Addy Osmani, fourteen years at Google, the man who taught half of us how to make a web page load fast, announces he is joining Anthropic to work on Claude Code. Big move. But what actually woke me up was the repost above it: "new transfer from Google to Anthropic. We seriously need an engineer transfer market. Track their value, transfer history, time at each company... who is building this?" So you guys are waiting for a José Mourinho or a Pep Guardiola of tech hiring? What's wrong with you?
Here we go: the AI mercato
Summer 2025. Mark Zuckerberg decides to build Meta Superintelligence Labs and starts recruiting personally, like a club president who fired his scouts. Sam Altman goes on a podcast and says Meta is offering his researchers signing bonuses of up to 100 million dollars. Signing bonus. Not salary. Most of us get a hoodie and a Slack account.
Ruoming Pang leaves Apple's foundation models team for a package reportedly above 200 million. Meta puts 14.3 billion into Scale AI for 49% of the company, and everybody understands the real product is Alexandr Wang. Google pays 2.4 billion in "licensing fees" to Windsurf, and by pure coincidence its CEO Varun Mohan walks into DeepMind the same week, right after OpenAI tried to buy Windsurf for three billion. In football that's a hijacked transfer. In tech it's a non-exclusive licensing agreement. Same move, worse name.
And my favorite: Andrew Tulloch was reportedly offered up to 1.5 billion over six years. More than Real Madrid ever paid for anyone. He said no. Lolz. A few months later he joined Meta anyway. Double lolz.
LinkedIn is the new Sky Sports
The money is crazy, but honestly that's not the worst part. You know what I think? The worst part is that we watch it.
A few years ago an engineer changing jobs was an "I'm happy to share" post with forty likes, mostly from their old manager. Now a researcher moving from OpenAI to Meta is breaking news. There are threads breaking down the package, "sources close to the deal", reaction memes, and people who have never met the person arguing in the comments about whether it's a good signing for the club. Euronews literally called these people "the equivalent of franchise athletes".
Engineers have fans now. So I did the only reasonable thing. I made it a pack opening.
- Alexandr Wang: Scale AI to Meta, $14.3B for 49% of the club.
- Ruoming Pang: Apple to Meta, ~$200M package.
- Varun Mohan: Windsurf to DeepMind, $2.4B "licensing".
- Andrew Tulloch: Thinking Machines to Meta, said no to $1.5B, signed anyway.
- Addy Osmani: Google to Anthropic, undisclosed · 14 years at the club.
Meanwhile, in the Sunday league
Screen, online assessment, take-home, live coding, system design, behavioral, "culture" call. Then AI cheating showed up: 38.5% of 19,368 technical interviews had signs of AI assistance, so Google, Cisco and McKinsey brought back in-person rounds. Fair enough. But guess who pays: the honest candidate, who now does the take-home, then the flight, then the whiteboard, then the "defend your code" session. I did eight rounds at Canonical this year. Eight.
LTC: ??
Nobody asked Ruoming Pang to reverse a linked list. Nobody sent Addy Osmani a 90 minute HackerRank. They got scouted on what they built and shipped. The only card in the pack with a LeetCode rating is yours.
The only thing that predicted anything was the number of problems solved. Translation: grinding LeetCode makes you better at LeetCode interviews. Don't try to change my mind :)
Did Real Madrid make Mbappé dribble around cones before signing him?
So watch people play. Give them a slow service to debug. A system to design with a budget, a deadline and a legacy database nobody is allowed to touch. A messy bug, and watch how they investigate.
So, are you drafted?
Probably not. Me neither. Our card is the bronze one with a cute husky on it and LTC 99.
The people with nine figure contracts got hired exactly the way everyone should be hired. Somebody looked at what they built. So until recruiters start scouting like clubs, do the scouting for them. Put your work somewhere they can open it in a tab.
Frequently Asked Questions
- What is the "tech mercato"?
- Mercato is the word football fans use for the transfer window, the period where clubs buy, sell and loan players. I use it for what is happening to a very small tier of AI engineers and researchers in 2026: labs like Meta, OpenAI, Google DeepMind and Anthropic are trading them with signing bonuses, licensing deals used as transfer fees, and retention bonuses used as release clauses. It looks exactly like football, minus the shirts.
- How much are top AI researchers paid in 2026?
- For the S+ tier, numbers that used to be reserved for athletes. Sam Altman said Meta offered his researchers signing bonuses of up to 100 million dollars. Apple's former head of foundation models reportedly moved to Meta for a package above 200 million. Google paid 2.4 billion dollars in licensing fees mostly to bring the Windsurf founders to DeepMind. These are a few dozen people. The median engineer is not living in that market.
- Is the software engineering job market really bad in 2026?
- It is split in two. AI and machine learning roles are growing fast, but general software engineering postings are still far below their 2020 level according to Indeed, and the Stanford Digital Economy Lab found that employment of 22 to 25 year olds in AI-exposed jobs is about 19% below where it should be, up from 15% a year earlier. The mechanism is not mass firing, it is that companies stopped hiring juniors.
- Why did tech interviews become harder and longer?
- Two reasons stacked on each other. Fewer open roles means more candidates per role, so companies add rounds to filter. And AI cheating exploded: one analysis of over 19,000 technical interviews found signs of AI assistance in more than a third of them, so Google, Cisco and McKinsey brought back in-person interviews and most companies added take-home tests. Candidates pay for both trends with their time.
- Is LeetCode a good way to evaluate a software engineer?
- It measures how you think under time pressure, and that part is real. But it mostly measures how good you are at LeetCode. Interviewing.io analyzed over 100,000 interviews and found no correlation between a LeetCode rating and interview performance. If you want to test problem solving, give candidates an actual engineering problem: a slow service to debug, a system to design under real constraints, a messy bug to investigate.