Job candidates at OpenAI have started asking a new question in interviews: how much AI compute will I get?
Thibault Sottiaux, who leads engineering on OpenAI's Codex coding tool, says he's fielding this question more and more. "I am increasingly asked during candidate interviews how much dedicated inference compute they will have," he said. OpenAI President Greg Brockman put a finer point on it: "The inference compute available to you is increasingly going to drive overall software productivity."
The idea gaining traction in Silicon Valley circles is that AI inference - the cost of actually running AI models to generate code, text, and answers - should be treated as a formal part of engineer compensation, right next to salary, bonus, and equity.
The Math Behind the Hype
Tomasz Tunguz of Theory Ventures is the loudest voice pushing this framing. He predicts engineers "will start to be paid in tokens" in 2026 and advocates measuring "productive work generated per dollar of inference cost."
Here's how the numbers break down: a 75th-percentile software engineer earns roughly $375,000. Add an estimated $100,000 in annual AI inference costs, and you're at $475,000 in total compensation - with over 20% attributed to AI usage. Tunguz himself says he spends about $12,000 a year automating 31 daily tasks through AI tools.
Peter Gostev from Arena took it further, suggesting that OpenAI and Anthropic should build recruitment sites listing "token budget for the job alongside the salary range."
Meanwhile, at least one compensation submission on Levels.fyi already lists a Copilot subscription as a workplace benefit. GitHub Copilot's premium tier doles out monthly credits - one user reported 300 credits per month at roughly $0.04 per request.
A Perk Is Not Compensation
Here's where I push back: calling compute access "compensation" is a stretch. Giving engineers AI tools is no different from giving them a laptop, a good monitor, or air conditioning. These are things you need to do the job. Nobody calls office WiFi a compensation pillar.
The real story is simpler and more interesting. AI inference is becoming a significant line item for engineering teams, and companies are starting to budget for it per-engineer rather than as a blanket infrastructure cost. That shift matters for finance teams, hiring managers, and anyone negotiating a job offer at a company that heavily uses AI coding tools.
But framing inference budgets as the "fourth component of compensation" feels like VC narrative construction. If your employer hands you $100,000 worth of compute to do your job, that's an operating expense - not your pay. The test is simple: can you take it home, sell it, or spend it on rent? No? Then it's a tool, not compensation.
The part worth paying attention to is the signal underneath the noise. Usage per user at OpenAI is growing faster than user growth. Engineers are competing internally for GPU access based on project priority. Compute scarcity is real, and having more of it does translate to higher output. Companies that give engineers generous AI budgets will attract better talent - not because compute is compensation, but because good tools make for better workplaces.
That's always been true. We just didn't call a Herman Miller chair a compensation pillar.