Invest Like The Best on AI: 12 Essential Episodes, Ranked

The 12 Invest Like The Best AI episodes worth your time, ranked and summarized: Sam Altman, Gavin Baker, Dylan Patel, Dan Sundheim, and Brian Chesky on compute shortages, token demand, model moats, and what the data supports.

The best Invest Like The Best episode on AI is Inside Dan Sundheim's Bets on Anthropic, OpenAI, and SpaceX (February 2026), where the D1 Capital founder explains the clarity-of-thought test that put him into Anthropic and then argues that hyperscalers are quietly trading a fragmented customer base for a concentrated one. Eleven more conversations follow it below, and together they add up to a working education in how professional investors underwrite artificial intelligence.

Patrick O'Shaughnessy has spent the past two years pointing the show directly at the AI trade, and the guest list reflects it: Sam Altman on compute bottlenecks, Gavin Baker across three separate episodes on chips and orbital data centers, Dylan Patel on token supply, Brian Chesky and Dara Khosrowshahi on what AI does to a large operating company. We summarized every episode of the show in our library, then ranked the twelve that deliver the most usable ideas per hour of listening.

One scope note before you dive in: these summaries cover the show's recent run, mostly the past year plus a few from late 2025, so treat this as the current state of the AI conversation on Invest Like The Best while the back catalogue runs far deeper. Each entry explains why the episode earns its rank, the specific takeaways worth acting on, and a quote that captures the argument. The three themes that run across all twelve are unpacked after the list.

1. Inside Dan Sundheim's Bets on Anthropic, OpenAI, and SpaceX

Invest Like The Best · Patrick O'Shaughnessy · Dan Sundheim · 1h 26m · February 2026

The rare AI conversation that starts with the underwriting process and only then gets to the companies. Sundheim, founder and CIO of D1 Capital, explains the clarity-of-thought test that put him into Anthropic: he weighs how precisely a CEO can write down what they intend to build, and Dario Amodei's writing reminded him of Bezos in 1997. His read on hyperscaler customer concentration is the most contrarian argument on this entire list.

Key takeaways

  • The clarity-of-thought test: a CEO's written articulation of strategy has predicted extraordinary businesses better for Sundheim than any income statement. Amazon's 1997 shareholder letter is his canonical example.
  • LLM economics resemble Netflix crossed with Spotify: heavy upfront spend to create the asset, then high-margin distribution, with personalization supplying the pricing power on top of a near-commodity product.
  • Hyperscalers are shifting from a fragmented customer base to a concentrated one, as 4-5 LLM labs become the dominant buyers and insource their own compute once cash flow allows.
  • Private markets have fewer competitors but everyone runs the same intrinsic-value analysis, while public markets are crowded with participants playing entirely different games, which is where the mispricing lives.
  • OpenAI's everything strategy (consumer, enterprise, hardware, robotics, science) spreads one fixed asset across many revenue lines, while Anthropic concentrates on enterprise and coding. Sundheim leans toward focus.
I place a lot of weight, rightly or wrongly, on clarity of thought and the ability to communicate as a CEO what you want to achieve and how you're going to achieve it. And especially in written form because taking the time to write something down, you actually really have to go through everything you plan to do and express it in a way that makes sense to everybody else. — Dan Sundheim

Read the full episode summary

2. Why the AI Boom Is Just Getting Started

Invest Like The Best · Patrick O'Shaughnessy · 1h 20m · June 2026

The best episode here for the arithmetic. Alex walks through his firm's Anthropic position and the bottom-up math behind it: power users inside the company burning $100 a day in tokens, roughly 20 million coders worldwide at $20-30K of annual spending potential, and a half-trillion dollar market from coding alone before anyone outside engineering touches the product. His L-curve framing explains why browser-native software adoption goes vertical.

Key takeaways

  • Coding was AI's first killer app, and the shift from autocomplete assistants to agentic systems is what turned it into a half-trillion dollar addressable market.
  • Adoption follows an L-curve when friction disappears: software that requires only a browser skips the slow installation phase that shaped every prior enterprise S-curve.
  • Enterprise AI is under 1% penetrated even after the growth of the past two years, which frames how early the revenue curve still is.
  • Foundation models held their moats: models specialize by task (finance and PE work versus PDF ingestion), harnesses and ecosystems create lock-in, and the compute gap keeps open-source a step behind the frontier.
  • Investing in the picks and shovels first (chips and infrastructure) bought time to identify the model-layer winners over the following 2-3 years.
The enterprise AI or enterprise application AI market is less than 1% penetrated and we've never seen, you know, we talk about S-curves, we call this an L curve, just straight up. — Alex

Read the full episode summary

3. Sam Altman on AGI, Compute, and Human Agency

Invest Like The Best · Patrick O'Shaughnessy · Sam Altman · 55m · July 2026

The only episode on this list featuring the person actually allocating the compute. Altman's shifting-bottleneck model (research ideas, then compute, then data, then compute again with research reasserting itself) is the clearest public description of how a frontier lab chooses what to solve. His remark that today's de-risking runs match the size of entire compute runs from 18 months ago is the single most useful stat for sizing the capex cycle.

Key takeaways

  • The binding constraint on AI progress rotates. Recognizing which one binds right now, then attacking it directly, is the operating discipline Altman describes.
  • Compute and research compound each other: more compute buys more experiments, which produces better use of the next tranche of compute.
  • OpenAI narrowed back to its core mission of abundant, low-cost intelligence after hedging into consumer apps and media, and credits the refocus for the past year of progress.
  • Altman expects roles to move up a level of abstraction while human accountability stays valuable, because people care who stands behind a decision or a work of art.
  • He treats concentration of power as the sharper risk: safety arguments can be used to justify restricting AI access to a small group, and preserving broad access is his stated priority.
An amazing statistic I heard recently is our biggest de-risks now for our upcoming runs are as big as like the entire compute run from 18 months ago or something. So, compute and research ideas are not as separate as they sound. — Sam Altman

Read the full episode summary

4. GPUs, TPUs, and the Economics of AI Explained

Invest Like The Best · Patrick O'Shaughnessy · Gavin Baker · 1h 28m · December 2025

The infrastructure primer to hand anyone who wants one episode on why the chip war matters to portfolios. Baker covers the Hopper to Blackwell transition, explains how reasoning models carried scaling through an 18 month hardware delay, and makes the case that AI is the first computing wave where the lowest-cost token producer holds a structural advantage. That is the lens he uses to describe Google pulling economic oxygen out of the market.

Key takeaways

  • Evaluate models on the $200 per month premium tiers. Baker's analogy: judging a free tier is like assessing a 30-year-old by watching their 10-year-old self.
  • Reasoning models supplied a new post-training scaling law that kept progress moving through the 18 month Blackwell delay.
  • Cost leadership is strategically decisive in AI for the first time in computing, because the cheapest token producer can price competitors out of their own economics.
  • The strongest bear case is edge inference: phones running pruned frontier models at 30-60 tokens per second would pull demand away from the cloud buildout.
  • The AI ROI debate is already settled in the filings, with GPU-buying public companies showing higher returns on invested capital after their AI spending ramped.
With software, anything you can specify, you can automate. With AI, anything you can verify, you can automate. — Gavin Baker

Read the full episode summary

5. The Supply and Demand of AI Tokens: Dylan Patel Interview

Invest Like The Best · Patrick O'Shaughnessy · Dylan Patel · 45m · April 2026

Forty five minutes with close to zero filler. Patel turns his own P&L into the argument: AI spend at his firm climbed from negligible to $7 million a year, about 25% of salary costs, and non-technical employees now ship in weeks what once took hundreds of people years. The Anthropic margin data he cites (roughly 30% to 72%+ gross margins as revenue scaled from $9B toward $40B+ ARR) is the cleanest evidence that demand outruns compute supply at current prices.

Key takeaways

  • The great inversion: execution used to be hard and ideas cheap. Implementation is now easy and metered in tokens, so the scarce skill is choosing which ideas justify the spend.
  • Anthropic's gross margin expansion alongside flat compute capacity implies pricing power, with demand exceeding supply at today's rates.
  • Access to frontier models is becoming a competitive moat of its own, concentrating advantage among enterprises that can pay for and secure early access.
  • Release cadence compressed from roughly six months to two, because AI accelerates the internal experiment loop at the labs themselves.
  • Robotics is the next demand wave: once software work is trivially cheap, attention moves to the physical economy where most value sits.
What used to matter a lot was execution was very very difficult and ideas were cheap. Now ideas are cheap and plentiful but execution is very easy. So really only the good ideas are the ones that can justify the spend on super cheap implementation. — Dylan Patel

Read the full episode summary

6. The AI Selloff Doesn't Match the Data: Top AI Investor Explains

Invest Like The Best · Patrick O'Shaughnessy · Gavin Baker · 1h 18m · August 2026

The episode to reach for when AI equities fall 40-60% while the fundamentals accelerate. Baker walks the drawdown line by line and lands on the number that settles it: spot pricing for GPU clusters rose 50-60% over six to seven months, with buyers paying close to $4 per GPU hour for Blackwell capacity against $2.50 earlier. His contract repricing thesis is the part most models still leave out.

Key takeaways

  • GPU spot pricing rose while the market priced in a glut, which points to an acute compute shortage rather than easing demand.
  • Hyperscalers hold long-term GPU contracts struck below current spot, so the roll-off will reprice compute higher and accelerate operating cash flow.
  • Open-source model growth shifts margin dollars between providers while leaving total compute demand intact: a token costs the same flops, memory, and watts wherever it comes from.
  • Roughly 250,000 to 500,000 people worldwide use generative AI meaningfully, under 0.01% of the population, which frames how much demand a scale-up to even 1% would add.
  • Investment professionals routing every headline through the same model has compressed market cycles, with Japanese capacitor stocks completing a three-year cycle in six weeks.
A token is a token, and you need the exact same amount of compute to make a token all else equal. Takes the same amount of flops, the same amount of memory, the same amount of watts. — Gavin Baker

Read the full episode summary

7. Gavin Baker on Orbital Compute, TSMC, and Frontier Models

Invest Like The Best · Patrick O'Shaughnessy · Gavin Baker · 1h 22m · May 2026

The most forward-looking episode here, and the clearest answer to who controls the pace of the buildout. Baker's argument is that TSMC's wafer discipline acts as the brake keeping AI capex from turning into a debt-fueled bubble, since Nvidia could sell $2-3 trillion of GPUs annually if the supply existed. His orbital compute section (GPU racks in space, laser interconnects, vast solar arrays, radiators for cooling) is the most concrete version of that idea on tape.

Key takeaways

  • Anthropic added $11 billion of ARR in a single month, comparable to what Palantir, Snowflake, and Databricks built together over a decade.
  • TSMC's controlled wafer expansion forces disciplined capital allocation across the whole AI supply chain and limits overbuilding.
  • Economic value at the model layer concentrates at the frontier, and models that felt remarkable months earlier become unusable as the frontier moves.
  • Orbital compute solves the long-run power constraint with unlimited solar and radiative cooling, building on SpaceX's proven work in laser links and satellite operations.
  • Cheap US natural gas has become a real manufacturing advantage for energy-intensive AI infrastructure.
Anthropic added their combined businesses in one month. That's just nothing like that has ever happened in the history of capitalism. Forget my career. Just the flatout history of capitalism, the history of business. — Gavin Baker

Read the full episode summary

8. The Two Harvard Dropouts Who Raised $800M to Take On NVIDIA

Invest Like The Best · Patrick O'Shaughnessy · Gavin Uberti and Rob Munro · 1h 32m · June 2026

A pure supply-side episode and a useful counterweight to the Nvidia consensus running through the rest of this list. The Etched founders explain low-voltage inference (power scales with the square of voltage, so cutting voltage buys large thermal headroom) and cluster-scale memory built on custom interconnects with roughly 5x lower latency than GPUs. Their claim that inference becomes the largest market in the world is the bet worth arguing with.

Key takeaways

  • General-purpose silicon carries buffer for IoT, edge, and data center use at once. Designing for inference alone let Etched delete defaults the rest of the industry treats as fixed.
  • Thermal limits cap GPU compute density, and low-voltage design attacks that ceiling directly because power scales with the square of voltage.
  • Bandwidth across the cluster matters more than bandwidth per chip, which is why Etched built its own interconnect and its own racks.
  • Vertical integration is a velocity strategy: thermal test chips with predicted hot spots and overpressurized cold plates let them validate before silicon returned.
  • Token production today resembles handcrafted manufacturing, and the founders expect industrial-scale economics to arrive the way they did for consumer hardware.
We know inference is going to be the biggest market in the world. Whoever produces the most tokens is going to be the most valuable company in the world. — Rob Munro

Read the full episode summary

9. World's Top Researcher on AI, LLMs, and Robot Intelligence

Invest Like The Best · Patrick O'Shaughnessy · Sergey Levine · 1h 15m · March 2026

The episode that tells you where token demand travels after software. Levine, co-founder of Physical Intelligence, argues that general-purpose robot foundation models will arrive faster than narrow task-specific machines, mirroring how language models beat purpose-built translation systems. The detail worth keeping: one model already controls radically different robot bodies with varying degrees of freedom, which suggests the hard part is physical understanding.

Key takeaways

  • Solving physical interaction broadly beats targeting narrow tasks, because training across many robots and tasks builds transferable understanding of physics.
  • Multimodal language models supply the common sense that robot learning always lacked, letting a robot reason through situations it has yet to encounter.
  • Pre-trained robotic foundation models need far less task-specific data, so new skills come from small amounts of demonstration.
  • The same model generalizes across multi-fingered hands and entirely different form factors with no explicit description of the hardware it drives.
  • A Cambrian explosion of robot applications becomes possible once the intelligence layer exists as a platform anyone can build on.
We believe that doing it at the full level of generality might actually in the long run be easier than trying to special case very specific narrow application domains. — Sergey Levine

Read the full episode summary

10. How Brian Chesky Is Redesigning Airbnb for the AI Era

Invest Like The Best · Patrick O'Shaughnessy · Brian Chesky · 1h 23m · May 2026

The strongest operator episode for understanding how AI reshapes an org chart. Chesky is blunt about which roles survive: everyone becomes a hybrid who manages people and does the work, so engineering leaders keep coding, lawyers keep reading case law, and design leaders keep designing. His Hawaii system (perfect one city with a team of 10-12, expand to ten, then go global) is a repeatable launch method inside a 7,000-person company.

Key takeaways

  • Founding and running a company draw on different skills: founder instincts are innate, while the CEO job is counterintuitive and has to be learned.
  • Start hands-on with a new team and give ground grudgingly, auditing the work first so that empowerment rests on knowledge of what is actually happening.
  • The Hawaii system makes problems small on purpose: one city, an elite team of 10-12, full excellence, then expansion.
  • AI pressures two groups hardest: pure people managers with no hands-on craft, and anyone rigid about changing how they work.
  • Consumer AI is the next wave after today's enterprise concentration, and it demands design, marketing, and culture excellence that enterprise playbooks skip.
No one is born a good CEO. I think people are basically born good founders or said differently it's innate. You don't have to learn how to be a good founder. The job of CEO is completely counterintuitive and almost all of your intuition about what to do is wrong. — Brian Chesky

Read the full episode summary

11. Why The Laws of Startup Physics Have Changed: Ben Horowitz Interview

Invest Like The Best · Patrick O'Shaughnessy · Ben Horowitz · 1h 2m · February 2026

The macro frame for the whole list. Horowitz's core point is that AI deploys over infrastructure that already exists, skipping the roads-and-fiber phase every prior platform required, which is why Cursor reached $1 billion in revenue faster than any company before it. His estimate that roughly 40 people globally have built large-scale models explains valuations that look absurd until you price the constraint honestly.

Key takeaways

  • AI ships instantly over the existing internet, compressing the gap between a working model and a revenue curve to a degree earlier platforms could only envy.
  • Frontier model builders number in the dozens and the skill comes from hands-on work at Google, OpenAI, or Anthropic, so scarcity drives the pay and the valuations.
  • Horowitz favors technology solutions over policy solutions, arguing that shipped products solve problems at scale while policy tends to shift power to administrators.
  • Confrontational management (firing executives, reorganizing power, delivering hard feedback) is the line he draws between competent and great founders.
  • He frames AI as democratizing at the access layer, putting superintelligence on every smartphone even as it widens the gap between creators.
If you really want to change the world, if you really want to make it a better place, I think you can build a solution for darn near anything. If you want to change the world for the better, it's never been a better time to be an entrepreneur. — Ben Horowitz

Read the full episode summary

12. Uber CEO on AI, Autonomous Vehicles, and the Future of Transportation

Invest Like The Best · Patrick O'Shaughnessy · Dara Khosrowshahi · 1h 11m · June 2026

The best episode here on AI meeting atoms. Khosrowshahi treats autonomy as a supply problem inside a supply-first marketplace, and his observation that magic normalizes within minutes (a five minute pickup felt like sorcery until 6.5 minutes started feeling slow) is the most useful warning against pricing novelty into AV forecasts. His note that developers in India drove 10x commit volume with agents shows where adoption actually surfaces.

Key takeaways

  • Uber is a supply company: more drivers, restaurants, and retailers create latent demand that materializes on its own, which shapes how autonomy gets underwritten.
  • Novelty decays within minutes, so autonomous vehicles will compete on safety, efficiency, and price the way every mature transport product does.
  • AI adoption inside a large company clusters in unpredictable places, and the job is to find the internal rebels racing ahead and promote what they build.
  • Break chaos into independent dimensions, solve each one separately, then recombine. Khosrowshahi credits this for steering Uber through its most turbulent stretch.
  • Daniel Ek's advice to optimize for impact over comfort is what moved him from Expedia into the Uber job.
We don't just exist in the digital sphere. Our experiences encompass digital first in terms of your interaction with our platform, but then they are delivered or fulfilled in the real world. — Dara Khosrowshahi

Read the full episode summary

What these episodes have in common

Compute is the binding constraint, and its price keeps climbing

The most durable disagreement in AI investing is whether compute stays scarce, and these episodes come down hard on one side. In The AI Selloff Doesn't Match the Data, Gavin Baker points at spot pricing for GPU clusters rising 50-60% over six to seven months, with buyers paying close to $4 per GPU hour for Blackwell capacity against $2.50 earlier. Prices that move up while equities fall 40-60% describe a shortage, and Baker's contract repricing thesis (hyperscalers locked in rates below spot, and the roll-off reprices higher) is the mechanism most models leave out.

Sam Altman describes the same constraint from inside the lab. His framing is that the bottleneck rotates between research ideas, compute, and data, and that the two are less separable than they sound: today's de-risking runs at OpenAI are the size of entire compute runs from 18 months ago. Dylan Patel supplies the financial fingerprint in The Supply and Demand of AI Tokens, citing Anthropic gross margins moving from roughly 30% to over 72% as revenue scaled from $9B toward $40B+ ARR with capacity roughly flat. Margin expansion without capacity expansion is what excess demand looks like on a P&L.

Where the episodes diverge is on the ceiling. Baker's orbital compute episode argues TSMC's wafer discipline is the brake keeping capex sane, with Nvidia capable of selling $2-3 trillion of GPUs annually if supply allowed, and points to GPU racks in orbit as the long-run answer to the power constraint. Sundheim takes the other side of the demand chain, warning that hyperscalers are becoming dependent on 4-5 labs who will insource compute as cash flow improves. Both can hold: the tightness is real today, and the buyer concentration is the risk on the other end of it.

The commoditization thesis lost, and the moats moved somewhere surprising

Two years ago the consensus held that foundation models would converge into an undifferentiated utility. Four episodes here document the reversal. In Why the AI Boom Is Just Getting Started, Alex lists the sources of durable differentiation his firm found: models that specialize by workload (one for finance and private equity work, another for PDF ingestion), harnesses and ecosystems that accumulate switching costs around an API, and a benchmark gap the open-source tier struggles to close because the compute sits with the labs.

Sundheim reaches the same conclusion through a different analogy. He describes LLM businesses as Netflix crossed with Spotify: heavy upfront spend to create the asset, then high-margin distribution, with personalization creating pricing power on top of what looks like a commodity. Baker's orbital compute episode sharpens it into a rule about where returns land, observing that value at the model layer concentrates at the frontier and that models which felt remarkable months earlier become unusable once the frontier moves.

The supply side of that argument gets its rebuttal from Etched. Gavin Uberti and Rob Munro are building inference silicon on the premise that whoever produces the most tokens becomes the most valuable company in the world, which implies the fight eventually migrates from model quality to cost per token. Baker's point in the selloff episode sits underneath all of it: a token costs the same flops, memory, and watts regardless of which model emits it. Open-source growth moves margin dollars around the industry while leaving the compute bill exactly where it was.

Operators are rebuilding the org chart faster than the market is pricing it

The three operator episodes on this list describe an organizational change already in motion. Brian Chesky is the most direct about it: the roles under pressure are pure people managers with no hands-on craft, and the survivors are hybrids who manage a team and still do the work, so engineering leaders keep coding and design leaders keep designing. His Hawaii system, which starts a new business in one city with a team of 10-12 before expanding to ten cities, is how he gets startup behavior out of a 7,000-person company.

Dara Khosrowshahi supplies the field data. Developers in India at Uber drove roughly 10x commit volume using autonomous agents, an adoption pattern that showed up with no relation to seniority or role. His instruction is to go find those internal rebels and promote what they built. Dylan Patel turns the same shift into a budget line, with AI spend at his firm reaching $7 million a year, roughly 25% of salary costs, and non-technical employees shipping work that used to require large teams and long timelines.

Ben Horowitz explains why the market keeps underestimating the speed. AI deploys over infrastructure that already exists, skipping the roads-and-fiber phase every earlier platform needed, which is how Cursor reached $1 billion in revenue faster than any company before it. Sergey Levine points at the next leg: once software work becomes cheap, robotic foundation models move the same dynamic into the physical economy, where most of the value has always been. Read together, these four make a single claim. The adoption curve is running ahead of the org charts, and the org charts are running ahead of the models analysts use to forecast the revenue.

Every episode referenced

Frequently Asked Questions

Which Invest Like The Best episodes are about AI?

Twelve episodes in our library focus squarely on artificial intelligence: Dan Sundheim on Anthropic and OpenAI, Sam Altman on AGI and compute, three Gavin Baker conversations (GPUs and TPUs, the AI selloff, orbital compute and TSMC), Dylan Patel on token supply and demand, Alex on why the AI boom is early, Etched founders Gavin Uberti and Rob Munro on inference chips, Sergey Levine on robot intelligence, Brian Chesky on Airbnb in the AI era, Ben Horowitz on startup physics, and Dara Khosrowshahi on autonomous vehicles. All twelve are ranked and summarized above.

What is the best Invest Like The Best episode about AI?

Start with Inside Dan Sundheim's Bets on Anthropic, OpenAI, and SpaceX (February 2026). It pairs a complete investing framework (the clarity-of-thought test for evaluating CEOs, the Netflix-plus-Spotify model of LLM economics) with a genuinely contrarian call on hyperscaler customer concentration. For infrastructure and chip economics, Gavin Baker's GPUs, TPUs, and the Economics of AI Explained is the better first listen. For the view from inside a frontier lab, choose the Sam Altman episode.

Is there an Invest Like The Best episode about Airbnb?

Yes. How Brian Chesky Is Redesigning Airbnb for the AI Era (May 2026, 1h 23m) covers the founder-to-CEO transition, what Chesky learned scaling to 7,000 employees, and his Hawaii system for launching new businesses: perfect one city with an elite team of 10-12 people, expand to ten cities, then go global. He also makes a pointed prediction that pure people managers face the most pressure from AI, while hybrid leaders who still practice their craft gain ground.

What do Invest Like The Best guests say about which AI companies to invest in?

Anthropic comes up most often. Sundheim owns it and credits Dario Amodei's written clarity; Alex describes his firm's position and the coding-market math behind it; Gavin Baker cites $11 billion of ARR added in a single month. Baker also makes the bull case for Nvidia and TSMC, arguing TSMC's wafer discipline limits overbuilding, and flags SpaceX for orbital compute. These are guest opinions relayed from the conversations, so treat them as the starting point for your own research.

Where can I read Invest Like The Best summaries for free?

Right here. Every episode listed on this page links to a full 1% Better summary with the core takeaway, four to eight detailed insights, and notable quotes, all free and with no account required. Our library covers the Invest Like The Best catalogue alongside more than a thousand episodes from shows including Huberman Lab, The Tim Ferriss Show, All-In, and Modern Wisdom, so you can preview a two-hour conversation in about three minutes.

Keep exploring