The past few weeks have shown one thing: interest in artificial intelligence has not only not waned, it has actually increased. In addition to new language models such as ChatGPT-5 and the flood of examples of where AI is used today, there is now also a discussion about the salaries of AI experts.

The catalyst was Facebook’s parent company Meta, which, like Apple, has fallen somewhat behind in the development of competitive language models. And now Meta CEO Mark Zuckerberg has pulled the emergency brake, or rather opened the locks on his treasure chest wide.

He offered a hundred million dollars in sign-up bonuses to a handful of AI experts who worked for competitors and whom he wanted to have at Meta. In the case of Matt Deitke, the 24-year-old founder of an AI startup, Zuckerberg even increased the offer. After Deitke rejected a $125 million offer, Zuckerberg met with him personally and doubled the offer to $250 million. One hundred million in the first year, the rest over the next four years. Deitke was so perplexed that he asked friends and colleagues what he should do. Ultimately, he accepted the offer.

Now, this may seem extreme to us. If even the starting salaries of AI doctoral students in the low single-digit million dollar range are incomprehensible to many of us, then salaries in the tens or hundreds of millions are even more so. Most of us don’t earn that much in our entire lives.

But then again, we are not only used to such sums in the world of sport, we also approve of them. Lionel Messi, for example, signed a seven-year contract with FC Barcelona in 2009 that guaranteed him €12.5 million a year and earned his old club a transfer fee of €250 million. In 2022, Cristiano Ronaldo moved to Saudi Arabian club Al-Nassr for two and a half years for a whopping €500 million. For what? So that the two of them can kick a leather ball around on the field?

We may still find these sums crazy when it comes to athletes, but we accept them for “our soccer heroes.” So I think it’s only fair that intellectual athletes should also be able to earn such high incomes. And it makes sense for companies. Here’s why:

For GPT 3.0/3.5, the training alone cost over $60 million in electricity costs. The training was carried out on a GPU cluster consisting of 25,000 Nvidia H100s at a unit price of $10,000, for a total value of $250 million. New language models are trained on significantly more GPU clusters, using the latest Nvidia Blackwell GPUs, which can cost between $30,000 and $40,000 each. A cluster can therefore easily cost one or two billion dollars in hardware costs. If a training run lasting several months is then carried out on it, this adds up to hundreds of millions of dollars more. One mistake, one lapse in concentration, and the costs explode, or, what is much more serious, you fall behind in the race for the top spot in language models. Possibly even irretrievably behind.

If you have the top AI talent in-house who can prevent such mistakes in advance or come up with more efficient algorithms and thus get to market faster, then the costs for these top talents are quickly recouped. Apart from that, it is currently important to occupy the emerging market for AI applications with its AI language models as quickly as possible. The future business potential runs into the tens of trillions of euros, so a hesitant approach and saving in the wrong place would be a serious economic mistake. And Meta, like Apple, is currently struggling against heavyweights such as OpenAI and Microsoft with ChatGPT, Google with Gemini, and Anthropic with Claude. Here, too, as with smartphone operating systems such as iOS and Android, there may soon come a point where there is no room for a third, fourth, or fifth language model alongside the market leaders. And with that, no business.

The competition for the best talent is causing collateral damage. Tesla has been hit hard, with its AI team working on Tesla Dojo—the supercomputer cluster for Full Self Driving (FSD)—being almost completely poached by the competition. Musk had no choice but to dissolve Tesla Dojo and merge it with X.ai.

In other words: yes, the sums offered for top AI talent are crazy, but somehow they make more sense than the sums football players receive when they switch teams.