LLenny's Podcast
← All episodes
Brian Balfour (Reforge)17 August 2025

Why ChatGPT will be the next big growth channel (and how to capitalize on it)

4Frameworks
15Insights

Frameworks in this episode

Leadership5 steps

Driving AI Adoption Through Hard Constraints

AI transformation moves when you impose hard constraints, sort people into three groups, and fix the slowest part of the system.

Read →
Marketing5 steps

Four Criteria for Choosing a Distribution Platform (Enter and Exit)

Score a new platform on retention, monetizability, value exchange, and scale — then plan your exit before you enter.

Read →
Strategy4 steps

Platform Betting Strategy by Company Stage

Late-stage companies spread chips across platforms; startups make one focused bet and go all in.

Read →
Strategy4 steps

The Four-Step Distribution Platform Cycle

Every new growth channel opens then closes in the same four predictable steps — learn to see where you are.

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 4

Hot Take05:00

A Great Product Is Necessary But Not Sufficient — Distribution Is the Separator

Balfour pushes back on the common advice that winning is mostly about building a great product. In his experience the real separation between winners and losers is who builds truly great distribution, and startups are playing a race to reach escape velocity before an incumbent can copy them.

  • Most advice reduces winning to 'build a great product' — Balfour calls this necessary but not sufficient.
  • The actual separator between companies is who builds really great distribution.
  • He cites Alex Rampell's 2015 a16z post: startups win by getting distribution before the incumbent can copy them (escape velocity).

“building a great product is one of those things that's necessary um but not sufficient and actually the separation is between those that build really…”

Brian Balfour · 05:00
#distribution#startups#growth#product
Hot Take28:00

There Is No Opting Out — Adopting the New Platform Is a Prisoner's Dilemma

Even when integrating with a new platform looks bad in isolation (e.g. exposing your data to ChatGPT), Balfour argues you don't operate in isolation. If you don't move, competitors will, and customer expectations shift toward the new experience — making it a prisoner's dilemma with no opt-out.

  • In isolation, exposing your data through ChatGPT connectors looks like a bad idea.
  • But competitors will go to the new platform and customers will start expecting you there.
  • It becomes a prisoner's dilemma: you have to play, and it's better to be early than late — especially for startups.

“if you don't do it, your competitors are going to certainly go to the new platform and your customer expectations change”

Brian Balfour · 28:00

“it's better to be early than than to be super late uh to this game especially especially if you are a startup”

Brian Balfour · 28:30
#distribution#chatgpt#strategy#competition
Hot Take35:30

Apple Is Best Positioned on Context — But Isn't Executing

Asked for the backup if ChatGPT isn't the winner, Balfour says Apple is theoretically best positioned because its devices see everything about you, giving the ultimate view into your context — but there are no external signals it's executing. He puts Google next, given it owns email context and distribution via Search, Chrome and Android.

  • Apple's devices give it the ultimate view into a user's context.
  • But Balfour sees no external signals Apple is executing on it.
  • Google ranks next thanks to email context and distribution through Search, Chrome and Android.
  • He suspects a big portion of Gemini's MAU are accidental 'flyby' users.

“my hypothesis of who's best positioned but is not executing on it right now would actually be Apple”

Brian Balfour · 35:30

“through the the devices they basically can see everything so they have the ultimate view into your context”

Brian Balfour · 35:30
#apple#google#ai#prediction
Hot Take38:00

Anthropic Is Winning by Leaning Hard Into Coding

Relaying a conversation with Anthropic CPO Mike Krieger, Lenny explains that rather than chase ChatGPT head-on, Anthropic is doubling down on what Claude is exceptionally good at — developer tools, coding and backend work — and that the strategy is paying off, with revenue reportedly approaching $10 billion a year.

  • Krieger said ChatGPT has 'caught lightning in a bottle' and is going to win the consumer race.
  • Anthropic is deliberately focusing on developer tools, coding and backend work.
  • The differentiated bet is working — Claude's revenue is reportedly approaching ~$10B a year.

“we are specifically focusing on what is Enthropic and Colad incredibly good at which is developer tools, coding, backend stuff.”

Lenny · 38:30

“if you've seen their revenue recently, they're making I don't know like approaching 10 billion a year or some crazy amount of money.”

Lenny · 38:30
#anthropic#claude#coding#strategy

Explainer· 6

Explainer06:00

Why the Startup Escape-Velocity Game Just Got Much Harder

Balfour explains three forces that have made it harder for startups to out-run incumbents: incumbents can now copy faster (shrinking the escape-velocity window), organic distribution channels are shrinking, and AI is very good at writing software, flooding every market with competitors.

  • Incumbents copy faster now, shrinking the window to reach escape velocity.
  • Organic distribution has declined — SEO clicks, LinkedIn's algorithm change, social platforms sending less traffic.
  • AI code generation means infinite new competition; YC pumps out multiple copies of the same idea each cohort.
  • It has gotten easier only in rare cases like Cursor, where AI acted as a spark for early adopters.

“everybody's kind of feeling this infinite increase of competition, especially at the startup level. And you know, YC is pumping out six of the same…”

Brian Balfour · 07:00
#distribution#seo#ai#competition
Explainer30:30

Why the Real Moat in AI Is Context and Memory

Balfour argues the base models are increasingly interchangeable, so the differentiator is which one holds more of your context and memory. Usage compounds into a flywheel — more use means more stored memory, which feeds more personalized context and better outputs — and ChatGPT is furthest along on this.

  • Side-by-side, the models produce roughly the same result; context is the difference-maker.
  • It's the context PLUS the model that produces the best output.
  • Memory creates a flywheel: more use → more stored memory → more personalized context → better outputs.
  • ChatGPT was first to memory and has invested heavily in data/context connectors.

“the mode is really about is about context and memory.”

Brian Balfour · 30:30

“The more you use it, the more it's able to store memory around you, which kind of feeds more personalized context, which produces better outputs,…”

Brian Balfour · 31:00
#ai#moat#chatgpt#memory
Explainer31:30

Retention and Engagement Beat Raw Distribution Every Time

Balfour notes that historically the winner was never the player with the biggest distribution at a given moment — it was the one with the best retention and engagement. He points to Menlo Ventures data showing ChatGPT's retention curves level off higher than rivals and display the rare 'smile curve' that signals escape velocity.

  • Google and Facebook both won with better retention/engagement, not the biggest starting distribution.
  • Menlo Ventures' data shows ChatGPT retention curves leveling off higher and shifting up over time.
  • The 'smile curve' — retention that dips then climbs back — is an early indicator of escape velocity, seen in winners like Slack.

“It was the one that had the best retention and engagement. Google had the best retention engagement over the others.”

Brian Balfour · 32:00

“they have the very elusive smile curve, right?”

Brian Balfour · 32:30
#retention#engagement#chatgpt#metrics
Explainer57:30

How to Evaluate Which New Platform to Bet On

Balfour gives the signals he'd weigh when choosing a new distribution platform: retention and depth of engagement over vanity metrics like signups/MAU, the quality and monetizability of that platform's users, the value exchange being offered, and finally raw scale. He uses iOS vs Android to show why user quality can beat user count.

  • Prioritize retention and depth of engagement over vanity metrics like MAU or signups.
  • Weigh user quality and ability to monetize those users.
  • Analyze the value exchange — these platforms reward whoever best arbitrages the rules.
  • iOS/Android: Android has ~70%+ of devices but only ~30% of market share by dollars — betting Android-only lost, iOS-only could win.

“the better signal is retention and depth of engagement of the users on this platform than it is like pure kind of user level like…”

Brian Balfour · 57:30

“even today it's something like Android has 70 70ome percent of devices but only 30% of the market share by dollars and it's the exact…”

Brian Balfour · 58:00
#platforms#evaluation#ios#android
Explainer1:08:30

Hard Constraints — Not Memos — Drive Real AI Adoption

Balfour says the single most impactful thing a company can do to become AI-native is impose hard constraints, not issue manifestos. Examples include capping each function's headcount at a benchmarked fraction of peers, refusing new headcount until you prove AI can't do the job, and executives refusing to review a PRD unless it ships with three prototypes.

  • Hard constraints move the needle far more than communication, owners, incentives or career-ladder changes.
  • One company set a benchmark that each function would be a fraction of peer-company size, forcing AI adoption.
  • Another (possibly Shopify) blocks new headcount until you prove AI can't accomplish the work.
  • An exec won't review a PRD unless it comes with three prototypes.

“the most impactful thing um that you can do is form really hard constraints”

Brian Balfour · 1:08:30

“this might have been Shopify or another who was like you are not allowed new headcount until you prove to us that you're not able…”

Brian Balfour · 1:10:00
#ai-adoption#management#constraints#transformation
Explainer1:10:30

Catalysts, Converts and Anchors — Why the Best Companies Exit People

Balfour describes three groups in any AI transformation: catalysts who lead the charge, converts who adapt but need structure and permission, and anchors who quietly drag their feet. The companies furthest along treat AI as a fundamental culture change and set a hard deadline to exit anchors — fewer than 10% take that stance, but they're getting the most results.

  • Catalysts lead on their own; converts adapt but need structure, permission and a clear plan; anchors create silent friction.
  • Leading companies treat AI as a fundamental culture change, not a light tool adoption.
  • Cultures thrive on density — you can't have 20-30% operating in a different culture.
  • Fewer than 10% set a hard deadline to exit anchors, but those companies see the most adoption and results.

“in every transformation what we see is essentially three groups of folks.”

Brian Balfour · 1:10:30

“less than 10% of companies we see are taking this hard stance but I would say they are probably the ones that are farthest along…”

Brian Balfour · 1:13:00
#ai-adoption#culture#management#transformation

Story· 2

Story39:00

How Udemy Cut Creator Rev-Share From 80% to 15-20%

As a concrete example of the platform-closing pattern, Balfour describes Udemy: it launched with a roughly 80% rev-share to course creators to attract a marketplace, then about a year ago pushed that share down to somewhere between 15-20%. It's the same close-the-platform-to-monetize move he expects to repeat in AI.

  • Udemy opened with an ~80% rev-share to creators to bootstrap its marketplace.
  • About a year ago it cut creator rev-share to roughly 15-20% (paying out ~25-30%).
  • Same pattern of suppressing creator economics to monetize will play out in AI platforms and smaller agent ecosystems like Cursor's.

“I believe it was about a year ago they announced that they're essentially pushing that rev share down to something like somewhere between 15 and…”

Brian Balfour · 39:30
#udemy#platforms#monetization#creators
Story1:14:30

Executives Are Wildly Disconnected From Real AI Adoption

Balfour says most executives who issued AI decrees assume adoption is happening naturally — but it usually isn't. When his team asks end-users how many teammates use a prototyping tool, ~90% say just them and one other person. He recounts a principal PM at a famous 'AI-native' company whose stalled experiment only moved after they happened to tell the CEO at a happy hour.

  • Most CEOs and executives are disconnected from actual AI adoption inside their companies.
  • ~90% of the time, an end-user says only they and one other person actually use a new tool.
  • A stalled prototyping experiment at a major 'AI-native' company only unblocked after the PM told the CEO at a happy hour.
  • Leaders must get to the ground floor; companies like Shopify measure actual adoption and usage.

“most CEOs or most executives are incredibly disconnected from the actual AI adoption taking place in in inside their companies.”

Brian Balfour · 1:14:30

“almost 90% of the time it's like ah it's like me and this one other person and everybody else hasn't uh like taken it up…”

Brian Balfour · 1:15:00
#ai-adoption#leadership#measurement#founder-mode

Takeaway· 3

Takeaway50:00

Startups Can't Spread Their Chips — Make One Focused Bet

Balfour distinguishes late-stage companies, which can afford to place multiple bets and wait to see the winner, from startups, which have scarce resources and must go all-in on one platform. His simplified advice: play the game, don't trick yourself into opting out, and whoever you bet on, make it a single focused bet.

  • Late-stage companies can spread bets and wait; the risk is waiting too long.
  • Startups must choose one platform and go all-in — higher risk, higher reward.
  • Historic failures came from playing multiple games at once with scarce resources.
  • Two rules: play the game, and make it a focused bet.

“You don't have the luxury to spread your chips. Like you have to go all in. You have to choose one and go all in.”

Brian Balfour · 50:00

“Don't opt out of the game. Don't don't trick yourself into thinking that you can't play the game.”

Brian Balfour · 51:30
#startups#strategy#betting#focus
Takeaway56:00

It Always Feels Too Late — And It Almost Never Is

Reflecting on Lenny joining Substack when it felt too late, Balfour argues that new opportunities always feel over-crowded to the people considering joining. He cites Marc Andreessen thinking he'd missed Silicon Valley in the 80s, and notes the bubble of Twitter/podcast chatter tricks you — only ~1% of people actually know about what feels saturated.

  • New waves always feel too late to newcomers — but usually aren't.
  • Marc Andreessen thought he'd missed Silicon Valley when he arrived in the 80s.
  • Living in the Twitter/podcast bubble distorts perception; ~1% of people know about the thing you hear about daily.

“Mark Andre has this famous quote. He's like, I came to Silicon Valley in the 80s. I thought it was over. It was too late.…”

Brian Balfour · 56:00

“A lot of times when people think that it's too late. It's it's definitely not too late and it's always only just getting started.”

Brian Balfour · 56:30
#timing#opportunity#substack#adoption
Takeaway1:17:00

Your Output Is Only as Fast as the Slowest Part of the System

Quoting Fared Masavat, Balfour frames AI adoption as a system whose output is capped by its slowest part — often IT, legal or procurement rather than the tools themselves. He warns that speeding up only engineering just moves the bottleneck: product is the output of design, PM and engineering together, so accelerating one part doesn't accelerate shipped product.

  • Output is constrained by the slowest part of the system.
  • The slowest parts are often IT, legal and procurement — not the tooling.
  • Accelerating only engineering shifts the bottleneck to PM or design.
  • The system exists to ship product, not produce code — product is a function of design, PM and engineering.

“the slowest your output is uh constrained by the slowest part of your system.”

Brian Balfour · 1:17:00

“product is an output of design, PMS and engineering. The system is there not to produce code. It's to ship product, right?”

Brian Balfour · 1:18:00
#bottleneck#systems#productivity#ai-adoption