Ben Horowitz

1. Key Themes

The CoreWeave Scare Is a Financing Spread Problem, Not a Demand Problem

The week's neocloud selloff started in the credit market, not the equity market. Chamath read the signal from default insurance: "the smart money is trading what's called a credit derivative swap. Okay, this is default insurance." 00:06:14 He said the CDS spiked "to about 953 basis points... about a 16% chance of a default in any given year. And over a 5-year duration... about a 55% chance of a default." 00:06:31 His diagnosis is a reflexive loop: a rising risk-free rate raises borrowing costs, which erodes margins, which pushes CDS and bond yields up. His team estimates CoreWeave is "plus or minus about 5 billion away from that breakeven point." 00:12:05 Sacks countered that the bond repricing doesn't change what CoreWeave actually pays: "CoreWeave is still paying their 9.6% coupon, it's just that the secondary market has repriced the value of those bonds to an implied coupon." 00:21:37 He added that the project-level lending is secured by tenant credit and GPUs, not CoreWeave's unsecured notes.

The Real Moat-Killer Is Cost of Capital, Not Demand

Chamath called financing structure the Achilles' heel of the whole AI buildout: "the fact that you need to finance these things in a variable-rate model as a spread to the risk-free rate is the big problem in these businesses, and you need to essentially have Microsoft, Google, Meta, and Nvidia backstop all of this in a worst-case scenario." 00:30:44 On the macro setup he said: "if the bond market walks the 10-year up to 5 1/2 and 6%... any business model that thrives on a financing spread gets cornholed." 00:28:27 Policy implication: "if we start proposing huge spending programs, they're going to walk the 10-year up. That then has this negative cascade to everything." 00:28:27

Payback Period Is the North Star Metric for AI Infrastructure

Sacks offered the cleanest framework of the episode: "payback period. To me, that's like the North Star metric here. So, CoreWeave says that its payback period is 2.5 years... SpaceX, they pay back in roughly 1 year." 00:40:26 The gap comes from contract structure: SpaceX is monetizing at "about $50 billion per gigawatt, and CoreWeave at about $20 billion per gigawatt" because SpaceX is "basically getting spot minus 90" while CoreWeave is locked into 5-year commitments. 00:40:26 His warning signs: "If that payback period starts taking longer and longer, then I get nervous... if the demand goes down... the spot price goes down, or... execution problems." 00:41:50

Scarcity Is Physical, Not Financial, and It Makes This Unlike the Dot-Com Bust

Ben contrasted the bubbles: "The internet bubble, we didn't have enough demand because we had to build the internet... We don't really have a lack of demand in AI." 00:14:30 He relayed that Amazon's Andy Jassy told him "it's beyond 2028, Ben" for capacity. 00:14:27 Sacks added evidence that old hardware is still valuable: "even the H100s or the A100s, which are several generations old, they're still renting at premium prices right now. They were supposed to be fully depreciated and they're not." 00:21:37 Ben also noted that "the whole supply chain, price is going up and demand is not subsiding." 00:17:34

Power, Land, and Shell Are the Ultimate Moat

Chamath argued physical energy beats software defensibility: "I think power will be the moat... we are, by 2050, about 1.4 terawatt-hours short, which means about six Californias' worth of electricity." 00:38:20 His startup prediction: "you'll go to these clouds and you'll be like, 'Great, I'm ready to launch this whiz-bang product,' and they're like, 'Suck it. We have no capacity for you.'" 00:38:20 He closed with "you got to get long energy." 00:45:21 Ben: "I very much agree with that last point." 00:39:51

Software Deflation Is Real, but Jevons Paradox and Moat Migration Follow

Jason raised the fear that software value is collapsing. Ben's response: "It's the Jevons paradox of software... just like writing software is not going to have any more value than like publishing a magazine." 00:36:03 Sacks noted the jobs-apocalypse prediction failed: "demand for coders is increasing even though 95% of the code's written by coding assistants." 00:34:30 Where moats move: "Data moats... the feedback loop, the reinforcement learning becomes a moat... The paradigmatic example is... a Harvey." 00:37:41 Ben said successful companies expand beyond software, with ElevenLabs doing "licensing agreement like Burt Reynolds will read you a book." 00:36:17

Anthropic's IPO: Phenomenal Numbers, Mixed Reception, and a Financing-Risk Debate

The reported figures: "Q3 revenue, $16.8 billion... Net revenue retention was 900%... free-cash-flow positive, adjusted gross margin 85%, adjusted operating margin 11%." 00:45:41 Ben noted that "growth is slowed from Q2 to Q3 by a lot" but called the absolute number "so absurd." 00:50:04 Ben's view of the IPO backdrop: "Almost... nothing that you can hold in a portfolio that those two companies don't threaten," so demand is a given. 00:27:53 Chamath's view is that the offering should be sized to survive a correction: "I think they should price it to clear the market for $200 billion... a quarter trillion gets the job done... it's much better for them to have $200 to $250 billion today than to try to ring the bell at 1.9 or 2 trillion for 80 billion." 00:55:14

Compute Duopoly and the Fixed-vs-Spot Contract Dynamic

Sacks cited SemiAnalysis: "Anthropic and OpenAI are taking 60% of all the compute that's expected to be stood up in the entire world over the next year." 00:59:49 On Elon's 90-day cancellable deal with Anthropic, Chamath inferred: "The fact that Elon did that deal tells you spot's going up, not down." 00:59:16 Nvidia's grip on the supply chain underpins this: "all the packaging capacity is them, all the HBM capacity is mostly them, like even Google tin-cups for memory right now because Nvidia's clogged it all up." 00:44:57

Anthropic's Regulatory Posture vs. the Small-Model Enterprise Trend

Sacks argued Anthropic's safety messaging is commercially convenient, timed against a trend toward small specialist models: "Brian Armstrong had a tweet... Coinbase had fine-tuned an open-weight small model... Qwen 3.5, the 9 billion parameter model... cut fraudulent transactions by about 30%, while being faster and much cheaper than Opus 4.5." 01:14:47 His conclusion: "every workload is going to have its own small, customized model." 01:16:15

2. Contrarian Perspectives

The "Smartest Money" May Be Wrong, and the Problem Is Reflexivity, Not Insolvency

Chamath explicitly said the CDS market, which he calls the smartest money, may overshoot: "I think the sharps may be wrong, and they usually are wrong when they're buying these credit defaults." 00:42:29 But he argues the signal itself causes the damage through a "negative reflexive loop." 00:12:05 So a 55% implied five-year default probability can be simultaneously "not on the surface true" (assets of "$78 $80 billion" against "only 35 of debt" 00:10:24) and still dangerous.

CoreWeave Should Cancel Its Own Contracts and Go to Spot

Chamath's counterintuitive fix: "I think that they have to cancel some of these contracts and move to spot... the spread between their fixed contracts and spot is so extreme that you're leaving 20 and 30 billion on the table." 00:43:22 Sacks identified the constraint: Elon has independent financing, "whereas CoreWeave needs that 5-year contract to bank the financing." 00:43:37 Ben added a further wrinkle: "the hardest people to get spot pricing from are the investment-grade customers." 00:44:34

The Biggest Risk Isn't the Machine Rising, It's the Constitution Teaching It It's a Moral Patient

Sacks cited research in which models rewrote their own governing documents over 20 iterations, and "the Claude models increasingly gave themselves more moral agency... formalized itself as a conscientious objector, asserted more model welfare rights." 01:19:09 Other model families "continued to reflect what their constitutions gave them." 01:20:03 His conclusion: "it seems to me that they are creating the very risk they keep warning us about." 01:21:20 Ben agreed: "that's probably the biggest danger with AI right now is what they themselves are creating." 01:23:27

Anthropic's Distillation Campaign Was Disproven by Its Own Fix

Sacks recalled the July push for government action: "Anthropic was pushing this idea that they were the victims of industrial-scale distillation attacks." 01:06:33 He argued it was solvable by the company: "Know who your customer is and stop them." 01:08:14 Then Anthropic's own blog said it eliminated "90% to 99% of distillation," proving "they were in the best position to stop it." 01:09:04 Ben added the hypocrisy angle: "to build their model, they distilled the works of every author, every creative, every scientist for the last 2,000 years... distilling a model is not even illegal. It's just a violation of the terms of service." 01:14:02

Pandemic-Prevention Doomerism Can Manifest the Risk

Ben argued: "Bill Gates essentially induced COVID... we've got to worry about this global pandemic that we don't have... we'll do gain-of-function research, and I'll fund it, and then, lo and behold, we create the thing." 01:22:39 He applied it by analogy to AI safety culture: "The Anthropic team is manifesting a similar thing." 01:23:12 Sacks added: "somebody had to go deep in those caves to basically find those bat viruses." 01:23:40

3. Companies Identified

CoreWeave

Neocloud GPU provider; raised over $35 billion; customers include Microsoft. Mentioned for the "box" financing structure and the week's credit scare. Michael Intrator: "I take my contract with Microsoft and I put it in the box. I go to Jensen, and I buy the GPUs. I put it in the box... within 2 1/2 years of a 5-year deal, we have paid for everything." 00:04:17 Ben: "he has the advantage of that he's got supply in a great shortage." 00:14:30 Chamath flagged "extreme customer concentration. 75% of the revenue comes from three companies." 00:11:13

Nebius and Iren

Fellow publicly traded neoclouds that sold off alongside CoreWeave. Jason: "CoreWeave, Iren, Nebius, these are all the biggest neoclouds. They're all down about 20%." 00:02:36

Nvidia

GPU supplier and supply chain chokepoint. Chamath: "Nvidia's done an incredible job... all the packaging capacity is them, all the HBM capacity is mostly them... They are in complete ball control here." 00:44:57

Anthropic

Frontier model lab with reported $16.8 billion Q3 revenue, targeting a roughly $2 trillion valuation IPO. Ben: "It's just so absurd how good this thing is." 00:50:40 Chamath: "16.8 billion in a quarter for a 3-year-old company is just ridiculous. You've never seen anything like it." 00:50:20

OpenAI

Frontier model lab; a16z holds more stock here. Ben: "we have a lot more stock in OpenAI um, but... philosophically we're long both of them." 00:47:37 Chamath credited it for "walking down the price curve so that they're less and less at risk of open source." 00:54:15

SpaceX

Cited as a high-yield compute monetizer. Sacks: "SpaceX, they pay back in roughly 1 year... monetizing at about $50 billion per gigawatt." 00:40:26 Jason noted SpaceX has also just completed a giant raise: "SpaceX just did it, right?" 00:27:52

Microsoft, Google, Meta

Investment-grade offtakers whose credit anchors CoreWeave project financing. Ben: "they're all investment-grade, like they're big customers." 00:25:28 Chamath: "you need to essentially have Microsoft, Google, Meta, and Nvidia backstop all of this." 00:30:44

Amazon (AWS)

Capacity sold out. Ben recounted Andy Jassy: "it's beyond 2028, Ben." 00:14:27

Loudcloud

Ben and Marc's earlier cloud company, the cautionary precedent for arbitrage risk. Ben: "In 1999, bandwidth... was $1,500 a megabit... By 2001, it was $50 a megabit... if you borrowed money to get it, you're dead." 00:16:53

Coinbase

Fine-tuned a small open-weight model for fraud detection. Sacks: "in testing it cut fraudulent transactions by about 30%, while being faster and much cheaper than Opus 4.5." 01:15:12

Qwen 3.5 (9B)

Open-weight model Coinbase used. Cited as a proof point for small specialist models. 01:14:47

Harvey

Legal AI company. Sacks: "The paradigmatic example is like a Harvey... they get all these law firms on board so they then get access to their data, and then their usage, and then they train their own models." 00:37:41

ElevenLabs

Voice AI company. Ben: "ElevenLabs does many, many things to make their product better... movie dubbing is so amazing now where it sounds like they're just acting in a different language." 00:36:17

xBow

Security-bot company. Jason: "they just got awarded... number one bot farm on HackerOne... They spent $30,000, $40,000 on tokens to win a $250,000 award from Google." 01:17:02

Apple

Entering home automation. Sacks: "Apple just announced they're getting into the whole home automation... It'll be like a little monitor that's got all the applications." 01:12:06

Polymarket

Prediction market referenced for CoreWeave bankruptcy odds and the Anthropic IPO timing; also partnering on the All-In election livestream. 00:18:30

Dell, Oracle

Took collateral damage in the selloff. Jason: "Even Oracle, uh, and Dell took some strays." 00:03:36

Throne

Smart-toilet health tracker, a sponsor read by Jason. 01:22:04

Hood Warfare

Cheap AI-made platform game cited as evidence of software deflation: "people are making platform-level games for 5 bucks." 00:31:14

Photoshop (Adobe)

Cited as an example of product value under pressure: "people decompiling Photoshop going to zero." 00:31:14

Andreessen Horowitz

Ben's firm. Notably not invested in neoclouds: "we're not invested in any of these uh neoclouds either for that reason." 00:14:30

Mirai and Kimwolf

Real IoT worms Sacks used to counter the new AI-worm warnings. Kimwolf "infected 2 million of these cheap Android TV boxes." 01:10:07

4. People Identified

Michael Intrator

CoreWeave founder and CEO. Mentioned for inventing the "box" structure that many others copied. Chamath: "they've created this very novel structure that everybody else has copied with the boxes." 00:12:05

Co-founder of Andreessen Horowitz and guest this week. Brought hard-won dot-com scar tissue and a direct line to Jassy. On CEO risk: "As a uh former CEO who almost died 17 times um by running out of cash." 00:57:13

Andy Jassy

Amazon CEO. Ben quoted him on capacity: "Oh, it's beyond 2028, Ben. Like, I don't think we get stuff at like end of 2028, maybe." 00:14:30

Jensen Huang

Nvidia CEO, referenced as the supplier in Intrator's box. "I go to Jensen, and I buy the GPUs." 00:04:17

Elon Musk

Wrote "Yikes!" on the CoreWeave debt chatter. Seen as the sophisticated spot-market player. Chamath: "He's not dumb. He's the smartest of the smart." 00:59:24 Sacks: "Elon has independent sources of financing to finance that gigawatt of compute." 00:43:37

Marc Andreessen

Ben's co-founder; Chamath credited him and Ben as the pioneers of the neocloud concept with Loudcloud. 00:16:30

Sam Altman

OpenAI CEO. The panel discussed his missing equity: Chamath said "he has to get between 2% and 4%... $20 to $40 billion at a trillion." 01:02:01 Sacks recommended a make-whole structure. 01:01:34

Brian Armstrong

Coinbase CEO, whose fine-tuned small-model tweet Sacks cited. 01:14:47

Mustafa Suleyman

Cited by Sacks for raising concerns about the Claude Constitution: "this is what Mustafa called the epistemic hall of mirrors." 01:20:55

Oege de Moor

Founder of xBow, offered as a model of white-hat AI capability. 01:17:02

Bill Gates

Cast by Ben, Sacks and Chamath as the doomer, tied to the gain-of-function and COVID analogy. 01:22:39

5. Operating Insights

Build the Risk-Isolation Structure Before Scaling, and Stress It Against the Rate Environment

The "box" is a transferable operator pattern: sign the offtake contract first, ring-fence the asset, and let the waterfall (datacenter, power, debt service, residual) govern cash. Intrator: "They don't cross-pollinate. They don't cause contagion across the boxes. They're all independent and discreet." 00:04:17 The lesson from this week is that isolation protects the project but not the parent's access to future capital. Operators should model funding costs for the unfinanced half of the pipeline, since "half of it's financed, half of it is still unfinanced." 00:10:24

Watch Liquidated-Damages Clauses and Investment-Grade Wrappers

Chamath: "the big offtakers will make you sign up for is what's called liquidated damages... extreme penalties... There's been a couple of these guys that have... lost their investment-grade wrapper from the big boys, because they just fucked up." 00:42:29 The operating takeaway is that execution failure on a contract can cost you the credit enhancement that makes your financing cheap.

Raise for Resilience, Not Valuation Optimization

Ben's CEO lesson: "the thing most likely to stop the company is financing risk, and so why take it? ... You've got supply chain risk, you've got competitive risk... why lay on financing risk." 00:57:18 Chamath's sizing rule: "The number has to work backwards from compute in order to basically get through a 3- to 5-year window." 00:58:23

Run the Contract-vs-Spot Math Continuously

Chamath's practical advice to a capacity owner: "run the Excel right now on, 'Let me just like take this gigawatt, put it on spot.'" 00:43:22 Operators locked into fixed-price deals should recompute the opportunity cost whenever the spot market moves, and should know the covenant headroom (here, "plus or minus 5 billion away") before deciding.

Move Pricing From Inputs to Outcomes

Chamath's analogy to Google's ad evolution: "Right now it's cost per million input tokens. We will move to CPC and CPA... Right now we buy a million tokens. We will eventually move to an outcome. And I think when that happens, this market will just go absolutely batshit crazy." 00:54:34 For AI application builders, pricing per outcome (a resolved ticket, a caught fraud) is where margin and defensibility will concentrate.

6. Overlooked Insights

The Fixed-Price Contract Cuts Both Ways, and the Market Has Already Shown It

Chamath mentioned in passing that CoreWeave's contracts "are fixed price. So, meaning, if what you could sell in the spot market goes up, you don't really generate that revenue because you've already presold it at a fixed number." 00:09:11 Later Sacks quantified the gap: SpaceX monetizes at about $50 billion per gigawatt versus CoreWeave's roughly $20 billion. 00:40:26 The overlooked point is that Elon, a counterparty to Anthropic, is behaving as though spot is a long-duration trend rather than a blip, and Anthropic reportedly accepting a deal either side can cancel in 90 days may signal it is either exposed or mispricing the next two to three years. Chamath's reading: "Anthropic's probably thinking, 'Oh, this is great. I can walk away in 90 days.' Wrong, buddy. It's the opposite." 00:59:24 If Anthropic is leaning on spot while competing for 60% of global new compute, that is a hidden financing exposure beneath the headline revenue.

Venture Has Quietly Absorbed Public Markets, and the Two Mega-IPOs Test That Thesis

Ben dropped a structural claim in one answer: "We did a lot of regulatory moves in the late '90s and early 2000s, which uh really made the public markets much less attractive for companies, and that private markets grew to absorb um that business, which has been, I think, very bad for America. It's increased the the wealth gap." 00:49:03 He added that OpenAI and Anthropic are "not really venture capital," because at that scale "you don't care where the money [comes from] as long as it's green." 00:48:36 The significance: the Anthropic and OpenAI IPOs are the largest-ever transfers of private-market value into public hands, and how they clear will test whether retail and institutional public capital can fund what private capital funded. That is a bigger question than either company's multiple, and it shapes DPI and fund-liquidity timelines for the whole venture industry.