Synthesis
Model commoditisation and switching
The substitutability argument: frontier models have near-perfect substitutes and customers can switch with little effort, so there is no lock-in or lead that lasts and providers c…
Across 25 episodes the recurring claim, mostly Sam Lessin’s, is that switching is easy and leads do not last. He would swap Claude ‘tomorrow’ , calls a perfect-replacement model business very hard , and says ‘it’s too easy to switch. Leads don’t matter’ . Dave Morin says he switches all the time and the labs look like infrastructure providers and reports DeepSeek offering a Fable-level model at roughly a hundredth of the cost . Guest Adrian says algorithms and ideas are copied within days . Jess Lessin broadly agrees that her team keeps switching and that optionality matters , but reports anecdotally that OpenAI struggles to move customers off Anthropic and that frontier models were gaining share, not losing it, to open source . Brit Morin expects models to flip-flop for months rather than commoditise and thinks many services keep stickiness . Sam concedes premium positioning may be an exception .
Key points
- Sam says all frontier models are substitutes and he could switch immediately, so there is no differentiation or lock-in — 9 episodes. (Sam · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 14:41)
- Dave and Jess also describe switching between models, and enterprises optimising on cost — 4 episodes. (Dave · Why SpaceX’s $1.75T IPO Makes (Jun 2026) ▶ 20:43)
- Cheaper open or Chinese models are the reported pressure on closed-lab pricing — 5 episodes. (Dave · Anthropic Wants to Slow Down. (Sep 2026) ▶ 06:19)
- Ideas and algorithms are copied within days, so value built on them collapses — 3 episodes. (Adrian · Adrian (Oct 2026) ▶ 51:09) (Ben Broca · Henrik Werdelin & Ben Broca (Mar 2026) ▶ 44:30) (Sam · Why SpaceX’s $1.75T IPO Makes (Jun 2026) ▶ 30:11)
- The practical advice is to keep the model swappable and compete elsewhere (infrastructure, proprietary data, secret algorithms) — 4 episodes. (Sam · Silicon Valley Is Great at Tec (Oct 2026) ▶ 11:50) (Adrian · Adrian (Oct 2026) ▶ 51:28)
Where guests disagree
Will models commoditise or keep differentiated stickiness? ⚠️
- Models are a commodity with no lock-in; leads do not last (Sam) (Sam · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 14:41)
- Models will flip-flop rather than commoditise, and normal users want good-enough and do not switch (Brit); frontier models were gaining share and OpenAI cannot easily win customers from Anthropic (Jess) (Brit · Why the Met Gala Hates Tech, E (May 2026) ▶ 40:13)
Frameworks
- Three places left to build real AI businesses — Adrian’s answer to commoditisation: own infrastructure (energy, GPUs), proprietary data, or algorithms you keep secret. (Adrian · Adrian (Oct 2026) ▶ 51:28)
- Infrastructure: energy, GPUs and compute
- Proprietary data you do not share
- Algorithms, currently a commodity copied within days
Stories & examples
- Sam’s model stack: Sam runs Gemini for security-sensitive finance and health, Grok for images and a central Claude setup that controls many services; if Claude disappeared he would switch to OpenAI in a day. Lesson: With easy switching, none of the stack is sticky and it’s all plumbing. (Sam · Why SpaceX’s $1.75T IPO Makes (Jun 2026) ▶ 19:04)
- The Jev classifier copied within 24 hours: A single-pass LLM classifier that works really well was copied by everyone within 24 hours and its makers saw no value, per Adrian. Lesson: Algorithms are a commodity; a good idea in AI collapses to zero value quickly. (Adrian · Adrian (Oct 2026) ▶ 52:22)
- Sam re-implements Granola: Annoyed with Granola’s limits, Sam built his own transcription and summarisation tool; when his first API limited him, Claude pointed him to AssemblyAI, in which he invested 100k in the first round. Lesson: With no implementation cost, switching to a better service is instant, so nobody earns real money. (Sam · OpenAI’s $852 Billion Valuatio (Apr 2026) ▶ 28:29)
Numbers
- Dave reports OpenRouter data now shows a rival beating Anthropic for the first time in over a year and DeepSeek has a Fable-level model for roughly 100x less cost. (Dave · Anthropic Wants to Slow Down. (Sep 2026) ▶ 06:19)
Caveats & warnings
- Sam concedes that exceptions exist and that premium positioning, like Apple for rich people, could be defensible, though he thinks most intelligence will be an arbitrage game. (Sam · Anthropic’s Bet on Coding Is W (Mar 2026) ▶ 22:53)
- Sam says we are not there yet on commodity software, only in that direction. (Sam · Apple, OpenAI & Why Nobody Mak (Feb 2026) ▶ 14:16)
- Sam admits his own Fable upgrade is ‘not exciting’ but ‘fine’, the vibe he thinks matters to the revenue ramp. (Sam · Open Source AI Wins. Now Labs (Jul 2026) ▶ 10:14)
- Sam says is the stickiest UX known to man, and you could have a model that’s 50% as good, so a free model 50 percent as good could win consumer. Attribution uncertain (Sam, Dave or Jess). (one of the hosts · Google I/O Just Killed SEO? Ag (May 2026) ▶ 38:52)
Also covered here: Cost structure wins commodity markets
Nine episodes, again led by Sam Lessin, with one unattributed host in the clearest statement. In a commoditised market only the best cost structure wins, so judge the frontier labs against open models on cost (L2). Sam predicts the labs will lose to open source on cost for almost all agent work , says Anthropic is very expensive relative to everything else and most token spend will go to cheap models , and relays two public-company CTOs who use all models and switch to minimise cost, so a lab that cannot hold an efficiency edge is worthless . He also argues that bots compete everything to the marginal dollar, so capitalism ‘outmodes itself’ . Dave relays that four labs at a dinner said they had raised a billion and were spending about 900 million on inference , and cites DeepSeek at about a hundredth of the cost of a Fable-level model . Guest Ben Broca thinks Cursor’s Composer 2 was an attempt to make margins . Jess adds the labs lose money on training and look profitable on inference .
Key points
- In a commoditised market the best cost structure wins; compare labs with open models on cost, not quality — 3 episodes. (one of the hosts · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 35:58)
- Sam expects open models to take most agent workloads on cost, with premium tokens a niche — 3 episodes. (Sam · Stripe Is Trying To Buy PayPal (Jul 2026) ▶ 34:26)
- Competition and automation drive margin out of everything — 1 episode. (Sam · Apple, OpenAI & Why Nobody Mak (Feb 2026) ▶ 19:08)
- Reported lab spending on inference is very high relative to funds raised — 2 episodes. (Dave · Anthropic’s Bet on Coding Is W (Mar 2026) ▶ 20:49)
- Non-frontier players should expect to sell or rent infrastructure rather than a destination product — 2 episodes. (Dave · Apple’s New Siri, SpaceX’s $30 (Jun 2026) ▶ 05:51) (Ben Broca · Henrik Werdelin & Ben Broca (Mar 2026) ▶ 42:02)
Stories & examples
- Cursor’s Composer 2: Ben says Cursor dropped Composer 2, which he believes was RL’d on a Chinese open model (Kimi 2.5), as a way to make margins. Lesson: Custom models on open source can get 90% of the value. (Ben Broca · Henrik Werdelin & Ben Broca (Mar 2026) ▶ 42:02)
Numbers
- Raised versus spent on inference by each of four labs, as relayed by Dave: 900 million dollars spent of one billion raised (spoken ‘900’) (host’s framing, approx.) (Dave · Anthropic’s Bet on Coding Is W (Mar 2026) ▶ 20:49)
- Cost of DeepSeek’s Fable-level model relative to Fable: 100 times cheaper (host’s framing, approx.) (Dave · Anthropic Wants to Slow Down. (Sep 2026) ▶ 06:27)
Caveats & warnings
- Dave relays that four labs at a dinner said they raised a billion dollars and were spending 900 of it on inference. (Dave · Anthropic’s Bet on Coding Is W (Mar 2026) ▶ 20:49)
Also covered here: No return on being first
Nine episodes carry this, nearly all from Sam Lessin. His headline is ‘there is no return on being first, which means you’re in a brutal competition to zero’ . His reasoning: following is much cheaper than blazing the trail, and he illustrates it with a rough 100,000 tokens to figure something out against 10,000 to clone it (N4, N5), with a Slow portfolio company, Polsia, as the cloneable example . He reads Anthropic overtaking OpenAI as proof there is no winner-take-all AGI race and says leads do not matter , though he notes Anthropic’s coding lead may not last . The idea of a zero-sum race is traced in the corpus to what hosts relay as Eric Schmidt’s ‘San Francisco Consensus’, which Schmidt is reported to doubt . Guest Ben Broca’s only dissent from pure copying is that speed is the single moat, which model companies hold by seeing new models about three months early . A relayed anecdote says only lab researchers at a dinner believed one company would be left standing .
Key points
- Sam argues there is no return on being first in AI, so the labs are in competition to zero — 3 episodes. (Sam · Anthropic Wants to Slow Down. (Sep 2026) ▶ 06:49)
- Following is far cheaper than discovering, so clonable businesses see margins collapse — 2 episodes. (Sam · Apple, OpenAI & Why Nobody Mak (Feb 2026) ▶ 17:59)
- The ‘whoever reaches AGI first wins’ story is disputed by the hosts as a fundraising narrative — 4 episodes. (Sam · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 14:24)
- If copying is instant, speed is the only moat (Ben Broca) — 1 episode. (Ben Broca · Henrik Werdelin & Ben Broca (Mar 2026) ▶ 44:24)
Stories & examples
- Polsia: a sole founder and agents running the business: Polsia, a Slow portfolio company, has one founder and no team; AI agents run the business in real time and ARR went from about 100K to about 700K in under a week. Lesson: Agentic businesses can scale fast, but a clone can re-implement them cheaply, which Sam uses to argue margins collapse. (Sam · Apple, OpenAI & Why Nobody Mak (Feb 2026) ▶ 15:03)
Numbers
- Tokens to figure it out vs clone it: 100000 tokens (to figure out) (host’s framing, approx.) (Sam · Apple, OpenAI & Why Nobody Mak (Feb 2026) ▶ 18:12)
- Tokens needed to clone it: 10000 tokens (to clone) (host’s framing, approx.) (Sam · Apple, OpenAI & Why Nobody Mak (Feb 2026) ▶ 18:14)
Caveats & warnings
- Sam says Anthropic’s coding lead may not last: technical people now prefer Codex, though it does not matter who leads right now. (Sam · Stripe Buys OpenRouter, OpenAI (Aug 2026) ▶ 30:38)
- Relayed (Dave or Brit): at a dinner, only the lab researchers raised their hands when asked who believed one company would be left standing. (one of the hosts · Silicon Valley Is Great at Tec (Oct 2026) ▶ 18:13)
How much to trust this
- Evidence: 27 episodes, 4 independent speakers (the four hosts and unidentified host speech count as one voice, the show; guests counted individually).
Blueprints that use this
Sources
- (01)
Is Google About to Kill Canva? │Claude Goes Rogue, Tesla vs. Waymo
23:27 into “Is Google About to Kill Canva? │Claude Goes Rogue, Tesla vs. Waymo”
- (02)
Why the Met Gala Hates Tech, Elon vs OpenAI Drama, and the Rise of Chinese AI Models
40:13 into “Why the Met Gala Hates Tech, Elon vs OpenAI Drama, and the Rise of Chinese AI Models”
- (03)
Stripe Is Trying To Buy PayPal For $150b | OpenAI Cyber Incident, Frontier Model Economics, Peptides
- (04)
- (05)
Open Source AI Wins. Now Labs Are Running to Washington |Thinking Machines, Johnny Ive, Realtime API
- (06)
Elon Musk vs OpenAI Trial, Google Cloud Surge, Meta’s Blocked Acquisition, Anthropic Winning
- (07)
Allbirds Pivoted to AI Data Centers | SpaceX $1.5T IPO, Anthropic $800B, $175M Seed Rounds
- (08)
Apple, OpenAI & Why Nobody Makes Money in AI | The SaaS Apocalypse
13:03 into “Apple, OpenAI & Why Nobody Makes Money in AI | The SaaS Apocalypse”
- (09)
OpenAI's $852 Billion Valuation Makes Zero Sense (ClawCon Tokyo, Anthropic's Leak, Goodbye Internet)
- (10)
The Internet Is Fighting Back Against Your AI Agent
11:20 into “The Internet Is Fighting Back Against Your AI Agent”
- (11)
Stripe Buys OpenRouter, OpenAI vs. Anthropic & the AI Backlash
26:11 into “Stripe Buys OpenRouter, OpenAI vs. Anthropic & the AI Backlash”
- (12)
The Risk Of Securitizing GPUs Like Long-Lived Infrastructure & More - A MOL Special Team Edition
- (13)
Apple’s New Siri, SpaceX’s $300B IPO & the AI Company Killing Its Own Brand
5:51 into “Apple’s New Siri, SpaceX’s $300B IPO & the AI Company Killing Its Own Brand”
- (14)
SpaceX Hit $2.5 Trillion. Now Watch What Elon Buys Next.
25:45 into “SpaceX Hit $2.5 Trillion. Now Watch What Elon Buys Next.”
- (15)
Google I/O Just Killed SEO? Agentic Search, Universal Cart, and the New Web
38:52 into “Google I/O Just Killed SEO? Agentic Search, Universal Cart, and the New Web”
- (16)
Chinese AI Model GLM 5.2 Beating Frontier Models | Meta Glasses, Polymarket Scandal, AI Talent War
- (17)
Why SpaceX's $1.75T IPO Makes No Financial Sense (And Doesn't Have To)
19:04 into “Why SpaceX's $1.75T IPO Makes No Financial Sense (And Doesn't Have To)”
- (18)
Meta Muse vs Instinct vs OpenAI Dots: Do AI Agents Have Any Moat?
51:09 into “Meta Muse vs Instinct vs OpenAI Dots: Do AI Agents Have Any Moat?”
- (19)
Anthropic Wants to Slow Down. Is It Because They're Losing?
6:19 into “Anthropic Wants to Slow Down. Is It Because They're Losing?”
- (20)
Anthropic’s Bet on Coding Is Working (OpenAI Shopping Pivot, A16Z’s Top 50 List, $1B Tennis Channel)
- (21)
Silicon Valley Is Great at Tech and Terrible at Business
11:20 into “Silicon Valley Is Great at Tech and Terrible at Business”
- (22)
AI Frontier Lab Is Essentially Doomed | Open AI Offered U.S. Equity, Xbox Is In Trouble, Post-AGI
- (23)
The New Startup Stack: One Founder + Agents | Henrik Werdelin (Audos) and Ben Broca (Polsia)
- (24)
Every CTO Is Quietly Switching Off Anthropic and OpenAI | GLM 5.2, Vox Americanus, Claude Fable 5