In March 2026, Jensen Huang said something that should have stopped every high performer in their tracks. He was talking about how he intended to compensate engineers at Nvidia. The base salary, he said, would be in the few hundred thousand dollar range. On top of that, he would give them token budgets worth roughly half of that so that they could, in his words, "be amplified 10X." He went further: if a $500,000 engineer wasn't consuming at least $250,000 worth of tokens, he would be, and this is a direct quote, "deeply alarmed." Most people heard this as a headline about AI spending. They missed what it actually was: the first serious public articulation of a new compensation philosophy for high performers in the age of AI agents. Let's think through what that actually means.
For decades, total compensation has had three components: base salary, bonus and equity. The specific weighting varies by company, stage and role. But the structure has been remarkably stable. It is about to gain a fourth component. Tomasz Tunguz of Theory Ventures has called it the "fourth component of engineering pay," being AI compute in the form of token budgets, API credits or access to agentic AI infrastructure. The framing is still early and the category doesn't have standard names yet, but the logic is compelling and the direction is clear. Here is why. If AI agents genuinely multiply individual output, and if access to a well-orchestrated fleet of agents means that one person can do the work of ten, then the value of that access is not a perk. It is core to what a high performer can produce. Restricting it, or failing to provide it, is the equivalent of giving a surgeon a toolkit from 1970. The talent is the same. The output capacity is categorically different. High performers understand this. The question is whether they are demanding it yet.
A brief clarification for anyone who hasn't spent time in the weeds of this. AI model usage is measured in tokens, which are the units of text that large language models process. When you interact with an AI system, every word you input and every word it outputs consumes tokens. Tokens cost money. Running complex, multi-step AI agent workflows, particularly the kind that involve multiple agents working in parallel across long contexts, can consume substantial quantities of tokens at non-trivial cost. This means that a person's ability to use AI at maximum leverage is constrained, in practice, by the token budget available to them. An organisation that provides generous token budgets can run AI workflows at a scale and complexity that organisations with restricted budgets simply cannot. The leverage gap this creates is real, measurable and growing. This is why Huang's framing matters. He's not talking about giving engineers a shiny tool. He's talking about ensuring they have the computational resource to operate at maximum leverage. The token budget is, in that sense, a performance-enabling capital allocation. And like all capital, the people who have access to it can do things that the people who don't, cannot.
Before we go further, it's worth being precise about the difference between how most people use AI and how high performers will use it. They are not the same thing, and the gap between them is going to define the next decade of the performance distribution. Most people use AI in what you might call assistance mode. They interact with it conversationally. They ask it to draft things, summarise things, explain things. They are the operator; the AI is a responsive tool. This is useful. It is not transformative. The ceiling on this mode of AI use is roughly the level of a very capable, very fast junior assistant. For average performers, access to this assistant is significant uplift. For high performers, it is table stakes. The mode of AI use that genuinely transforms output is agentic. In agent mode, you are not the operator of a single tool but the orchestrator of a system. You design workflows, deploy parallel agents, define objectives and constraints, monitor outputs, iterate on architectures and synthesise results that no single agent could produce alone. You are, in effect, a manager of a team that never sleeps, never complains and can be replicated at near-zero marginal cost. The difference in leverage between these two modes is not incremental. It is structural. A person operating in assistance mode gets marginally faster at their existing work. A person operating in agent-orchestration mode can compress what previously required a team of people into a workflow they manage alone. Here is the uncomfortable part. The capability to operate in agent mode is not evenly distributed. The research on AI Interaction Competence, which is the ability to elicit, filter and verify model outputs effectively, shows that this skill is predicted not by prior knowledge or academic credentials but by a specific set of meta-cognitive abilities: the capacity to decompose complex problems, evaluate probabilistic outputs and iterate on system design under uncertainty. These are, not coincidentally, the same abilities that differentiate high performers from average performers in almost every professional domain. AI assistance narrows the performance gap. AI agentic capability widens it. This is the mechanism underlying the Moses Curve, and it's moving faster than most people have registered.
Go back to the recognition loop framework from our earlier piece. The signals that matter to high performers are costly ones: signals that require genuine underlying ability or performance to obtain and therefore carry real information. Token budgets are becoming one of those signals. Here's why. A generous, unrestricted token budget is only valuable to someone who can actually use it. The computation required to run complex multi-agent workflows, to design the architectures, orchestrate the agents, evaluate the outputs and iterate on the system, is a skill that most people don't have and can't fake. Giving that budget to a mediocre performer produces nothing. Giving it to a high performer with genuine agent-orchestration capability produces a 10x output multiplier. Companies that offer substantial token budgets as part of compensation are, whether they know it or not, sending a signal about what kind of people they want and what kind of output they expect. It is a costly signal in exactly the Spence sense: it requires real capital to maintain and it can only be justified if the recipient can convert it into disproportionate output. If a company is offering you a generous token budget, pay attention. They believe you are the kind of person who can turn computation into disproportionate output. If they're not, if tokens aren't even part of the conversation, ask yourself whether they understand how work is about to change, because the evidence suggests they don't.
Let's be honest about both sides of this, because the critical counterargument is serious. The case for tokens is straightforward: compute is leverage. Leverage that is currently underpriced and undervalued in most compensation discussions. High performers who can convert token budgets into disproportionate output are worth dramatically more to organisations that give them the compute to do so. The appropriate response, from a compensation design standpoint, is to make token allocation an explicit and significant component of total compensation rather than an incidental expense buried in an IT line item. The case against is also real. Critics have described token budgets as "company scrip," a form of compensation that sounds valuable but is actually rebranded operational expenditure. The tokens are consumed in service of the company's objectives, not the employee's. They don't vest. They don't appreciate. They can't be taken to the next job. If a company substitutes token budgets for cash or equity, it is offloading compute costs onto the employee's compensation package without giving them anything that compounds or transfers. Both arguments are correct. They're addressing different things. Token budgets are valuable compensation when they are additive to cash and equity, representing genuine leverage expansion. They are compensation substitutes, and therefore problematic, when they replace cash or equity. The high performer's job in the negotiation that's coming is to hold that distinction clearly and demand both.
If you are a high performer, and if you're reading this on TalentGiants, the probability is high, here is what the above means in concrete terms.
Learn agent orchestration now, before it becomes table stakes. The window in which agentic AI is a differentiating capability rather than an expected one is closing faster than most people think. Right now, the ability to design and run multi-agent workflows is rare enough to create genuine leverage. In two to three years, it will be the equivalent of being able to use Google. The returns to early expertise in this domain are asymmetric.
Start tracking your token consumption. You cannot negotiate for something you haven't measured. Begin understanding what your actual AI compute usage looks like, how many tokens your workflows consume, at what cost and what output that produces. That quantification becomes the foundation of a compensation argument: "I consumed X in compute and produced Y in output. That's a leverage ratio of Z. Here's what that budget needs to look like."
Make token budgets an explicit line in compensation discussions. Not a hypothetical future item but a present one. The conversation Huang is having publicly about compute as compensation is the conversation you should be having privately in your next offer negotiation. What is the token allocation? Is it unrestricted? What happens if you consume more than the baseline? These are legitimate questions, and the answers tell you a great deal about how seriously a company is thinking about AI-era performance.
Evaluate employers on their AI infrastructure. Companies that have invested seriously in AI infrastructure, that have access to frontier models, that have built internal tools on top of them and that have genuine agent capability, are categorically more interesting environments for high performers than companies still deciding whether to allow ChatGPT access. The compute environment you work in is going to determine what you're capable of producing. Choose accordingly.
Watch the Moses Curve. The divergence between high performers and average performers is accelerating. The people who learn to operate AI agents at maximum leverage will compound their advantage in ways that are genuinely difficult for average performers to close. The recognition loops, the compensation structures and the environments that will reward this capability are forming now, at exactly the kind of hypergrowth companies that appear on the TalentGiants leaderboard.
Here is what this is really about, underneath all the detail. We are at an inflection point in what it means to be a high performer. The attributes that defined excellence for the previous generation, being deep individual expertise, superior execution speed and exceptional communication, remain valuable. They are not sufficient. The high performer of the next decade is someone who combines those attributes with the ability to orchestrate AI systems at scale. Someone who can design the workflow, manage the agents, evaluate the outputs and synthesise the results into something that no individual person and no unorchestrated AI could produce alone. That person is going to be extraordinarily rare for the next several years. And they are going to be extraordinarily valuable. The organisations that recognise this now, that build the AI infrastructure, provide the compute budgets and create the environments where this capability can be exercised, will attract those people. The organisations that don't will watch them go somewhere else. Compensation in tokens is not a gimmick. It's an early signal of a much larger reorientation in how the most valuable work gets done, how it gets recognised and what it pays. Pay attention to who's already offering it.
This article is part of the High Performance in the AI Era series. Read also: The Recognition Loop and Software That Makes You Better.