There is a moment that is consistently described in various ways by various people, but it always involves the same fundamental surprise. When a software engineer answers the phone or opens their email, they may find a billionaire on the other end. They may be mid-career, just past their PhD, or even still enrolled in graduate school. not a hiring manager. not in charge of HR. The CEO. For myself. making a job offer to them. Speaking to Business Insider, a tech employee recalled how peculiar it was to get a direct call from Sam Altman while conducting an interview at OpenAI.
Another said they saw Mark Zuckerberg’s name in a chain of emails regarding a position they were thinking about at Meta. In the end, both accepted the positions. Board members and executive hires used to be the only candidates for that type of personal recruitment used at the top of the corporate ladder. It’s happening to engineers now.
| Field | Details |
|---|---|
| Topic | AI Talent War — Compensation Race Among Tech Giants |
| Key Companies Involved | Meta, OpenAI, Google DeepMind, Microsoft AI, xAI |
| Average ML Engineer Salary (U.S., 2025) | $175,000 (up to ~$300,000 at senior levels) |
| Top Reported Offer | Up to $300 million over 4 years (Meta to OpenAI researchers) |
| Notable Case | Matt Deitke, 24 — offered ~$250 million by Meta; dropped out of UW doctoral program |
| Meta’s Superintelligence Lab Investment | $14 billion (including Scale AI deal with Alexander Wang) |
| Cost to Build GPT-4 (2023) | ~$79 million (Stanford AI Institute estimate) |
| Cost to Build Google Gemini 1.0 Ultra | ~$192 million |
| London ML Engineer Salary Range | £140,000–£300,000 |
| Key Industry Risk | Healthcare, insurance, logistics priced out of AI talent market |
| Reference Website | Stanford HAI – AI Index Report |
It is now truly challenging to handle the numbers associated with these discussions on a regular basis. According to reports, Zuckerberg offered senior OpenAI researchers compensation packages of up to $300 million over four years, with more than $100 million in the first year alone. Matt Deitke, a 24-year-old student at the University of Washington pursuing a doctorate in computer science, reportedly received an offer of about $250 million to leave academia and join Meta’s new AI Superintelligence Labs. He accepted it. Meta gained whatever it is that someone like that produces, while the university lost a PhD student. The conversation reveals that value isn’t currently being assigned in classrooms but rather in the technology economy.
Even though it may appear that way from the outside, irrationality is not what is causing this. The cost of developing a frontier AI model is extremely high; according to Stanford’s AI Institute, Google’s Gemini 1.0 Ultra cost about $192 million to train, while OpenAI’s GPT-4 cost about $79 million in 2023, and those figures have been rising ever since.
Paying $10 million or even $50 million for one engineer who can significantly improve the result begins to seem like a reasonable line item when you’re already investing that much money in a single model. Alexandru Voica, an AI policy specialist at Synthesia and former Meta employee, put it simply: if you’re spending $1 billion to create a model, it makes sense to spend tens of millions to hire the experts who create it. Even though the math is uncomfortable, it’s still math.
The volatility of the entire situation stems from the supply side of this equation. There are only a few thousand researchers worldwide, and that’s being generous, who can truly train, refine, and significantly enhance large language models at the frontier level. The number of graduates with machine learning credentials from universities is increasing annually, but those who have worked specifically on model architecture, scaling behavior, and training dynamics at the cutting edge for five or ten years are not easily replaced.
The demand has skyrocketed. The supply hasn’t changed much. Even though the resulting salaries feel more like the GDP of a small nation than a paycheck, the wage inflation that follows is fundamental to economics.
The speed of the shift is almost dizzying to watch. A senior software engineer at a large tech company might have made $300,000 in total compensation fifteen years ago, which was already exceptional by most professional standards. For some AI roles today, that figure represents the floor rather than the ceiling. Prior to bonuses, equity, or signing packages, base salaries at top AI labs have reached $440,000. According to recruitment firm Robert Walters, machine learning engineers in London are earning between £140,000 and £300,000, which was previously the salary of managing directors of investment banks and partners at law firms.
All of this collateral damage is ending up somewhere that isn’t given enough attention. Businesses that actually require AI expertise to update outdated infrastructure, such as insurance, healthcare, and logistics, are unable to match what Meta or Google can offer. It is a “massive opportunity gap” that is actively unsustainable, with entire industries unable to access talent while a small number of technology companies absorb it, according to Mark Miller, the founder of an insurance AI startup.
That framing seems appropriate. It’s possible that the current concentration of AI expertise within a small number of large companies is causing a structural issue for the economy as a whole. This issue won’t become apparent for several years until hospitals and logistics networks realize their aspirations for AI are essentially locked behind compensation packages that they will never be able to match.
The decision is genuine but difficult for the engineers themselves. The unique experience of working inside a fast-moving machine that doesn’t always pause to ask questions comes with a $50 million package at Meta. Startups pay less, but they give employees a greater sense of ownership and a direct connection between their work and the future. Although it may be an overly literal metaphor, Voica’s description of it as a cog versus a driver captures something real about the trade-off. Some individuals steal the cash. The impact is what some people desire. It seems that some are offered enough money to make the difference irrelevant.
Whether the current economics hold true is still up for debate. The salary pressure should theoretically lessen if the cost of training frontier models drastically decreases, either through improved hardware, more effective architectures, or some other development that researchers are most likely already working on somewhere.
However, the engineers who know how to create these systems are currently in a position that virtually no technical workforce has ever held, as model costs continue to rise and the competition between Meta, OpenAI, Google, and Elon Musk’s xAI shows no signs of abating. They are currently the world’s most expensive professionals. And no one seems to understand that more than the CEOs who are on the phone.
