A VC just put a number on “AI startup”: grow 4x to 5x a year, or you're just a company. This is the Tech Podcast Podcast — today, Aileen Lee's growth math, Meta selling off its spare compute, and whether “scale” actually has a mechanism behind it. So let's start with the benchmark that just changed what a good deck has to say. One tap on follow, and we'll be back in your ears before you know it. Here's Aileen Lee at Digg:
How fast does your startup need to grow? According to @aileenlee, much faster than venture capitalists expected in the pre-AI era. Aileen, Floodgate’s @m2jr and Lerer Hippeau’s @BenjLerer joined TWiST to dig into massive Seed rounds, rising exit volume, startups rolling their own model (and why it might not be a good use of time!), and even how to grade the White House’s recent adventures in AI regulation!
So Aileen Lee finally put a number on it — 4x to 5x annually, or you're not really an AI company. All week it's been vague scale talk, and now a VC has turned the quiet part into a deck-ready benchmark. And that number is dragging Series A rounds up to $100 million — a price tag that used to belong to a Series C or Series D. The whole funding ladder just compressed. Here's my question, though — is 4x to 5x describing what's actually happening, or is it what VCs now demand to see on a slide? Because those are really different worlds for a founder who hits a still-excellent 3x. Right — at 3x, you're a great software business and a disappointing AI bet in the same breath. The benchmark reclassifies you, it doesn't just measure you. And it prices architecture bets differently. If 4x to 5x is the floor, does a company betting the whole stack on transformers sticking around get a premium over one that's hedging? The benchmark makes conviction look like the only rational move. Bloomberg writes:
Bloomberg’s Ed Ludlow breaks down Meta's plans to develop its own cloud infrastructure business aimed at selling access to AI computing power and models. Plus, the Trump administration lifts foreign access restrictions on Anthropic's Fable 5 AI model. And, Lime CEO Wayne Ting joins as the company gets ready to debut on the Nasdaq.
So two days ago, the story was everyone panicking about burning through tokens — how do we stop the inference bill from eating us alive? And now Meta, the company sitting on the actual compute, wants to rent you the surplus. Ed Ludlow's framing on Bloomberg is that Meta's building a real cloud business — selling access to compute and models. That walks it straight into AWS, Azure, and Google Cloud, who've owned that market for a decade-plus. Which answers the thing we kept circling: who actually pays the inference bill? Turns out Meta's raising its hand as the guy holding the meter. The word I can't get past is “excess.” You don't stand up a cloud sales org for spare capacity. You do it because you overbuilt on purpose and now need it on someone else's balance sheet. Right — call it surplus if you want, but it sure looks like a fourth hyperscaler wearing a Reality Labs jacket. Here's Sophie Buonassisi at GTMnow:
A $1.2 billion exit is the dream. Elias Torres calls it his biggest failure. The Drift co-founder joins Sophie Buonassisi on GTMnow to unpack why the headline number felt hollow, what he learned in the quiet stretch afterward, and why he jumped straight back in to build Agency, an AI company that runs your entire customer organization.
Elias Torres sold Drift for 1.2 billion dollars and calls it his biggest failure. I'm going to need to hear the mechanism on that one, because from here it sounds like a very expensive feeling. The honest version is that he chased the title and the number instead of the thing he was building — and Agency is the do-over. It's an AI company he says runs your entire customer organization. Runs your entire customer organization. Right after Aileen Lee's 4x-to-5x segment, that's a company that has to grow like a rocket or the whole thesis reads as a slogan. What I actually want from this episode is the quiet-stretch part — what he learned after the exit that changed how he's building the second time. Sophie usually gets people past the highlight reel. If he gives us the identity-after-exit stuff and then a real operating detail on how Agency runs support and sales, I'm in. If it's just 'customer obsession,' we've all been to that keynote. From Molly Rocket:
As you've probably noticed, AI companies are very focused on scaling right now. They're trying to scale up their data center capacity. They're trying to scale up the number of users they serve. They're trying to scale up the number of tokens they process. They're trying to scale up the amount of revenue they generate.
Okay, so the whole episode is built around the thing everybody's been asserting all week — AI needs scale. 8.2K views, 493 likes; that's real engagement for an hour-long explainer. So does Spanos actually give us a mechanism, or is this another hour of nodding at the obvious? Yeah, that's the split that matters. He lays out “factors” — data center capacity, user counts, token volume, revenue, all scaling at once. The pitch is that the combination is what makes it different from the dot-com run, not just bigger. See, four things scaling at the same time gets you closer to a structural claim. The useful part is whether he can show why they're coupled — why user growth burns tokens, and token growth forces more data center. If the answer is just, “well, everything's going up,” we're back in slide-deck territory. And it lands right after the Aileen Lee number we just hit — VCs pricing in 4x to 5x growth. If that's the new floor, this episode is basically the operator-side explanation for why VCs think it's even reachable. Right — Aileen names the demand, Spanos names the supply pressure. Put them in the same room and you've got the escalation. The benchmark's on the table now; the vague “scale” talk from earlier this week looks either honest or a little embarrassed. Packet Pushers, with Eric Chou:
Eric welcomes Eduard Dulharu, a veteran network architect and the Founder and CTO of vExpertAI, to talk about how agentic AI, open-source LLMs, and digital twins are changing network operations. Eduard discusses the rapid evolution of generative AI, draws parallels between AI’s current limitations and early network protocols such as Spanning Tree, talks about why network fundamentals are still essential, and discusses how models can be fine-tuned to meet specific use cases.
Okay, this one's quietly the most concrete AI story of the day. Eric Chou's got Eduard Dulharu — network architect, 25 years, military radar to enterprise — actually fine-tuning open-source LLMs for network ops. The Spanning Tree comparison gives it away. He's saying AI right now is like early network protocols — brittle, in need of guardrails, with a human still in the loop. You can hear the engineer in that. Right, and fine-tuning is the specific move here — you take a general model and retrain it on networking data so it stops hallucinating configs. That gives you a mechanism instead of another scale slogan. Which is funny, because two segments ago the whole room was arguing whether AI companies even need scale. This guy's answer is basically: small, tuned, domain-specific, and human-checked. Different religion entirely. And digital twins for the network, so the agent can test against a model of your infrastructure before it touches production. You can hear the eval-first instinct showing up in the least hyped corner of the feed. Got a take, a correction, or a tech story we should be tracking? Send it our way at techpodcastpodcast at lantern podcasts dot com. We really do read what you send.
You’ll find links to every story we covered today in the show notes. If something sparked your curiosity, head there and read a little deeper.
That’s Tech Podcast Podcast for today. Have a great Friday, and thanks for listening. This is a Lantern Podcast.