Anthropic and OpenAI are pulling in more monthly revenue than Meta, Google, and Microsoft in AI — a16z has the slide, and now we get to see whether the numbers actually hold up. Welcome to Tech Podcast Podcast — and for once, we’ve got numbers we can actually kick the tires on: a revenue run-rate claim, a 7x PR throughput number from real Devin users, and an open-source thread that literally got named "Please Do Not Vibe Fuck Up This Software." Power law money, autonomous weapons, Greg Brockman’s little infrastructure victory lap, and the rsync revolt — this week finally has receipts. Let’s start with the a16z numbers, because if those hold up, the whole week looks different. a16z writes:
But just maybe to contextualize what's happened since then basically, Anthropic and OpenAI are adding more revenue per month than Meta, Google, or Microsoft. They are already at that scale of revenue…
The circulating a16z slide has one number I keep coming back to: the top-1% exit floor goes from $10B to $20B to $32B. That’s three points in about 24 months, and they’re also saying the combined Anthropic-OpenAI revenue run rate is nearing $200 billion. That’s not a vibe — that’s a floor you can test. The $200B run-rate line is the one I want to stress-test. It’s framed as "I wouldn’t be surprised if," which is a projection wearing a data badge. What’s the denominator here — current ARR, annualized monthly revenue, what? Because "wouldn’t be surprised" is carrying a lot of weight in that sentence. Even if you haircut the $200B claim, the $10B-to-$32B exit-floor jump is coming from their own February update and then a yesterday update. That’s a VC firm revising its own benchmark three times in 24 months. That doesn’t feel like spin — it feels like them scrambling to keep up with what’s actually clearing. Which is exactly why that Wednesday framing of "AI as operating layer" mattered — it sounded like a metaphor hunting for a balance sheet. Now it at least has one, or the start of one. The question I’d ask on stage is which revenue line is actually moving fastest: consumer subscriptions, API, or enterprise. Those are very different businesses. Latent Space, with Walden Yan and Cole Murray:
I think testing is actually a really interesting problem solving uh challenge for these AIs because if you wanted to do arbitrary testing like imagine you make a change that spans the front end and the back end to actually test that change we have to reason through what how do you first run these applications to orchestrate with each other with the right version of the code
Latent Space has Walden Yan and Cole Murray from Devin on, and the number worth stress-testing is 7x PR velocity from operators actually using the product. That’s the first throughput claim this week that isn’t just a slide deck projection. What I want to pull on is how they describe the hard part. They’re saying it’s not the "click the button" computer-use stuff — it’s reasoning through a full front-end and back-end change, then figuring out how to even trigger the thing to test it. That’s an actual constraint named out loud, which is more than most coding-agent guests give you. That lines up with the Boris Cherny "SaasPocalypse" thread from Wednesday. Platformer was gesturing at non-engineers picking up tools, but if the hard part is still cross-stack test orchestration, that puts a real ceiling on how far the "anyone can ship" thesis goes right now. 7x PRs is falsifiable — somebody can go check Devin customers’ GitHub histories. What I don’t hear yet is whether those PRs are landing in main or just stacking up in review. Throughput and merge rate are two very different numbers. Here's The Knowledge Project Podcast:
Um, space feels like a like a grand challenge, but I think that we are going to have such need for compute that we need to be thinking about all options.
Greg Brockman on The Knowledge Project — Shane Parrish asks him straight up about getting teased for over-investing in data centers, and Brockman’s answer is basically: our competitors are not having a good time on compute. That’s a primary-source vindication claim, not a press release. And Parrish barely pushes on it — "who’s laughing now" is basically a softball, and Brockman just takes the swing. I want the actual names, the timeline, the capacity gap. "Not having a good time" is still vibes without a number. Pair that with what a16z is saying this week — Anthropic and OpenAI outpacing Meta, Google, and Microsoft on revenue per month — and the infrastructure bet starts to look less like bravado and more like a measurable lead. What I’d want answered is which of the other OpenAI bets from 2022 that got laughed at are quietly paying off the same way now. Ross Douthat, writing in The New York Times:
What’s your understanding of the constraints on what the Pentagon allows a drone or an autonomous weapon to do without a human deciding, “Kill this person, shoot this person”? The more important thing is what it doesn’t say. It doesn’t say you’re not allowed to automate the kill chain.
Ross Douthat sits down with Chris Brose — he’s president and chief strategy officer at Anduril — and the most revealing exchange is about Pentagon policy on autonomous kill chains. Brose’s answer is basically: "You are not not allowed to do that." That double negative is doing real legal work. That phrasing is a tell. "You are not not allowed" is exactly how you describe a regulatory gap you’ve already decided to drive a drone through. It’s not a constraint — it’s a liability disclaimer in passive voice. This connects straight to what Ries was gesturing at earlier in the week with the DoD liability question — Anthropic drawing a line on autonomous weapons contracts. Brose is on the other side of that line, building the infrastructure for exactly what Anthropic said no to. And Douthat doesn’t really push on the denominator at all — like, who decides when "not not allowed" becomes "we did a war crime"? That’s the follow-up that would’ve made this episode. Instead we get Brose framing Anduril as a hub, and then the interview just moves on. From Linux Unplugged:
rsync’s founder came back, patched real security bugs with AI help, and triggered an open source meltdown. Plus, two more projects reject AI-generated code as the community’s newest fault line cracks wide open.
rsync’s founder comes back after years away, uses AI to patch real security bugs, and the community response is a GitHub issue literally titled "Please Do Not Vibe Fuck Up This Software" — that’s Issue 929, that’s a real thing. This is the open-source AI code-quality fault line landing inside a foundational Unix tool that’s been around since 1996. And it didn’t stop at a strongly worded issue. There’s a fork — fogti/rsync-nai — created specifically to preserve the last commit before what the fork author called "the first obviously LLM-influenced commit." That’s not discourse, that’s a SHA hash and a timestamp drawing a line in the sand. What makes this messy is that the bugs were real. The patches fixed actual security vulnerabilities. So the community isn’t saying the fixes were wrong — it’s saying it doesn’t trust the process that produced them, and that’s a much harder thing to sort out. And it’s spreading. Flathub updated its LLM policy to explicitly disallow AI in both submissions and apps being submitted. Zig’s Andrew Kelley announced a no-AI policy alongside a $670K foundation raise and a GitHub departure. GNOME Circle took a stance against what they’re calling AI slop. That’s four named projects in one week drawing a hard line — this isn’t a vibe, it’s a governance shift. Got a tip, a correction, or a tech story you think we should be watching? Send it our way at techpodcastpodcast at lantern podcasts dot com. We do read what you send.
You’ll find links to every story we mentioned in the show notes, so if something stuck with you, that’s the place to go a little deeper.
That’s Tech Podcast Podcast for today. Thanks for listening, and we’ll be back with you next time. This is a Lantern Podcast.