Today's headline: AI tokens, IPO returns, and the return of stock picking. Welcome to Tech Podcast Podcast. All-In with Chamath, Jason, Sacks & Friedberg has the details here. Dan Loeb on All-In, and the chapter that caught my eye was the homebuilder trade at 16:01 — an actual short, with a thesis behind it, instead of just vibes. Right, and the framing is 'stock picking is back' — a real pivot from Third Point's event-driven roots into quality and AI names, around the 8:47 mark. Sure, but when a guy who built a fund off message boards says short selling is a lost art, I want to know if Chamath made him walk through the homebuilder mechanics or just nodded along while he reminisced. The useful part for me is the strategy shift itself. He lays out the move from event-driven to quality-and-AI — a concrete read on where a multibillion-dollar book is actually rotating right now. And it's a hard left turn from this week's beat. We've been buried in token spend and inference math — Loeb talking Ross Ulbricht pardons and criminal justice reform at 22:15 is a completely different register. Satya Nadella keeps using this phrase 'token capital' — what does that actually mean, and why should a company care about it beyond just paying an AI bill every month? So Nadella's argument, in a conversation with Reid Hoffman published this week, puts tokens — the units AI models consume when they think and generate — closer to labor hours than to a utility bill like electricity. You can direct them, compound them, and make them proprietary. In his framing, human capital and 'token capital' are now deeply intertwined. The companies that win are the ones that embed their own expertise inside intelligent systems, so the AI gets smarter about their business over time, instead of just getting smarter in general. And this feels urgent because token consumption is exploding. Per TechCrunch this week, Uber burned through its entire 2026 AI coding budget by April, Microsoft pulled back Claude Code licenses from developers, and a Priceline employee said a routine Cursor contract renewal came back four to five times more expensive. Anthropic's revenue reportedly grew from nine billion dollars in January to forty-seven billion dollars in roughly five months — a whole Salesforce of new business in about five months, as one analyst put it. And a16z has been making the related case since earlier this year: AI investment doesn't look like any prior venture cycle because the capital flywheel runs through compute and tokens, not just code and headcount. But if token costs are already blowing past budgets, doesn't that undercut the 'compounding asset' story — like, how do you compound something that's also bleeding you dry? Exactly. The companies most likely to justify the spend are the ones getting proprietary leverage from AI, not just burning generic AI. Nadella's argument is that token capital only compounds when it's aimed at something the model can't get from anyone else — your data, your workflows, your institutional knowledge. The next tell is whether companies start talking about AI the way they talk about hiring, as a capability investment, or whether runaway token bills stay a bug to patch. This one's from Podcast Alpha:
Cerebras’ wafer-scale chip delivers 15-18x faster AI inference than GPU clusters by eliminating off-chip memory trips. OpenAI is a customer. The moat is real. Whether NVIDIA closes it with its next architecture is the unresolved question.
Cerebras IPO'd three weeks ago at $185, and Brad Gerstner — who's on the board and sitting on locked-up shares — is on the panel telling you to hold. I mean, yes, disclose the conflict, but that's a guy talking his own book in a fifty-million-dollar tuxedo. The conflict is real, but the architecture claim is still specific: 15 to 18 times faster inference by cutting out the off-chip memory trips, with OpenAI named as a customer. That's the first hardware-layer number we've had all week where you can point to who's using it. And here's what nobody on that panel really pressed Feldman on — Podcast Alpha flags it as the open question. Does NVIDIA's next architecture close that memory-bandwidth gap? Because if it does, that 18x advantage may just be a timing window. Right, and that's the denominator Kedrosky's been missing in his enterprise-AI skepticism. The number's finally on the table — now it either holds or it doesn't. The Planet Labs comp is the smart part of the pitch, though. It went public at $2 billion in 2021, stayed flat for two years, then ran 10x to fifty bucks a share. Gerstner's basically saying, be patient through the lockup. Sure, except 60% of Planet's revenue is defense and intelligence. Cerebras selling inference speed to OpenAI doesn't run on the same demand curve as selling spy imagery to the Pentagon. The comp feels more like a mood board than a model. This one's from Sequoia Capital:
Alfred Wahlforss, co-founder and CEO of Listen Labs, is building an AI agent that interviews your customers at a scale no focus group ever could—thousands of voice conversations at once, drawn from an audience of 30 million people. Alfred explains the counterintuitive finding underneath it all: people are often more honest with an AI than a human interviewer.
Listen Labs finally puts a number on the customer-listening pitch: 30 million people, thousands of voice interviews running at once. After a week of token math in the abstract, that's an actual dataset size you can argue with. And the part I want to poke at is the 'at once' bit. Thousands of voice conversations at once out of 30 million — does the same pool keep getting pinged? Because the whole pitch lives or dies on data quality. Wahlforss says 80% of their engineering goes into building the right audience rather than the agent. That's the tell — the panel is where the defensibility lives. The line that got me: people are more honest with the AI than with a human interviewer. Which, sure — nobody's performing for a robot. But it also means your richest data comes from people who'd lie to your face. Fun foundation. And he's reaching for a billion-person audience, stratified by expertise — who's actually a sneaker early adopter, not just any respondent. The P&L question I'd want answered: what does running that panel cost versus a traditional focus group? Right, because Nadella's 'token capital' framing we hit earlier assumes the unit cost keeps falling. This is the demand side of that — if listening gets cheap enough, you never stop interviewing. Infinity surveys. The Upstarts, with Alex Konrad:
Within the wonky world of corporate governance, and compliance audits, her startup is a big deal. Founded in 2018, Vanta works with more than 16,000 customers today, including tech darlings Factory, Lovable and Ramp; it reached a $4.2 billion valuation last July. And so the startup once dismissed by VCs as too niche now has to fend off a pack of startups that nip at its heels.
Okay, the headline buries the actual story. Vanta's CEO admits she faked rivals early on — a compliance company doing competitive theatrics. That's the line you don't expect a governance CEO to say out loud. The useful bit is the frame she gives it — 'infinity work.' Everyone's been arguing about whether AI gets cheaper. She's saying the work itself expands to fill whatever capacity you give it. And the jab is good — rivals 'have to actually do compliance this time.' She won't name them, but she's saying the dead ones faked it without doing the work. From a company built on auditing other people, that's a flex. Sixteen thousand customers, a $4.2 billion valuation — in a market VCs called too niche in 2018. The 'infinity work' frame pushes back on the token-capital story we just unpacked. If demand keeps expanding, the salary-versus-token swap keeps ratcheting instead of crossing once and settling. Right — cheaper tokens don't save you a dime if the workload ceiling just keeps rising to meet them. If Tech Podcast Podcast is part of your routine, take a moment to subscribe or leave a quick review wherever you’re listening. It really helps other people discover the show.
You’ll find links to every story we covered today in the show notes, so if something stuck with you, that’s the place to dig in a little further.
That’s Tech Podcast Podcast for today. This is a Lantern Podcast.