I had this conversation with an employee: in the near future, companies may not recruit people by offering only salary and working hours.
Until now, the conditions of a good job were familiar: salary, title, stock options, benefits, working hours, remote work. What a company could give one person was mostly described in the language of money and time.
But when AI moves into the center of knowledge work, one more condition gets added.
Which models can this person use? How many agents can they run in a day? How far can they reach into company data and internal tools?
This is not a perk. It is a question about the ceiling of productivity.
The Numbers Are Already Here
In March 2026, Jensen Huang said at GTC that engineers may need annual token budgets. The important point was not the exact wording, but the framing: compute would sit beside compensation as a way to amplify human output.
He also described token allocation as a recruiting tool. That is far beyond education reimbursement or a software allowance. It suggests the grammar of hiring is changing.
A week later, Microsoft’s Charles Lamanna shared a concrete example at a Seattle event. A candidate had effectively asked for a guaranteed daily token budget for the team as a condition of joining. The amount was described as ranging from about $100 per day to several hundred dollars.
The interesting detail is that the candidate did not ask for a personal allowance. The request was for a team-level token budget. Tokens are becoming an organizational design variable, not just an individual benefit.
At the leading edge of the market, this is already explicit. Anthropic’s research fellowship offers a weekly stipend and, separately, compute budget of about $15,000 per month. Pay on one line, compute on another.
The Conductor and the Orchestra
Think about an orchestra. The conductor does not personally play every instrument. The conductor interprets the piece, distributes parts, and takes responsibility for the whole sound. The orchestra performs.
The knowledge worker using AI agents is moving into this position. Research can be delegated. Code review can be assigned. Report drafts can be generated in parallel across several directions. The human takes responsibility for interpretation and final judgment. The industry already has a word for this: orchestration.
The old knowledge worker was a performer. Eight hours, one person’s attention, one person’s stamina: that was the ceiling. For the worker who becomes a conductor, the ceiling is no longer only personal time. It is the size and quality of the orchestra they are allowed to lead.
Tokens and compute determine that orchestra. Does the budget allow a quartet or a hundred-piece ensemble? What level are the players, meaning which models are available? The same conductor can produce very different sound with a different orchestra.
| Old condition | Emerging condition | Meaning |
|---|---|---|
| Salary | Annual AI budget | Living standard and the ceiling of compute-enabled output become part of the same negotiation. |
| Working hours | Concurrent agent capacity | The number of tasks one person can delegate in parallel starts to matter as much as direct time. |
| Rank and authority | Data and tool access | Performance differs by who can connect AI to internal knowledge and operating systems. |
| Benefits | Model tier and secure environment | Access to strong models in a safe environment becomes production infrastructure, not a perk. |
The labor contract expands from “how much will you pay me?” to “what orchestra will you let me conduct?”
The Strongest Objection: Won’t Tokens Become Basically Free?
The honest objection is obvious. Token prices are falling quickly. Gartner expects inference costs for one-trillion-parameter models to fall sharply by 2030 compared with 2025.
If so, maybe tokens become like electricity or internet access: ordinary infrastructure, no longer something to negotiate.
But the same Gartner discussion carries the opposite warning: even as unit costs fall, enterprise AI spending can rise. Agentic work consumes far more tokens per task than ordinary chat, and usage can grow faster than prices fall. Goldman Sachs has projected a large increase in token consumption by 2030 as agents spread.
So even if the unit price commoditizes, the total allocation per person can remain scarce. Cheap electricity does not make data-center power budgets disappear. The negotiation unit simply shifts from “how many tokens?” to “how much annual compute?”
The Trap of Tokenmaxing
One thing has to be clear: token usage is not productivity. It is input.
The side effects are already showing up. Reports describe Uber burning through its 2026 AI budget in four months, Meta employees tracking heavy AI users through a “Claudeonomics” leaderboard, and Amazon encouraging employees to use as many tokens as possible. Usage itself starts to look like a performance indicator.
| Misleading metric | Better question |
|---|---|
| How many tokens were spent | What decisions, code, reports, or customer touchpoints did those tokens become |
| How long people used AI | Did human judgment move to a more valuable part of the workflow |
| How many prompts were sent | Did the team build repeatable workflows and review criteria |
| How much budget was burned | Did the same budget improve quality and speed next time |
Burning more is not a skill. The skill is converting tokens into output. Give the same orchestra to two conductors and you can get entirely different sound. As token budgets grow, the gap widens between people who can convert compute into results and people who cannot.
Where Korea Is Now
In 2026, large Korean companies are entering the stage of distributing access rights.
Samsung, after blocking external generative AI following an internal data leakage incident in May 2023, officially introduced ChatGPT, Gemini, and Claude across affiliates in June 2026. The company tested the tools with 2,500 employees beforehand, sent more than 50 top executives to an AI bootcamp, and is training thousands of executives.
LG CNS signed an enterprise-wide Claude agreement with Anthropic that can apply across the LG group. Hyundai AutoEver operates the internal generative AI service H Chat. GS reports that employees created ten thousand work tools through its internal platform MISO over a year.
This shows the time lag with Silicon Valley. The U.S. discussion has already moved toward individuals negotiating token budgets. Korea is still mostly in the stage where companies distribute access uniformly: same accounts, similar limits, no token budget in job postings yet.
That lag may not last. Once access is distributed evenly, usage differences become data. When a company can see who produces several times more output with the same tools, allocating more resources to that person becomes a matter of time. Equal access is the stage before differentiated allocation.
What Will Be Written in the Contract
It would be an exaggeration to say every worker will negotiate tokens like salary within three years. But it is plausible for roles where one person’s output directly affects revenue or product speed: engineers, researchers, designers, planners, consultants, marketers, and anyone whose work is made of information and judgment.
The labor conditions will look like this.
The name may not be tokens. It may be compute credits, AI budget, or agent seats. The label is not the point. The point is that the amount of compute assigned to a person becomes part of the labor condition.
This does not make human value disappear. It makes it sharper. As AI takes over more tasks, what remains with the person is the judgment of what to delegate, the responsibility to review the result, and the ability to connect output to the real world.
If salary and tokens stand side by side, the question is not “how much can this person burn?”
How large an orchestra can we trust this person to conduct?
Discussion
0 comments