Semafor · Oct 22, 2026 · 36m · 29m saved
With Satya Nadella, Chairman and CEO of Microsoft
Photo: Brian Smale and Microsoft, CC BY-SA 4.0 via Wikipedia
Microsoft runs its own frontier AI models about six months behind the rest of the industry, and spends roughly a fifth of what the leading labs spend to do it. Satya Nadella, Chairman and CEO of Microsoft, told Reed Albergotti of Semafor's Future Proof at Microsoft's CISO Summit in Seattle that the gap is deliberate. While rivals race to ship the newest model first, he would rather bet on judgment and governance than raw frontier capability.
"When you think Microsoft, you think trust, you think, 'Oh, this is the place where my data will be secure. This is where I do my taxes.'" — Satya Nadella
Satya has run Microsoft since 2014, staked the company's AI strategy on a partnership with OpenAI that runs through 2032, and is now building Microsoft's own in-house "superintelligence" models alongside it.
The full interview is covered here so you can skip it. 36 minutes of audio, 7 minutes of reading.
Here are the 5 takeaways that matter.
Key Takeaways
Microsoft now treats its own AI models as insider threats it can't fully trust
Alignment and "mechanistic understanding" haven't kept pace with scaling-law gains
Satya's biggest fear is "nested black boxes" — stacked opaque AI systems nobody can audit
Consumer trust has to be earned, not bought with flashier products than Meta's Muse
He calls Excel "the biggest hit consumer product Microsoft had"
Microsoft runs its frontier models about six months behind, at roughly a fifth of the spend
The OpenAI partnership, with IP access through 2032, covers the gap
Satya won't name a Microsoft successor and dislikes the term "AGI"
Asked who the next CEO might be: "All of them"
1. The Insider-Threat Metaphor
Satya said he has largely given up hope that near-term AI models can be made interpretable or reliably aligned, even as the raw capability keeps climbing.
Scaling laws beat his expectations; alignment research did not
"Alignment seems elusive... or even mechanistic understanding is tough. And I'm not saying these are forever going to be elusive, but they are today." — Satya Nadella
Since he can't guarantee a model's internals are safe, his conclusion is to build the safety outside the model instead. That is the thinking behind a Microsoft post Satya wrote recommending enterprises manage AI models with the same playbook — least-privileged access, identity, observability — they already use on human insider risk.
Models get the same label a company gives a risky employee
"You got to treat... these intelligences that are inside, that have all the capability to be an insider risk. And so, therefore, that's why I use that metaphor." — Satya Nadella
