Tokens are becoming the new unit of compensation in the era of artificial intelligence.

When it comes to stock trading, just look at analysis reports by the Golden Qilin analyst team—authoritative, professional, timely, and comprehensive—helping you uncover high-potential theme opportunities!

In the past, when engineers talked about compensation, they talked about annual salary, bonuses, stocks, and options. Today, the winds in Silicon Valley are quietly changing. At GTC 2026, NVIDIA CEO Jensen Huang offered a striking assessment: in the future, Tokens may become part of engineers’ compensation. Almost at the same time, an OpenAI frontline manager also revealed that an increasing number of job seekers are starting to care about how much dedicated inference compute they’ll actually receive after joining the company. Compensation data platform Levels.fyi observed that some engineers have already included AI subscription services like Copilot in their compensation and benefits.

A new line of evidence is becoming clearer: Tokens are moving from technical units of measurement, to commercial pricing units, and then onward to social distribution units. This means AI is no longer just a new tool—it is fostering a new yardstick for value. This shift is not happening only in Silicon Valley. Public data shows that China’s average daily Token calls have grown by more than 1,000 times over the past two years. The National Data Bureau has also made it clear that Tokens are the “settlement unit” connecting technological supply and commercial demand. This indicates that Tokens are accelerating their transformation from technological language into industry language.

Many people’s understanding of Tokens still stays at the level of technical terms. They are the basic unit for measuring the information processed by models; today, mainstream large-model services already widely charge by Tokens. But what is truly worth paying attention to is not how they are charged—rather, once a unit of measurement leaves the back office and enters the market, it does not stop at “technical convenience.” It embeds itself in companies’ cost accounting, enters product pricing logic, seeps into organizational incentive mechanisms, and ultimately even touches the order of social distribution.

In the industrial era, labor was measured by hours worked; in the internet era, platform value was measured by traffic. In the AI era, value is increasingly likely to be converted in terms of Tokens.

Jensen Huang compares large-scale data centers to “Token factories,” and this metaphor is key. It reminds us that data centers are no longer just warehouses that provide compute power; they continuously transform electricity, chips, data, and algorithms into standardized outputs that are consumable, billable, and tradable: Tokens. In other words, a data center is not only infrastructure—it is more like a production workshop of the new industrial era. And Tokens are the brand-new products coming off the line nonstop inside the workshop.

Once you understand it this way, many phenomena connect. Why do companies care more and more about model call quotas? Why, during hiring, do people start asking how many Tokens—and how much inference budget—can be allocated to a position? Why is what each side is competing for not just the ranks on model leaderboards, but chips, compute power, and inference capability? Because in this system, whoever holds the production capacity for Tokens controls the new channel for value export; whoever has the authority to configure Tokens controls the new power of allocation.

In the past few years, people were more concerned about whether models would become smarter, whether they would replace humans, whether they could write articles, or whether they could draw pictures. But for enterprises and capital, the more critical questions have already become: how will intelligence be measured, priced, configured, and allocated? Tokens matter not because they are mysterious, but because they are becoming that new yardstick.

And a new value scale only becomes an organizational reality once it is truly written into budget spreadsheets and paycheck stubs.

If the changes above mainly took place in server rooms and capital markets, then the new developments in Silicon Valley’s hiring market show that Token logic has started to enter companies from within. In the past, when technology companies competed for talent, they mainly relied on three things: salary, bonuses, and equity. Today, compute power is becoming the fourth pillar of compensation in Silicon Valley. OpenAI President Greg Brockman said bluntly that available compute power will directly affect software development efficiency. Venture capital firm Theory Ventures further predicts that by 2026, AI inference costs may become the fourth component of an engineer’s compensation, alongside salary, bonuses, and equity.

The significance of this is not that companies are issuing one more benefit. It is that enterprises have started allocating part of their means of production directly to core knowledge workers. Free lunches, gyms, and insurance are life benefits; Copilot, GPT quotas, the Cursor enterprise edition, and dedicated inference quotas are production benefits. Salary addresses current income; bonuses are tied to short-term performance; equity is tied to long-term expectations; and a compute budget directly affects output in the present. In the traditional office era, companies gave you a computer, a desk, and an email account; in the AI era, companies also have to give you a “second brain,” a “code co-pilot,” and “inference fuel.”

It’s also worth noting that leading Chinese companies have begun restructuring internal organizations around Tokens. Alibaba recently established Alibaba Token Hub, integrating core segments such as models, MaaS, and applications. What this reflects behind the scenes is the same kind of change: the basic unit for allocating resources within organizations is gradually shifting from “products” to “Tokens.” After all, in the AI era, the most important employee benefit is no longer just helping you live more comfortably—it’s helping you get stronger at what you do.

This change will also quickly raise the overall cost of top talent. According to related calculations, if an engineer incurs an additional $100,000 in annual inference costs, total labor costs could reach $475,000, meaning that in the future, more than 20% of compensation costs may come from AI usage expenditures. This shows that AI may not necessarily make high-end engineers cheaper; it may instead make them more expensive. Because generative AI is not simply replacing engineers—it is amplifying the leverage of top engineers. The better someone is, the more effectively they can turn high-quality models into higher output, and the more willing the company is to stack even higher Token budgets on them. The result is likely that regular engineer roles get squeezed more severely, while the combined costs of top engineers and efficient teams are raised higher.

When companies tilt internal compute resources, it will ultimately spill over into the labor market and form a new stratification. In the future, what companies will compete for is not just a specific engineer, but the combination of “engineer + model capability + inference budget.” On the surface, everyone is doing the same type of work; but in reality, the “digital external brains” behind different employees are not on the same level. Some people have enterprise-grade Copilot, dedicated API quota, and the right to call high-performance models; others can only use restricted versions. The new divide in the future labor market may not just be whether someone can program, but whether they have the right to access stronger compute power.

Of course, when compute power enters the payroll, it brings not only incentives but also governance challenges. Once inference budgets become formal resources, companies have to answer a few new questions: Who deserves more Tokens? Should Tokens be distributed evenly, or tilted toward contribution? Who uses Tokens efficiently, and who is wasting Tokens? In future performance evaluations, what they look at may not just be human productivity, but also “Token output per unit.” This means Tokens are moving from procurement questions to management questions. Whoever can allocate limited Tokens to the highest-value tasks, and whoever can reserve high-cost inference for high-return scenarios, is more likely to win in the next round of competition.

And once companies have already begun allocating compute power internally, it is not surprising that society will also start discussing compute power. For a long time, people have debated fallback mechanisms after technological substitution, focusing on “universal basic income.” Now, Altman is trying to rewrite the question: rather than give money, give compute power; rather than only ensuring consumption ability, empower production ability.

This may not become reality quickly, but it already reveals a change worth paying attention to: in the future society, what may truly be scarce is not only income, but generative capability. Whether a person can write efficiently, program, learn, and start a business increasingly depends on whether they can call powerful model capabilities at low cost. In industrial society, core fairness mainly manifests as income fairness and opportunity fairness. Society is more concerned about whether someone has a job, has income, and has basic security. But in an intelligent society, the meaning of fairness may quietly change, increasingly taking the form of fairness in compute accessibility, fairness in rights to use models, and fairness in digital productive capacity. So-called inclusiveness may not only mean handing out some money—it may more likely mean giving ordinary people a ticket to enter a new mode of production.

However grand the vision may be, it still has to be tested by reality in the end. A reminder from Microsoft CEO Satya Nadella hits the crux of the issue. Artificial intelligence must bring genuine improvements in fields such as healthcare and education, otherwise society will not accept such high-energy-consumption Token production for the long term. The real question is what these Tokens ultimately turn into: more precise assistive diagnostics, more widely accessible educational services, more efficient R&D collaboration, or merely the manufacturing of platform bills and capital stories. Society will not automatically recognize this kind of technological progress just because you have produced massive volumes of Tokens.

This is also where the AI industry needs the most vigilance today: Tokens are increasingly like a “new electricity bill.” On the surface, pay-as-you-go charging is reasonable and transparent; but once companies truly embed AI into key scenarios such as R&D, customer service, office workflows, marketing, and code generation, Token consumption will quickly expand from a controllable expense into variable costs that continuously devour budgets. Many companies think they’re integrating “intelligence,” but what they install first is a continuous charging system. Without task tiering, model routing, cache reuse, prompt governance, and cost monitoring, a so-called full embrace of AI may very well turn into full exposure to the AI cost curve.

So, what is truly worth asking is not only whether Tokens will become part of wages, and not only whether they can replace “universal income.” The more fundamental question is: who defines the value of Tokens, who decides how Tokens are allocated, and who constrains Token costs? And who ensures that Tokens ultimately translate into broad public benefits, rather than just the profit-and-loss statements of a few platforms?

In future social competition, on the surface it may look like model competition; deeper down, it is compute power competition; and even deeper, it is governance competition centered on Tokens. A technical term is worth writing about in a column not because it is new, but because it is penetrating systems across different layers: from the server room to the company, from the company to the market, and from the market to society. Tokens first began as units of measurement inside algorithms; later they became settlement units in the commercial world; and now there is faintly a trend toward moving toward social distribution units as well. In a few years, when we look back on today’s discussions by Jensen Huang and Altman about Tokens, inference budgets, and the “fourth pillar of compensation,” we may find that they were never only discussing a technical unit. Instead, society has begun to use a new yardstick—reassessing capabilities, reallocating resources, and rearranging distribution.

(Author: Hu Yiwei. A data practitioner. Author of the book “The Future Is Promising: Walking with Artificial Intelligence”)

A massive flow of information and precise insights—everything on the Sina Finance APP

Responsible editor: Yang Ci