The $SPCX IPO prices todays at $1.77 trillion. Brad Gerstner (@altcap) and Gavin Baker (@GavinSBaker) just spent 80 minutes on @BG2Pod making the case for why that's the cheapest it's going to get.
Here are the 10 takeaways worth saving:
1. In 30 days, SpaceX added $29 billion in AI compute revenue from the Anthropic and Google deals and jumped from not being an AI hyperscaler at all to being the #4 hyperscaler globally, passing Oracle. The trailing multiple compressed from ~100x to 39x in the same window. CoreWeave, Nebius, Iron and the 50 other neoclouds VCs are funding in Silicon Valley are now competing for a smaller slice of what's left.
2. xAI's Google deal generates more operating profit per gigawatt than Anthropic, Meta, Google or OpenAI's own infrastructure. Freda at Altimeter calculated a 55% IRR on Colossus 1. Borrow at 7%, invest at 55%, the math maths. Jensen called the build itself an "N of 1": 100,000 GPUs is normally a 3-year planning cycle plus 1-year deployment. xAI did it in 19 days. Speed is literally cost.
3. The premium Google is paying SpaceX for terrestrial compute is partly a call option on orbital. They want first-in-line when space data centres go live, so they're overpaying today to lock in the relationship.
4. Orbital compute costs about $5 billion per gigawatt of capex to put GPUs in space, vs $20-25 billion per gigawatt for the non-silicon half on the ground (land, power, cooling, switchgear). Space, power and cooling are effectively free up there. The unlock is rapid two-stage reusability of Starship, which takes launch from $1,500 per kg on Falcon to $250 per kg and eventually asymptotes to the cost of fuel. Elon says 3 years, Bezos says 6, truth is probably 4-5.
5. Starlink is at less than 1% global household penetration. The forecast going from $10bn to $50bn in connectivity revenue by 2028 is still only 0.3% of the global telecom market. TAM is not the constraint.
6. The most underrated piece of the IPO is the Cursor acquisition. Cursor and Anthropic each hold more proprietary coding tokens than exist on the public internet combined. xAI bought 700-800 people and a frontier-quality coding dataset, dropped it into Colossus 2 for 3 weeks of training, and Composer 2.5 went pareto-dominant on coding 12 days ago. Grok 4.3, a 1.5 trillion parameter model, is also currently on the pareto frontier. There are now four frontier labs, not three: xAI alongside Google (Gemini 3.1 Pro), Anthropic and OpenAI.
7. Anthropic just shipped Fable 5 (Mythos with safety classifiers). Karpathy says it's SOTA on benchmarks but the real unlock is long-running tasks. Stripe refactored a 50 million line Ruby codebase in a day. Used to take many weeks with many engineers. Noam Brown's corollary: snapshot benchmarks are dead. The x-axis now has to be time, tokens or compute, because frontier models can solve most problems if you let them run long enough. Nobody has ever run Mythos for a year continuously. We may never actually know how smart any given generation is.
8. The cleanest framing of the long-running thesis: imagine Albert Einstein, no need to eat, sleep, or age, thinking about one problem in fundamental physics for one straight year. That's the case for spending $1.5 trillion a year on compute.
9. The frontier captures ~90% of AI revenue. Open source captures ~80% of tokens. Both are true. The market priced one and missed the other. The bear case last year was that cheap open-source tokens would close the gap. Six months in, the frontier is extending its lead instead.
10. Capex for the hyperscalers is moving from $1.1T to ~$1.5T by 2027 (Morgan Stanley revised up). Inference revenue is projected at $300B by 2027 with 60-70% gross margins and roughly 35% of capex going to non-revenue training runs. Brad thinks $300B is low and we end this year at $200B+. Meanwhile Nvidia has not been losing share to ASICs. Once you adjust for Anthropic on TPUs, Nvidia has held or expanded against Broadcom, AMD, Cerebras, MTIA and OpenAI's Jalapeno. Tokens-per-watt is revenue-per-watt in a power-constrained world.
Bonus: The MAG 7 added $1 trillion of revenue over the last 7 years and that produced $17 trillion of market cap. The forecast: SpaceX, Anthropic and OpenAI add the next $1 trillion in revenue in 4-5 years. Three companies, half the time.
The $SPCX IPO prices todays at $1.77 trillion. Brad Gerstner (@altcap) and Gavin Baker (@GavinSBaker) just spent 80 minutes on @BG2Pod making the case for why that's the cheapest it's going to get.
Here are the 10 takeaways worth saving:
1. In 30 days, SpaceX added $29 billion in AI compute revenue from the Anthropic and Google deals and jumped from not being an AI hyperscaler at all to being the #4 hyperscaler globally, passing Oracle. The trailing multiple compressed from ~100x to 39x in the same window. CoreWeave, Nebius, Iron and the 50 other neoclouds VCs are funding in Silicon Valley are now competing for a smaller slice of what's left.
2. xAI's Google deal generates more operating profit per gigawatt than Anthropic, Meta, Google or OpenAI's own infrastructure. Freda at Altimeter calculated a 55% IRR on Colossus 1. Borrow at 7%, invest at 55%, the math maths. Jensen called the build itself an "N of 1": 100,000 GPUs is normally a 3-year planning cycle plus 1-year deployment. xAI did it in 19 days. Speed is literally cost.
3. The premium Google is paying SpaceX for terrestrial compute is partly a call option on orbital. They want first-in-line when space data centres go live, so they're overpaying today to lock in the relationship.
4. Orbital compute costs about $5 billion per gigawatt of capex to put GPUs in space, vs $20-25 billion per gigawatt for the non-silicon half on the ground (land, power, cooling, switchgear). Space, power and cooling are effectively free up there. The unlock is rapid two-stage reusability of Starship, which takes launch from $1,500 per kg on Falcon to $250 per kg and eventually asymptotes to the cost of fuel. Elon says 3 years, Bezos says 6, truth is probably 4-5.
5. Starlink is at less than 1% global household penetration. The forecast going from $10bn to $50bn in connectivity revenue by 2028 is still only 0.3% of the global telecom market. TAM is not the constraint.
6. The most underrated piece of the IPO is the Cursor acquisition. Cursor and Anthropic each hold more proprietary coding tokens than exist on the public internet combined. xAI bought 700-800 people and a frontier-quality coding dataset, dropped it into Colossus 2 for 3 weeks of training, and Composer 2.5 went pareto-dominant on coding 12 days ago. Grok 4.3, a 1.5 trillion parameter model, is also currently on the pareto frontier. There are now four frontier labs, not three: xAI alongside Google (Gemini 3.1 Pro), Anthropic and OpenAI.
7. Anthropic just shipped Fable 5 (Mythos with safety classifiers). Karpathy says it's SOTA on benchmarks but the real unlock is long-running tasks. Stripe refactored a 50 million line Ruby codebase in a day. Used to take many weeks with many engineers. Noam Brown's corollary: snapshot benchmarks are dead. The x-axis now has to be time, tokens or compute, because frontier models can solve most problems if you let them run long enough. Nobody has ever run Mythos for a year continuously. We may never actually know how smart any given generation is.
8. The cleanest framing of the long-running thesis: imagine Albert Einstein, no need to eat, sleep, or age, thinking about one problem in fundamental physics for one straight year. That's the case for spending $1.5 trillion a year on compute.
9. The frontier captures ~90% of AI revenue. Open source captures ~80% of tokens. Both are true. The market priced one and missed the other. The bear case last year was that cheap open-source tokens would close the gap. Six months in, the frontier is extending its lead instead.
10. Capex for the hyperscalers is moving from $1.1T to ~$1.5T by 2027 (Morgan Stanley revised up). Inference revenue is projected at $300B by 2027 with 60-70% gross margins and roughly 35% of capex going to non-revenue training runs. Brad thinks $300B is low and we end this year at $200B+. Meanwhile Nvidia has not been losing share to ASICs. Once you adjust for Anthropic on TPUs, Nvidia has held or expanded against Broadcom, AMD, Cerebras, MTIA and OpenAI's Jalapeno. Tokens-per-watt is revenue-per-watt in a power-constrained world.
Bonus: The MAG 7 added $1 trillion of revenue over the last 7 years and that produced $17 trillion of market cap. The forecast: SpaceX, Anthropic and OpenAI add the next $1 trillion in revenue in 4-5 years. Three companies, half the time.
The $SPCX IPO prices todays at $1.77 trillion. Brad Gerstner (@altcap) and Gavin Baker (@GavinSBaker) just spent 80 minutes on @BG2Pod making the case for why that's the cheapest it's going to get.
Here are the 10 takeaways worth saving:
1. In 30 days, SpaceX added $29 billion in AI compute revenue from the Anthropic and Google deals and jumped from not being an AI hyperscaler at all to being the #4 hyperscaler globally, passing Oracle. The trailing multiple compressed from ~100x to 39x in the same window. CoreWeave, Nebius, Iron and the 50 other neoclouds VCs are funding in Silicon Valley are now competing for a smaller slice of what's left.
2. xAI's Google deal generates more operating profit per gigawatt than Anthropic, Meta, Google or OpenAI's own infrastructure. Freda at Altimeter calculated a 55% IRR on Colossus 1. Borrow at 7%, invest at 55%, the math maths. Jensen called the build itself an "N of 1": 100,000 GPUs is normally a 3-year planning cycle plus 1-year deployment. xAI did it in 19 days. Speed is literally cost.
3. The premium Google is paying SpaceX for terrestrial compute is partly a call option on orbital. They want first-in-line when space data centres go live, so they're overpaying today to lock in the relationship.
4. Orbital compute costs about $5 billion per gigawatt of capex to put GPUs in space, vs $20-25 billion per gigawatt for the non-silicon half on the ground (land, power, cooling, switchgear). Space, power and cooling are effectively free up there. The unlock is rapid two-stage reusability of Starship, which takes launch from $1,500 per kg on Falcon to $250 per kg and eventually asymptotes to the cost of fuel. Elon says 3 years, Bezos says 6, truth is probably 4-5.
5. Starlink is at less than 1% global household penetration. The forecast going from $10bn to $50bn in connectivity revenue by 2028 is still only 0.3% of the global telecom market. TAM is not the constraint.
6. The most underrated piece of the IPO is the Cursor acquisition. Cursor and Anthropic each hold more proprietary coding tokens than exist on the public internet combined. xAI bought 700-800 people and a frontier-quality coding dataset, dropped it into Colossus 2 for 3 weeks of training, and Composer 2.5 went pareto-dominant on coding 12 days ago. Grok 4.3, a 1.5 trillion parameter model, is also currently on the pareto frontier. There are now four frontier labs, not three: xAI alongside Google (Gemini 3.1 Pro), Anthropic and OpenAI.
7. Anthropic just shipped Fable 5 (Mythos with safety classifiers). Karpathy says it's SOTA on benchmarks but the real unlock is long-running tasks. Stripe refactored a 50 million line Ruby codebase in a day. Used to take many weeks with many engineers. Noam Brown's corollary: snapshot benchmarks are dead. The x-axis now has to be time, tokens or compute, because frontier models can solve most problems if you let them run long enough. Nobody has ever run Mythos for a year continuously. We may never actually know how smart any given generation is.
8. The cleanest framing of the long-running thesis: imagine Albert Einstein, no need to eat, sleep, or age, thinking about one problem in fundamental physics for one straight year. That's the case for spending $1.5 trillion a year on compute.
9. The frontier captures ~90% of AI revenue. Open source captures ~80% of tokens. Both are true. The market priced one and missed the other. The bear case last year was that cheap open-source tokens would close the gap. Six months in, the frontier is extending its lead instead.
10. Capex for the hyperscalers is moving from $1.1T to ~$1.5T by 2027 (Morgan Stanley revised up). Inference revenue is projected at $300B by 2027 with 60-70% gross margins and roughly 35% of capex going to non-revenue training runs. Brad thinks $300B is low and we end this year at $200B+. Meanwhile Nvidia has not been losing share to ASICs. Once you adjust for Anthropic on TPUs, Nvidia has held or expanded against Broadcom, AMD, Cerebras, MTIA and OpenAI's Jalapeno. Tokens-per-watt is revenue-per-watt in a power-constrained world.
Bonus: The MAG 7 added $1 trillion of revenue over the last 7 years and that produced $17 trillion of market cap. The forecast: SpaceX, Anthropic and OpenAI add the next $1 trillion in revenue in 4-5 years. Three companies, half the time.
Bezos built AWS because he had to overbuild capacity for Black Friday and then figured out how to monetise the idle compute the other 364 days. Elon is running the same playbook in reverse. He built compute to train Grok, then turned the surplus into the fastest-growing cloud business in history.
10 more takeways from the pod here:
https://t.co/Zy9VBNNMoC
The $SPCX IPO prices todays at $1.77 trillion. Brad Gerstner (@altcap) and Gavin Baker (@GavinSBaker) just spent 80 minutes on @BG2Pod making the case for why that's the cheapest it's going to get.
Here are the 10 takeaways worth saving:
1. In 30 days, SpaceX added $29 billion in AI compute revenue from the Anthropic and Google deals and jumped from not being an AI hyperscaler at all to being the #4 hyperscaler globally, passing Oracle. The trailing multiple compressed from ~100x to 39x in the same window. CoreWeave, Nebius, Iron and the 50 other neoclouds VCs are funding in Silicon Valley are now competing for a smaller slice of what's left.
2. xAI's Google deal generates more operating profit per gigawatt than Anthropic, Meta, Google or OpenAI's own infrastructure. Freda at Altimeter calculated a 55% IRR on Colossus 1. Borrow at 7%, invest at 55%, the math maths. Jensen called the build itself an "N of 1": 100,000 GPUs is normally a 3-year planning cycle plus 1-year deployment. xAI did it in 19 days. Speed is literally cost.
3. The premium Google is paying SpaceX for terrestrial compute is partly a call option on orbital. They want first-in-line when space data centres go live, so they're overpaying today to lock in the relationship.
4. Orbital compute costs about $5 billion per gigawatt of capex to put GPUs in space, vs $20-25 billion per gigawatt for the non-silicon half on the ground (land, power, cooling, switchgear). Space, power and cooling are effectively free up there. The unlock is rapid two-stage reusability of Starship, which takes launch from $1,500 per kg on Falcon to $250 per kg and eventually asymptotes to the cost of fuel. Elon says 3 years, Bezos says 6, truth is probably 4-5.
5. Starlink is at less than 1% global household penetration. The forecast going from $10bn to $50bn in connectivity revenue by 2028 is still only 0.3% of the global telecom market. TAM is not the constraint.
6. The most underrated piece of the IPO is the Cursor acquisition. Cursor and Anthropic each hold more proprietary coding tokens than exist on the public internet combined. xAI bought 700-800 people and a frontier-quality coding dataset, dropped it into Colossus 2 for 3 weeks of training, and Composer 2.5 went pareto-dominant on coding 12 days ago. Grok 4.3, a 1.5 trillion parameter model, is also currently on the pareto frontier. There are now four frontier labs, not three: xAI alongside Google (Gemini 3.1 Pro), Anthropic and OpenAI.
7. Anthropic just shipped Fable 5 (Mythos with safety classifiers). Karpathy says it's SOTA on benchmarks but the real unlock is long-running tasks. Stripe refactored a 50 million line Ruby codebase in a day. Used to take many weeks with many engineers. Noam Brown's corollary: snapshot benchmarks are dead. The x-axis now has to be time, tokens or compute, because frontier models can solve most problems if you let them run long enough. Nobody has ever run Mythos for a year continuously. We may never actually know how smart any given generation is.
8. The cleanest framing of the long-running thesis: imagine Albert Einstein, no need to eat, sleep, or age, thinking about one problem in fundamental physics for one straight year. That's the case for spending $1.5 trillion a year on compute.
9. The frontier captures ~90% of AI revenue. Open source captures ~80% of tokens. Both are true. The market priced one and missed the other. The bear case last year was that cheap open-source tokens would close the gap. Six months in, the frontier is extending its lead instead.
10. Capex for the hyperscalers is moving from $1.1T to ~$1.5T by 2027 (Morgan Stanley revised up). Inference revenue is projected at $300B by 2027 with 60-70% gross margins and roughly 35% of capex going to non-revenue training runs. Brad thinks $300B is low and we end this year at $200B+. Meanwhile Nvidia has not been losing share to ASICs. Once you adjust for Anthropic on TPUs, Nvidia has held or expanded against Broadcom, AMD, Cerebras, MTIA and OpenAI's Jalapeno. Tokens-per-watt is revenue-per-watt in a power-constrained world.
Bonus: The MAG 7 added $1 trillion of revenue over the last 7 years and that produced $17 trillion of market cap. The forecast: SpaceX, Anthropic and OpenAI add the next $1 trillion in revenue in 4-5 years. Three companies, half the time.
Nvidia agreed in September to put up to $100bn into OpenAI, which will turn around and spend much of it on Nvidia chips. The structure has a name and a history: vendor financing, the same play that turned the late-1990s telecom boom into a 95% collapse. We have run this experiment before, and it didn't end great.
Vendor financing is simple. A supplier funds its own customer so the customer can buy the supplier's product. Nvidia releases its money to OpenAI as each of 10 gigawatts of data centres gets built, and by early December its own CFO admitted the deal still wasn't a definitive agreement, months after it had helped fuel a market rally. The cash makes a loop.
The loop is everywhere now. Nvidia bought $2bn of CoreWeave's stock. Jane Street committed about $6bn of compute and another $1bn of equity. Meta signed roughly $21bn of CoreWeave capacity through 2032. And the buildout is increasingly funded by term loans with the GPUs themselves pledged as collateral, $8.5bn and $3.1bn facilities closed this year alone.
In the late 90s, Lucent Technologies spun out of AT&T with Bell Labs behind it and ran this exact play. To win orders from new carriers with no cash, it lent them the money to buy Lucent equipment. The loans landed on the income statement as revenue while the shaky debt sat on the balance sheet as an asset. Lucent committed $7bn of customer financing. One borrower, WinStar, drew a $2bn commitment, went bankrupt, and cost Lucent a $700m write-off. Lucent lost $16.2bn in 2001 and $11.8bn in 2002. The stock fell from around $84 to near zero, about $258bn of market value gone, roughly 95%. Nortel ran the same playbook into the same wall.
Why does it end the same way every time? Hyman Minsky gave the mechanism: stable booms breed instability because borrowers climb a ladder. On the first rung, hedge finance, your cash flow covers interest and principal. On the second, speculative finance, it covers the interest and you roll the principal over. On the third, Ponzi finance, it covers neither, and you survive only on the asset rising or on someone handing you fresh money. Vendor financing is the engine that walks a whole sector up that ladder, because the supplier has every reason to keep lending to keep booking the sale.
Look at where the buildout sits on that ladder. The four hyperscalers will spend about $700bn on capex in 2026: Alphabet $185bn, Amazon $200bn, Meta $135bn, Microsoft $190bn. Their combined free cash flow is falling to its lowest since 2014. Amazon's is going negative, somewhere between minus $17bn and minus $28bn on current estimates, after dropping 95% over the past year to $1.2bn. The most profitable companies in history are turning their profits into concrete faster than the revenue to justify it is arriving. That is the speculative rung, by definition.
The strongest objection is that the analogy breaks on quality. Lucent lent to fragile startups with no revenue. Today the buyers are Microsoft, Amazon and Google, the most profitable companies on earth, and they can fund this from operations.
But the fragile borrower was never the hyperscaler. It sits one layer down. The hyperscalers are profitable, yet their cash flow is going negative under the weight of the capex. The Ponzi risk lives in the layer they fund; the model labs committing to $100bn plus of chips against revenue that doesn't exist yet, and the neoclouds like CoreWeave borrowing against the chips as collateral. Dario Amodei said a gigawatt runs about $10bn over five years, the vendor fronts it, the lab repays as revenue grows, and that is reasonable, right up until you stack the deals to where you need $200bn a year by 2027 or 2028, at which point you can overextend. The current commitments need precisely that. The fragility has just moved down the stack.
So there is a clean test for which way this goes. This is hedge finance, and the bears are wrong, if the end demand revenue shows up to retire the advances. Watch whether OpenAI, Anthropic and xAI revenue is tracking toward that collective $200bn a year bar by 2027-28. If it is, the loop closes and the buildout was real. If the gap widens while the financing keeps flowing, coverage ratios thinning and lenders sweetening terms, which Morgan Stanley is already flagging, that is the climb up Minsky's ladder. And the trigger will be upstream of any stock crash; the first vendor that declines to fund the next gigawatt.
AI demand is real. The open question is whether the revenue arrives fast enough to retire the financing before the financing stops.
The shape of the unwind matters too. @bgurley reference point is the dot com nuclear winter; three to four years before Amazon climbed out while most of its peers died. Circular financing pulls the boom higher and stretches it longer, so the correction it sets up arrives deeper and later, not gentler.
The names that walk out the other end will be the ones whose revenue was real before the financing paused, not the ones whose revenue was just another node in the loop.
Ask yourself, is this revenue or is it the same dollar coming back around?
bill gurley (@bgurley) just spent an hour with shane parrish (@shaneparrish) talking AI, markets and how venture really works.
here are my 10 key takeaways:
1. china's biggest AI edge might be structural rather than technical. they've open-sourced around 10 models, so their labs train, test and learn off each other. picture 2 farming villages, one where farmers just trade goods and go home, the other where everyone shares every technique they find. the second compounds knowledge far faster, and that's the system china has built around AI.
2. plenty of US startups are already building on those chinese open-source models, even as washington and silicon valley treat china as the enemy. it rarely makes headlines. some western labs are also lobbying for heavy AI regulation, because expensive, complex rules are a moat that walls off the cheap open-source rivals they can't beat on price.
3. those 'circular' AI deals get written off as fake demand, a cloud provider hands anthropic billions, anthropic spends it straight back on that same cloud. anthropic's ceo dario amodei argued at dealbook the money would never have been spent otherwise, so it's real, and it inflates everyone's growth. that cuts both ways, it makes a correction more likely as valuations detach from reality, but pushes it further out, since the money keeps flowing until it suddenly stops.
4. venture has grown far more comfortable with risk now that everyone believes in power laws, the idea that one giant winner pays for all the losers. amazon lost roughly $2-3bn before turning cash-flow positive, uber around $15bn, today's AI leaders will blow past both. the danger of burning $100m+ a month is that the spending hides whether the business actually works. grow that fast and the cash floods the numbers, so you can't tell if each customer is profitable or you're just renting revenue that vanishes the moment you stop paying.
5. stablecoins are a real threat to the 2-3% visa and mastercard skim off every card payment. those two run ~60% operating margins as a duopoly the banks built and still profit from, so the industry won't change it willingly. most of the developed world has instant, near-free bank transfers (the UK for 20 years), the US doesn't, because the banks lobbied to block it. gurley's bet is stablecoins route around the whole system before washington ever fixes payments.
6. gurley calls the traditional IPO a 'greedy power grab'. bankers set the price and pick which clients get shares, so the day-one pop lands with their favoured clients, not the company, which just sold itself too cheap. a finance student would run an open auction, matching buyers and sellers anonymously at a clearing price, basically a crypto ICO. wall street had this with direct listings and pushed the market back, because the old way pays them far more.
7. proxy advisors like ISS, the firms that tell big institutional shareholders how to vote, are 'more of a heist'. they grade your governance with a scoring system they won't reveal, then sell you the consulting to raise the score, paid by both sides. their incentives also drift from shareholders, they reflexively oppose deals like elon's tesla package even though it paid him nothing unless the stock rose many times over, about as aligned with shareholders as pay gets. gurley would happily run that structure at any company he's backed.
8. an underrated edge in any field is properly knowing its history. john lasseter once served a 10-course dinner with each course tied to a classic cartoon that shaped animation. magnus carlsen won the chess-history trivia on a break at a world championship. picasso was a master realist by 14, long before the abstract work. people at the top of a craft know exactly where it came from, and that depth is rare enough to set you apart in an interview or a pitch.
9. we ask AI for far less than it can do. instead of getting the top 10 of something then sorting it by hand, push the whole job into one prompt, like 'top 10, list pros and cons, rank by X, then again by Y, and total it up'. gurley does the same for restaurants in gemini, which sits on all the google reviews, rather than 'is this place good' he asks 'what 3 dishes do people rave about, and what do they warn against'.
10. benchmark, the venture firm behind early bets on ebay and uber, runs on 5 fully equal partners, no managing partner, no founder taking the biggest slice. that makes it easy to recruit top investors out of hierarchical firms, kills the annual fight over pay, and pushes senior partners to develop the juniors, since a junior's win counts as much as their own. the catch is that with no boss, no one owns new initiatives, which is why benchmark's site has been a single splash page for about 15 years.
My conversation with @bgurley on thinking, making decisions, and the future
0:00 Systems Thinking & Mental Models
05:21 The Power of Knowing Industry Bedrock
08:50 Traits in Founders
11:44 Surprising AI Use
13:13 The Future of AI
23:04 Is Tesla Self-driving THAT good?
24:15 Non-Consensus Opinions
24:53 The AI Buildout (Bubble?)
29:40 The Role of Retail Investors
34:26 Stablecoin
39:55 AI and Debt Analysis
45:05 Storytelling as a Superpower
50:12 Lessons from Uber
52:10 Inside the Benchmark Structure
59:42 Success
Listen and Learn!
(Includes paid partnerships.)
Nvidia agreed in September to put up to $100bn into OpenAI, which will turn around and spend much of it on Nvidia chips. The structure has a name and a history: vendor financing, the same play that turned the late-1990s telecom boom into a 95% collapse. We have run this experiment before, and it didn't end great.
Vendor financing is simple. A supplier funds its own customer so the customer can buy the supplier's product. Nvidia releases its money to OpenAI as each of 10 gigawatts of data centres gets built, and by early December its own CFO admitted the deal still wasn't a definitive agreement, months after it had helped fuel a market rally. The cash makes a loop.
The loop is everywhere now. Nvidia bought $2bn of CoreWeave's stock. Jane Street committed about $6bn of compute and another $1bn of equity. Meta signed roughly $21bn of CoreWeave capacity through 2032. And the buildout is increasingly funded by term loans with the GPUs themselves pledged as collateral, $8.5bn and $3.1bn facilities closed this year alone.
In the late 90s, Lucent Technologies spun out of AT&T with Bell Labs behind it and ran this exact play. To win orders from new carriers with no cash, it lent them the money to buy Lucent equipment. The loans landed on the income statement as revenue while the shaky debt sat on the balance sheet as an asset. Lucent committed $7bn of customer financing. One borrower, WinStar, drew a $2bn commitment, went bankrupt, and cost Lucent a $700m write-off. Lucent lost $16.2bn in 2001 and $11.8bn in 2002. The stock fell from around $84 to near zero, about $258bn of market value gone, roughly 95%. Nortel ran the same playbook into the same wall.
Why does it end the same way every time? Hyman Minsky gave the mechanism: stable booms breed instability because borrowers climb a ladder. On the first rung, hedge finance, your cash flow covers interest and principal. On the second, speculative finance, it covers the interest and you roll the principal over. On the third, Ponzi finance, it covers neither, and you survive only on the asset rising or on someone handing you fresh money. Vendor financing is the engine that walks a whole sector up that ladder, because the supplier has every reason to keep lending to keep booking the sale.
Look at where the buildout sits on that ladder. The four hyperscalers will spend about $700bn on capex in 2026: Alphabet $185bn, Amazon $200bn, Meta $135bn, Microsoft $190bn. Their combined free cash flow is falling to its lowest since 2014. Amazon's is going negative, somewhere between minus $17bn and minus $28bn on current estimates, after dropping 95% over the past year to $1.2bn. The most profitable companies in history are turning their profits into concrete faster than the revenue to justify it is arriving. That is the speculative rung, by definition.
The strongest objection is that the analogy breaks on quality. Lucent lent to fragile startups with no revenue. Today the buyers are Microsoft, Amazon and Google, the most profitable companies on earth, and they can fund this from operations.
But the fragile borrower was never the hyperscaler. It sits one layer down. The hyperscalers are profitable, yet their cash flow is going negative under the weight of the capex. The Ponzi risk lives in the layer they fund; the model labs committing to $100bn plus of chips against revenue that doesn't exist yet, and the neoclouds like CoreWeave borrowing against the chips as collateral. Dario Amodei said a gigawatt runs about $10bn over five years, the vendor fronts it, the lab repays as revenue grows, and that is reasonable, right up until you stack the deals to where you need $200bn a year by 2027 or 2028, at which point you can overextend. The current commitments need precisely that. The fragility has just moved down the stack.
So there is a clean test for which way this goes. This is hedge finance, and the bears are wrong, if the end demand revenue shows up to retire the advances. Watch whether OpenAI, Anthropic and xAI revenue is tracking toward that collective $200bn a year bar by 2027-28. If it is, the loop closes and the buildout was real. If the gap widens while the financing keeps flowing, coverage ratios thinning and lenders sweetening terms, which Morgan Stanley is already flagging, that is the climb up Minsky's ladder. And the trigger will be upstream of any stock crash; the first vendor that declines to fund the next gigawatt.
AI demand is real. The open question is whether the revenue arrives fast enough to retire the financing before the financing stops.
The shape of the unwind matters too. @bgurley reference point is the dot com nuclear winter; three to four years before Amazon climbed out while most of its peers died. Circular financing pulls the boom higher and stretches it longer, so the correction it sets up arrives deeper and later, not gentler.
The names that walk out the other end will be the ones whose revenue was real before the financing paused, not the ones whose revenue was just another node in the loop.
Ask yourself, is this revenue or is it the same dollar coming back around?
bill gurley (@bgurley) just spent an hour with shane parrish (@shaneparrish) talking AI, markets and how venture really works.
here are my 10 key takeaways:
1. china's biggest AI edge might be structural rather than technical. they've open-sourced around 10 models, so their labs train, test and learn off each other. picture 2 farming villages, one where farmers just trade goods and go home, the other where everyone shares every technique they find. the second compounds knowledge far faster, and that's the system china has built around AI.
2. plenty of US startups are already building on those chinese open-source models, even as washington and silicon valley treat china as the enemy. it rarely makes headlines. some western labs are also lobbying for heavy AI regulation, because expensive, complex rules are a moat that walls off the cheap open-source rivals they can't beat on price.
3. those 'circular' AI deals get written off as fake demand, a cloud provider hands anthropic billions, anthropic spends it straight back on that same cloud. anthropic's ceo dario amodei argued at dealbook the money would never have been spent otherwise, so it's real, and it inflates everyone's growth. that cuts both ways, it makes a correction more likely as valuations detach from reality, but pushes it further out, since the money keeps flowing until it suddenly stops.
4. venture has grown far more comfortable with risk now that everyone believes in power laws, the idea that one giant winner pays for all the losers. amazon lost roughly $2-3bn before turning cash-flow positive, uber around $15bn, today's AI leaders will blow past both. the danger of burning $100m+ a month is that the spending hides whether the business actually works. grow that fast and the cash floods the numbers, so you can't tell if each customer is profitable or you're just renting revenue that vanishes the moment you stop paying.
5. stablecoins are a real threat to the 2-3% visa and mastercard skim off every card payment. those two run ~60% operating margins as a duopoly the banks built and still profit from, so the industry won't change it willingly. most of the developed world has instant, near-free bank transfers (the UK for 20 years), the US doesn't, because the banks lobbied to block it. gurley's bet is stablecoins route around the whole system before washington ever fixes payments.
6. gurley calls the traditional IPO a 'greedy power grab'. bankers set the price and pick which clients get shares, so the day-one pop lands with their favoured clients, not the company, which just sold itself too cheap. a finance student would run an open auction, matching buyers and sellers anonymously at a clearing price, basically a crypto ICO. wall street had this with direct listings and pushed the market back, because the old way pays them far more.
7. proxy advisors like ISS, the firms that tell big institutional shareholders how to vote, are 'more of a heist'. they grade your governance with a scoring system they won't reveal, then sell you the consulting to raise the score, paid by both sides. their incentives also drift from shareholders, they reflexively oppose deals like elon's tesla package even though it paid him nothing unless the stock rose many times over, about as aligned with shareholders as pay gets. gurley would happily run that structure at any company he's backed.
8. an underrated edge in any field is properly knowing its history. john lasseter once served a 10-course dinner with each course tied to a classic cartoon that shaped animation. magnus carlsen won the chess-history trivia on a break at a world championship. picasso was a master realist by 14, long before the abstract work. people at the top of a craft know exactly where it came from, and that depth is rare enough to set you apart in an interview or a pitch.
9. we ask AI for far less than it can do. instead of getting the top 10 of something then sorting it by hand, push the whole job into one prompt, like 'top 10, list pros and cons, rank by X, then again by Y, and total it up'. gurley does the same for restaurants in gemini, which sits on all the google reviews, rather than 'is this place good' he asks 'what 3 dishes do people rave about, and what do they warn against'.
10. benchmark, the venture firm behind early bets on ebay and uber, runs on 5 fully equal partners, no managing partner, no founder taking the biggest slice. that makes it easy to recruit top investors out of hierarchical firms, kills the annual fight over pay, and pushes senior partners to develop the juniors, since a junior's win counts as much as their own. the catch is that with no boss, no one owns new initiatives, which is why benchmark's site has been a single splash page for about 15 years.
bill gurley (@bgurley) just spent an hour with shane parrish (@shaneparrish) talking AI, markets and how venture really works.
here are my 10 key takeaways:
1. china's biggest AI edge might be structural rather than technical. they've open-sourced around 10 models, so their labs train, test and learn off each other. picture 2 farming villages, one where farmers just trade goods and go home, the other where everyone shares every technique they find. the second compounds knowledge far faster, and that's the system china has built around AI.
2. plenty of US startups are already building on those chinese open-source models, even as washington and silicon valley treat china as the enemy. it rarely makes headlines. some western labs are also lobbying for heavy AI regulation, because expensive, complex rules are a moat that walls off the cheap open-source rivals they can't beat on price.
3. those 'circular' AI deals get written off as fake demand, a cloud provider hands anthropic billions, anthropic spends it straight back on that same cloud. anthropic's ceo dario amodei argued at dealbook the money would never have been spent otherwise, so it's real, and it inflates everyone's growth. that cuts both ways, it makes a correction more likely as valuations detach from reality, but pushes it further out, since the money keeps flowing until it suddenly stops.
4. venture has grown far more comfortable with risk now that everyone believes in power laws, the idea that one giant winner pays for all the losers. amazon lost roughly $2-3bn before turning cash-flow positive, uber around $15bn, today's AI leaders will blow past both. the danger of burning $100m+ a month is that the spending hides whether the business actually works. grow that fast and the cash floods the numbers, so you can't tell if each customer is profitable or you're just renting revenue that vanishes the moment you stop paying.
5. stablecoins are a real threat to the 2-3% visa and mastercard skim off every card payment. those two run ~60% operating margins as a duopoly the banks built and still profit from, so the industry won't change it willingly. most of the developed world has instant, near-free bank transfers (the UK for 20 years), the US doesn't, because the banks lobbied to block it. gurley's bet is stablecoins route around the whole system before washington ever fixes payments.
6. gurley calls the traditional IPO a 'greedy power grab'. bankers set the price and pick which clients get shares, so the day-one pop lands with their favoured clients, not the company, which just sold itself too cheap. a finance student would run an open auction, matching buyers and sellers anonymously at a clearing price, basically a crypto ICO. wall street had this with direct listings and pushed the market back, because the old way pays them far more.
7. proxy advisors like ISS, the firms that tell big institutional shareholders how to vote, are 'more of a heist'. they grade your governance with a scoring system they won't reveal, then sell you the consulting to raise the score, paid by both sides. their incentives also drift from shareholders, they reflexively oppose deals like elon's tesla package even though it paid him nothing unless the stock rose many times over, about as aligned with shareholders as pay gets. gurley would happily run that structure at any company he's backed.
8. an underrated edge in any field is properly knowing its history. john lasseter once served a 10-course dinner with each course tied to a classic cartoon that shaped animation. magnus carlsen won the chess-history trivia on a break at a world championship. picasso was a master realist by 14, long before the abstract work. people at the top of a craft know exactly where it came from, and that depth is rare enough to set you apart in an interview or a pitch.
9. we ask AI for far less than it can do. instead of getting the top 10 of something then sorting it by hand, push the whole job into one prompt, like 'top 10, list pros and cons, rank by X, then again by Y, and total it up'. gurley does the same for restaurants in gemini, which sits on all the google reviews, rather than 'is this place good' he asks 'what 3 dishes do people rave about, and what do they warn against'.
10. benchmark, the venture firm behind early bets on ebay and uber, runs on 5 fully equal partners, no managing partner, no founder taking the biggest slice. that makes it easy to recruit top investors out of hierarchical firms, kills the annual fight over pay, and pushes senior partners to develop the juniors, since a junior's win counts as much as their own. the catch is that with no boss, no one owns new initiatives, which is why benchmark's site has been a single splash page for about 15 years.
bill gurley (@bgurley) just spent an hour with shane parrish (@shaneparrish) talking AI, markets and how venture really works.
here are my 10 key takeaways:
1. china's biggest AI edge might be structural rather than technical. they've open-sourced around 10 models, so their labs train, test and learn off each other. picture 2 farming villages, one where farmers just trade goods and go home, the other where everyone shares every technique they find. the second compounds knowledge far faster, and that's the system china has built around AI.
2. plenty of US startups are already building on those chinese open-source models, even as washington and silicon valley treat china as the enemy. it rarely makes headlines. some western labs are also lobbying for heavy AI regulation, because expensive, complex rules are a moat that walls off the cheap open-source rivals they can't beat on price.
3. those 'circular' AI deals get written off as fake demand, a cloud provider hands anthropic billions, anthropic spends it straight back on that same cloud. anthropic's ceo dario amodei argued at dealbook the money would never have been spent otherwise, so it's real, and it inflates everyone's growth. that cuts both ways, it makes a correction more likely as valuations detach from reality, but pushes it further out, since the money keeps flowing until it suddenly stops.
4. venture has grown far more comfortable with risk now that everyone believes in power laws, the idea that one giant winner pays for all the losers. amazon lost roughly $2-3bn before turning cash-flow positive, uber around $15bn, today's AI leaders will blow past both. the danger of burning $100m+ a month is that the spending hides whether the business actually works. grow that fast and the cash floods the numbers, so you can't tell if each customer is profitable or you're just renting revenue that vanishes the moment you stop paying.
5. stablecoins are a real threat to the 2-3% visa and mastercard skim off every card payment. those two run ~60% operating margins as a duopoly the banks built and still profit from, so the industry won't change it willingly. most of the developed world has instant, near-free bank transfers (the UK for 20 years), the US doesn't, because the banks lobbied to block it. gurley's bet is stablecoins route around the whole system before washington ever fixes payments.
6. gurley calls the traditional IPO a 'greedy power grab'. bankers set the price and pick which clients get shares, so the day-one pop lands with their favoured clients, not the company, which just sold itself too cheap. a finance student would run an open auction, matching buyers and sellers anonymously at a clearing price, basically a crypto ICO. wall street had this with direct listings and pushed the market back, because the old way pays them far more.
7. proxy advisors like ISS, the firms that tell big institutional shareholders how to vote, are 'more of a heist'. they grade your governance with a scoring system they won't reveal, then sell you the consulting to raise the score, paid by both sides. their incentives also drift from shareholders, they reflexively oppose deals like elon's tesla package even though it paid him nothing unless the stock rose many times over, about as aligned with shareholders as pay gets. gurley would happily run that structure at any company he's backed.
8. an underrated edge in any field is properly knowing its history. john lasseter once served a 10-course dinner with each course tied to a classic cartoon that shaped animation. magnus carlsen won the chess-history trivia on a break at a world championship. picasso was a master realist by 14, long before the abstract work. people at the top of a craft know exactly where it came from, and that depth is rare enough to set you apart in an interview or a pitch.
9. we ask AI for far less than it can do. instead of getting the top 10 of something then sorting it by hand, push the whole job into one prompt, like 'top 10, list pros and cons, rank by X, then again by Y, and total it up'. gurley does the same for restaurants in gemini, which sits on all the google reviews, rather than 'is this place good' he asks 'what 3 dishes do people rave about, and what do they warn against'.
10. benchmark, the venture firm behind early bets on ebay and uber, runs on 5 fully equal partners, no managing partner, no founder taking the biggest slice. that makes it easy to recruit top investors out of hierarchical firms, kills the annual fight over pay, and pushes senior partners to develop the juniors, since a junior's win counts as much as their own. the catch is that with no boss, no one owns new initiatives, which is why benchmark's site has been a single splash page for about 15 years.
AI is trained to produce the average. That means the average is about to become free. And once the average is free, every edge a person or a company has, in business, in product, in their own work, moves out to the tail.
The deviation is where the value now lives. This argument dates back to 1948, where mathematician, Claude Shannon, defined the information contained in an event as its surprisal, which is simply the negative log of its probability.
Shannon concluded a near certain event carries almost no information. A rare event carries a great deal.
It follows that the least informative point in any distribution is its centre. The average is, by construction, the most predictable and so the most boring thing a person can know.
That has always been true. It just never mattered much, because producing the average used to be expensive. A competent average essay, an average deck, an average analysis, an average driver each took real work, so the average was scarce enough to carry a price.
AI now removes that cost. Models are trained to predict the most likely next token, which means they regress to the mean by design. Their output variance sits below the real world's.
Mode collapse and Reinforcement Learning from Human Feedback (RLFH) narrow it further, with rare words and unusual constructions among the first things to disappear. A large language model is, in effect, a factory for the statistical centre of everything it read.
The argument runs in five short steps.
1. Information lives in the surprise, and the average carries almost none of it.
2. The average used to be scarce, which is why it was valuable.
3. AI is about to make the average infinite and free.
4. When something becomes free, the value moves to whatever is still scarce alongside it.
5. The thing still scarce is the tail.
In this clip, Uber CEO @dkhos covers average vs edge.
His mentor Barry Diller would refuse the summarised version of anything and insist on going to the source, the raw model, the analyst who actually built it. The reasoning is that everyone has the same reaction to the average, so the edge is always in the unusual 20%, the P95 case, the part a summary strips out. Going to the source is a way of catching the tail before someone averages it away.
This principle is what makes Uber hard to run. A P95 bug, the kind of failure that hits 1 time in 20, reaches a casual rider perhaps once a month. The same bug reaches a driver who is in the app 8 hours a day every single week. The average experience is fine. The average is always fine. People judge the service on its tail.
The same thing happens on the upside. Dara talks about how quickly magic turns to normal. You push a button, a car appears, and by the next day you're annoyed it took 6 and a half minutes.
We adjust to whatever becomes the new baseline and stop noticing it. Psychologists Brickman and Campbell named this hedonic adaptation in 1971. A company's average service simply disappears from view.
Customers barely register the mean. What they feel is the tail, the delight at the top end and the failure at the bottom. The worst 5% is what people remember, and the average goes unnoticed.
Even the troublemakers fit the same shape. Evolution acts on the rare outliers in a population, the mutants, and that is where every meaningful change starts. Antibiotic resistance is the clearest case. Roughly 1 bacterium in a billion carries a mutation that survives the drug, and within days it has repopulated the entire colony.
It is the same law each time. Information, money, attention and feeling all collect in the tail and drain away from the mean.
There is a serious objection here. A large part of the economy runs on killing the tail. McDonald's sells a burger that tastes identical every time. Six Sigma is a 40 year management discipline devoted entirely to stamping out variance. Reliability is sameness, so surely the value sits in the average.
That is true, and it reveals an important distinction. There are two tails, and they pull in opposite directions. A company kills its downside tail, the defect, the P95 bug, the late delivery, through obsessive operational sameness. It competes and wins on its upside tail, the surprise that no rival is delivering.
The mean in the middle is table stakes, and table stakes are exactly what AI now hands out for nothing. Uber is fanatical about killing its left tail and grows on its right. The flat average in between earns almost nothing.
There is a condition under which this is wrong. If AI's average output stays clearly worse than a competent human's for long enough, then the average stays scarce, keeps its price, and the argument softens. The thing to watch is the gap between median human and median model work. It has been closing fast, and in many domains the model is already past the human median.
So the move is easy to state and hard to do. Stop competing on the average, because the average has just been nationalised. The average post, the average deck, the average model, the average memo are all sliding towards zero.
A simple test works on anything you make. Would a model trained on the internet have produced roughly this? If the answer is yes, then it's the average, and the average is now worth nothing.
The value is the distance between what was made and what the model would have guessed. That distance has a name. It is surprisal, and surprisal is the whole game.
The mean is only ever a story about a crowd. Every person, and every thing worth building, is a single point out in the tail. That is where to aim.
There's a speed limit on how fast a company can change. Push past it, and change stops compounding and starts dissolving the very thing that made the company work.
Biologists have a name for that failure mode. They call it error catastrophe.
AI has pushed every company's rate of change straight up against that limit, which is why it's worth understanding what the limit actually is.
On a recent podcast, Uber's Dara Khosrowshahi described his own company in biological terms. Companies are organisms, and organisms evolve by mutating. The ones that sit still, running a single process with a single information flow, are the ones that die.
So he goes looking for "troublemakers," the people whose deviation from the norm might be the useful mutation. And by his own estimate, AI has sped up the rate of change inside companies by 5x.
So what happens to a system when you raise its mutation rate that fast? Biology has studied this for decades, and the answer is surprising.
In 1971, the chemist Manfred Eigen worked out the mathematics of how replicating things evolve. A healthy population sits around a "master sequence," which is the working blueprint, surrounded by a cloud of slightly different mutants. Holding it together is a critical mutation rate called the error threshold.
Below the threshold, the master sequence stays intact and good mutations slowly accumulate. That is ordinary, healthy evolution.
Above the threshold, the system tips into error catastrophe. Copies start mutating faster than the blueprint can be passed on accurately, the population smears out across every possible variant, and the master sequence is lost.
We have learned to use this deliberately, and it is how some antivirals work. Molnupiravir, one of the COVID drugs, is built on this exact idea. It raises the virus's own mutation rate just past its error threshold, and the virus copies itself into nonsense.
Favipiravir and ribavirin do something similar. Past that point, the very force that powers evolution is what kills the virus.
A company behaves the same way, with its own version of each piece.
The master sequence is the core that actually works. It's the dispatch logic that 10bn-plus trips a year depend on, the payment rails, the safety process, and the unwritten "this is how we do it" that lives in people's heads.
Mutation is every troublemaker, every rewritten workflow, every process an agent redesigns from scratch. As in biology, most of them are neutral or harmful, and a few are gold.
Below the threshold, the good mutations stick and compound. This is the version Dara is describing, and he is right to want it.
Above the threshold, the company can't copy its own core faster than it's changing it. 6 teams end up running 6 versions of the same process, and no one can say which is canonical. That is error catastrophe inside an organisation.
AI is the mutagen that raises the rate. 5x, by Dara's own number, with developers shipping 10x the code commits, agents rewriting internal tooling, and every team told to rebuild from first principles at once.
This reverses the usual job of a leader. For 50 years, the hard part was creating change at all, because change was slow, expensive, and fought at every level.
Now change is cheap and fast. The rare skill is the opposite one, knowing when to slow it down and hold the company below its error threshold while everyone around is pushing to go faster.
There's a fair objection to all of this. Companies are not viruses. A firm can hold its core completely stable while it mutates only the edges, so there may be no single threshold to fall off. An ad can be rewritten daily, while the system that handles people's payments should barely change at all.
That objection is right. It also shows exactly how to manage the risk. Biology already runs different mutation rates in different places. The genome protects its core genes with proofreading and very low error rates, while the immune system's hypervariable regions, the ones that have to keep pace with new pathogens, are allowed to run hot.
Dara talks about hunting for troublemakers. You have to decide, in advance, which parts of the company are allowed to change fast and which have to stay almost frozen. Safety, payments, and the core dispatch system have to stay near zero mutation. Internal tooling, marketing copy, and experiments can run hot. The troublemakers belong in the second group, and far away from the first.
Bacteria even solved the scheduling problem. They don't keep a constant mutation rate. Under stress, such as starvation or antibiotics, they trigger the SOS response, switch on error-prone enzymes, and push their mutation rate up.
Once they have adapted, they bring it back down. The rate is a dial they adjust to conditions, turned high when cornered and low when settled.
Running it this way has a deeper payoff. When the mutation rate is high, evolution stops rewarding the single fittest peak and starts rewarding the flattest, most mutation-tolerant ground. Biologists call this survival of the flattest. In a world changing at 5x, the most adaptable operator tends to beat the highest-performing but fragile one.
There is a clean test for whether the analogy actually holds. If the firms that pushed AI-driven process change hardest through 2025 and 2026 come out with the best reliability, the best margins, and no rise in incidents, then the threshold doesn't really bind and the biology doesn't transfer. The signal to watch is incident rates, well ahead of adoption rates.
That points to a number almost no team is tracking yet. Most companies are measuring "AI adoption," which is mostly a vanity metric. The number that will matter by 2027 is the internal error rate, meaning how much the core degrades while the edges mutate. You can see it in concrete things. Production incidents, failed payments, safety events per million trips, and how often a team has to roll back something it just shipped.
There is an uncomfortable implication in all of this. The AI era's first casualties may well be its fastest movers. A company that mutates its own core faster than it can recopy it can wake up one quarter unable to say what it is.
Everyone is busy pricing the risk of moving too slowly. Almost no one is pricing the risk of moving too fast and melting the core.
AI is trained to produce the average. That means the average is about to become free. And once the average is free, every edge a person or a company has, in business, in product, in their own work, moves out to the tail.
The deviation is where the value now lives. This argument dates back to 1948, where mathematician, Claude Shannon, defined the information contained in an event as its surprisal, which is simply the negative log of its probability.
Shannon concluded a near certain event carries almost no information. A rare event carries a great deal.
It follows that the least informative point in any distribution is its centre. The average is, by construction, the most predictable and so the most boring thing a person can know.
That has always been true. It just never mattered much, because producing the average used to be expensive. A competent average essay, an average deck, an average analysis, an average driver each took real work, so the average was scarce enough to carry a price.
AI now removes that cost. Models are trained to predict the most likely next token, which means they regress to the mean by design. Their output variance sits below the real world's.
Mode collapse and Reinforcement Learning from Human Feedback (RLFH) narrow it further, with rare words and unusual constructions among the first things to disappear. A large language model is, in effect, a factory for the statistical centre of everything it read.
The argument runs in five short steps.
1. Information lives in the surprise, and the average carries almost none of it.
2. The average used to be scarce, which is why it was valuable.
3. AI is about to make the average infinite and free.
4. When something becomes free, the value moves to whatever is still scarce alongside it.
5. The thing still scarce is the tail.
In this clip, Uber CEO @dkhos covers average vs edge.
His mentor Barry Diller would refuse the summarised version of anything and insist on going to the source, the raw model, the analyst who actually built it. The reasoning is that everyone has the same reaction to the average, so the edge is always in the unusual 20%, the P95 case, the part a summary strips out. Going to the source is a way of catching the tail before someone averages it away.
This principle is what makes Uber hard to run. A P95 bug, the kind of failure that hits 1 time in 20, reaches a casual rider perhaps once a month. The same bug reaches a driver who is in the app 8 hours a day every single week. The average experience is fine. The average is always fine. People judge the service on its tail.
The same thing happens on the upside. Dara talks about how quickly magic turns to normal. You push a button, a car appears, and by the next day you're annoyed it took 6 and a half minutes.
We adjust to whatever becomes the new baseline and stop noticing it. Psychologists Brickman and Campbell named this hedonic adaptation in 1971. A company's average service simply disappears from view.
Customers barely register the mean. What they feel is the tail, the delight at the top end and the failure at the bottom. The worst 5% is what people remember, and the average goes unnoticed.
Even the troublemakers fit the same shape. Evolution acts on the rare outliers in a population, the mutants, and that is where every meaningful change starts. Antibiotic resistance is the clearest case. Roughly 1 bacterium in a billion carries a mutation that survives the drug, and within days it has repopulated the entire colony.
It is the same law each time. Information, money, attention and feeling all collect in the tail and drain away from the mean.
There is a serious objection here. A large part of the economy runs on killing the tail. McDonald's sells a burger that tastes identical every time. Six Sigma is a 40 year management discipline devoted entirely to stamping out variance. Reliability is sameness, so surely the value sits in the average.
That is true, and it reveals an important distinction. There are two tails, and they pull in opposite directions. A company kills its downside tail, the defect, the P95 bug, the late delivery, through obsessive operational sameness. It competes and wins on its upside tail, the surprise that no rival is delivering.
The mean in the middle is table stakes, and table stakes are exactly what AI now hands out for nothing. Uber is fanatical about killing its left tail and grows on its right. The flat average in between earns almost nothing.
There is a condition under which this is wrong. If AI's average output stays clearly worse than a competent human's for long enough, then the average stays scarce, keeps its price, and the argument softens. The thing to watch is the gap between median human and median model work. It has been closing fast, and in many domains the model is already past the human median.
So the move is easy to state and hard to do. Stop competing on the average, because the average has just been nationalised. The average post, the average deck, the average model, the average memo are all sliding towards zero.
A simple test works on anything you make. Would a model trained on the internet have produced roughly this? If the answer is yes, then it's the average, and the average is now worth nothing.
The value is the distance between what was made and what the model would have guessed. That distance has a name. It is surprisal, and surprisal is the whole game.
The mean is only ever a story about a crowd. Every person, and every thing worth building, is a single point out in the tail. That is where to aim.
It's not a coincidence Jamie Dimon, Elon, Jensen, and Barry Diller all share the same management philosophy: getting the truth from the source
Dara on what he learned from Barry Diller in his 20+ years working with him:
"I was an analyst at Allen & Company working on the LBO model for Paramount.
Barry didn't want to talk to the MD or the VP or the associate.
He'd say, 'Who built the model? I'm going to talk to that guy.'
He wanted to hear straight from the source.
It's the filtering that gets the edge out of the situation, and it's often the edge that gives you an edge. It's not the average."
@GavinSBaker in our conversation from 2024 explained how Elon, Jensen, and Jamie work:
"Wherever in the company the problem is, that is who Jensen and Elon go to work with — the subject matter expert, whether they're 23 or 50, there's no hierarchy.
When JPMorgan was buying Bear Stearns, the best modeler in the company was 24 years old.
They set up a desk for him side-by-side with Jamie. Jamie would say, "Change that, change this."
He didn't ask for that guy's boss. He said, 'This guy is the best, I want to work with him.'"
@patrick_oshag@dkhos Everyone is busy pricing the risk of moving too slowly.
Almost no one is pricing the risk of moving too fast.
https://t.co/tyhau2hOQf
There's a speed limit on how fast a company can change. Push past it, and change stops compounding and starts dissolving the very thing that made the company work.
Biologists have a name for that failure mode. They call it error catastrophe.
AI has pushed every company's rate of change straight up against that limit, which is why it's worth understanding what the limit actually is.
On a recent podcast, Uber's Dara Khosrowshahi described his own company in biological terms. Companies are organisms, and organisms evolve by mutating. The ones that sit still, running a single process with a single information flow, are the ones that die.
So he goes looking for "troublemakers," the people whose deviation from the norm might be the useful mutation. And by his own estimate, AI has sped up the rate of change inside companies by 5x.
So what happens to a system when you raise its mutation rate that fast? Biology has studied this for decades, and the answer is surprising.
In 1971, the chemist Manfred Eigen worked out the mathematics of how replicating things evolve. A healthy population sits around a "master sequence," which is the working blueprint, surrounded by a cloud of slightly different mutants. Holding it together is a critical mutation rate called the error threshold.
Below the threshold, the master sequence stays intact and good mutations slowly accumulate. That is ordinary, healthy evolution.
Above the threshold, the system tips into error catastrophe. Copies start mutating faster than the blueprint can be passed on accurately, the population smears out across every possible variant, and the master sequence is lost.
We have learned to use this deliberately, and it is how some antivirals work. Molnupiravir, one of the COVID drugs, is built on this exact idea. It raises the virus's own mutation rate just past its error threshold, and the virus copies itself into nonsense.
Favipiravir and ribavirin do something similar. Past that point, the very force that powers evolution is what kills the virus.
A company behaves the same way, with its own version of each piece.
The master sequence is the core that actually works. It's the dispatch logic that 10bn-plus trips a year depend on, the payment rails, the safety process, and the unwritten "this is how we do it" that lives in people's heads.
Mutation is every troublemaker, every rewritten workflow, every process an agent redesigns from scratch. As in biology, most of them are neutral or harmful, and a few are gold.
Below the threshold, the good mutations stick and compound. This is the version Dara is describing, and he is right to want it.
Above the threshold, the company can't copy its own core faster than it's changing it. 6 teams end up running 6 versions of the same process, and no one can say which is canonical. That is error catastrophe inside an organisation.
AI is the mutagen that raises the rate. 5x, by Dara's own number, with developers shipping 10x the code commits, agents rewriting internal tooling, and every team told to rebuild from first principles at once.
This reverses the usual job of a leader. For 50 years, the hard part was creating change at all, because change was slow, expensive, and fought at every level.
Now change is cheap and fast. The rare skill is the opposite one, knowing when to slow it down and hold the company below its error threshold while everyone around is pushing to go faster.
There's a fair objection to all of this. Companies are not viruses. A firm can hold its core completely stable while it mutates only the edges, so there may be no single threshold to fall off. An ad can be rewritten daily, while the system that handles people's payments should barely change at all.
That objection is right. It also shows exactly how to manage the risk. Biology already runs different mutation rates in different places. The genome protects its core genes with proofreading and very low error rates, while the immune system's hypervariable regions, the ones that have to keep pace with new pathogens, are allowed to run hot.
Dara talks about hunting for troublemakers. You have to decide, in advance, which parts of the company are allowed to change fast and which have to stay almost frozen. Safety, payments, and the core dispatch system have to stay near zero mutation. Internal tooling, marketing copy, and experiments can run hot. The troublemakers belong in the second group, and far away from the first.
Bacteria even solved the scheduling problem. They don't keep a constant mutation rate. Under stress, such as starvation or antibiotics, they trigger the SOS response, switch on error-prone enzymes, and push their mutation rate up.
Once they have adapted, they bring it back down. The rate is a dial they adjust to conditions, turned high when cornered and low when settled.
Running it this way has a deeper payoff. When the mutation rate is high, evolution stops rewarding the single fittest peak and starts rewarding the flattest, most mutation-tolerant ground. Biologists call this survival of the flattest. In a world changing at 5x, the most adaptable operator tends to beat the highest-performing but fragile one.
There is a clean test for whether the analogy actually holds. If the firms that pushed AI-driven process change hardest through 2025 and 2026 come out with the best reliability, the best margins, and no rise in incidents, then the threshold doesn't really bind and the biology doesn't transfer. The signal to watch is incident rates, well ahead of adoption rates.
That points to a number almost no team is tracking yet. Most companies are measuring "AI adoption," which is mostly a vanity metric. The number that will matter by 2027 is the internal error rate, meaning how much the core degrades while the edges mutate. You can see it in concrete things. Production incidents, failed payments, safety events per million trips, and how often a team has to roll back something it just shipped.
There is an uncomfortable implication in all of this. The AI era's first casualties may well be its fastest movers. A company that mutates its own core faster than it can recopy it can wake up one quarter unable to say what it is.
Everyone is busy pricing the risk of moving too slowly. Almost no one is pricing the risk of moving too fast and melting the core.
There's a speed limit on how fast a company can change. Push past it, and change stops compounding and starts dissolving the very thing that made the company work.
Biologists have a name for that failure mode. They call it error catastrophe.
AI has pushed every company's rate of change straight up against that limit, which is why it's worth understanding what the limit actually is.
On a recent podcast, Uber's Dara Khosrowshahi described his own company in biological terms. Companies are organisms, and organisms evolve by mutating. The ones that sit still, running a single process with a single information flow, are the ones that die.
So he goes looking for "troublemakers," the people whose deviation from the norm might be the useful mutation. And by his own estimate, AI has sped up the rate of change inside companies by 5x.
So what happens to a system when you raise its mutation rate that fast? Biology has studied this for decades, and the answer is surprising.
In 1971, the chemist Manfred Eigen worked out the mathematics of how replicating things evolve. A healthy population sits around a "master sequence," which is the working blueprint, surrounded by a cloud of slightly different mutants. Holding it together is a critical mutation rate called the error threshold.
Below the threshold, the master sequence stays intact and good mutations slowly accumulate. That is ordinary, healthy evolution.
Above the threshold, the system tips into error catastrophe. Copies start mutating faster than the blueprint can be passed on accurately, the population smears out across every possible variant, and the master sequence is lost.
We have learned to use this deliberately, and it is how some antivirals work. Molnupiravir, one of the COVID drugs, is built on this exact idea. It raises the virus's own mutation rate just past its error threshold, and the virus copies itself into nonsense.
Favipiravir and ribavirin do something similar. Past that point, the very force that powers evolution is what kills the virus.
A company behaves the same way, with its own version of each piece.
The master sequence is the core that actually works. It's the dispatch logic that 10bn-plus trips a year depend on, the payment rails, the safety process, and the unwritten "this is how we do it" that lives in people's heads.
Mutation is every troublemaker, every rewritten workflow, every process an agent redesigns from scratch. As in biology, most of them are neutral or harmful, and a few are gold.
Below the threshold, the good mutations stick and compound. This is the version Dara is describing, and he is right to want it.
Above the threshold, the company can't copy its own core faster than it's changing it. 6 teams end up running 6 versions of the same process, and no one can say which is canonical. That is error catastrophe inside an organisation.
AI is the mutagen that raises the rate. 5x, by Dara's own number, with developers shipping 10x the code commits, agents rewriting internal tooling, and every team told to rebuild from first principles at once.
This reverses the usual job of a leader. For 50 years, the hard part was creating change at all, because change was slow, expensive, and fought at every level.
Now change is cheap and fast. The rare skill is the opposite one, knowing when to slow it down and hold the company below its error threshold while everyone around is pushing to go faster.
There's a fair objection to all of this. Companies are not viruses. A firm can hold its core completely stable while it mutates only the edges, so there may be no single threshold to fall off. An ad can be rewritten daily, while the system that handles people's payments should barely change at all.
That objection is right. It also shows exactly how to manage the risk. Biology already runs different mutation rates in different places. The genome protects its core genes with proofreading and very low error rates, while the immune system's hypervariable regions, the ones that have to keep pace with new pathogens, are allowed to run hot.
Dara talks about hunting for troublemakers. You have to decide, in advance, which parts of the company are allowed to change fast and which have to stay almost frozen. Safety, payments, and the core dispatch system have to stay near zero mutation. Internal tooling, marketing copy, and experiments can run hot. The troublemakers belong in the second group, and far away from the first.
Bacteria even solved the scheduling problem. They don't keep a constant mutation rate. Under stress, such as starvation or antibiotics, they trigger the SOS response, switch on error-prone enzymes, and push their mutation rate up.
Once they have adapted, they bring it back down. The rate is a dial they adjust to conditions, turned high when cornered and low when settled.
Running it this way has a deeper payoff. When the mutation rate is high, evolution stops rewarding the single fittest peak and starts rewarding the flattest, most mutation-tolerant ground. Biologists call this survival of the flattest. In a world changing at 5x, the most adaptable operator tends to beat the highest-performing but fragile one.
There is a clean test for whether the analogy actually holds. If the firms that pushed AI-driven process change hardest through 2025 and 2026 come out with the best reliability, the best margins, and no rise in incidents, then the threshold doesn't really bind and the biology doesn't transfer. The signal to watch is incident rates, well ahead of adoption rates.
That points to a number almost no team is tracking yet. Most companies are measuring "AI adoption," which is mostly a vanity metric. The number that will matter by 2027 is the internal error rate, meaning how much the core degrades while the edges mutate. You can see it in concrete things. Production incidents, failed payments, safety events per million trips, and how often a team has to roll back something it just shipped.
There is an uncomfortable implication in all of this. The AI era's first casualties may well be its fastest movers. A company that mutates its own core faster than it can recopy it can wake up one quarter unable to say what it is.
Everyone is busy pricing the risk of moving too slowly. Almost no one is pricing the risk of moving too fast and melting the core.
My conversation with @dkhos, CEO of Uber.
Dara took over in 2017, when Uber was losing roughly $4.5B a year.
Today the company generates $10B in free cash flow and is worth about $150B.
We discuss:
- How Daniel Ek convinced him to take the job
- How Uber spent a full year of its AI budget in a single quarter
- Uber's approach to autonomous vehicles
- Drones, hotels, and building a superapp
- Lessons from Allen & Co, Barry Diller, and Reed Hastings
Enjoy!
Timestamps
0:00 Intro
3:44 Bringing Order to Uber’s Chaos
7:22 Managing Stress and Going All In
14:28 Why Uber Is at the Center of AI and Physical
22:39 How to Win in Autonomous Vehicles
32:25 The Trillion-Dollar AV Opportunity
37:05 Drones, Robotaxis, and Global Adoption
38:20 Uber Eats, Uber One, and Aggregating Supply
47:00 The Future of the Uber App
55:55 Lessons from Barry Diller