The hottest perk in Silicon Valley isn't a bigger equity grant or a corner office with a view. It's compute. Specifically, AI tokens the computational currency that powers tools like Claude, ChatGPT, and Gemini are being folded into engineering compensation packages as though they were stock options. And depending on who you ask, this is either the most logical evolution in tech pay since RSUs, or a masterful bit of corporate sleight of hand.
The Pitch
The argument for tokens-as-compensation is intuitive enough. Agentic AI has changed what a single engineer can accomplish. Where a developer once typed queries into a chatbot, they now orchestrate swarms of autonomous agents that write code, debug systems, and churn through tasks around the clock. The bottleneck is no longer talent alone it's the compute that talent can access.
Jensen Huang made this case at Nvidia's GTC conference this week, suggesting that top engineers should receive around half their base salary again in token budgets. By his math, a senior engineer might burn through $250,000 a year in AI compute. Tomasz Tunguz of Theory Ventures had been tracking the trend even earlier, noting that startups were already treating inference costs as a fourth pillar of compensation alongside salary, equity, and bonuses. Using data from Levels.fyi, he estimated that roughly one dollar in five of a top-quartile engineer's total package now goes to compute.
The logic is circular, in the best sense: more tokens make an engineer more productive, and more productive engineers justify the cost of the tokens.
The Boom Behind It
This shift didn't happen in a vacuum. The explosion of agentic AI systems that don't just answer questions but take independent, multi-step actions has sent token consumption through the roof. A prompt-and-response conversation might use a few thousand tokens. An engineer running a fleet of background agents can blow through millions in a single day without touching a keyboard.
The New York Times recently reported that engineers at Meta and OpenAI are competing on internal leaderboards that track token usage, turning compute consumption into a kind of sport. One Ericsson engineer in Stockholm reportedly spends more on Claude than he earns in salary, with his employer footing the bill. Token budgets are quietly becoming a standard job perk, the way free lunch or gym memberships once were.
The Catch
But not everyone is celebrating. There's a darker reading of the trend that deserves attention.
First, there's the productivity trap. If a company gives you $200,000 worth of compute on top of your salary, the unstated expectation is that you will produce at a dramatically higher rate. That token budget isn't a gift it's a bet on your output. And bets come with accountability.
Second, there's a headcount question lurking underneath the numbers. When a company's token spend per engineer approaches or exceeds that engineer's salary, the math around how many humans are needed to coordinate the work starts to shift. The same compute that empowers an engineer today could, in theory, reduce the need for their colleagues tomorrow.
And then there's the compensation illusion. Jamaal Glenn, a former VC and current financial services CFO, has argued that tokens may be a clever mechanism for inflating the perceived value of a package without actually increasing the things that matter most to an employee's financial future. Token budgets don't vest. They don't appreciate. They don't follow you to your next job or show up in a competing offer. If companies successfully normalize compute as pay, they could hold cash and equity flat while pointing to swelling token allowances as proof they're investing in their people.
The Verdict — For Now
The honest answer is that it's too early to tell. AI tokens could genuinely become the fourth pillar of compensation in tech, a meaningful resource that separates the engineers who thrive from those who stagnate. Or they could become the corporate equivalent of in-store credit: useful, but not the kind of value that builds wealth.
What's clear is that the framing matters. Tokens-as-tools and tokens-as-pay are fundamentally different propositions. Engineers would be wise to know which one they're being offered and to negotiate accordingly.
