AI Engineers Turn Down 40% More Money to Stay at OpenAI

OpenAI posted four machine learning engineer roles last Tuesday at $265,000 base. By Friday, 312 people had applied. Meanwhile, a Series B startup down the street in SOMA can't fill the same role at $370,000.

This is the paradox gripping AI hiring in May 2026: demand outstrips supply by 3.2 to 1, yet engineers routinely reject offers that pay 40% more. The reason isn't what most operators think.

The Numbers Don't Lie

Q1 2026 saw $300 billion flow into startups globally, almost entirely driven by AI ventures, according to jobsbyculture.com. These companies need engineers. Badly. The typical Series A or B AI startup now offers senior engineers $350,000 to $400,000 base salary, plus equity that could be worth millions.

Major labs pay less in cash. TechInterview's compensation analysis puts OpenAI's senior engineer band at $250,000 to $300,000 base. Anthropic runs similar. DeepMind comes in slightly lower, offset by Alphabet RSUs.

Yet the talent flows one direction. OpenAI and Anthropic see acceptance rates above 85% for engineering offers. Smaller AI startups struggle to break 30%.

What Engineers Actually Want

The answer emerged from exit interviews with 47 engineers who turned down startup offers in Q1. Three factors dominated their decisions:

First, compute access. An engineer at OpenAI gets priority access to GPU clusters that would cost a startup $2 million monthly. "I can run experiments in hours that would take weeks elsewhere," one engineer told Job Lobster's career comparison guide. Startups promise GPU budgets. Labs deliver them.

Second, proximity to frontier research. The major labs employ the researchers who invented the architectures everyone else uses. Working alongside them accelerates learning in ways a 40% salary bump can't match. One Anthropic engineer described it as "the difference between reading papers and writing them."

Third, equity structure matters more than amount. OpenAI's Profit Participation Units and Anthropic's pre-IPO stock come with complex terms, but they're tied to companies with clear paths to liquidity. A Series B startup might offer 0.5% equity. But 0.5% of what? The median AI startup valuation dropped 31% in Q1 2026.

The Hiring Playbook That Doesn't Work

Smaller AI startups keep making the same mistake: they lead with money. High-Signal Hiring's Neil Matthams calls this "the FAANG trap." Startups see big tech salary bands and add 30%. They assume engineers optimize for total comp.

They don't. Not in AI. Not in 2026.

The successful AI startups stealing talent from major labs take a different approach. They offer specific technical advantages: a novel architecture to explore, a unique dataset, or a problem the big labs won't touch. Cohere pulled three senior engineers from OpenAI by giving them ownership of the entire multilingual model roadmap. Adept got two from Anthropic by offering first authorship on papers.

What Changes the Math

Two scenarios flip the equation:

Late-stage startups with imminent liquidity events suddenly become attractive. When Inflection AI announced its Microsoft partnership in March, applications from major lab engineers jumped 400%. The combination of high comp and near-term liquidity beats even OpenAI's PPUs.

Specialized roles also see different dynamics. MLOps engineers, data pipeline specialists, and deployment engineers care less about frontier research. They optimize for scope and compensation. A startup offering $400,000 to own the entire inference stack beats a narrow role at DeepMind.

The Reality for Operators

If you're hiring AI engineers in May 2026, here's what works:

Stop competing on base salary. You'll lose. OpenAI engineers making $265,000 aren't leaving for $370,000. They might leave for technical ownership, published research, or a liquidity event in 18 months.

Be specific about compute budgets. "Generous GPU allocation" means nothing. "$200,000 monthly compute budget per engineer" means something. Put hard numbers in the job post.

Show the technical depth. List the papers your team has published. Name the models you've shipped. Specify the scale of data you work with. Engineers evaluating roles scan for these signals.

The 3.2x demand ratio is real. But it measures job posts, not viable opportunities. The engineers rejecting 40% pay increases aren't irrational. They're optimizing for variables that matter more than cash in a field moving this fast.

The winning move isn't a bigger number. It's a better story.