Satya Nadella, CEO at Microsoft, says “20% to 30% of our code is written by software.”
This came during a fireside chat with Meta’s Mark Zuckerberg at the LlamaCon conference on Tuesday. For those of us watching, it was a new phase of software development, where humans and machines work side by side, not in theory, but in the process of producing code.
This is beyond speeding things up. It’s about what’s being built, who’s building it, and how much control engineers still have.
Nadella pointed out that Python has been more responsive to machine-written code, while C++ — a language notorious for its strictness — is trailing. That tells us something important: the technology isn’t universal, and the shift isn’t clean.
Interestingly, when Nadella turned the same question back to Zuckerberg — how much of Meta’s code is machine-generated — Zuck didn’t offer numbers. “I don’t know,” he admitted, though he did confirm that Meta is investing heavily in automated development and expects machines to handle “half of our coding work” before long. The silence on current figures wasn’t lost on anyone.
Meanwhile, Google’s CEO Sundar Pichai, on an earnings call last week, claimed more than 30% of Google’s code is now machine-generated. But here’s the catch: no one seems quite sure what counts as “generated.” Are we talking about boilerplate scripts suggested by autocomplete tools or fully functional modules? Pichai didn’t say, and Nadella didn’t clarify either. That makes it hard to compare — and easy to overhype.
Microsoft’s Chief Technology Officer Kevin Scott added another layer of boldness to the mix. “I expect 95% of all code to be machine-generated by 2030,” he said recently. It’s a huge number. But these projections gloss over a real tension developers feel: where does assistance end and overreliance begin?
The tools themselves — those AI-driven co-pilots, reviewers, and debuggers — are becoming essential. At Microsoft, Nadella says they’re now baked into the workflow, not just for writing code, but for catching bugs and managing quality. It’s not about coding faster. It’s about trusting what the system builds — and knowing what to do when it breaks.
But even with all this automation, we’re not seeing uniform results. Some languages benefit more than others. Some teams resist the change entirely. And while executives trade stats and predictions, there’s still little clarity on what “AI-generated” actually means in practice.
The bottom line? We’re not in a world where machines are just helping. They’re now co-authors of the digital infrastructure that powers everything — and not everyone is ready for that level of partnership.
