token consumption is the only signal, there is no mechanism to connect spend to value, either before or after the budget runs out. In research I conducted at McKinsey, I interviewed more than 30 CxOs about how their organizations were budgeting for AI. The answer, consistently, was improvisation. Companies are drawing from wherever they can, including layoffs, to fund AI adoption.
Tokenmaxxing is a double-edged sword. On one side, it represents a genuine leap in engineering leverage. You’re seeing humans acting as high-level conductors of a massive digital workforce, using tokens as raw material to accelerate innovation and eliminate drudgery. On the other hand, raw consumption is mistaken for progress, incentivizing bloated, inefficient code rather than elegant problem-solving.
What Should You Track Instead of Token Count?
Tokenmaxxing as a short-term adoption push isn't entirely without merit. If the goal is to build AI fluency across an organization, a blunt consumption signal can spark experimentation, tool adoption, and the development of real AI competence. But it only works if the organization matures the measurement over time.
That maturation means moving from "more AI usage is good" to tracking what I'd call “impact density”. Impact density is the value you get from AI spend rather than the fuel you consume — the outcome produced per token, not the tokens consumed. In practice, that means tracking:
- Token-to-outcome ratio: Which engineers and workflows are solving problems with surgical efficiency, achieving more with less
- Autonomous throughput: How much toil is being handled entirely by agents, without human intervention
- Cycle time reduction: Whether AI spend is translating to actual business velocity, not just activity
- Maintainability index: Whether AI-generated code is building long-term value or accumulating technical debt
- Autonomy rate: The decline of human intervention costs as automation scales; it distinguishes high burn that drives margin-positive growth from high burn that accumulates agent debt, the hidden cost of AI decisions your financial controls never caught
- Cost per conversion: Every dollar tied to a tangible business result: a loan approval, a support ticket resolved, a code commit that shipped
These metrics all connect spend to outcomes and can't be gamed through tokenmaxxing.
How Do You Track the Right Metrics?
Tracking the metrics that describe impact density requires visibility that most organizations don't have yet. That's exactly the problem Revenium's AI Economic Control System is built to solve. Revenium captures every AI transaction and traces it back to the workflow and outcome that triggered it, giving engineering and finance teams:
- Full-stack attribution: Every cost connected to the user, workflow, and business outcome that drove it, from the first model call through every downstream tool invocation, third-party API hit, and infrastructure event
- A shared view: Engineering and finance see the same numbers for the first time
- Real-time economic control: Circuit breakers halt runaway agents before they exceed configured thresholds, replacing after-the-fact budget shock with proactive cost management
If you’re ready to move beyond token count, check out “The Financial Blind Spots in Autonomous AI”. This report goes deeper into the economics of agentic AI and what financial control actually looks like in practice.
If you're ready to see Revenium in action, book a demo with our experts today.
