Reed Albergotti on Anthropic's numbers, open source AI bans, and teasing his Satya Nadella interview

Reed Albergotti, Semafor's San Francisco bureau chief, covers three threads worth tracking: a coming Satya Nadella interview, the shape of the open source AI debate, and where the next bottleneck in AI investment lands.

Satya Nadella interview

Albergotti's interview with Nadella, recorded at Microsoft's CISO Summit, drops next week on his Future Proof podcast. The teaser question is pointed: given that Nvidia now out-earns Microsoft per quarter, and with the hyperscalers and SpaceX AI all growing fast, is there a world where Microsoft remains a great business but falls out of the Mag Seven? He says Microsoft's comms team audibly gasped.

The more substantive detail from the Nadella conversation concerns AI alignment. Talking to CISOs, Nadella expressed something close to disappointment with current models — he doesn't think alignment is working and has written off mechanistic interpretability as a near-term solution. His practical response is to wrap deterministic, guardrailed software around the models and work within those constraints. That's a notably sober posture from someone with as much riding on frontier AI as Nadella does.

“Is there a world in which Microsoft is still a great business a few years from now, but they're not in the Mag Seven? I heard a little gasp from the Microsoft comms team when I asked that question... The hyperscalers would love to ban open source. The open-fill funded groups are having dinners talking about open source is gonna be banned.”

Open source AI bans

Washington-funded groups are reportedly holding dinners to discuss banning open source AI. Albergotti is skeptical it goes anywhere. His argument is less about technical enforceability and more about the absence of a workable legal theory — there's no existing framework to hang a ban on, and every tech regulation with real teeth has created downstream headaches (the DMCA being his sharpest example, still complicating legitimate security research today).

The counterargument, raised in the conversation, is that banning enforcement is easier than it sounds — the file-sharing analogy holds to a point: torrents still exist, but Napster-style centralized hosting was killed and casual infringers get letters. The practical challenge is making a data center hosting open weights legally liable, which is a narrower ask than banning the concept outright.

Albergotti's read on why it probably doesn't happen anyway is market weight. He puts roughly $25 trillion of combined market cap behind the pro-open-source signatories on recent open letters. Whether or not Washington is "coin operated" (his phrase), that kind of economic gravity tends to determine outcomes.

He also broke a small scoop: Perplexity has open-sourced its sandboxing technology, which he frames as the more productive direction — crowdsourcing safety tooling rather than concentrating it in two or three Silicon Valley labs.

Capital as the next bottleneck

Albergotti's forward-looking concern is credit. He argues tech companies have been borrowing opportunistically — cheap rates, investor appetite — but banks are approaching their lending limits. The Mag Seven can self-fund through cash flows, but for everyone else, debt capacity starts to bind. He was planning to write up the "bottleneck economy" thesis for his newsletter that day.

His related observation on consumer AI is worth noting: the reason consumer applications haven't materialized at scale isn't product or demand — it's compute. There simply aren't enough data centers running frontier models for consumer workloads. His guess is that open weights and on-device inference get there first, a kind of Netscape moment before the cloud-native second wave arrives.