Podcast TLDR

Searchable index of every TLDR page. Open one to get the actionable takeaways.

Add a link

Paste a YouTube or podcast URL to save it on this device as Requested (watchlist). This static site can’t generate a TLDR by itself — paste the same link in the Podcast TLDR chat (or email it) and a new episode page will land in the library.

Requested / to summarize

No links saved yet. Add one above.

Click a card to open its TLDR

YouTube thumbnail: Inference Will Be $10TRN Market: What No One Sees | Naveen Rao | 20VC

[20] Ready ~73 min Interview

Naveen Rao: Inference will be a $10T market, energy is the bottleneck, and pure software investors are in trouble

20VC (20VC with Harry Stebbings) · host Harry Stebbings · guest Naveen Rao (founder & CEO, Unconventional AI)

Use case: pressure-test AI infrastructure bets: why ~10 GW a year is the realistic US build rate (~$50B per GW), why the efficiency gains now come from model–hardware co-design, and why Rao expects 80% of tokens to run on open models but only ~50% of spend.

Open source on YouTube ↗

[01] Ready ~22 min YouTube

How to scale intent, quality, and artistry with AI | Katie Dill (Stripe)

Lenny’s Podcast · Katie Dill (Stripe)

Use case: keep AI-speed shipping from becoming zombie UI — encode POV into tools, edit past done, spend savings on originality.

Open source on YouTube ↗

[02] Ready ~19 min YouTube

What product looks like when coding is solved

Lenny’s Podcast · Geoff Charles (Ramp CPO)

Use case: design an AI software factory — find/remove each bottleneck as coding, review, QA, and coordination get automated.

Open source on YouTube ↗

[03] Ready 45 min YouTube

We Built Grok Bot. Here Are Our 14 Best Bots

Peter Yang · Peng Zheng & Lauren Tan

Use case: specialized bot fleets for design, eng swarms, and life ops — escalate trust skill by skill.

Open source on YouTube ↗

[04] Ready ~82 min YouTube

How we built Grok Bot in a month

Lenny’s Podcast · Roman Ugarte

Use case: ship a colleague-like agent fast — isolation, cloud computers, manual onboarding, unship aggressively.

Open source on YouTube ↗

[05] Ready ~18 min YouTube

How to run your product team like a research lab

Lenny’s Podcast · Dan Shipper (Every)

Use case: split exploration labs from execution product teams so you welcome new model drops instead of dreading them.

Open source on YouTube ↗

[06] Ready ~74 min YouTube

Bill Gates’s Blunt Warning on A.I.

The Ezra Klein Show · Bill Gates

Use case: treat cyber/bio misuse as already-crossed thresholds — demand monitoring, don’t bet on self-regulation or Jevons paradox for jobs.

Open source on YouTube ↗

[07] Ready ~53 min YouTube

Why wars are so difficult to end

Dwarkesh Patel · Sarah Paine

Use case: diagnose war termination before you fight — culminating points, limited vs unlimited aims, and when “victory” turns into endless occupation.

Open source on YouTube ↗

YouTube thumbnail: GTM Roundtable: The State of AI-Native Sales

[08] Ready ~91 min YouTube

GTM Roundtable: The State of AI-Native Sales w/ Clay, Parallel, Avoca & Profound

The Crew Podcast · Liam Mulcahy, Jack Gashi, Mark Ebert, Todd Busler

Use case: run AI-native sales ops when plans are pencil — deal caps, honest “worst year” hiring, and W2 over equity for most AEs.

Open source on YouTube ↗

[09] Ready ~36 min YouTube

The State of Startups in 2026

The Lightcone (Y Combinator) · Garry, Jared, Diana & Harj

Use case: build for the atoms+agents shift — hard tech, end-to-end job automation, data/RL envs; taste and “what to build” beat pure coding skill.

Open source on YouTube ↗

[10] Ready ~21 min YouTube

Paul Graham On Startups, Ambition, and Great Founders

Y Combinator · Paul Graham · Vivian Shen

Use case: calibrate founder ambition — day-to-day fear beats billionaire dreams; hire for “gets what they want”; ship pace still wins in AI.

Open source on YouTube ↗

[11] Ready ~49 min YouTube

This Is What It Takes to Get a Data Center Financed

Odd Lots · Travis Wofford (Baker Botts)

Use case: stage data-center capital to permits and power — finance the shell, not the GPU; haircut the interconnection queue before you buy “powered land.”

Open source on YouTube ↗

[12] Ready ~68 min Apple Podcasts

Inside Ramp’s Sales Playbook: How to Build a Sales Machine

20VC / 20Sales · Max Freeman (SVP Sales, Ramp)

Use case: build an early sales team the Ramp way — overpay for high-horsepower reps against 4–6× output, engineer outbound, and run 100% OTE until you have data.

Open source on Apple Podcasts ↗

[13] Ready ~66 min Talk + Q&A

Tackling the Grid Reliability and Affordability Challenge in the Age of AI

Belfer Center / HKS Energy Policy Seminar · Manu Asthana (former CEO, PJM)

Use case: understand why power, not chips, is the binding AI constraint — and the nine flawed-but-necessary levers (deferred retirements, data-center flexibility, long-term contracts, permitting) for adding supply without wrecking affordability.

Open source on YouTube ↗

[14] Ready ~37 min Interview

What It Costs to Make Power Cheap

Critical Capital (Crux) · host Alfred Johnson · guest Zach Dell (co-founder & CEO, Base Power)

Use case: size Base Power’s scale-up claims (≈500 MW this year, “a couple GW” next, 100→300→500 installs/day, 20 GWh factories) and see how cost of capital and tax-credit sourcing rules set the customer’s price.

Open source on YouTube ↗

[15] Ready ~76 min Interview

Building a Power Company for the Next 50 Years

David Senra (podcast) · host David Senra · guest Zach Dell (co-founder & CEO, Base Power)

Use case: learn Base’s counter-positioned model (own the battery and sell power, not hardware), its utility “megawatts-as-a-service” business, and the operating system behind it (north stars, turtles, hot potatoes, monthly updates).

Open source on YouTube ↗

[16] Ready ~26 min (of 52) Panel segment

Monitoring the Situation #3: Who Is Nick Land & the State of Energy (Zach Dell segment)

a16z (The a16z Show / a16z Deep Dives) · hosts Erik Torenberg, Katherine Boyle, Erin Price-Wright · guest Zach Dell

Use case: hear Zach Dell’s Series C-era pitch: batteries as “non-wires alternatives,” why Texas price signals make it the energy lab, and how China supply risk pulled the Austin factory forward.

Open source on YouTube ↗

[17] Ready ~66 min Interview

Former CIA Spy: How to Read Anyone, Think Clearly, and Protect Yourself

The Mel Robbins Podcast · host Mel Robbins · guest Jonna Mendez (former CIA Chief of Disguise)

Use case: borrow a 30-year CIA officer’s playbook: MICE for why people act, how disguise changes how you’re seen, cues that someone is lying, intuition vs. anxiety, and courage built from preparation.

Open source on YouTube ↗

[18] Ready ~66 min Interview

Neil Movva: 1MW datacenters, commoditizing compute, and why 95% of inference could be asynchronous

Substrate (Cat Tang) · host Cat Tang · guest Neil Movva (co-founder & CEO, Sail Research)

Use case: understand why slow inference can be 5–10× cheaper: bigger batches, cheaper non-HBM chips, sub-megawatt modular sites, and software that routes around 90%-uptime hardware.

Open source on YouTube ↗

[19] Ready ~26 min Live interview

Don Wilson on the GPU Market’s Trillion Dollar Question

Odd Lots (Bloomberg Podcasts) · hosts Joe Weisenthal & Tracy Alloway · guest Don Wilson (founder & CEO, DRW)

Use case: understand how GPU futures (possibly listable Nov 9, out to 3 years at CME) could cut neocloud cost of capital, plus DRW’s “crunch spread,” token futures, and the case for perps.

Open source on YouTube ↗

No episodes match that search.

← Back to library Open source on YouTube ↗

Episode 45 min Sep 27, 2026

We Built Grok Bot. Here Are Our 14 Best Bots

Peter Yang · guests Peng Zheng & Lauren Tan (Grok Bot design & eng leads, SpaceXAI)
youtu.be/xZ5TEaleUdg

Context: The people building Grok Bot show the 14 bots they actually run for work and life — design production, eng swarms, bot-design audits, and personal ops — plus how they escalate trust so bots ship without babysitting.

Key insights

Do the first 5%, then let the bot scale the rest

Peng’s design bot connects to Figma via Figma MCP and a skill that encodes file layout plus design-system colors, type, and spacing. He builds the system and one keyframe; the bot extends that into every screen of the flow.

“I will do the first 5%.” — Peng Zheng
Use an eng-lead bot that manages bots — it shouldn’t do the work

Lauren hands big projects to Matcha. Matcha breaks work down for specialist eng bots; those bots spin cloud coding agents. Net effect: “really massive agent swarms” without Lauren being the router.
Trust ladder: watch → correct → skill → routine (one skill at a time)

Don’t dump autonomy on day one. Supervise, encode what worked as a skill, and only promote to a standing routine once the bot nails the task in one shot. Escalate scope skill-by-skill, not all at once.

“First, watch your bot work and correct it. Turn what worked into a skill. Once it nails the task in one shot, make it a routine.”
Aim for a Michelin kitchen, not a software factory

High throughput is fine; the framing that matters is craft quality. “Software factory” implies mass-manufactured slop. Specialized bots + human taste on the first slice is how they keep output sharp while volume rises.

“I like to call it the Michelin kitchen… when you say software factory, it has this connotation of mass manufactured slop.”
Make bots verify before merge — then you can read PRs after they land

Lauren’s eng bots test their own work before merging. With enough earned trust, she sometimes doesn’t look at a PR until after it’s landed and still finds it acceptable — the verification loop is the unlock, not blind faith.

“Sometimes I actually don’t even look at the PR until after it’s landed and then I’m like, ‘Oh, okay. Yeah, that looks good.’”
Give each bot one job — and explicit anti-jobs

Lauren’s Dr. Eggbot designs/audits bots: coding bots get a tight “poteto-mode” bar (one job, unslopped, verified); non-coding bots get one voice, one job, and explicit things they must not do (e.g. a scout doesn’t post; a drafter doesn’t send).
Put PM, design, and eng bots in the same room

Cross-functional bot collaboration (not only siloed specialists) shows up as a dedicated segment — treat multi-role bot chat as a product surface, not just parallel one-off agents.
Life ops count: supplies, listings, multi-city travel

Peng’s chief-of-staff bot buys supplies and lists gear. Lauren’s assistant bot booked a multi-city trip. Same delegation mindset as work: anything you repeatedly touch with keyboard/mouse is a candidate.

“Everything I touch with my keyboard and mouse, I try to delegate to my bots.”

Key use case

Job to be done

Build a small fleet of specialized AI bots (not one general chat) that reliably takes over repetitive keyboard/mouse work — design production, engineering execution, and life logistics — while you keep judgment and taste.

Who it’s for

Builders already using agents who want patterns from people shipping Grok Bot itself: designers, eng leads, and operators who need scale without “mass-manufactured slop.”

Primary outcome

You do the hard first slice (taste + system + trust), then bots expand, test, and eventually land work — including PRs — with less live supervision.

Concrete demos (lean)

Photo → site diorama: A bot turns a street photo into a diorama published on Peng’s website — personal creative pipeline, not just work tickets.
Design bot: One keyframe → full user flow via Figma MCP + design-system skill.
Eng swarm: Matcha orchestrates; IC eng bots execute via cloud agents; self-test before merge.
Bot factory: Dr. Eggbot designs and audits other bots so new agents start with clear scope.

Episode chapters

0:00 — The bots that the Grok Bot team actually uses
1:23 — Peng’s chief of staff bot buys supplies and lists gear
6:31 — Putting PM, design, and eng bots in the same room
10:07 — Turning a street photo into a diorama on Peng’s website
14:16 — How Lauren’s bots test their own work before merging
17:59 — Dr. Eggbot, Lauren’s bot that designs and audits your bots
21:27 — The assistant bot that booked Lauren’s multi-city trip
25:22 — Eng bots that sometimes land PRs before Lauren reads them
38:59 — How to trust your bots with more work, one skill at a time

← Back to library Open source on YouTube ↗

Episode ~82 min Sep 8, 2026 Actionable

How we built Grok Bot in a month

Lenny’s Podcast · guest Roman Ugarte (product, Grok Bot / SpaceXAI; previously Growth at Cursor)
youtu.be/maSdsTLaMuU · show notes

Overview: A tiny isolated team took Grok Bot from blank page to internal prototype in ~4 weeks, then to public launch in ~3 more. Roman’s playbook: start fresh (don’t bolt onto Cursor), cloud-native bots with their own computers, hide mechanics, personally onboard hundreds of users, unship aggressively, and decide every product call with a colleague-pilled lens — “what would you want from a human teammate?”

Most actionable learnings

Isolate a tiny “cave” team for micro-decisions

Handful of people, physically separate, private channels, sole objective: an amazing knowledge-work agent. Roman’s hindsight: a bigger group with a 6–12 month vision would not have landed the same product — they needed daily non-obvious micro-calls without org drag.
Start from scratch when the vision is different

Bolting knowledge work into Cursor was debated hard. They chose a new surface to control every pixel and avoid “shipping your org chart” (multiple visions sharing one cluttered UI). Coding agents already did non-coding tasks — but paper cuts, intimidation, and brand association hurt non-technical users.

“…three different visions that all kind of share a screen… shipping your org chart style…” — Roman on the multi-tab retrofit path
Two early bets that weren’t obvious: cloud + own computer

(1) Cloud-first — never make users think about local/cloud, sleeping laptops, or phone↔home tethering; same state everywhere (text, call, kick off from phone).
(2) Each bot gets its own computer — humans don’t work only via MCPs/APIs; click pixels, type into boxes. Sharing one laptop with a “colleague” is absurd — don’t force AI into that model either.

“An AI that does 100% of the job feels categorically different from one that gets you 90% there.” — Roman (pinned framing)
Personally onboard hundreds — and sit through the pain

~2 weeks, ~200–300 live onboardings by the core team. First sessions were rough on purpose: 20 minutes of a computer not spinning up or a confused user → “that can never happen again tomorrow.” Mix taste-makers with unconventional users (e.g. a coffee-shop owner stress-testing Shopify copy/workflows) to escape the Silicon Valley AI bubble.
Don’t lead the witness — let usage patterns emerge

Internally, week 1–2 was “5–10 scoped bots.” Then people promoted a standout bot to chief of staff that fans work out to the others. Team noted it, did not prescribe it in early access, watched externals rediscover it, then lightly encouraged (not a one-way door).
Hide mechanics; show presence

Default: message → bot works → progressive updates + Slack-like active indicator — no live tool-call / chain-of-thought dump. Teammates don’t narrate every click. Users asked for to-do / prioritization views; almost nobody wanted the raw stream. Strong stance first; add selective visibility later.
Unship before you scale — then hill-climb “just works”

Pre-launch: strip experimental features and developer observability that leaked into the user surface. Parallel push wasn’t feature menus — it was backend hill climbs (computer-use precision, logins, Salesforce clicks) unlocked by concrete failing tasks from sales/recruiting, then immediate love when a workflow unblocked.

Frame launches as “Grok Bot can now…” not “Grok Bot now has…” — capabilities without new pixels.
Be colleague-pilled on every product debate

When both SaaS answers feel sensible, ask: what would you want from a human teammate here? Answers usually converge. North star: a team of autonomous AI bots for job + life — steerable, not micromanaged, with the tools to do ambitious work. Example target UX: Slack huddle energy (5-min voice + screen share) with bots, not just chat threads.
Long-lived roles > one-off chats; automations in plain language

Bots are persistent swim lanes with memory — not a new chat per task (and the copy-paste tax that creates). Automations: say “remind me at 8 a.m. every day” — no sidebar wizard; Roman says ~99% of platform automations are built that way. Skills/slash commands should be background primitives users never need to learn.
Culture that preserves startup speed: delete + just do

Deleting the product — as models improve, remove scaffolding/overhang even if it upsets a few users. Just do the thing — not an ask-permission culture; see it, fix it, pull resources. Moats (data, distribution) are discovered by obsessing on “useful today” and pulling the 3–6 month future forward — not by planning moats on a whiteboard.

Key use case

Job to be done

Ship a knowledge-work AI agent that feels like a team of colleagues (not chat + plugins): always-on in the cloud, each bot with its own computer, long-lived memory, and enough reliability that you can truly delegate — a “no-look pass.”

Who it’s for

PMs, founders, and operators building agents — especially anyone tempted to bolt a new surface onto an existing coding/chat product instead of starting clean.

Primary outcome

Concrete build/process moves: isolation, manual onboarding, cloud+computer bets, “can now” framing, and culture habits (“delete the product,” “just do the thing”).

Speed timeline

~4 weeks First line of code → usable internal prototype (handful of people, separate office area + private Slack).
~3 weeks Internal all-hands / company dogfood → public GA (unship jank + “make it just work”).
~3 weeks Post-launch at recording — already standing-room meetups; Lenny: “hottest AI product” moment.
~2 weeks Manual onboarding window: core team walked ~200–300 early users through live sessions.

Steal-these playbooks

New-user setup (Roman’s tip)

Connect the tools a new teammate would need (email, Slack, systems of record).
Ask the bot what it can take off your plate after reading those sources.
Spin specialist bots for the 1–2 suggestions that are real chunks of work — not “draft an email.”

Power-user scaffold

Design how bots collaborate and where shared artifacts live.
Push digests to one legible store/DB you actually read daily.
Infovore pattern: firehose → notify / daily roundup / (careful) paging for true urgents.

Concrete examples from the episode

Sales GTM: Tools without good MCPs/APIs — computer use turned “stuck elsewhere” workflows into assistant-with-a-laptop work; pixel-click precision was a repeated hill climb.
Recruiting always-on sourcing: Nightly conference PDFs → extract new co-author names → research → match SpaceXAI connections → Slack ask for intros (not résumé sorting).
QA bot eating Grok Bot: Install Grok Bot inside a bot’s computer; regression-test workflows release-to-release and write results to Notion.
Launch filter: If you can’t write a compelling launch tweet users will feel, maybe don’t build it.
OpenClaw → product: Keep “tools + colleague mental model”; drop Mac mini/VPN hackiness; sand edges so mainstream/business users never need slash skills.

SpaceXAI pillars (as stated)

Coding surfaces: Cursor + Grok Build — every pixel for shipping production software.
General knowledge work: Grok Bot as AI teammates inside companies (single-player → team/org memory next).
Models: Train capable models with a practical “useful AI” mandate — applied, not vague superintelligence chase.

Episode chapters (from show notes)

0:00 — Introduction
2:09 — Origin: building from scratch in one month
8:40 — Why a separate product (not inside Cursor)
11:20 — Manually onboarding a couple hundred people
14:29 — Hiding internal mechanics from users
18:41 — Timeline: beta → public launch
19:14 — Unshipping features and simplifying
23:50 — Early use cases and feedback
26:50 — Product philosophy: “Grok Bot can now”
30:02 — Cloud-first architecture
33:12 — The fresh-start advantage
35:54 — Vision: a true team of AI colleagues
39:20 — The “colleague-pilled” framework
42:36 — Work vs personal: one product or two?
47:14 — Long-lived agents, memory, computer abstraction
51:04 — Always-on infovore / chief of staff
53:35 — Speed and preserving startup feeling
58:20 — SpaceXAI pillars
1:00:44 — First 90% vs last 10%
1:03:30 — Moving fast at scale
1:06:40 — How Cursor kept winning
1:10:04 — Values: “deleting the product” & “just do the thing”
1:11:45 — Moats: discovered, not planned
1:15:11 — Tips for new and power users
1:18:00 — Lightning round

← Back to library Open source on YouTube ↗

Episode 18 min Sep 24, 2026 Lenny & Friends Summit

How to run your product team like a research lab

Lenny’s Podcast · Dan Shipper (co-founder & CEO, Every)
youtu.be/DqF08Dz3nok

Context: A ~18‑minute summit talk on building product when frontier models (Fable 5.1, Astra 6) keep rewriting what’s possible. Shipper’s fix: separate exploration from execution with a tiny labs team, a staged research pipeline, and hard promotion criteria — so you scale today’s product without going crazy every model drop.

Key insights

Exploration and execution pull opposite ways — stop asking one team to do both

2:07 Exploring the frontier is divergent (try models, demos, discard most). Executing a roadmap is convergent (focus, say no). Forcing both on the same people tears the org; the roadmap usually wins because it has deadlines.

“Never make any major life decisions within 30 days of a meditation retreat, a psychedelic experience, or your first encounter with a frontier model.”
Your early adopters already live in the future — isolate them so they don’t derail shipping

4:36 Some people explore far more than others. They’re the best sensor for what the product might become — and a massive distraction if left embedded in delivery. A labs team (even of one) separates concerns cheaply because AI multiplies what one person can try.

“How do I harness the early adopters on my team… without distracting everybody else?”
Two-slice teams beat two-pizza teams; pair pirate + architect

7:35 Bezos’s 8–10 person two-pizza team is too big for AI-speed exploration. One or two people go farther, faster; more adds coordination overhead and competing visions. Composition that works: pirate finds messy value; architect shapes it into something valuable, beautiful, and extensible.

“In the AI age, I call it a two-slice team. You want one or two people max.”
Compress the make → know-if-good loop; dogfood real work

8:35 The single most important labs practice is a tight feedback loop. Build for yourself when possible; otherwise use early-adopter customers. Experiments should serve actual work so you can tell useful from merely new.
Parallel competing experiments map an unknown frontier

9:22 Unlike product, labs should run many approaches at once — even to the same problem. It looks messy; that’s the point when capabilities move and the frontier is unknown.
Make the 90% ROI-positive (content, early adopters, capability briefings)

9:53 Disposal is the labs success metric — but discarded work can still pay: Every turns tries into external content customers love; feeds early-adopter programs; and shares “what’s now possible” so product doesn’t have to explore itself.

“On a labs team, you are going to expect to dispose of like 90% of what you make.”
Promote with rigor: usage + 10× better + affordable — treat advances as events

16:06 Weekly Notion pipeline review at all-hands keeps product in sync without scrambling into their own experiments. Gates: (1) are people using it and coming back? (2) is it still 10× better a month later? (3) can you afford it at customer scale?

Key use case

Job to be done

Harness early adopters (and early-adopter customers) to probe new model capabilities — without distracting the product team that has to ship a coherent roadmap.

Who it’s for

Product leaders and founders whose teams are torn between “try every new model” and “hit the roadmap.” Works even as a labs team of one.

Success signal

Your org welcomes new model releases instead of dreading them — while still scaling the current product.

Do this week

→Name your early adopters — people already running frontier models on weekends. Put exploration on them; protect everyone else from the distraction.
→Staff a two-slice lab (1–2 people max): a pirate (messy “slop cannon” hunting value) + an architect (turns vibe-coded wins into extensible systems).
→Dogfood for real work — tightest loop is build-for-yourself; else recruit a few early-adopter customers. Ask: useful, or just new?
→Run competing experiments in parallel on the same problem — map an unknown frontier instead of converging too early.
→Make the discarded 90% pay rent — publish experiment writeups, feed an early-adopter program, and brief the product team on new capabilities.
→Stand up a left→right research pipeline with weekly review + explicit promotion gates (usage, 10× better, affordable at scale).

Labs vs product

Labs team

Explore what’s newly possible on model drops
Many parallel / competing experiments
Expect to throw away ~90%
Mode: divergent

Product team

Improve & scale what already works
Keep the product coherent for existing customers
Expect to adopt ~10% of lab tries
Mode: convergent (focus, say no)

Research pipeline (lab → product)

1Many lab experiments — most die; a few show promise.

2Internal adoption gate — do teammates start using it for real work?

3Architect pass — turn the messy prototype into a measurable system.

4Early customers + product roadmap fit — not everything survives.

5Scale / release / merge winners into the main product (self-disruption that actually works at AI speed).

Concrete examples

Kate bench → Kate pass (Every): 12:20 After ~3 years of trying, Shipper fed Kate’s historical copy edits into Fable; now Everywhere’s “Kate pass” suggests Kate-like edits. Architect Yanek added a per-doc dashboard; Kate does 12% less work on those edits vs. the prior month. Next: early customers + generalize beyond copy edit.
Anthropic Labs: 6:43 Claude Code, MCP, Skills, Claude Design came from a small experiment group; winners get more investment; ~thousands of unseen experiments.
OpenAI Codex: 17:18 Small team outside the main app; parallel form-factor bets (IDE vs CLI); desktop launched Feb 2026, grew fast, merged into ChatGPT — small team effectively taking over an ~800M DAU surface.
Capability shock demos: Waterloo 3D from one prompt (~4 hrs); ~1,000-agent sim from a paper; Astra editing in Premiere; deck animations made by Astra (he didn’t touch Animate).

Rough chapters (from captions)

0:06 — Building on a moving frontier; Fable 5.1 / Astra 6 demos
1:50 — Rule #1: don’t make big decisions after first frontier-model high
2:07 — Core tension: explore vs execute
3:12 — Why make a lab; find early adopters
4:51 — Labs vs product; labs team of one; 90% / 10%
6:43 — Anthropic Labs precedent
7:31 — How to run a lab: two-slice, pirates & architects
8:35 — Dogfooding, parallel experiments, ROI on the 90%
11:10 — Research pipeline design
12:20 — Kate bench / Kate pass case study
15:23 — Weekly pipeline review + promotion criteria
17:18 — Codex merge into ChatGPT; closing test

← Back to library Open source on YouTube ↗

Episode ~22 min Sep 25, 2026 Lenny & Friends Summit

How to scale intent, quality, and artistry with AI | Katie Dill (Stripe)

Lenny’s Podcast · Katie Dill (Stripe)
youtu.be/GLvFTMtw4Jk

Context: A ~22‑minute Lenny & Friends Summit talk (Sep 10, 2026, SF). Dill warns that the AI building boom can repeat post‑war modernism’s failure mode — copied patterns without intentionality → zombie UI. She names three watch points, then four practices to keep care, brand, and originality in the work while shipping faster.

Most actionable learnings

The AI boom echoes post‑war modernism: copy without intentionality → zombie buildings / zombie UI

0:06 Post‑WWII builders reused modernist geometries without the founders’ principles; patterns thinned until banks signaled nothing. Teams of three now ship what needed thirty — same watch point: proliferation of patterns ill‑suited to context.

“Zombie UI that’s monotonous, vacant, or uncared for.”
If you don’t have a point of view, AI will supply a generic past

5:44 Define what the brand is for you and for users, who you want to be to them, and what they care about — that becomes standards. More important when building is distributed and ownership is diffuse.

“First one, you have to have a point of view. If you don’t, AI will give it for you.”
Quality standard is about the output, not how it was made — go one level deeper than the customer can see

7:20 In a Stripe ad crit, a partner asked what the bar should be for AI‑made work. Dill’s answer: users don’t care about the pipeline; they care if it’s good. They left with 17 craft fixes (“Pepsi bubbling”) — meticulous detail that still shows up for customers. Not perfection; one level deeper.

“Why should that matter how it was made? … It’s the output at the end of the day that matters.”
Design the system that scales intent — Gutenberg‑level opinionated tooling, not blank‑filling docs

9:02 Agents and generative UI remove the designer‑in‑the‑room who used to fill blanks. Stripe’s docs MCP was too vague (same prompt → three results). A CLI on the design system + full templates/flows (not just atoms) makes AI “far more obedient,” consumes docs at the right time to avoid context rot, and embeds standards in production.

“The old system scaled consistency, but the new system needs to scale intent.”

“The object of design is no longer the screen. It is the system itself.”
A rule‑perfect system can still be dead — refuse to confuse done with good; install editors

12:22 Staffing used to filter ideas before build; now you can build 20 ideas in a week and react to the real thing — but saying no is harder post‑build. Someone must judge the end‑to‑end whole. Experience work as a user; watch for journey incoherence; push incomplete work to completion. AI is weak at unexpected details, deeper meaning, and continuous themes — editors fill those gaps and own every pixel.

“A system can satisfy every rule and still be dead.” — Christopher Alexander (via Dill)

“We have to unlearn that built means done and that done means good.”
Iterate past first AI polish — 56 passes is care, and AI expands the possibility space

15:05 Summit opening animation: 3D scene → AI animation help → first cut felt jilted → 56 iterations to realism and life. Without AI he might not have attempted the complexity or viewpoints; with scrutiny, AI opened the space. Same pattern on Built to Grow cover: human marbler + AI fine‑tune of color/line.

“56 iterations later, he got to something truly beautiful… you can sense the care.”
Use AI as creative catalyst: specific inputs, stressed outputs, protected strangeness

16:28 Differentiation gets harder when everyone can build. Invent new interfaces and aesthetics — best practices aren’t written yet. Improve prompts with brand beliefs and source material; stress outputs (no burrito complacency); use adversarial agents but stay the quality driver. Leaders: don’t only mandate AI — give room to explore. “Protect the strange.” Raise the ceiling, not just the floor.

“AI lowers the cost to create. Let’s spend some of that savings on making something truly special.”

“AI can help us raise the ceiling, not just the floor.”

Key use case

Job to be done

Ship AI‑accelerated product without sliding into generic “most probable” interfaces — encode a real point of view into tools, edit past the first polished draft, and spend the time savings on originality.

Who it’s for

Design / product leaders and builders whose teams can now go prompt→UI in minutes, and who need coherent brand standards when ownership is diffuse (humans + agents).

Success signal

Users feel the product was cared for — personality, context‑fit, clever details — not monotonous zombie UI. AI raises the ceiling, not only the floor.

Do this week

→Write the POV — brand for you + for users; who you want to be to them; what they care about. Without it, AI invents a generic one.
→Encode standards into the means of production — not just a docs MCP: opinionated templates/flows where builders actually build (Stripe’s CLI harness), so same prompt → coherent outcomes.
→Name the editor — who owns end‑to‑end “is this a whole?” post‑build. Unlearn built = done = good.
→Experience it like a user — solve the problem? attuned to how they think? coherent across the journey? (Isolation polish lies.)
→Stress outputs past the burrito — push one more iteration; use adversarial critique; don’t ship because speed enamored you.
→Spend the AI savings on strange — protect exploration time; invent interfaces/aesthetics; raise the ceiling, not just ship the old thing faster.

Three watch points → zombie UI

1Most probable ≠ original — LLMs excel at what’s popular / has worked; weak at what’s specific to your brand and user context (Korean‑BBQ site that looks like SaaS).

2Temptation of done — AI makes work feel finished too fast (the microwave‑burrito problem). Polished corners ≠ solves the problem or differentiates.

3Disposable → unowned — when work feels cheap/quick, long‑term care and maintenance drop; responsibility for decisions slides away.

Four recommendations

1. Point of view

Define brand + user standards first
Diffuse ownership → easy to abdicate to AI
Stripe example: optimism in color, copy, products

2. Encode into the machine

Old systems scaled consistency
New systems must scale intent
Object of design = the system (templates + flows), not only the screen

3. Don’t confuse done with good

Filter moved post‑build (harder to say no)
Need an editor of the whole journey
Accountability for every decision / pixel

4. Unleash artistry

AI is also the greatest creative catalyst
Best practices aren’t written yet — invent
Amplify human judgment; don’t only accelerate monotony

Concrete examples

Zombie banks / Turkish copy‑paste skyline: 1:11 4:19 Trust‑signaling bank architecture → generic boxes; literal place in Turkey as cautionary visual for pattern proliferation.
Korean BBQ site that reads as SaaS: 2:38 LLM‑probable UI with no character or use‑context.
Care signals: 4:48 Grok Bot’s animation; calendar tab showing the date; Link agent wallet with built‑in troubleshooting for agent shopping failures.
Stripe ad “Pepsi bubbling”: 7:20 Cross‑functional crit → alignment that craft bar ≠ “AI discount” → 17 concrete fixes; verb for meticulous craft.
Docs MCP → opinionated CLI: 11:04 Same prompt, three results → harness on design system with templates/flows, docs injected to avoid context rot.
Gutenberg’s 290 characters: 10:02 Extensible + opinionated enough that machine‑set type could feel handmade (no rivers/lakes of whitespace).
Summit open + Built to Grow cover: 15:05 18:05 56 animation iterations; human marbler + AI fine‑tuning to scale judgment.
Gothic vs zombie modernism (close): 20:34 Ruskin: no two Gothic columns alike — unique hand of the maker. Users want anticipated needs + brand character, not “we animated it in 30 minutes.”

Rough chapters (from captions)

0:06 — Post‑war building boom → modernism without intentionality
1:11 — Zombie buildings; AI building boom watch points
2:17 — Watch point 1: most probable / Korean BBQ SaaS
2:55 — Watch point 2: temptation of done (burrito)
3:54 — Watch point 3: disposable work → zombie UI
4:48 — Care examples (Grok Bot, calendar tab, Link wallet)
5:44 — Rec 1: have a point of view; Stripe optimism
7:20 — AI quality bar; Pepsi bubbling
9:02 — Rec 2: encode standards; Gutenberg; scale intent
11:04 — Stripe MCP → CLI templates/flows
12:22 — Rec 3: don’t confuse done with good; editor role
15:05 — 56‑iteration summit animation
16:28 — Rec 4: unleash creativity; invent interfaces
18:38 — Better inputs / stress outputs / protect the strange
20:34 — Gothic craft vs zombie boom; close

← Back to library Open source on YouTube ↗

Episode ~19 min Sep 25, 2026 Lenny & Friends Summit

What product looks like when coding is solved

Lenny’s Podcast · Geoff Charles (CPO, Ramp)
youtu.be/ZG8Mf3P9xzI · Recorded live at Lenny and Friends Summit (Sep 10, 2026, SF)

Context: A ~19‑minute summit talk arguing that once coding agents exist, the race is no longer “code faster” — it’s find and remove the next bottleneck in your product factory. Geoff walks Ramp’s loop (identify → define → build/review/test → coordinate → autonomous improve), names the agents (Glass, Inspect, Review Buddy, Testo, Gadget), and closes with three PM tracks plus an invite to copy the stack.

Most actionable learnings

Winning isn’t driving harder — it’s removing bottlenecks around the driver

0:39 Geoff’s LeMons crash frames the thesis: in pro racing the driver is ~15% of outcome; the car + team dominate. Pit stops went 67s → 1.8s by finding bottlenecks (specialization, tech, practice) — not by asking mechanics to work 37× harder. F1 bar: ~90% of parts change yearly.

“Winning isn’t just about driving faster. The best drivers in the world can only perform up to the level of the system.”
AI removes a bottleneck and moves it — race the find→remove cycle

2:54 Clock starts at customer pain and stops when a product solves it. Engineers automated coding; the bottleneck shifted to define / collaborate / coordinate / ship / test / release. Tip: “be as lazy as engineers” — invest in the PM-side factory. Speed of the loop decides the race.

“AI simply removes the bottleneck, but moves it. And the best team… is the team that can find the bottleneck faster, remove it, and move on to the next one.”
Identify: don’t paste Gong into a 1M-token window — build a real insight agent

3:54 Pain lives in Gong, Zendesk, LogRocket, surveys, angry exec email. A 1M-token window is <0.5% of Ramp’s Gong transcripts. First “hate channel” of daily complaints “got out of hand”; they rebuilt with ETL + vector search + clustering that knows product/teams/features. Surfaces: Slack Q&A, HTML dashboard, and a “hate podcast” of ~100 customers. Goal: access + traceable picks of whom to talk to — not replace customer conversations.
Define: connect AI to systems; the next eng contract is evidence + requirements + prototype

6:29 Blank “what do you want to build?” is a bad start. Glass connects Snowflake + user research + strategy + specs + codebase + design system — “AI is your tech lead.” Eng doesn’t want a lonely prototype or a lonely long spec; they want qual+quant proof the problem is real, requirements coding agents can use, and a prototype inside the real product for inspiration.

“They don’t want your prototype. It’s useless. They don’t want your long spec. That’s useless. They want the combination of qualitative and quantitative data… the actual requirements… and a prototype that they can actually get inspired by. That is the next contract.”
Build: own the coding harness, then attack review and browser QA as the next bottlenecks

7:54 Inspect (Slack-native, <5s, deploy preview): built in-house for harness + codebase depth; ~1M sessions; 75% of PRs built by Inspect; ~1,000 PRs last month from non-engineers. Works best with strong architecture/codebase. Review Buddy: quality/security, routes humans, audits generating prompts — 93% of PRs auto-handled so seniors spend time on the risky 7%. Testo: browser QA across ~100 production-data combos; blocking bugs + design-system feedback; 425 bugs in 30 days.
Coordinate: “every question is an API” — make the org legible to agents

11:14 After you ship a lot, human attention is the bottleneck. Gadget maps intent → formal record (Notion roadmap/specs/calls, Slack, Linear). It answers status with evidence, updates roadmaps, pings late owners, handles sales Qs (availability, price, Brazil…), routes unknowns, and drafts help-center / blog / customer email at launch. 85% of questions to PMs fully answered by AI; misses get answered then fed back into the system.

“For you to empower an agent, the agent needs to be able to understand and read the organization. And the organization needs to be legible to your agents.”
Automate small reactive loops so PMs can stay ambitious

13:27 PMs gravitate to high-confidence visible dopamine fixes — “a huge mistake.” For small work, AI runs route → Linear match/dedupe/count/rank → plan → Slack “ready to code?” → code → test/CI → knowledge-base update. Humans “in the loop only just barely.” Result: 60% of UX issues flagged by customers, sales, CX, or internals fixed within 24 hours.
Embrace constraints, not bottlenecks — pick one dimension to be world-class

15:33 Against unlimited budgets: Audi (2006) couldn’t go fastest, so they chose fuel efficiency → fewer stops → three straight Le Mans wins. Find the one bottleneck that can 10× the company and start there. On “are we fast enough?”: hire a driver who knows speed and let them challenge you (Niki Lauda / Enzo story) — culture proxy when velocity isn’t a clean lap time.

“Constraints force you to choose a dimension in which you can be world-class. So embrace your constraints, but not your bottlenecks.”
PM role splits three ways: technical factory-builder, tastemaker, GM

16:25 Being out of the loop on what engineers can drive is leverage, not a threat. Tracks: (1) Technical PM — ships the factory that helps build (under-invested); (2) Tastemaker / driver — holds the bar AI can’t; (3) GM — expands into marketing/sales/growth/ops and owns business outcomes. Recap: speed = find/remove bottlenecks; next bottleneck always appears; obsess a bit more about the factory than the shipped product. Everything shown is “outdated” — copy and surpass; best product is the next one launched.

“The best product you’ll ever build is the next product that you’ll launch. That begins in the software factory that you build today.”

Key use case

Job to be done

Design an AI-era “software factory” so your clock — pain → shipped fix for the customer — keeps shrinking as each bottleneck (coding, review, QA, coordination, small UX loops) moves after you automate it.

Who it’s for

Product leaders and PMs who already have coding agents (or should) and feel the drag shifting onto definition, review, testing, and human attention.

Primary outcome

A concrete stage map + named patterns to copy: insight agent, systems-connected define agent, Slack coding + review + browser QA, “every question is an API,” and autonomous small-issue loops.

Do this week

→Name your current bottleneck — where does work wait longest after coding got cheaper? (define, review, QA, coordination, or small reactive tickets.)
→Connect AI to systems before “what do you want to build?” — wire one quant + one qual source, plus codebase/spec format, so define isn’t a blank chat.
→Change the eng handoff to the three-part contract: qual+quant evidence, coding-agent-ready requirements, and an in-product prototype for inspiration.
→Make the org legible to agents — roadmaps, specs, and tickets in systems of record so a coordination agent can answer with evidence.
→Automate one small loop (route → dedupe/rank → Slack “ready to code?” → test/CI → update docs) so PMs stop living on dopamine fixes.

Ramp factory stages (as stated)

1Identify — customer insight agent (ETL + vectors + clustering); Slack / dashboard / hate podcast.

2Define — Glass: systems-connected spec + JTBD + in-product prototype; AI as tech lead.

3Build / review / test — Inspect → Review Buddy → Testo (coding → review → browser QA).

4Coordinate — Gadget; every question is an API; org must be agent-legible.

5Improve — autonomous small UX loops; humans barely in the loop; 60% fixed in 24h.

Numbers Geoff cited

Inspect: ~1M sessions; 75% of PRs built by Inspect; ~1,000 PRs last month from non-engineers; runs in Slack with deploy preview (<5s).
Review Buddy: 93% of PRs automatically handled; seniors focus on the remaining 7%.
Testo: ~100 production-data combinations; 425 bugs caught in the last 30 days.
Gadget: 85% of questions asked to PMs fully answered by AI.
Autonomous UX loop: 60% of identified UX issues fixed within 24 hours.
Context scale: 1M-token window < 0.5% of Ramp Gong transcripts.

Rough chapters (from captions)

0:06 — LeMons story; driver vs system
1:28 — Pit stops 67s → 1.8s; F1 90% parts turnover
2:54 — AI moves the bottleneck; be lazy like engineers
3:54 — Identify: hate channel → customer insight agent
5:58 — Define: Glass; AI as tech lead; next contract
7:54 — Build: Inspect coding agent
9:16 — Review Buddy (93%) and Testo (425 bugs)
11:14 — Coordinate: Gadget; every question is an API
13:27 — Autonomous small loops; 60% UX in 24h
14:35 — Measuring speed; Audi constraints; three PM tracks
18:06 — Three takeaways; “copy us”; Enzo close

← Back to library Open source on YouTube ↗

Episode ~74 min Sep 29, 2026 The Ezra Klein Show

Bill Gates’s Blunt Warning on A.I.

The Ezra Klein Show · Bill Gates (Microsoft co‑founder; chair, Gates Foundation)
youtu.be/A_156w0aYtU

Context: A ~74‑minute conversation (posted Sep 29, 2026). Gates argues AI alarmism hasn’t gone far enough: cyber and bio misuse thresholds are already crossed, industry self‑regulation is “insane,” and large job displacement follows once reliability thresholds fall — while the good stuff (health, education, ag, languages) arrives slower unless society forces safeguards and accelerates beneficial uses. Foundation priority: make AI a force for equality, not inequality.

Most actionable learnings

Cyber + bio misuse thresholds are already crossed — imminent risk is humans with tools, not sci‑fi RSI

11:20 Gates: models can already empower catastrophic cyber (banks/grid) and bio design that used to need nation‑state resources. Coding agents find 20‑year‑missed security bugs in minutes. The control / recursive‑self‑improvement problem is real later; the near‑term supervisory gap is bad actors using today’s models.

“We crossed the cyber threshold and we crossed the bio threshold early this year.”
Don’t bet on self‑regulation, liability, or CEO goodwill — require monitoring in all models

15:34 Compare to FDA / seat belts / airline safety: society does not keep people safe by “just sue them after the catastrophe.” Open‑source and unfiltered copies defeat product‑liability incentives. Ask: put safeguards in before millions of deaths, or after the first gigantic events?

“You can’t rely on the industry to self‑regulate here. I mean, it’s just insane.”
Safeguards need not “lose the China race” — keep customization, force monitored platforms

24:58 Gates’s concrete shape: models can stay free/customized but must remain on a platform where a sovereign can verify monitoring wasn’t stripped. “China wins if we regulate” is a get‑out‑of‑jail‑free card; both sides already opened Pandora’s box — test whether China will join humanity‑level protections.

“Putting in the safeguards and the monitoring will not meaningfully slow things down.”
Jobs: reliability thresholds, not vibes — coding crossed; white‑collar next; Jevons dies when the whole task is automated

27:30 So far AI created more jobs than it destroyed. Only coding has clearly crossed the “more reliable than humans” bar for managers; accounting, legal, telesales/support are next. Demand elasticity (Jevons) helps only while humans still own a necessary slice. Once the full task is AI, spend moves to the token budget — radial tires / warehouse robots didn’t create 4× demand.

“As soon as you complete the entire task… that goes into the token budget. It doesn’t go into the human salary budget.”
Policy levers: tax AI labor like humans (FICA‑equivalent) + pre‑commit “human‑reserved” roles

35:50 If a firm fires a worker and hires a robot/AI for the same unit of labor, why shouldn’t that unit still pay into pensions? Society can choose not to follow the cheapest economic signal. Human‑reserved domains (childcare, eldercare, portions of medical care / education) are deliberate “nature reserves” — keep human employment even when AI could do it.

“Just because the economic signals say that it’d be lower cost to use an AI doesn’t mean society has to do that.”
Accelerate the good path deliberately — languages, health, ag; benefits lag digital harms

38:36 Foundation bet: bring AI to under‑served languages (e.g. Yoruba medical advice at night) and to farmers (crop disease, fertilizer, varieties — already live at scale in India/Africa). Low‑income countries lacking doctors get outsized upside; middle‑income job ladders get hit later. Physical‑world benefits (drugs, housing, grid) are rate‑limited by regulation/trials; cyber/bio/job shocks are not — so push good uses and cut friction on beneficial regulation.

“The good stuff if we’re not careful arrives more slowly than the bio/cyber risk, the psychosocial risk and the jobs risk.”
Education & kids: design for the learn modality (not answer‑key), with real monitoring

59:00 AI is bimodal: students use it to skip struggle (cognitive offloading) or to learn more. Prefer systems that coach reasoning step‑by‑step and block “just give the answer” escape hatches. NYC example (KDUm): <10 min end‑of‑day checks → teacher gets which concepts / which kids are stuck. For children: age‑graded parental visibility + monitoring beats a blunt ban; assume parents are not technical (second‑phone problem).

“You clearly would get by modality where a student can use AI to be lazy… or a student can use AI to help them learn more.”
Societal engagement metrics: safeguards first; avoid climate‑style for/against polarization

22:22 Gates’s alarm jumped because thresholds were crossed and outside response was silence relative to the danger. Two success metrics for the next political cycle: (1) real cyber/bio safeguards; (2) a middle‑ground dialogue — if one party is “for AI” and the other “against,” you get climate‑change politics with no room to manage. Blocking data centers won’t slow capability; focus on model monitoring and governance.

“We crossed those thresholds and there was complete silence.”

Key use case

Job to be done

Wake builders, policymakers, and operators to already‑crossed misuse thresholds — install supervisory/monitoring layers, plan for reliability‑threshold job loss, and deliberately accelerate beneficial AI (languages, health, education) so upside isn’t permanently slower than downside.

Who it’s for

Product/eng leaders shipping AI features; anyone advising on policy, safety, or workforce planning; foundation / global‑dev builders targeting under‑served languages and last‑mile health/ag.

Success signal

Monitoring required across closed and open models; concrete FICA‑style / human‑reserved job proposals in debate; education tools that coach instead of completing homework; less “race with China = no safeguards” fatalism.

Do this week

→Threat‑model today’s models for misuse — cyber exploit assist + dual‑use bio/chem. Assume open weights can strip filters; design products and org policy accordingly.
→Reject “liability is enough” in any internal AI‑governance doc — write the monitoring / logging / abuse‑response layer you’d want before a catastrophic incident.
→Map roles by reliability threshold — which workflows already prefer AI (coding)? which cross next (support, claims, accounting)? plan reskilling before the pricing competitor with “very few human employees” arrives.
→Don’t lean on Jevons for fully automated tasks — if humans no longer own a necessary slice, demand growth funds tokens, not headcount.
→Pilot “coach not complete” for learning tools — step‑by‑step reasoning, block answer dumps; add teacher/parent visibility where kids are users.
→If you serve non‑English users — treat language quality as a product/equity feature; English‑first models leave remote health/ag users behind.

Gates’s risk stack

1Imminent: humans using AI for cyber + bio attacks (already capable).

2Near‑term: white‑collar + eventually blue‑collar job destruction after reliability thresholds.

3Psychosocial / kids: cognitive offloading, companionship, unmonitored use.

4Later: control problem / systems acting against interests (RL without hard supervisory absolutes).

5If control held: abundance → purpose/meaning problems (philosophical), not malaria/doctor scarcity.

Policy ideas he pushes

Require

Supervisory / monitoring layer on all models (closed + open)
Stay on platforms that can verify safeguards weren’t removed
Engage society now — don’t wait for mass‑casualty proof

Labor & tax

Tax AI/robot “units of labor” like human FICA into pensions
Pre‑declare human‑reserved jobs (care, parts of medicine/education)
Expect ~20‑year+ transition; minimize damage while retaining control

Don’t waste effort on

Blocking data centers as an AI slowdown (builds elsewhere)
“Industry charter against profit max” when many labs compete
Binary for‑AI / against‑AI party politics

Concrete examples

AP Bio shock → coding shock: 4:30 Pre‑ChatGPT demo: perfect AP Biology five on novel questions. Later: Claude/coding models match his lifelong coding strength and find flaws — still weak at higher‑level “what to do” strategy (but foundation strategy reviews now use models as peers).
Glass Wing / Mythos‑class cyber: 13:10 Agents find injected exploit chains humans missed for decades; limited early‑access “fix bugs first” can’t cover the code surface — period of extreme vulnerability.
Filter overtune vs no filter: 20:40 Heavy safety filters can block legitimate cancer research questions (push users to weaker models or ungated open weights) — hard dual‑use separation is why monitoring + requirements matter more than vibe‑based refusals.
Mechanical Turk shutdown: 44:22 Economist skepticism (“how can MTurk still exist?”) → shut down a week later — “the more you know, the more concerned you are.”
Ag AI at scale: 50:45 Million+ farmers in India; Africa weather/disease/fertilizer guidance — one domain moving “full speed” while health/education lag on tokens, phones, language, regulation.
KDUm NYC classroom loop: 62:06 Short daily checks → immediate teacher dashboard of stuck concepts; small‑scale stunning results, classic education scale‑up risk when moving past volunteer teachers.

Rough chapters (from video description + captions)

0:00 — Intro: why Gates’s alarm stands out
1:59 — Letterman 1996 → what he used to think about AI
4:30 — When AI became superhuman (AP Bio, coding)
10:28 — Race vs safety (Jensen Huang counterpoint)
11:20 — Cyber and bio thresholds
14:47 — Why self‑regulation fails
24:57 — Race with China; monitored platforms
27:03 — How AI will disrupt jobs
34:23 — No‑scarcity job market
35:48 — Taxing AI labor
37:32 — Human‑reserved jobs
38:36 — AI for global languages
40:16 — Economic impact on developing nations
42:53 — Jobpocalypse debate / inequality vs equality
52:22 — Epstein questions
56:32 — Foundation spend‑down & AI
59:08 — AI and education
1:00:52 — Learning loss in rich countries
1:06:19 — AI regulations for children
1:10:08 — Political future of AI
1:11:08 — Rebuilding foreign aid
1:12:36 — Book recommendations

← Back to library Open source on YouTube ↗

Episode ~53 min Sep 25, 2026 Lecture series finale

Sarah Paine — Why wars are so difficult to end

Dwarkesh Patel · Sarah Paine (naval war college / strategy historian)
youtu.be/f6cjAEr08qk · Final lecture in the Sarah Paine series

Context: A ~53‑minute lecture on war termination — why wars are easy to start, hard to end, and harder still to end on the terms you wanted. Paine uses the Russo‑Japanese War as a model exit, then maps two failure modes: regional wars that go global, and conventional wins that turn into endless insurgencies. Frameworks from Clausewitz (culminating points) and Sun Tzu (death ground) applied to Gulf War, Korea, Iraq, and today’s conflicts.

Most actionable learnings

Exit strategy before first shot — Japan’s model war termination

20:23 Russia out‑multiplied Japan on territory, population, industry, army, navy — yet Japan won a regional war and got out. Key: before attacking Port Arthur they had already lined up Theodore Roosevelt (via Harvard alumni channels) to mediate Portsmouth. At Mukden, with officers and manpower exhausted, Yamagata “called the America card” rather than chase Moscow — knowing Russia would never request peace without an invasion Japan could never mount.

“If you start a fire, you’ve got to put it out. So it’s time to call the fire department over in the United States.”
Know your culminating point of victory — stop when the porridge is “just right”

8:19 Clausewitz pairs culminating point of attack (battle/campaign) with culminating point of victory (whole war). Stop short → you leave gains on the table; push past → costs/risks exceed benefits and you’re weaker. A tell you’ve overshot: a major third‑party intervention that changes the war’s nature. Japan’s Portsmouth timing is the Goldilocks case; Nogi’s 45,000 dead at Port Arthur (a whole Japanese army of four) shows how easy it is to burn past the attack peak for one fortress.

“Beyond that point… the scale turns and the reaction follows with a force that is usually much stronger than that of the original attack.”
Limited aims are hard to define, easy to execute; unlimited aims are the reverse — death ground

34:48 Unlimited objective ≈ regime change: conceptually simple, brutally hard to execute because you put the other side on Sun Tzu’s death ground — they fight to annihilation rather than bargain (Ukraine facing existential stakes). Limited objectives leave the defeated government in place so you must negotiate — conceptually messy (where on the continuum?), but far cheaper to finish. Gulf War I (explicit limited goals via NSD‑54; allies paid >$50B; ~258 US deaths) vs Korea’s costly waffling between status‑quo and whole‑peninsula aims.

“Unlimited wars are conceptually simple… but executing them is really difficult because you put people on death ground.”

“The political objective is always primary. It is never the military objective.” — Clausewitz (via Paine)
Two pathologies after a “win”: escalation to global war, or victory‑denial insurgency

39:32 Pathology 1: a regional fight stretches long enough for rivals to pile on (American Revolution → France/Spain/Dutch; WWI alliance misalignment; Pearl Harbor chain). Pathology 2: you defeat the conventional army but guerrillas deny stabilization — Japan in China, Washington’s army+militia cost‑imposition, later Vietnam/Iraq/Afghanistan. The math is value of the object: home teams treat it as existential; distant interveners eventually call it a day.

“If you live in theater, it’s home. It’s existential… If you’re intervening, the further away… the lower the value of the object.”
Regime change hands you the loaded diaper — mosaic societies don’t become states on command

48:58 Toppling Baghdad was the easy part; owning the aftermath was not. Mosaic societies (ethnic primary identities, neighbors with regional ambitions) resist the national‑identity glue Korea already had. Borders explode from ~158 miles (Kuwait) to 2,000+ miles; Saudi Arabia, Iran, and Turkey will meddle whether invited or not — sometimes a single shrine bombing lights a sectarian war. Know this before you choose unlimited aims.

“The problem when you do regime change is the loaded diaper is yours… now what?”

Key use case

Job to be done

Before (or during) a conflict, map each party’s limited vs unlimited aims, relative value of the object, geographic access/sanctuary, and alliance alignments — so you can stop at the culminating point instead of stumbling into global war or endless occupation.

Who it’s for

Analysts, policymakers, officers, and serious citizens trying to reason about Ukraine, Taiwan, Middle East interventions, or any “quick” regional fight that might not stay regional.

Success signal

You can state the political objective in one sentence, name what “enough” looks like, and spot when military momentum is about to drag you past it.

Do this week

→Write the exit before the entry — Japan lined up Roosevelt mediation before Port Arthur. If you can’t name who mediates and what you’ll trade, you don’t have a termination plan.
→Label every actor: limited vs unlimited — and vis‑à‑vis whom. Russia unlimited vs Ukraine; Ukraine limited vs Russia. Misreading that matrix is how you put someone on death ground by accident (or on purpose).
→Define the culminating point in advance — what battle/campaign result is “enough”? What third‑party intervention would prove you’ve gone too far?
→Score value of the object — existential for locals vs discretionary for interveners. Insurgency cost‑imposition works when that gap is wide.
→Refuse military‑objective drift — Korea’s “we’re on a roll, take the whole peninsula” subordinated politics to ops. Clausewitz: politics stays primary.
→Pre‑mortem mosaic occupation — if regime change is on the table, inventory ethnic fault lines, neighbor ambitions, and border miles you’ll suddenly own.

Clausewitz + Sun Tzu toolkit

1Culminating point of attack — battle/campaign peak; push past and the reaction hits harder than your original blow.

2Culminating point of victory — whole‑war peak; third‑party great‑power entry is a classic overshoot signal.

3Limited vs unlimited objectives — bargain with a surviving government vs regime change / no choices left for the defeated.

4Death ground (Sun Tzu) — annihilation stakes → fight‑to‑the‑death; transforms “failed states” into lethal adversaries.

5Value of the object — how much the political aim is worth to each side; drives who quits first under cost imposition.

Case contrast

Worked (model exits)

Russo‑Japanese 1904–05: exit planned pre‑war; Portsmouth before Russian rail/manpower advantage locked in
Gulf War I: NSD‑54 limited aims; coalition paid the bill; tiny coalition deaths; Kuwait restored — no “loaded diaper”

Broken (pathologies)

Korea: waffled limited↔unlimited; Chinese intervention after push toward Yalu; years of costly stalemate
Iraq 2003 / China theater for Japan: conventional win → insurgency victory‑denial; mosaic identity + neighbor meddling
Rev. War / WWI chains: regional war → global pile‑on as rivals exploit protracted fighting

Questions before any war

Who’s involved / interested? 51:06 Map primary adversaries and every neighbor who gains from prolonging or ending the fight.
Limited vs unlimited — and vs whom? Same actor can be unlimited against one neighbor and limited against another; alignments follow from that matrix.
Relative value of the object? Existential for whom? “Nice to have” for whom? Expect the high‑value side to outlast the low‑value intervener.
Geographic access & sanctuary? Uninterrupted supply lines (China↔Korea), invulnerable rear areas, and border length change what “victory” can mean.

Rough chapters (from video timestamps)

0:00 — How Japan beat a far bigger Russia
8:18 — Clausewitz’s culminating points of victory
14:24 — What finally made Russia give up?
26:54 — Gulf War: the war the US didn’t pay for
30:31 — Korea and the cost of shifting war aims
34:48 — Sun Tzu’s “death ground”
39:32 — How regional wars go global: WWI to Pearl Harbor
46:38 — Why insurgencies deny victory: China to Iraq
51:06 — Questions to ask before any war

← Back to library Open source on YouTube ↗

Episode ~91 min Sep 29, 2026 GTM roundtable

GTM Roundtable: The State of AI-Native Sales w/ Clay, Parallel, Avoca & Profound

The Crew Podcast · host Chris Balestras · guests Liam Mulcahy (Head of GTM, Parallel), Jack Gashi (VP Sales, Avoca), Mark Ebert (SVP Revenue, Profound), Todd Busler (Head of Sales, Americas, Clay)
youtu.be/d22ydIw6IJg

Context: Four revenue leaders at AI-native companies compare what the job looks like in years two–three: territory plans that get shredded mid-year, AE productivity math that nobody trusts, deal caps to stop turtling on inbound, and the hiring honesty that “the first year is going to hurt.” Closes on spouse negotiations, young kids as a recruiting edge, and keeping a few things written in pen while the rest stays pencil.

Most actionable learnings

Write most of the plan in pencil — protect a few things in pen

4:52 Annual territory and capacity plans collapse when you’re hiring this fast — mid-year reshuffles after people already dug into accounts. Leaders at the table now spend less time on planning because they throw the plan out two months later (often for good reason). Counter-move: let product priors, headcount, and goals stay erasable, but keep a few “semi-sacred” items harder to reverse — sales process and enterprise paper-process expectations — so the company doesn’t take everything down to the studs every month.

“Everything’s written in pencil… you do have to get the team to write some things in pen.”
Two-day evals, four-month paper — set founder expectations early

7:36 Product can finish a Fortune‑50 evaluation in ~two days; legal/MSA still takes ~four months. Founders experience that as “the deal was done” after the eval. Train the company that paper is a separate fight — and keep that expectation less reversible than weekly product narrative updates.

“We’ve only just begun to fight… we’re just getting into the MSA.”
Cap open deals / cut inbound when reps turtle — load until win rate hurts

33:51 Traditional capacity math says these teams could double headcount; per‑AE productivity looks “made up” vs. three years ago. Practical ops: watch rolling 30/60 win rate and open ops per AE like a hawk. Past a VO / open‑op threshold, shut off inbound so parking-lot reps can’t sit on 30 deals while a hungrier AE with eight would crush the same lead. Thesis: load until win rate suffers — then remember variance is huge (some reps handle 30; some are best at three), so expect Slack pushback 10 minutes after any team-wide deal cap.

“I’d rather give an op to a good AE who has eight open opportunities than someone who’s got 30.”
Forget the equity for most AEs — optimize W2 and career acceleration

46:40 Jack’s blunt hiring line: for ~90% of AEs, earning power is dwarfed by W2/commission. A rep riding 75→200 ARR can grab a huge share of the commission pool over a few years and look at equity later as secondary. For sharp mid‑20s candidates, sell the seat as a career accelerator toward the VP role where equity finally matters — not as a lottery ticket.

“The first thing I’d say is forget the equity.”
Want coin-operated + long view — not mercenaries maximizing year‑one W2

49:55 “Coin-operated” = show me the scoreboard and comp plan and I’ll work hard — money matters. That is different from one-year extractors who only maximize earnings or Series‑B candidates who lead with equity/moat questions. Hire people who will take a hard job because the pay is good and stay long enough to learn the game. Preview the pain on purpose: early-stage glory is not equity cosplay — “this is going to be the worst 12 months of your life.”

“This is going to be the worst 12 months of your life… you are going to get the [stuffing] kicked out of you.”
No points for Lovable pages or more Clay tables — ask how many people they called

21:16 AI-native reps can rabbit-hole on tools. Culture counterweight (Clay / Becca): in pipeline review, ignore the shiny app demo and ask “how many people did you call today?” OGs who still prospect daily remain extremely valuable; tool fluency without outbound volume doesn’t score.
Put post-sale into sales with quota + club — and negotiate nuclear keys at home

0:42 Year‑two/three motion: move AM into the sales org — redo comp, carry quota, eligible for club (“you’re a slinger now”). On the personal side: agree a one-year “helmet on” window with your spouse and give them nuclear keys to eject if family structure breaks; hiring parents with young kids can be a deliberate recruiting advantage when GTM leaders already live that game.

“Like you have the nuclear keys… if it starts with the actual stuff that matters, you will eject.”

Key use case

Job to be done

Operate an AI-native revenue org when headcount, product, and quotas move faster than annual planning — without letting reps turtle on inbound or candidates romanticize early-stage equity.

Who it’s for

VP Sales / Heads of GTM at fast AI startups, and AEs deciding whether to join them.

Success signal

Plans stay pencil where they should; paper-process and sales-process norms stick; inbound routes to reps who can work it; hires know the first year will hurt.

Do this week

→Label pencil vs pen — list what can be rewritten monthly (territories, headcount bets) vs. what needs friction to change (sales stages, enterprise paper SLAs).
→Instrument open ops + win rate — rolling 30/60 close rate from SQL and open opportunities per AE; draft an inbound shutoff threshold.
→Rewrite the AE pitch — lead with W2/comp path and “worst 12 months” preview; demote equity for IC roles.
→Pipeline ritual — one question that kills tool theater: “How many people did you call / meetings booked?”
→If moving AM → sales — announce quota + club eligibility as a deliberate identity shift, not a silent comp tweak.

Around the table

Liam Mulcahy — Head of Go-To-Market, Parallel
Jack Gashi — VP of Sales, Avoca
Mark Ebert — SVP of Revenue, Profound
Todd Busler — Head of Sales, Americas, Clay
Chris Balestras — Host; co-founder, Crew (GTM recruiting / media / investing)

Rough chapters (from show notes)

0:00 — The worst 12 months of your life
0:51 — Welcome to the Crew podcast
2:05 — Headcount planning heading into next year
2:50 — Why most of the plan is written in pencil
3:55 — The four-month paper process nobody can explain
11:53 — When your best rep is also shipping code
20:57 — Moving the AM function into sales
21:17 — What you actually tell a candidate about quota and club
24:47 — Interviewing AMs and writing the rules as you go
29:33 — How do you do capacity planning right now
29:55 — Why the productivity math doesn’t add up
31:11 — Reps turtling on deals
32:33 — Setting a deal cap ten minutes after the meeting
34:14 — It’s a race, and what that does to coverage
44:55 — Forget the equity
48:50 — Coin-operated vs. money-motivated
56:36 — Hiring someone with a young kid at home
61:09 — Quality of time, and what the family actually sees
75:04 — The nuclear keys

← Back to library Open source on YouTube ↗

Episode ~36 min Sep 17, 2026 The Lightcone · YC

The State of Startups in 2026

The Lightcone (Y Combinator) · Garry Tan, Jared Friedman, Diana Hu & Harj Taggar
youtu.be/yslXlV2BP_Y

Context: A ~36‑minute Lightcone roundtable (posted Sep 17, 2026). YC partners share batch data from the last ~12–18 months: hard‑tech share jumped 8%→20%, median end‑of‑batch revenue ~$8K→~$20K MRR, and almost one in five accepted companies is now solo‑founded. Thesis: startups are moving from bits to atoms, software is becoming an AI harness that does the whole job, and the scarce skill is knowing what to build.

Most actionable learnings

Bits → atoms is the batch story: hard tech 8%→20% across robotics, manufacturing, defense, compute & power

1:06 Diana’s breakdown of recent accepts: robotics 1%→6–7%, industrial manufacturing 4%→10%, defense 1.5%→~5%, semiconductor/photonics ~1%→~4%, power infrastructure ~1%→~3%. Physical stacks roughly tripled–quintupled. One in six founders in the current summer batch has a PhD — expertise matters again when you’re touching silicon photonics or metal supply chains (e.g. Nox Metal rebuilding US metal manufacturing at “software growth rates” off defense demand).

“Things that actually touch atoms and not just bits.”
AI is making hard tech fundable again — and VCs flipped fast when SaaS looked soft

3:29 Historically hard tech was hard because of supply chains + a huge software surface; AGI compounds the opposite way — AI helps with the software / planning / ops layer so technical founders can ship atoms. Investor appetite snapped back mid‑Winter batch when SaaS stocks dipped and Claude Code surged; by Demo Day hard tech was suddenly “interesting” again. Silicon Valley venture was built for hard tech; the SaaS decade was the detour.

“Make something people want — and people sure do want those things right now.”
Software isn’t dead — become the harness agents use (or get MCP’d and lose the moat)

15:49 Systems of record (Salesforce/Slack) still work when agents use them more than humans. Path: treat agents as customers; be the place work happens, not a read/write API. Release an MCP without becoming the harness and your data moat walks. Batch signal: full‑stack “does the job” companies (not point SaaS) rose from ~10%→25%+ of accepts — insurance broker, clinical intake, medical billing end‑to‑end. Juicebox: LLM people‑search → agent that contacts/schedules → 2–3× revenue per account.

“If you are a system of record you either will be preyed upon… or you kind of have to be a harness.”
Speed is real: median batch exit MRR ~$8K→~$20K; some go $0→seven figures in 3 months

17:59 Median company still enters at $0. End‑of‑batch median jumped hard; averages higher. New: companies hitting seven‑figure revenue during the ~3‑month batch (used to take ~18 months). Drivers: agentic coding (~year since Opus‑class agents worked) → more mature products faster; buyers pay more when software does the job vs tracks it. Separately: YC has funded 12+ companies each doing >$10M/yr (some hundreds of millions) selling data / RL environments to labs — a stealthy category labs reportedly spend ~$1B on.

“People want their job just be done and are willing to buy software that just gets the job done.”
Robotics needs specialized / fine‑tuned models — foundation models are a starting point, not the product

27:08 Physical 3D space has far more degrees of freedom than language; robots need real‑time action plans (no “come back later”). YC robotics companies fine‑tune Physical Intelligence models for verticals (data‑center cabling, boxing SKUs) — none reportedly ship PI out of the box. Parallel: proprietary data + open‑weight fine‑tunes can beat frontier generalists for a niche (River AI / Tinker‑class tooling).

“You have to actually fine‑tune it for like your specific case… in order for it to work.”
Solo founders ~5%→18–19%; scarce skill is knowing what to build (and prompt), not being the CTO

29:38 Classic combo was hustler + world‑class technologist. Coding agents collapse the “must be exceptional at building” bar — so more people can start alone (Instacart/Coinbase/Rippling‑era solos were rare exceptions). Co‑founders still help (elite peers signal + later execution); trend is add co‑founders after traction, not 50/50 day one. YC is already “YC for solo founders.”

“Knowing what to prompt and knowing what to build is so much more difficult and valuable than… the CTO being able to code the thing.”
Experienced founders are back — taste + years managing people transfer to managing agents

32:54 Late‑30s/40s/50s founders who “know where the dragons are” are unusually powerful now (Peter Steinberger as canonical). Gatekeeping (must have cool investors / perfect co‑founder) matters less; high‑order bit is opinionated taste about what to build. Managing coding‑agent fleets rhymes with managing eng teams — career managers (Shopify’s Tobi, etc.) may outspin a brilliant 19‑year‑old on agent ops. Advice: just start prompting; wire up each new model and re‑attack the bugs last month’s model couldn’t crack.

“You need to know what to build. That’s like the high order bit now.”

“You wake up in the morning, you like wire up a new model… and then these things that even a month ago you’re just like why isn’t it working? It just starts working.”

Key use case

Job to be done

Decide what to build in 2026 with YC’s batch signal: lean into atoms (robotics/defense/mfg/compute), ship end‑to‑end agent products that do the job, or sell scarce data/RL envs — and bet on taste + iteration speed over classic co‑founder/coding gatekeeping.

Who it’s for

Founders choosing a category; operators at systems‑of‑record companies who must become harnesses; experienced ICs/managers wondering if now is their window.

Success signal

You can name whether you’re selling atoms, a job‑completing agent, a harness, or lab data — and your weekly loop includes wiring new models + deepening domain taste, not just shipping more CRUD SaaS.

Do this week

→Pick a stack lane: atoms (robotics/defense/mfg/power/photonics), end‑to‑end job agent, system‑of‑record→harness, or data/RL envs for labs.
→Audit your product: point solution operators use vs agent that completes the workflow — design the latter (Juicebox‑style upsell).
→If you own a system of record: plan harness features before you ship a bare MCP that lets someone else own the workflow.
→Robotics / vertical AI: budget for domain fine‑tunes + proprietary data; don’t expect general foundation models to ship your use case raw.
→Solo / experienced builders: start without waiting for the perfect co‑founder; add elite peers after traction; your edge is knowing what to build.
→Weekly model upgrade ritual: wire the newest model, re‑run last month’s unsolved bug list, ship whatever suddenly works.

Batch stats they cited

1Hard tech share: 8% → 20% of recent accepts.

2Category lifts: robotics 1→6–7%; mfg 4→10%; defense 1.5→5%; semis/photonics ~1→4%; power ~1→3%.

3Revenue: median end‑of‑batch MRR ~$8K → ~$20K; some $0→7‑figure in 3 months.

4Product mix: full end‑to‑end “does the job” companies ~10% → 25%+ of batch.

5Founders: solo ~5% → 18–19%; ~1 in 6 summer‑batch founders hold a PhD.

6Hidden boom: 12+ YC cos each >$10M/yr (some $100M+) on data / RL environments.

Concrete examples

Nox Metal: 8:26 Rebuild US metal supply chain in empty Detroit factories; grow at software rates by serving new defense startups that legacy sleepy suppliers can’t keep pace with (Stripe‑for‑defense‑ecosystem analogy).
Dipole Labs / Lamb Labs / Bot: 10:29 Optical data‑center switches (photon GPU↔GPU), alternative processors, ternary low‑precision silicon — compute demand so high even older A100 hours appreciate.
Juicebox: 20:24 From LLM candidate search → agent that outreaches and schedules; recruiters keep culture‑fit judgment; revenue per account expected to 2–3×.
Scale → AfterQuery / DataCurve era: 22:52 Data/RL‑env sellers often stay stealthy (incentive not to advertise); labs buy egocentric / real‑world robotics data in 8–9‑figure deals.
Peter Steinberger: 32:54 Early‑40s ex‑dev‑manager, AI‑pilled early, taste from “around the block” → canonical experienced‑founder win in the agent era.

Rough chapters (from video description)

0:00 — Intro
1:06 — Startups Are Moving From Bits to Atoms
3:29 — Why AI Is Making Hard Tech Easier
5:21 — Defense, Manufacturing, and the Return of Hard Tech
9:48 — AI Compute Is Becoming a Physical Infrastructure Problem
12:34 — Robotics Is Approaching Its ChatGPT Moment
15:40 — Software Isn’t Dead. It’s Becoming the Harness
17:59 — Why AI Startups Are Growing Faster
22:52 — The Hidden Boom in Data and RL Environments
27:08 — Why Robotics Will Need Specialized Models
29:38 — The Rise of the Solo Founder
32:54 — Why Experienced Founders Are Back
34:59 — What Founders Should Do Right Now

← Back to library Open source on YouTube ↗

Episode ~21 min Sep 3, 2026 Y Combinator

Paul Graham On Startups, Ambition, and Great Founders

Y Combinator · host Vivian Shen (YC Visiting Partner) · guest Paul Graham
youtu.be/5bxp78i96S8

Context: ~21‑minute sit‑down at the original Mountain View YC office (47th batch, 21st year). Vivian Shen presses PG on why today’s startups look more “serious,” what actually drives founders day‑to‑day, what “formidable” means, whether lean/cheap is dead in the AI era, what surprised him about AI, and what has and hasn’t changed about YC.

Most actionable learnings

Day-to-day fuel is fear of disaster — not the billionaire dream

6:11 Dutifulness isn’t enough for the obstacles. Ambition helps, but at any given moment founders are usually saving the “model train set” — servers crashing, looking like a fool — not calculating exit math. Wealth often arrives as a surprise after a decade of heads‑down work (sometimes PG is the first to do the last‑round math for them). Use this when coaching: design urgency around near‑term failure modes, not remote IPO fantasies.

“The engine of my model train set is falling off the edge of the table. I need to go and save it.”
Formidable = someone who gets what they want (that’s why investors care)

9:48 Private Jessica/PG language: formidable people get what they want in any situation. If you own stock alongside them, interests align — their wins are yours. Screen for evidence of getting outcomes (not just credentials or polish). Sam Altman as the archetype: extremely formidable before YC even accepted him; ambition is mostly inborn.

“Do you get what you want? … How formidable are you if you don’t get what you want?”
YC-as-resume-badge is a trap — startups have no easy majors

8:35 Early on people applied for prestige like Harvard. Difference: Harvard has soft majors you can coast through; a startup doesn’t. If it works it’s your life’s work (no more badges); if it fails it’s a weak credential. Starting a company is among the least efficient ways to “seem cool” — years of brutal hardness first. Don’t apply (or hire) for signaling.

“It’s like applying to Harvard and being forced to study theoretical physics.”
Frighteningly ambitious: don’t attack the incumbent head-on — wait until their model is obsolete

4:38 2012 essay ideas that got done: a “new Google” arrived not by cloning search, but when what users wanted (answers) no longer required web pages — OpenAI made search obsolete for PG personally. Pattern: wait for a paradigm shift that makes the incumbent’s model wrong, then build the new default. Today’s “serious” batch includes ICBM‑style cargo and multi‑approach cancer — death of a thousand cuts if a single cure doesn’t arrive.

“What I really wanted from searches was not web pages. I wanted information.”
Lean / little money isn’t dead — adjust scope to cash and sell the next milestone

11:51 PG hasn’t bought “lean startup” as a brand (never read the book) but rejects the idea that starting with little money is obsolete. Tokens/GPU bills look expensive now; inference prices at a given quality still crash. Even rockets: you can’t build the vehicle on nothing, but you can design, simulate, show experts, then raise for the next stage. Star Cloud‑class move: white paper + booked launch as commitment proof — works better when you already have domain cred.

“You just do what you can on the money you got. As long as you can get to some kind of milestone…”
AI surprise: human-shaped bullshitter first, then perfect — AGI is a smear, not a finish line

13:25 1980s expectation: perfect fly → mouse → human. Reality: full human capability with undergrad‑paper BS first, then climb toward reliability (jagged frontier: Riemann‑hypothesis vibes vs can’t get restaurant hours). From far away AGI looked like a thin finish line; standing on it, the line has width — some bits past, some mid‑smear. Practical: don’t wait for “AGI done”; ship against the parts of the smear that already work for your job.

“It turns out to be a smear and we’re on it.”
Shipping pace is still the best predictor — almost everything about startups is the same

16:32 Powerful AI tools haven’t erased variance: many batch companies still don’t ship fast enough. Bottleneck isn’t only generation speed — you have to invent what to ship. Nearly everything is unchanged; the weird new cost is giant AI/token bills replacing “salaries were the only big line item.” YC advantages still compound: less loneliness, peer problem‑solving, and “YC GDP” (sell into the batch’s early adopters who decide fast). Next trillion‑dollar company comes from formidable founders, not a prescribed idea — and founders still look like they did 20 years ago.

“So far almost everything is exactly the same.”

“Where the next giant company comes from is it comes from the right founders.”

Key use case

Job to be done

Choose and coach founders (or yourself) for serious, frighteningly ambitious work: optimize for near‑term failure aversion + getting what you want, pick ideas that ride paradigm shifts, keep capital lean by milestone, and keep shipping as the north star even when AI bills spike.

Who it’s for

Founders applying to / in YC; angels/partners screening ambition vs prestige; operators wondering if AI changed the startup playbook.

Success signal

You can point to evidence the founder gets outcomes; the company ships weekly; the idea waits for (or creates) an obsolete-incumbent moment — not a credential play.

Do this week

→Rewrite motivation talk: list this week’s “train set falling off the table” risks and assign owners — drop vague “think bigger” pep talks.
→Formidable screen: for each founder/hire, write one concrete story of them getting what they wanted against resistance.
→Ambition audit: if the idea is “new X,” ask whether X’s model is obsolete yet — or you’re attacking head‑on.
→Capital plan: define the next raisable milestone you can hit on current cash (design/sim/booking beats “need a full rocket”).
→Ship scoreboard: track releases this week vs last; AI tooling is not an excuse for slow invent→ship loops.
→Batch leverage: if you’re in a cohort, invent one “YC GDP” sale or help‑ask this week — early adopters who must listen.

Extra PG lines worth keeping

Serious vs old days: 1:32 “Supposedly good old days” is a perennial attack pattern since ~2008; Reddit‑era ideas aren’t “intercontinental ballistic cargo.”
Ambition suppressed, not absent: 7:44 Rare “became ambitious in batch” stories — more often pushy parents / obedience training hid ambition that was already there.
How YC began: 18:34 Angel firm + summer‑job alternative for undergrads → accidentally invented the batch; now “same thing but more of it,” pods feel like 2012‑scale YC.

Rough chapters (from video description)

0:00 — 21 Years of YC
1:07 — Why Today’s Startups Are More Ambitious
4:09 — Frighteningly Ambitious Ideas
5:13 — What Actually Motivates Founders
9:28 — What Makes a Founder Formidable
12:54 — What Surprised PG About AI
17:58 — How YC Started and What Hasn’t Changed

← Back to library Open source on YouTube ↗

Episode ~49 min Dec 11, 2025 Odd Lots

This Is What It Takes to Get a Data Center Financed

Odd Lots · Bloomberg Podcasts · hosts Joe Weisenthal & Tracy Alloway · guest Travis Wofford (partner and chair of the corporate department, Baker Botts, Houston)
youtu.be/cvNO9pZAhM0

Context: ~49‑minute studio conversation (recorded as of December 2025) on how data centers actually get financed. Wofford is a deal lawyer who staffs power, water, fiber, land, and the finance layer as one package. Hosts frame the scale: a Morgan Stanley note they cite forecasts about $2.9 trillion of global data-center spend through 2028, against ~$950 billion of S&P 500 capex in 2024 and generative-AI revenue they put around $16 billion. The usable part is the sequencing, the collateral split, and where the timeline actually breaks.

Most actionable learnings

Match the check to the phase — don’t securitize a site that isn’t powered

20:58 Three buckets, in order. Development capital is usually equity. Construction debt shows up only after permits and power. Takeout — securitization and the like — comes after the project is built. If a model shows a bond funding an unplugged site, the structure is ahead of the asset. Delay risk (a neo-cloud cutting growth because a third-party data center slipped) sits in the equity and construction window, not in a finished ABS.

“You have your development capital, that’s often equity funded. You have your construction capital, so you get a big construction loan once you’ve got your permits and your power. And then, after that’s been completed, then you do your takeout financing.”
Hyperscalers leave the shell off their balance sheet because tech ROIC beats infrastructure ROIC

9:14 Cash-rich AI companies still finance the building, fiber, and power connection through someone else. Infrastructure is lower-margin but long, predictable, and understood; the chip/software margin stays in-house. The legal tech is old: whole-business securitizations moved from cell towers to residential solar to data-center CMBS/ABS. Rating agencies already know the shape — SPV holds the assets, a separate entity bills and collects, another runs O&M, sized to stand through an anticipated repayment date.

“The return on their invested capital is much better on the tech side than it is on the infrastructure side.”

“It may not be exactly the same, but it rhymes.”
IG securitization underwrites the lease and the shell — not the GPU

12:30 ABS is the business of the powered shell or turnkey site (lease cash flows plus related power/services). CMBS is more mortgage-weighted; credit-tenant leases sit closer to the tenant. Risks that actually get priced: tenant quality, lease term, whether rent is below market (stickier), and whether the building survives technology change. For an investment-grade deal he puts advance rate around 40–50%. Rated final maturity can be 25–30 years, but principal is structured to be repayable in about 5–7 years inside a 10- or 15-year lease that can be re-upped — so the anticipated repayment date is covered before you need a re-lease. GPU residual value is a private-credit or equity question, not the ABS thesis. Accounting lives drifting from ~3 years to ~6 years change depreciation and EPS; they are not a lender’s repayment clock.

“The GPU life itself is more a question on a private credit story or an equity story. On the securitization side, they’re more focused on the turnkey data center itself.”

“The full principal amount of the bond would be able to be repaid within 5 to 7 years, but the lease itself is a 10-year lease or a 15-year lease.”
Underwrite chips on a 2–3 year product cycle, then a second and third use

15:07 Private-credit lenders are not ignoring aging. Nvidia’s H100/H200, then Blackwell (B100/B200), then the next generation he names (“Ruben’s”) land every two to three years. Financing a data hall of those for six or seven years is, in his analogy, financing the laptop on the table — except the check is tens or hundreds of millions. Economic life is a cascade, not one training job: training first, inference if the site is in a tier-one place like Virginia and near a city, then CPU halls for compute and analytics you knew were coming. Ask “when must this debt be repaid?” against the replacement cycle, and write the second use into the base case instead of hoping cloud-era depreciation still holds.

“Every 2 to 3 years you have these new chipsets… Do I want to be financing these laptops… for 6 years or 7 years?”
The choke point is interconnection — and most of the queue was never going to be built

18:38 Chips and financing are not the binding constraint he sees; power is, and power means permission. Load (the data center drawing from the grid) and generation are two separate interconnections you need at roughly the same time. At large-data-center scale, generation approval through FERC and the ISO/RTO (ERCOT in Texas) is on the order of five years — and that is only approval, then you connect and build. Sponsors who walked in quoting 18–36 months watch the date slip, which he compares to United boarding groups: a group-one ticket does not mean you board first. Virginia still wins on subsea cables and a deep cluster, but it is saturated; Texas/ERCOT is where speed and power are easier. Powered land — a REIT that has the interconnect and not the cooling design — is the trade, because decades of half-finished studies can be retargeted. About two and a half years before this AI wave he recalls ~2,600 GW in the interconnection queue, with an expectation that ~80% would never be built (and he calls that typical going forward). Adding a data-center layer to solar, batteries, or wind can make a previously uneconomic site work. Haircut the headline queue: five or ten applicants often file the same project. Strip that wash and what remains is hyperscalers, large enterprises, and other serious tenants; speculative names are what fail if there’s a shock. Behind-the-meter gas is not a full substitute: if the tenant walks (metaverse off, AI off, GPUs to quantum), generation with no grid tie is a stranded asset. Water is “not a problem unless it’s actually a problem” — his shop tracks all 98 Texas water districts for both supply and wastewater offtake, and picking the wrong one moves the timeline. Politically, December 2025 is hotter than January (moratorium talk in some states; not the Abbott/Texas posture). “Water-positive” in hyperscaler reports means money into watersheds, not a mood.

“If you can’t have both, then you’re probably not going to be able to support such a large facility.”

“There was, I want to say, 2,600 gigawatts in the interconnection queue and we were expecting 80% of that would never actually be constructed.”

“You want to have that generative capacity interconnected with the grid so that you don’t have a stranded asset.”
Pick the credit on purpose: non-dilutive private credit, one IG tenant, a name that lowers the spread

28:57 Private credit’s real attraction, when it offers it, is non-dilutive capital. Equity is dilutive and carries a risk premium; a lot of private-credit stories he sees target maybe a 15% IRR, sometimes 10–12% upfront plus a MOIC or IRR catch-up on takeout — still usually cheaper than the equity slice, and often looser than a bulge-bracket credit desk. Don’t underwrite “customizable” as a feature list. For investment-grade bonds, the market today is wholesale: one tenant, investment-grade. He does not expect Fitch, S&P, or Kroll to rate a pool of non-IG colo tenants investment grade until the book is thousands of tenants across hundreds of sites. A DOE Loan Programs Office guarantee, a DoD/“Department of War” program, or a Microsoft/Google project guarantee is “show me who you walk with”: at ~40% LTV with an IG tenant covering longer than the bond, you may not need the backstop, but it still cuts the spread you have to charge, and that spread becomes the next project. Operator quality is priced — a decade-long developer borrows cheaper than a two-year Johnny-come-lately — and outages (the hosts’ CyrusOne/CME cooling example) are underwritten, not ignored. Service-level cash for uninterruptible power sits in a reserve that is supposed to hit the SPV’s equity distribution, not the bond coupon. Unused contracted cash flows that don’t get an advance rate are a second capital sleeve, not free yield.

“One of the most attractive things that private credit can, but doesn’t always offer, is non-dilutive capital.”

“I don’t think that the expectation should be that you’re going to get an investment-grade rating on a pool of non-investment-grade tenants. At least not right now.”

“If you are a blue-chip operator and developer, you are going to have a lower cost of capital than you would otherwise.”

Key use case

Job to be done

Decide whether a data-center project is financeable, and with which instrument: stage equity → construction loan → takeout, collateralize the powered shell and the lease rather than the GPU, and don’t pay up for interconnection or land until the queue, the water district, and a grid tie for any on-site generation are real.

Who it’s for

Sponsors, infra investors, private-credit desks, and hyperscaler finance teams choosing ABS/CMBS vs bank vs private credit vs a guarantee.

Success signal

The model’s debt draw matches permits-and-power, the 5–7 year repayment fits inside the lease, GPU residual is in equity or private credit, and the megawatts survive a duplicate-application haircut.

Do this week

→Phase the capital: label every dollar as development equity, construction loan (only after permits + power), or takeout. Kill any term sheet that securitizes an unplugged site.
→Split the memo: shell/lease (tenant, term, below-market stickiness, 40–50% IG advance, 5–7 year ARD) in one section; GPU residual in another, marked private credit or equity.
→Reset chip life: replace a 6-year straight-line GPU cash-flow case with a 2–3 year Nvidia cycle plus a written second use (inference, then CPU analytics).
→Critical path: put load interconnection and generation interconnection on one timeline. Stress five years to approval, then connect and build — not 18–36 months.
→Haircut the queue: before buying powered land, drop projects that are duplicates (he sees five or ten filings for one site) and assume on the order of 80% of queued GW never gets built.
→Don’t strand the turbine: if the plan is behind-the-meter gas, require a path to interconnect that generation so a departed tenant doesn’t leave a useless plant. Name the water district (supply and offtake), not just “we have water.”

Extra lines worth keeping

Structures rhyme: 6:20 “There’s nothing new under the sun.” Cell-tower ABS → residential solar → data-center CMBS/ABS. Same technology, a few variables understood differently.
Guarantees are company you keep: 27:37 “Show me who you walk with, and I’ll show you who you are.” A US backstop or a Microsoft/Google guarantee means he doesn’t have to charge an extra increment for credit risk — “velocity of money” into the next project. At 40% LTV with an IG tenant longer than the bond, “do you really need it? Probably not, but it can help lower the cost of capital.”
Water is a district choice: 23:31 One Baker Botts water lawyer “knows all 98 water districts in Texas, and knows which one you’re supposed to go to if you want to get your project done quickly, and which one to avoid.” Potable is potable; the issue is supply plus chemically treated offtake. “Water-positive” means paying into watersheds.
Queue inflation: 44:04 “You have five or 10 different people making applications for the exact same project.” After that wash: infrastructure-grade tenants get built; speculative assets probably don’t, especially after an economic shock. Turbines are a separate, real bottleneck.
SLA cash hits equity: 35:12 Operators contract for “uninterruptible, non-intermittent power.” A reserve for those payments is meant to reduce the equity distribution from the SPV, not the bond payment. Squirrels still chew wires; that’s credit risk, which is why decade-long operators price tighter.

Rough chapters (from video description)

0:00 — Introduction
1:41 — Scale and financing requirements
3:54 — Guest introduction
5:18 — Market activity and historical context
7:28 — Securitization structures
8:51 — Off-balance-sheet rationale
10:24 — Risk assessment
12:19 — Deal structuring and credit protection
13:52 — GPU lifecycle and technology risk
17:07 — Location strategy and saturation
18:27 — Power and interconnection
20:51 — Project financing phases
23:48 — Water and regulation
24:48 — Politics and public-private partnerships
27:21 — Loan guarantees and credit enhancement
28:55 — Private credit
30:34 — Tenant diversification
31:39 — New instruments
32:59 — Operational risk and insurance
37:39 — Powered land
38:52 — Build-own-operate vs grid
40:59 — Crowding out
43:40 — Bottlenecks
44:54 — Wrap-up

← Back to library Open source on Apple Podcasts ↗

Episode ~68 min Oct 10, 2026 20VC · 20Sales

Inside Ramp’s Sales Playbook: How to Build a Sales Machine

The Twenty Minute VC (20VC) · 20Sales series · host Harry Stebbings · guest Max Freeman (SVP of Sales, Ramp; first sales hire, now leads 300+ sellers)
Apple Podcasts episode

Most actionable learnings

Pay 2× for horsepower, demand 4–6× output — and hire “mispriced” sellers

5:19 Ramp hired investment bankers into SDR seats (several now run large parts of the business): they work past nine-to-five, bring business acumen, and land with CFO buyers (“I just quit Goldman to join this credit card business”). Early teams have no time to teach basics, so this de-risks execution. The deal: about 2× standard SDR pay, but a target of 40–60 qualified meetings — 4–6× a normal SDR. Treat recruiting like Moneyball: the top 1–2 sellers at a company that grew past nine figures despite a low NPS, a premium price, and weak G2 reviews had to create their own demand — put them in a better product and they explode.

“We’ll pay 2x what a standard SDR is going to make. But guess what? You’ve got to book 40, 50, 60 meetings.”

“You have to treat go-to-market recruiting like Billy Beane.”
Interview by decomposing the number and reconstructing deals

11:22 “I did $2M” means nothing until you split it: inbound vs outbound vs expansion vs founder-sourced vs inherited pipeline. Then make them walk two won deals chronologically (how did you advance it, what business value, which workflow tied to which outcome) and one loss that keeps them up at night (single-threaded? no champion? didn’t qualify?). Also test account research and fundamentals — executive presence, second- and third-level discovery, quantifying pain. Final filter: would I buy from this person? Leave 10–15 minutes for their questions; the quality of those questions is the signal. His own misses came from over-indexing on big-brand pedigree instead of agency and all-in obsession.

“If they can’t unpack why they’ve lost an opportunity… testing for it is actually incredibly important.”
Outbound is an engineering problem: give sales engineers early

18:49 In late 2020 / early 2021, before the AI wave, cofounder Karim gave Freeman engineers to build OATS (Outbound Automation Team): codify the account signals reps used and automate coverage. With three to six outbounders working by hand you get a couple hundred emails and ~100 dials — not enough for a big TAM. Keep the system recursive: A/B test constantly and feed it what hand-written outreach learns. Ramp now runs a growth-engineering team of at least six. Measure meeting rate, not reply rate: just under 1% meetings per email sent is a strong standard.

“Sales is very much an engineering problem when you break it down. It’s a data infrastructure problem. It’s a math problem.”
Early comp: 100% OTE for two quarters, never hire just one, overpay for new products later

27:55 Skip quotas for at least the first two quarters for early AEs and SDRs: a guessed quota either blows out (you owe multiple six figures) or crushes confidence, and lost confidence is hard to win back. Set quotas once you have stage-progression data (sales cycles under six months). Hire at least two at a time so you can benchmark them against each other, not against the founder; if the gap is 10–20%, dig into structural causes (who got cap-table pipeline?) before cutting. Early on, sell one product to 20 customers who love it. At Ramp’s stage, over-incentivize new products — think 2–3× quota retirement — to build case studies and momentum.

“I usually advise to do 100% OTE for at least the first two quarters. You just don’t know what you don’t know.”
Shorten cycles by qualifying, not discounting; verticalize only once you hold 3–5% of the market

24:35 The fastest way to compress cycles is upstream qualification: can this person drive change, and will they commit within 12 months? If they aren’t C-suite or VP, learn who their boss is in the org chart. A buyer asking about price after call one is a bad sign. Up-market, co-author a business-value / TCO assessment (e.g., FTE leverage in accounting) instead of conceding on price. Verticalizing too early hurts land-grab growth. Do it once you own roughly 3–5% of the market and need conversion, not more coverage.

“The best way to compress sales cycles is qualifying with more rigor.”
Onboard hard, ramp by segment, and let frontline leaders ride shotgun

34:52 Bootcamp is 60–90 days and covers product, buyer personas, competition, and internal systems. Reps must pass discovery and demo certifications to get accounts, and are on live calls in 2–4 weeks after studying Gong recordings. After the first call, the rep’s leader owns them “start to finish” — relying only on enablement is fatal. Ramp periods: inside sales 60 days, commercial 90, mid-market 4 months, enterprise 5, strategic 6. Quota-to-OTE of 7–12× is healthy (old Snowflake/MongoDB 4–5× is too low; 20× means you need more reps). Early signal: the “annoying” new hires asking constant account-strategy questions are the great ones.

“If you’re running a big sales team and you’re just dependent on enablement, you’re going to die.”

Key use case

Job to be done

Stand up (or rebuild) a B2B sales org that scales: who to hire and how to test them, how to pay the first reps, how to automate outbound coverage, and how to onboard, ramp, and forecast without a big enablement team.

Who it’s for

Founders hiring their first AEs/SDRs, and sales leaders scaling from tens to hundreds of reps.

Success signal

Meetings per email near 1%, reps certified and on calls within 2–4 weeks, quota-to-OTE of 7–12×, and forecasts landing within a few percent.

Do this week

→Rewrite the interview loop: add a number-decomposition question (inbound/outbound/expansion/founder-sourced/inherited), two chronological won-deal walkthroughs, and one loss post-mortem.
→Build a Moneyball list: companies that grew past $100M despite low NPS and premium pricing; target their #1–2 reps.
→Assign an engineer to outbound: codify the account signals, automate coverage, A/B test, and track meetings per email sent (target ~1%).
→Fix early comp: 100% OTE for the first two quarters; hire in pairs; set quotas only once you have stage data.
→Put forecast accuracy in performance reviews (not comp) and check for single-threaded deals and status-quo losses in the weekly pipeline review.

Extra lines worth keeping

Pick the market first: 14:28 Freeman left a mid-market AE job to be an SDR again at Ramp because cards, bill pay, travel, and banking are each trillion-dollar-scale markets, and the engineering talent was exceptional. Index on market, talent density, and founders over equity and title.
CS as the NDR engine: 43:46 Ramp’s customer success team is technical (some are CPAs, ERP-fluent) and subsidized by product economics. A flawless first deployment “earns the right” to cross-sell.
Forecasting: 45:55 Finance-built models land within ~5% across the business. Ramp dropped separate forecast calls as duplicative and kept a weekly pipeline review. Most losses are to the status quo — “that’s on us,” a failure to create urgency.
AI capacity: 50:04 Pre-call research took 10–30 minutes per meeting. Ramp’s internal “Ramp Revenue” system cuts it to about 90 seconds, and the freed selling days support higher quotas.
Leadership: 53:19 His biggest weakness is jumping into every important deal, which teaches the org to wait for him. Be a “dual threat”: strategic in the boardroom, and still sending at least five cold emails a day.
New-rep advice: 60:04 Be obsessed, solve problems instead of just finding them, take onboarding seriously, accept blame and deflect credit, and bring a point of view when you ask for feedback.

Chapters (from show notes)

0:00 — Inside Ramp’s sales playbook
5:00 — Finding exceptional sellers, and why bankers make the best
18:00 — Building an outbound machine that scales beyond the SDR team
24:00 — Shortening sales cycles without slashing price
27:00 — Paying first reps and incentivizing new products
34:00 — Onboarding reps who can actually sell
40:00 — Quota levels and ramp time
44:00 — Customer expansion and accurate forecasting
54:00 — Growing as a leader without taking over every deal

← Back to library Open source on YouTube ↗

Talk + Q&A ~66 min Oct 5, 2026 Belfer Center

Tackling the Grid Reliability and Affordability Challenge in the Age of AI

Harvard Kennedy School Energy Policy Seminar (Belfer Center) · speaker Manu Asthana (former President & CEO, PJM Interconnection; board member, Alliant Energy) · moderated by seminar host “Joe” · Q&A includes Prof. William Hogan
YouTube video

Most actionable learnings

Power — not chips — is now the binding AI constraint, and the build gap is years wide

5:40 Asthana calls resource adequacy “perhaps the most critical enabler” of AI: analysis he’s seen says the chip constraint is no longer binding; power supply is. U.S. demand grew ~0.5%/yr from 2005–2025 (efficiency offset growth), eroding the industry’s muscle memory for big builds. EPRI projects ~39% more electricity by 2035 (~3.5% CAGR). A data center takes 18–24 months; a gas turbine ordered today gets a delivery slot in Q1 2031; transmission takes far longer. Plan around that mismatch.

“The constraint that’s binding is a supply of power.”
Count megawatts by accreditation, not nameplate — winter is where the risk lives

14:41 Under PJM’s numbers, 100 MW of nuclear counts as ~96 MW of reliable supply; 100 MW of solar counts as ~7 MW. After Winter Storms Uri and Elliott, the industry learned most risk is in winter from correlated outages tied to gas-supply disruptions, not the summer peak — so the grid had less effective supply than believed. Incremental solar’s accreditation falls toward zero as penetration rises (the risky hour is after sunset), which is why storage, especially long-duration, matters. Same lens shrinks the headline China gap (2025: ~60 GW added in the U.S. vs ~550 GW in China, ~350 GW of it solar).
The supply already exists on paper — the job is getting more of the queue built

22:25 ~2,000 GW sits in U.S. interconnection queues; ~549 GW has signed interconnection agreements. His point estimate of need is ~330 GW over 10 years (about a third from ~100 GW of scheduled retirements). Signed-queue capacity could cover roughly half the need — about five years of growth — but historically only 10–20% of queued projects get built. Retirements are partly discretionary: DOE has issued emergency orders keeping almost every retiring plant open, at an emissions cost.

“We’re gonna have to get better at building what is in the queue.”
Make data centers part of the supply solution: flexibility + bring-your-own-generation + long-term contracts

27:30 Most new data centers have backup diesels (PJM once saw ~4 GW switch to backup in microseconds after a voltage disturbance). You only need flexibility for “a sliver of hours in the year,” so price it through structured, market-based demand response rather than emergency orders. Pair that with incentives for large loads to sign long-term deals or bring their own generation — e.g., Microsoft’s 20-year Constellation deal to restart Three Mile Island, or Meta backing ~7+ GW of new generation for ~5 GW of load via an Entergy Louisiana tariff that Entergy says lowers other customers’ rates. Allocate incremental cost to whoever causes it.

“You don’t need it all the hours, you need it for a sliver of hours in the year.”
Higher prices only draw investment if policy is stable — beware of caps becoming permanent

25:47 New-build cost has roughly tripled (from ~$1,000/kW to close to ~$3,000/kW in recent market quotes), and retail rates are rising faster than inflation in markets exposed to a single clearing price. Short-term price caps made sense, but as long-term signals they deter investment and can hurt reliability. Policy “yo-yoing” (all-green to anti-green) adds a risk premium; longer-term consumer contracts, RTO-backed procurements (PJM’s up-to-15-year approach), and siting/permitting reform lower it.

“Higher prices plus higher uncertainty doesn’t always equal higher investment.”

Key use case

Job to be done

Reason clearly about AI-era power: how much reliable supply really exists, where the build bottlenecks are, and which policy and commercial levers (retirements, flexibility, contracts, permitting) trade off reliability, affordability, and emissions.

Who it’s for

Energy/infra investors, data-center and AI infra builders, utility and policy folks evaluating speed-to-power.

Success signal

You size projects on accredited (winter-risk) capacity, have a flexibility/BYOG story for large loads, and price policy uncertainty explicitly.

Do this week

→Re-underwrite on accreditation: convert any capacity assumptions from nameplate to ELCC-style accredited MW, stress-tested on a winter gas-disruption scenario.
→Map the lead times: turbine slots (~2031), transmission (10–15 years), data center (18–24 months) — and design a bridge plan for the gap.
→Mine the queue: look at projects with signed interconnection agreements that haven’t been built — that’s the fastest supply.
→Price flexibility: if you run or contract with large loads, model what curtailing to backup for a handful of hours a year is worth.
→Structure long-term offtake: use 15–20-year contracts or tariffs that put pricing risk on the load and protect existing ratepayers.

Extra lines worth keeping

History says it’s doable: 21:54 From 1950–1970 U.S. electricity demand grew 8.1%/yr during farm electrification. “Our society is capable of doing this.”
Every lever is flawed: 7:47 Nine partial solutions, each “deeply flawed” and each “probably necessary”; where to set the sliders is a societal choice for elected officials.
Markets under pressure: 44:35 Responding to William Hogan on FERC, Asthana says governors and regulators increasingly question whether a high single clearing price is “fair” — near-universal support for markets “has eroded tremendously.” He hopes it’s renewed.
Why off-grid is a bridge: 50:48 Tech firms want grid connection for its reliability; off-grid is driven by speed-to-power (even reports of scavenging airplane graveyards for jet engines).
China: 38:15 A long-term strategic, command-and-control approach that doesn’t map onto the U.S. societal compact; the gap is real but smaller on an accreditation basis.
AI for the grid: 31:51 PJM is working with Google’s Tapestry on AI-powered grid planning tools; AI could also optimize siting to minimize new transmission.

← Back to library Open source on YouTube ↗

Interview ~37 min Sep 15, 2026 Crux Climate

What It Costs to Make Power Cheap

Critical Capital (Crux) · host Alfred Johnson · guest Zach Dell (co-founder & CEO, Base Power)
YouTube video

Most actionable learnings

Base’s stated deployment pace: ~500 MW this year, “a couple gigawatts” next year

11:30 Dell says residential batteries are “the most scalable way to add capacity to the grid full stop,” and that Base has “added more capacity to the grid in Texas than just about anybody over the last 6 months.” Homes already have interconnection, so there is no queue wait.

“we’ll do on the order of, you know, 500 megawatts this year and, you know, a couple gigawatts next year”
Install throughput is the operating metric: 100+/day now, 300 and then 500 per day next

12:10 Base calls this a “factory in the field” alongside the Austin hardware factory. Every monthly company update opens with how many installs were done.

“over a 100 installs a day where we’re at today and soon we’ll be at 300 installs a day, then 500 installs a day”
Optimize fully landed $/kWh, not the bill of materials

27:12 Parts, manufacturing, logistics, install, financing, and tax credits are one equation. Since last summer’s OBBB rules, buying cheap Chinese parts can forfeit credits. Base designs almost everything and does final assembly, test, and pack in Austin.

“you’re solving for the lowest landed costs per kilowatt hour net of everything”
Raise equity big early, then drive the cost of capital down

18:04 Base has raised $2.5B in equity in about 3.5 years, most recently $1B in August at a $13B valuation (host intro). It started with a venture-debt line and is moving to tax-credit transfers, tax equity, and asset debt as returns rise.

“we have to go build a, you know, 20 gigawatt hour factory and then go build another one”
Pick markets for price signals plus single-family density

13:24 Texas (ERCOT) is ~80% competitive and 20% regulated, and Base works with co-ops, munis, and El Paso Electric there. Illinois (PJM) is next because PJM signals a capacity need. Dell also names Ohio and the Northeast. International expansion is on hold until roughly the second half of next year.

Key use case

Job to be done

Size Base Power’s scale-up claims (≈500 MW this year, “a couple GW” next, 100→300→500 installs/day, 20 GWh factories) and see how cost of capital and tax-credit sourcing rules set the customer’s price.

Who it’s for

Energy and climate investors, operators of grid-edge and DER businesses, and anyone diligencing Base Power.

Caveat

Every figure here is the company’s own claim from the conversation. None are audited, and they shift between interviews, so check the dates.

← Back to library Open source on YouTube ↗

Interview ~76 min Sep 6, 2026 David Senra

Building a Power Company for the Next 50 Years

David Senra (podcast) · host David Senra · guest Zach Dell (co-founder & CEO, Base Power)
YouTube video

Most actionable learnings

Counter-position the incumbents: sell electricity, not a $20K battery

14:22 Utility-scale storage is limited by interconnection and transmission congestion. Incumbent home batteries are “$20,000” premium one-time sales that sit idle about 99% of the time. Base owns the battery and trades it in wholesale markets. The customer pays “1/20th or even 1/40th” of the cost of owning it. To copy this, an incumbent would have to abandon its own model.

“I don’t sell batteries. I sell electricity.”
Utility business = “megawatts as a service” on ~10-year contracts

24:19 For CoServ, Base has built or is building a “100 megawatt and soon to be expanded” fleet. Base also works with Guadalupe Valley Electric Co-op and Farmers Electric. Utilities get a software console. Long contracts give Base financeable cash flows. Today Base serves residential homes only.

“We build 100 megawatts, 500 megawatts, thousand megawatts or gigawatt and they pay us for that.”
Run the company on three north stars and make the constraint visible

44:52 The three north stars are: the largest and fastest-growing distributed battery fleet, the lowest landed cost per kWh, and financial sustainability (unit economics). A physical turtle sits on the desk of whoever owns the critical-path constraint, and a “hot potato” squad meets daily until it is resolved. Dashboards hang over every team, and Dell has written a monthly update every month for 3.5 years.
Vertical integration is a supply problem: the supplier couldn’t keep up

66:44 Dell says Base Core is Base’s own battery, designed from three years of running “tens of thousands” of home batteries. It has a lower landed BOM, faster installs, and more power and energy. Base was “sold out through November and it’s August.” Factory 1 is running, and Factory 2 is a campus near the airport.

“we were buying all their stuff and they couldn’t make batteries”
Demand context: ~2% historical growth heading toward ~10%

15:39 Dell says US demand grew at roughly inflation, or ~2% CAGR, over 20 years, and he expects it to approach ~10% because of AI. Base plans to point its make–move–store–sell stack at AI load, both around central data centers and by bringing compute to its fleet.

Key use case

Job to be done

Learn Base’s counter-positioned model (own the battery and sell power, not hardware), its utility “megawatts-as-a-service” business, and the operating system behind it (north stars, turtles, hot potatoes, monthly updates).

Who it’s for

Energy and climate investors, operators of grid-edge and DER businesses, and anyone diligencing Base Power.

Caveat

Every figure here is the company’s own claim from the conversation. None are audited, and they shift between interviews, so check the dates.

← Back to library Open source on YouTube ↗

Panel segment ~26 min (of 52) Oct 12, 2025 a16z Deep Dives

Monitoring the Situation #3: Who Is Nick Land & the State of Energy (Zach Dell segment)

a16z (The a16z Show / a16z Deep Dives) · hosts Erik Torenberg, Katherine Boyle, Erin Price-Wright · guest Zach Dell
YouTube video

Most actionable learnings

Batteries move power through time, which makes them a cheaper “non-wires alternative”

02:21 Generation costs have fallen over the last 20 years, mainly because of solar, but delivered cost has risen because of transmission and distribution capex. Base puts batteries where the load and the interconnection already are.

“poles and wires move power through space and batteries move power through time”
The model at the time: Texas homes, 10–20% savings, now selling to utilities

00:47 The company was “almost three years” old. It started on off-the-shelf hardware with custom software and moved over two and a half years to its own batteries, installs, and operations. Customers save “on the order of 10 to 20% a month.” Base was starting a regulated-utility partnership model.
China supply risk pulled the Texas factory forward

14:32 Dell personally favors a global economy, but with tariffs and the critical-mineral cutoffs, “you better have a backup plan.” Base “accelerated our plans to build a factory in Texas.” He frames the $1B Series C as a planned build, not a reaction to tariffs.
Constraint is electrons, not chips, and utilities are newly open to new tech

06:27 Dell says the economy is now constrained by electrons. He calls the data-center build speculative, says it will be both overbuilt and underbuilt in boom-bust cycles, and still expects steady long-run demand.
Old grid plus no competition means no innovation

20:12 Dell says “on the order of 40–50%” of grid infrastructure was built in the 1960s. He positions Base as “outsourced R&D” for regulated utilities and as a competitor in deregulated markets, and he wants more competitors to show up.

Key use case

Job to be done

Hear Zach Dell’s Series C-era pitch: batteries as “non-wires alternatives,” why Texas price signals make it the energy lab, and how China supply risk pulled the Austin factory forward.

Who it’s for

Energy and climate investors, operators of grid-edge and DER businesses, and anyone diligencing Base Power.

Caveat

Every figure here is the company’s own claim from the conversation. None are audited, and they shift between interviews, so check the dates.

← Back to library Open source on YouTube ↗

Interview ~66 min Aug 17, 2026 Mel Robbins

Former CIA Spy: How to Read Anyone, Think Clearly, and Protect Yourself

The Mel Robbins Podcast · host Mel Robbins · guest Jonna Mendez (former CIA Chief of Disguise, 30-plus years at the agency)
YouTube video

Most actionable learnings

People act for four reasons: MICE (money, ideology, compromise, ego)

25:52 Mendez says the reasons people betray their government “boils down to four things.” In her telling, Americans do it for money and Russians typically for ideology. Compromise (leverage over a secret) “didn’t really work very well.” Ego is “the big one”: people passed over for promotion, people with grudges. A recruiter first has to find the flaw or niche.

“People don’t just randomly say, ‘I think I’m going to take all the secrets I know and give them to another government.’ They need a reason”
Courage isn’t inherent. It comes from preparation, then confidence

45:37 When she walked into about 20 armed, drunk young men in an alley in Latin America, her training told her not to turn back, because retreating would signal that something was wrong. Her chain is knowledge, then confidence, then courage. She went through a 3-day, 3-night hostile-interrogation course on half a cup of rice and a cup of water a day, and she says the agency trained constantly. To practice, start confronting what’s on your list, including hard conversations (1:00:44).

“I think courage comes from confidence. Confidence comes from knowledge.”
Intuition is one clear statement. Anxiety is a cloud of “what ifs”

50:49 Mendez follows an intuition and tracks it “like it’s a newspaper story” to see if she was right, especially about people, and that builds trust in it over time. Mel’s test, which Mendez agreed with: a clear direction (“move away,” “speak up”) is intuition, while scattered questioning is fear. Lying cues (50:04): eye contact drops first, the person tries too hard, the story bounces around, and they won’t let it go.

“They want you to believe it, so they’re trying a little too hard.”
Hit a ceiling? Make a change, and carry your skills with you

10:19 Mendez went from secretary to the director of a 1,000-person office (“that’s a dead end”), then to photo work, then switched to disguise mid-career to get an India posting, which led to Chief of Disguise. Excellence earned her access each time. Her advice to people who feel underestimated is “aim high” (13:32).

“If you can do the work in an excellent manner, they will let you do the work.”
Being seen is a set of details you can change

30:00 To make people less noticed, the CIA often made them older (“Nobody looks at old people too much”). Her method was to describe a person in fine detail, then change every element: hair, teeth and gumline (which changes the profile), glasses, contact lenses, ring, watch, cologne, a gold chain, one extra shirt button. Each officer got a written character sketch, then a test lunch in the CIA cafeteria among colleagues (38:12). She briefed President George H.W. Bush and got past the Secret Service while wearing a full-face mask that “dropped 20 years.”

“Someone in your family gave you a shirt in the last 10 years that you just hate… That’s your shirt for your disguise.”

Key use case

Job to be done

Borrow a 30-year CIA officer’s playbook for reading motives (MICE), spotting lies, telling intuition from anxiety, and building courage through preparation.

Who it’s for

Founders, sellers, and managers who negotiate, recruit, or read people under pressure, and anyone weighing a career pivot.

Caveat

This is a lifestyle-podcast interview of stories and personal heuristics, not research. The lying cues are her general impressions.

Do this week

→MICE your next negotiation: for one key counterpart, write down which of money, ideology, compromise, or ego is actually driving them. Ego is usually the big one.
→Rehearse the scary thing: pick one conversation you’ve been avoiding and prepare for it until you feel confident, then have it.
→Log your intuitions: write down one gut call about a person and track it “like a newspaper story” to see if it holds.
→Check for a ceiling: if your role is a dead end, list the skills you would carry into the next one.
→Build your “escape hatch”: choose a specific, vivid mental place for stressful moments.

Extra lines worth keeping

Teams: 28:06 “You don’t do anything alone… You’re always part of something bigger than yourself if you’re going to accomplish anything.”
Spy gear numbers: 07:15 A Mont Blanc pen camera “would take 100 pictures.” Microdots shrank a page of text “800 times” and were read through a lens that “looks like a grain of rice.”
Twins: 34:04 With realistic masks “we can make your twin.” In Moscow, the surveillance team would follow the twin while the real officer stayed in the kiosk.
Small talk beats a border: 41:28 A family of three hidden in a smuggler’s modified car got through a controlled border after the guard asked for the driver’s Billy Joel tape.
Happy place: 47:28 In the interrogation box she survived by going in her mind to Kansas’s Flint Hills, a technique she learned from Vietnam POWs. A generic beach “didn’t work.”
Overwhelmed by the world: 56:28 “Take a breath is first thing… if enough people take small steps, you can fix big pieces of it.”

← Back to library Open source on YouTube ↗

Interview ~66 min Oct 16, 2026 Substrate

Neil Movva: 1MW datacenters, commoditizing compute, and why 95% of inference could be asynchronous

Substrate (YouTube channel: Cat Tang) · host Cat Tang · guest Neil Movva (co-founder & CEO, Sail Research, inference and infrastructure for long-horizon agents)
YouTube video

Most actionable learnings

If your agent can wait, inference can be 5–10× cheaper

33:48 Movva quotes “around 5–10× savings” for latency-tolerant work, from two stacking gains of 2–3× each. First, packing larger batches so the engine gets “twice as many or three times as many tokens per GPU minute.” Second, running on alternative hardware that is weak on latency but “another maybe 2–3×” better on cost per FLOP. Sail’s Flex tier is built for work that can wait 30 minutes to an hour (31:33). That slack lets it max out batch size and wait for the machine that holds a request’s KV cache. A KV-cache miss can make a request take “easily three times as long” (37:00).

“There are no bad chips, only bad prices.”
Decode is the expensive part. Price your workload on output speed

04:40 Prefill (reading the prompt) runs in parallel and is compute-bound. Decode reads every parameter once per output token, so it is memory-bound. Movva calls decode “the only fundamentally expensive thing in transformers today.” NVIDIA GPUs need roughly 400–800 operations per byte fetched to be compute-bound (07:20). Serving at 100–200 tokens/sec requires speculators and PD disaggregation. Sail’s fastest tiers serve at 20–50 tokens/sec, which he says is where OpenAI and Anthropic frequently serve. At that speed chunked prefill is enough (50:05).

“Output TPS is the most expensive thing in inference.”
Sub-megawatt modular sites are untapped supply, live in about 16 weeks

23:36 The standard play is 10,000 chips (“something like 20 megawatts”) in one high-reliability facility. Sail instead targets sites “sub-megawatt in scale, maybe just a few hundred chips” in warehouses, abandoned steel mills or breweries, often with curtailable grid power. With power rights on site, “you can get something up in 16 weeks” (53:07), versus 12–24 month lead times for a 50–100 MW build (52:46). A contiguous 5 MW is “about 2,000 or so chips, 2,500.” The host cites a SemiAnalysis estimate that about 30% of US data center capacity will be modular by end of 2028, and NVIDIA piloting 5–20 MW modular sites (25:05). Sail’s first test machines will be online by end of year, with real scale “well into next year.”

“I actually believe that there are extreme diseconomies of scale to inference.”
Buy one nine of uptime and fix reliability in software

29:21 Following early Google, Sail buys cheap “one-nine-of-uptime machines and route[s] around the failures in software.” If a 90%-uptime site is half the cost, “I am 100% going to buy” it. This works because agents run over trajectories of minutes to hours. On a site failure Sail loses that turn but moves the state elsewhere and resumes “within 30 seconds” (31:08). The tradeoff is that it won’t be P99-latency controlled. The largest team, about 10 people, works on the control plane and distributed systems, not kernels (36:10).

“Agents don’t get work done in a single turn. They get work done over a trajectory.”
Work around scarce NVIDIA and HBM supply instead of trying to outbid for it

21:29 Movva says Anthropic seems willing to pay “something like $15 to $20 per hour” for a B300 hour, while open-source providers find it “hard to make a market above $10.” Sail leans on the RTX Pro Blackwell (SM120), which uses GDDR rather than HBM, plus AMD, Intel, TPUs and Trainium. On contiguity, Colossus “regularly rents for twice or three times” market price (41:17), and ClusterMAX platinum providers Nebius and CoreWeave charge the highest premiums. Sail wants to be “the most liquid buyer of compute in the world,” taking “any chip anywhere in the world for any duration” (42:05). Most contracts are still 3–5 years, even though operator payback is often 1–2 years (1:01:07).

“You cannot be using HBM and have a good time.”

Key use case

Job to be done

Decide which of your AI workloads can run asynchronously (background agents, deep research, RL rollouts) and buy that inference much more cheaply than interactive tokens.

Who it’s for

Founders and operators running agent products, AI infrastructure and data-center investors, and anyone underwriting compute, power or modular sites.

Caveat

This is the CEO’s own pitch. The 5–10× savings, 16-week builds and 95% async share are his claims or predictions, and the SemiAnalysis 30% figure was cited by the host.

Do this week

→Sort your token spend by latency need: tag each workload as interactive or able to wait 30 minutes or more. Price the second group on a flex or batch tier.
→Design agents to resume: checkpoint state per turn so a 30-second failover doesn’t lose the trajectory.
→Ask about cache hit rates: when comparing inference providers, ask for KV-cache hit rates and output tokens/sec targets, not just list price.
→Scout sub-megawatt power: list warehouses or industrial sites with existing power rights that could take a 1 MW pod in about 16 weeks.
→Try proactive compute: pick one “sleep-time” task an agent could run before the user asks, such as overnight deep research.

Extra lines worth keeping

The thesis: 1:05:00 Async inference “could easily be 95% or more of inference demand in total,” with Sail “following the sun” to run where solar makes the grid negative-priced, and older process nodes running redesigned chips for the lowest cost per FLOP.
Background agents: 03:04 OpenAI’s top internal users are spending “thousands of dollars per day on tokens.” Movva says “there is no bound on background agents.”
Best latency: 17:19 “The best latency of all is none.” That means predicting what the user wants and doing “sleep-time compute” ahead of time.
RL is inference: 43:20 “RL loops are almost 80% inference by compute or by cost.” Sail pairs with Tinker: new LoRA checkpoints every 5–10 minutes, with sampling from them “within a minute.”
Credit gap: 54:40 Lenders “want to see a Microsoft offtake,” meaning investment-grade buyers. Inference neoclouds sell to neolabs and small operators, so Sail tries to be “the trust arbiter” by vetting sites with NVIDIA and OEMs.
Centaur kernels: 38:04 Expert kernel engineers working with agents beat “just asking Claude, ‘Please improve the kernel. Make no mistakes.’”
Power spikes: 58:28 Synchronized training can swing GPUs “from drawing hundreds of watts to thousands of watts all at once” and brown out a site. GB300 racks add supercapacitor trays. Inference units are smaller (about 8 machines, 64 GPUs).

← Back to library Open source on YouTube ↗

Live interview ~26 min Oct 19, 2026 Odd Lots

Don Wilson on the GPU Market’s Trillion Dollar Question

Odd Lots (YouTube channel: Bloomberg Podcasts) · recorded live in Chicago · hosts Joe Weisenthal & Tracy Alloway · guest Don Wilson (founder & CEO, DRW)
YouTube video

Most actionable learnings

GPU futures turn a neocloud into a “giant basis trade” with cheaper capital

02:10 Tracy notes a CoreWeave bond yielding 15%. Wilson says the market is pricing in that neoclouds selling on-demand compute, or selling to less creditworthy buyers, carry a higher cost of capital. With futures, a neocloud can buy GPUs, “sell a strip of GPU futures,” then sell on-demand or six-month contracts to weaker credits who post “a big percentage of the contract value upfront.” Companies that hedge this way will see their cost of capital come down “much lower.”

“It’s highlighting the massive need for risk management products.”
Watch the calendar: listing could happen November 9, with contracts out 3 years

04:05 A comment period is open and the CFTC extended it. If regulators “don’t do anything,” the futures can list on November 9, which Wilson notes is conveniently after November 3. He says he can’t explain Washington’s concern and has heard “all sorts of different theories.” 07:20 Initially the CME contracts are expected to “go out three years,” extending further as liquidity builds. Each major chip gets its own contract, much like the “thousands” of electricity node contracts.

“Nobody knows what the value of a B200 will be four years from now, five years from now.”
GPU rental prices aren’t a one-way slide, so hedge both directions

07:29 Almost a year ago, hourly on-demand H100 prices had been “declining, declining, declining,” and skeptics asked why futures were needed. Around Thanksgiving prices “turned around and kind of went up and to the right” after Anthropic’s newest model drove a surge in inference demand. More recently B200, B300 and H100 prices surged again, while the Silicon Data token index price “has plummeted.” Wilson expects forward prices to sit lower because there is so much hedging demand to sell forward, which means investors will need to be paid to take the long side.

“These things are volatile. They can go up, they can go down, and a hedging instrument is badly needed.”
Track the “crunch spread”: GPU rental cost versus token revenue

13:33 Wilson expects futures on tokens themselves eventually (“we may be getting close”). The margin from renting GPUs at the on-demand price, loading an open-source model and selling tokens “through OpenRouter or something” is like an oil refinery’s crack spread or a soybean crush spread. DRW calls it the crunch spread. Token futures are at “the thought stage,” with the Silicon Data token index as a start. 10:46 He says the open-source inference ecosystem (Fireworks, Together, companies with “many billion dollar valuations”) is already rich enough for a futures market beyond bilateral offtakes.

“In oil we call it a crack spread. In soybeans, we call it a crush spread… At DRW, we call this the crunch spread.”
Cut token spend by routing, not rationing

17:14 DRW has more frontline AI adoption than a year ago, and token use is “like this exponential chart.” Asked about internal token allocation, Wilson says the key is routing inference requests thoughtfully, because “not every inference request needs to go to a frontier model.” DRW optimizes to “minimize token spend by optimally routing inference requests.” On indices (05:25): “it’s buyer beware.” Silicon Data’s index covers on-demand neocloud compute and will be IOSCO compliant, referring to the international regulators’ index best practices from about 2013. He says some other providers have never heard the phrase.

“Not every inference request needs to go to a frontier model.”

Key use case

Job to be done

Understand how listed GPU futures, a GPU-to-token “crunch spread” and perpetual contracts could change financing and risk management for neoclouds, inference providers and compute buyers.

Who it’s for

Compute buyers and sellers, neocloud and data-center financiers, traders watching CME and CFTC listings, and operators managing token budgets.

Caveat

Wilson runs DRW, which is backing these markets (Compute Exchange, Silicon Data), so his views are those of an interested party. No volumes, prices or DRW financials were given. The Nov 9 listing date depends on regulators not acting.

Do this week

→Mark November 9: follow the CFTC comment period on GPU futures and whether listing goes ahead.
→Map your compute exposure: list GPU contracts or rentals you hold by chip (H100, B200, B300) and term. Note what a price move up or down would cost you.
→Read the index rules: before relying on any GPU price index, check what it samples (on-demand neocloud versus bilateral offtake) and whether it claims IOSCO compliance.
→Compute your own crunch spread: on-demand GPU hour cost versus revenue from serving an open-source model’s tokens.
→Audit model routing: find requests going to frontier models that a cheaper model could handle.

Extra lines worth keeping

On the AI slowdown jitters: 01:36 Anything growing this fast won’t grow at a constant rate. Wilson calls the recent wobble “just kind of expected volatility… maybe people went on vacation in August.”
Overbuilding worry: 09:36 Futures can’t be sold at a made-up price: “there has to be a willing buyer and a willing seller.”
Not homogeneous enough? 12:30 “The amount of money going into just H100s or just B200s is large enough to support a single futures market.” Other chips are “very correlated,” so they trade as a basis, and large training clusters carry a premium.
Nvidia: 16:42 It is “unlikely that Nvidia will be able to maintain the market share” it has had, but Wilson is “pretty confident” it will keep growing as a company.
Perps for ETFs: 18:56 Oil ETFs lose to roll slippage. An ETF holding a perp tied to the Brent index would be “a much more efficient experience.” Producers still need dated futures far out the curve, so the two “augment one another.”
Perps and retail: 20:02 Hedge funds may pay 100% annualized to short a hard-to-borrow stock, which Wilson puts at about 30 basis points a day. Retail holders don’t collect that, but a perp would “mechanically” pass it through.
Futures vs. swaps: 22:43 “A swap that is cleared has the same economic risk… as a futures contract does.” A perp can be listed “with reasonable leverage and no auto-deleveraging.” This is aimed at “Terry,” who was in the audience and argues perps are swaps.
24/7 trading: 23:47 “It’s already here.” DRW has traded crypto since 2014. Its options market makers are “not happy” about making SOFR options markets over weekends, but “it’s just going to happen.”

← Back to library Open source on YouTube ↗

Interview ~73 min Oct 24, 2026 20VC

Naveen Rao: Inference will be a $10T market, energy is the bottleneck, and pure software investors are in trouble

20VC (YouTube channel: 20VC with Harry Stebbings) · host Harry Stebbings · guest Naveen Rao (founder & CEO, Unconventional AI; previously founded Nervana, acquired by Intel, and co-founded MosaicML, sold to Databricks for $1.3B)
YouTube video

Most actionable learnings

Plan around power: about 10 GW a year is realistic, at roughly $50B per gigawatt

01:44 Rao cites industry numbers of about $50B to build one gigawatt. A semiconductor fab costs $20–30B, so one gigawatt costs more than a fab. Data centers have mostly grown by soaking up spare grid capacity. The US grid averages about 900 GW of output but can peak at 1.2–1.3 TW. 04:59 At even conservative AI growth rates, that excess runs out “in the next few years.” Claims of building 40 GW in a year are “no way… we have never done that.” Building 10 GW “would be amazing,” and that alone is about half a trillion dollars. 03:11 China passed US total capacity around 2023 after building about 6× faster for 20 years.

“People are throwing around the word gigawatt like it’s easy… guys, this is not easy.”
Data centers can pay 30–40% more for electricity, so expect two-tier power pricing

07:35 Rao has no hard numbers because every behind-the-meter deal is bespoke. Anecdotally, though, data centers pay far more per kWh than grid customers, which creates “two-tiered pricing.” His rough math, which he flags as made up: compute sells for about 5× the cost of the electricity behind it, so a buyer “could pay 30%, 40% more for that electricity.” He mentions a Bloom Energy deal for “a couple of gigawatts in Arizona” as promising but unproven. He calls the current AI-debt wobble (CoreWeave came up) “a blip,” not a bubble bursting.

“We’re already seeing a two-tiered pricing occurring.”
Efficiency gains now come from co-designing model and hardware, not from transistors

09:16 Physical gains from Moore’s law and Dennard scaling have “largely stopped,” and recent GPU generations show little physical improvement. Most recent gains came from smaller data types (32→16→8→4-bit). The under-exploited lever is enforcing locality and cutting data movement at the model level. 15:14 Software alone on a standard GPU might yield about 5%, because off-chip memory caps gains from things like sparsity. Hardware built to exploit them can yield “massive gains.” 22:10 Unconventional is “energy first” and doesn’t treat the model as sacrosanct. Its first chip is a set of coupled oscillators. Rao sees chip specialization as “100%” the future and doesn’t expect chips to commoditize.

“We don’t treat the model as sacrosanct. We’re energy first in our mindset.”
Invest in verification and the spec, because AI moved the bottleneck from producing to checking

23:24 A noise-propagation model that once took his team about two weeks now takes Rao about an hour with an AI model. The catch is “AI slop”: too many artifacts, hard to verify at scale. To trust the result you have to own the spec too, and verify every artifact on the way to it. 28:29 Don’t trust agents as much as humans yet. The longer the chain of decisions, the more it can go “off the rails” while looking robust. Within five years, though, he expects we “will trust our agents more than we trust ourselves.”

“We’ve moved the problem from producing to verification.”
Route routine work to open models: 80% of tokens, about half of spend

1:06:21 In three years Rao expects about 80% of tokens to flow through open models. Spend would split roughly 50–50, with only 20% of tokens going to frontier models, like the iPhone’s ~20% of units and ~80% of margin. 41:41 Six-month-old models (“the Chinese open models”) are already “pretty darn good,” a hallmark of commoditization. 43:26 Inference operators (Fireworks, Modal, Baseten) face falling per-token prices and thinner margins. The hardware layer that turns energy into intelligence keeps the edge, which is why “the hardware is the place to be over the next 10 years.”

“There’ll be a rough 50-50 split on spend, but in terms of tokens, it’ll be only 20% to the frontier models.”
Back outcome businesses (AI plus humans), not classic SaaS or undifferentiated automation

45:04 “If you’re a pure software investor, you’re kind of f***ed.” The SaaS “sugar high” since about 2010–2011 was about chasing $100–200M markets, and Rao doesn’t see software as a huge market in 20 years. The new model is delivering an outcome with software plus services, like Harvey using AI and human verification to do legal work with fewer people. He would send “80% of my legal workloads” to a provider he trusted for NDAs and contract review. 1:05:21 Overrated: customer-support AI (“the worst place to invest right now”) and agentic coding outside the model providers, because value goes to the model and engineers switch tools “pretty fluidly.” 49:16 Owning your own model only makes sense at scale (a big firm like Cooley, not a “mom and pop shop in the middle of Nebraska”).

“I just don’t see software being a huge market in 20 years.”
For deep tech, judge technical progress instead of market validation

1:02:37 The idea Rao changed his mind on in the last 12 months: market validation is a SaaS-era metric. When the market is “true by inspection” (cure cancer, a better computer), validate progress on the technical problem instead. 1:10:11 The founders’ early story “100% will be wrong,” so you’re backing “someone to drive the bus, not the bus itself.” 1:00:04 With trillion-dollar markets, “the price is almost irrelevant.” He puts a percentage of his wealth into robotics regardless of entry price.

“It’s not about market validation. The market is there… It’s about progress of solving a problem.”

Key use case

Job to be done

Pressure-test where value lands in the AI stack: energy limits, the hardware that turns energy into intelligence, a commoditizing model and inference layer, and AI-plus-human outcome businesses. Then decide where to build, buy or invest.

Who it’s for

Infrastructure and energy investors, software and SaaS investors rethinking their focus, AI buyers setting a token budget, and deep-tech founders raising money.

Caveat

Rao runs an AI hardware startup, so “hardware is the place to be” is his own book talking. The 5× compute-to-electricity markup is, in his words, “making numbers up.” The $10T inference market and the 80/20 token split are forecasts, not data. No revenue or chip performance figures for Unconventional were given.

Do this week

→Sanity-check any gigawatt claim: multiply by about $50B per GW and compare it to the ~10 GW a year that Rao thinks the US can realistically build.
→Split your token budget: tag which workloads truly need a frontier model and move routine ones to open models. Rao’s 80/20 token split is the benchmark.
→Name your verification owner: for every AI-generated artifact stream (code, models, docs), write the spec and decide who or what checks each output against it.
→Re-screen the pipeline: discount customer-support automation and IDE or harness plays. Favor AI-plus-human outcome businesses and hardware or energy plays.
→For deep-tech diligence: replace “market validation” questions with technical-milestone questions, and judge whether the team can steer through pivots.

Extra lines worth keeping

The headline number: 56:48 “Just AI models themselves will be a trillion dollar market in the next few years. And the inference market is going to be a 10 trillion dollar market within 10 years.” Cutting token delivery cost by one order of magnitude “effectively 10X’s the market” (Jevons paradox).
Robotics: 55:42 Working prototypes for some tasks in 2–3 years and a real market in 5–7. 57:59 Robotics will be “100 times larger than agentic coding.” He invests in the tech stack, not the application layer, and expects model companies to end up shipping end-to-end devices. His one pick if he could put everything into a single other company: Physical Intelligence.
Developer jobs: 51:17 Benioff’s figure, as Harry quotes it: $300M of a $6B developer salary budget going to coding tools, about 5%. Rao thinks the share stays “sort of stable.” A company spending $10M a year on software still spends $10M, but ships fixes in a week, then in an hour. His skyscraper example: robots do 80% of the tasks, a 2-year build drops to 8 months, and you build twice as many.
RL: 35:02 RL is less efficient per unit of compute than pre-training but keeps models improving. It matters because “we cannot gather enough data to solve every task.” Verifiable rewards are the key.
Post-transformer: 39:10 Rao believes in a post-transformer world: reasoning by sampling traces and feeding them back in is “extremely inefficient,” and “our brains do not work in this way.”
Data companies: 32:52 They will specialize, not commoditize. The curriculum from pre-training through mid-training (RL, SFT) is “the secret sauce.”
Founders: 40:49 Vision matters more now. Being good at execution alone is “not going to get very far.”
Unconstrained: 1:04:20 He would build his own manufacturing and fabs, optimizing down to how the silicon ingot is sliced.
Investor meetings: 1:07:20 About 1,000 since 2014. In 2014, “no one had pitched a hardware startup in 10 years.”