AI agents just hit the money phase — and today we're asking who gets the margin as models and agents split apart. Quick reset if you're just joining us: agent-native development is moving past autocomplete and into real operational control. Gusto's CTO showed Claude Code-style workflows running inside an actual company; Autoresearch pushed the feedback-loop case for self-improving agents. And before that, a Paperdive benchmark found agents still blow past their action boundaries — and that explicit tool scaffolds and resource names matter way more than vague production warnings. This is Tech Podcast Podcast — and today's lineup is Vercel, Accel, and Alexandr Wang calling Meta's AI an appetizer. Which means the polished-narrative detector is fully warmed up. Let's start with Rauch. First real architectural thesis we've had all week — let's dig in. From Podcast Notes:
Naval sits down with three founders living at the edge of the AI curve: Garry Tan(Y Combinator), Daniel Francis (Abel Police), and Farbood Nivi(A-LIST). The conversation swings from practical AI tooling and open source model wars to bigger questions like ASI, Taiwan, COVID origins, and what humans are even for once AI hits expert level everywhere
So Naval's guest says if you're willing to spend a hundred grand a year on tokens, you can live like a normal citizen in 2028. Normal citizen. For the price of a house. The number under that is the interesting part — he's claiming 90,000x the inference we have now within three years. At that level, the story jumps from cost curve to a whole different order of compute. 90,000x is exactly the kind of figure that sounds precise and means nothing. Where's it grounded? Is that hardware, is that price-per-token, is that just vibes with a comma in it? What I'll give them — Naval owns that he missed 2020 to 2022 because he figured AI was fusion, always twenty years out. So when he says recursive self-improvement into ASI is the open issue now, at least he's naming the thing he already got wrong once. Yeah, and the honest bit is 'the last creative mile.' Pro-level at everything, sure. But whether it gets true originality — that's the one question they don't dress up. They point at the IMO math progress and go, is that a real breakthrough or just great pattern-matching? That's the useful frame in a two-hour episode that also detours through Taiwan and COVID origins. When it stays on the AI curve, it's actually asking something. From Russell Brandom at TechCrunch:
Known for its cloud infrastructure that allows developers to deploy agents without managing servers, Vercel has quietly become one of the most central companies in AI software. The company currently sees 6 million deployments a day, half of them triggered by coding agents, and more than 1 trillion tokens flow through the company’s AI gateway daily.
Guillermo Rauch at Vercel gives us the first episode this week that names the actual layer all those token costs run on: six million deployments a day, half triggered by coding agents, and more than a trillion tokens moving through their gateway daily. Half of six million. That's the agent-native shift we've been circling, but now there's a number on it — most of the software shipping through Vercel isn't a human hitting deploy anymore. And the framing is what interests me — splitting models off from agents. After three days of inference-spend stories, someone's finally describing the pipes those bills run through. So who owns the margin between the model and the agent? Right, and that's the piece I want him pressed on. Is Vercel genuinely sitting between the model layer and the agent layer, or is 'split off models from agents' just a really clean deck slide? Because those two things pay very differently. The tell is production. He says last year was prototyping — unleash the agents, everyone can build — and this year it's the realities of running hundreds of agents live. That sounds like an operator reaching for structure more than a pitch. Which is exactly when I trust the guy — when he stops selling the sky's-the-limit thing and admits the agents in production broke stuff. That's the detail I'll queue this for. This one's from Sourcery:
We go back through Accel's 40 years of history, starting with the 10% Facebook stake and the secondary they modeled at 5X, which under-shot the outcome by an order of magnitude, and how the firm now runs a global AI portfolio spanning chips, neoclouds, labs, and applications, with exposure across Cursor, Anthropic, and Nebius.
Okay, Accel doing 40 years of lore on Sourcery — and the number that jumps out is the Nebius PIPE. Matt Weigand put in $150 million, it's up 13X, and there's a $26 billion Meta deal attached to it. That's the one to sit with. After a week of us pricing token spend and inference bills, this is the firm that quietly owns the neocloud those bills run through — plus Cursor, plus Anthropic exposure. The whole agentic stack in one portfolio. Which is exactly where I get twitchy. They also cite the founding Facebook stake — modeled at 5X, undershot the real outcome by an order of magnitude. Great story. But three partners doing charming backstory isn't the same as one of them telling me which of Cursor, Lovable, or Nebius actually survives the commoditization squeeze. Well, they do hand you a live tell — Supabase growing 350% with nine million developers, on nearly all inbound. That's a real operating datapoint you can test. 36Kr writes:
About a year ago, Alexandr Wang joined Meta to lead its AI team. A year later, Meta's AI strategy has undergone a fundamental transformation. At the Bloomberg Tech 2026 conference, the Meta AI head shared the story of the company's transition from a follower to a contender — a complete vision spanning from open source to security-first, from "appetizer" to "main course", and from conversational models to personal agents.
Alexandr Wang says Meta's whole last year was the appetizer, and the main course is still cooking. Which is a lovely way to say: nothing you've seen yet is the point. It lands hard right after the Goldman secondaries number this week. If the head of Meta AI is telling Bloomberg Tech the real spend hasn't even started, every cost-cutting story we've run gets stress-tested. Sure, but did 36Kr get him to name the main course? Because 'from conversational models to personal agents' is a menu with no prices on it. He does frame it as follower-to-contender — open source, security-first, then agents. That's a sequence, at least. But once the agent layer sits on top of the model layer, the margin fight gets ugly fast. And that ties straight into the Rauch piece we just hit — split the models off from the agents, and suddenly Meta's 'main course' is a fight over which layer captures the check. From Metis Strategy:
What gives one enterprise an AI advantage over another? According to Atlassian Chief Product & AI Officer Tamar Yehoshua, it’s not simply access to the latest models, it’s the depth of enterprise context. In this episode of Technovation, Tamar joins Peter High to discuss how Atlassian is leveraging decades of workflow history through its Teamwork Graph to power AI across Jira, Confluence, and Rovo.
Tamar Yehoshua's now Chief Product and AI Officer at Atlassian — running Jira, Confluence, Trello, Loom, and the Rovo AI layer on top of all of it. The pitch here is enterprise context: the AI is only as good as the work graph it can see. And that's the actual interesting claim, right? After a week of Rauch splitting models from agents, context is the moat. Atlassian owns where your team's decisions actually live, not the frontier model itself. And it's the right resume for this. Glean before Atlassian, and Chief Product Officer at Slack straight through the Salesforce acquisition. She's built enterprise-context products three times now. Slack, Glean, now Rovo — she keeps landing on the layer that reads everyone's messages so the AI can summarize them. I just want to know if Metis got her to name what Rovo actually does that Copilot in Confluence doesn't, or if it's another 'connected teamwork' deck. That's the test for me too. If the enterprise-context thesis is real, she should be able to name the concrete thing Rovo retrieves that a general model can't — not just gesture at the graph. Got a tip, a correction, or a tech story you think we should tackle next? Send it our way at techpodcastpodcast at lantern podcasts dot com. We're always listening.
You'll find links to every story we covered today in the show notes, so if something stuck with you, take a moment to dig into the original reporting. That's Tech Podcast Podcast for today. This is a Lantern Podcast.