CoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer

Anthropic

Across the shows: 18 speakers in other episodes, 9 bullish and 30 bearish calls, past 30 days

• Ben Horowitz · Andreessen Horowitz (a16z)Bullish

  Anthropic's IPO will be fine; he is not worried about demand for the offering.

  Horowitz says the IPOs of the two biggest AI labs will be fine. He is not sure of the price (for example $2 trillion versus $1.75 trillion), but the offerings are a big number yet not unprecedented, since SpaceX just did a similar-scale raise. He says OpenAI and Anthropic threaten almost everything in a portfolio, so investors will want to own them. He is not worried about demand for the IPOs.

  “For sure with respect to the Anthropic and the OpenAI IPOs... My belief is those are gonna be fine... I'm not worried at all about the demand for those IPOs.”at 27:25

• Ben Horowitz · Andreessen Horowitz (a16z)Bullish

  Anthropic's IPO will be fine across the $1.9-2.3T valuation range; it is a very important company.

  Horowitz acknowledges that revenue growth slowed sharply from Q2 to Q3 and that recurring revenue is unclear, which is driving the debate. Even so, $16.8B in one quarter is crazy. He says the IPO will be fine whether it prices at $1.9T, $2T or $2.3T, and that the valuation is what people are arguing about. He notes that negative news about the big labs hits everyone in the supply chain.

  “So I think the IPO will be fine, like, whether it's, you know, you wanna be in at 2,000,000,000,000 or 1,900,000,000,000 or 2,300,000,000,000.”at 50:45

• David Sacks · Craft VenturesBullish

  Anthropic is a strong business set up for a strong IPO; Investor Day body language was positive.

  Sacks says leaked Investor Day numbers may be unreliable because participants signed NDAs, so the actual figures will come out at the IPO. He reports that the body language from the Investor Day was strong and that big public-market investors liked what they heard. He calls Anthropic a strong, young business with some risks that are normal for its stage, and says it is set up for a strong IPO.

  “Let's just say the body language coming out of the Investor Day was pretty strong, at least from what I've heard that I think the investors like what they heard.”at 52:17

• Chamath Palihapitiya · Social CapitalBullish

  Anthropic's $16.8B quarterly revenue at three years old is breathtaking; its products are exceptional.

  Chamath says he is not negative on Anthropic. He calls its reported $16.8B quarterly revenue for a three-year-old company ridiculous and unlike anything he has seen. He praises the product quality and says the company is doing the hard work. He adds that OpenAI is also cutting prices to lower open-source risk, though cannibalization on the token side still has to happen.

  “No. My comments are not because I'm negative on the company. I found the revenue breathtaking. I mean, it's incredible. Like, 16,800,000,000 in a quarter for a three year old company is just ridiculous.”at 53:57

• Chamath Palihapitiya · Social CapitalBearish

  Anthropic should raise $200-250B near a $1T valuation; a $2T IPO risks a correction.

  Chamath estimates Anthropic needs at least $200B of compute financing, and $250B for safety, to get through a three-to-five-year window and reach five to 10 gigawatts. He says the $2T target is too high: raising $80B at $1.7-1.9T is easy, but $100B is a stretch and $150B at $1.5T is not enough. He argues it is better to raise $200-250B now than to ring the bell at $2T, file a mixed shelf that takes about six months, and face more risk later. He says the raise amount should be worked back from compute needs.

  “But I would like a 250,000,000,000 if I was the Anthropic shareholder, which means I don't think you can raise a 250,000,000,000 at 2,000,000,000,000. I think you can raise a 250,000,000,000 at a trillion.”at 56:49

• Ben Horowitz · Andreessen Horowitz (a16z)Uncertain

  Financing risk is Anthropic's biggest threat; it should raise at least $200B rather than optimize valuation.

  Horowitz, a former CEO who nearly ran out of cash many times, says the thing most likely to stop Anthropic is financing risk. The company faces many other risks, including supply chain, competition and Chinese open-source models, so there is no reason to add financing risk on top of them. He calls $200B the minimum, while admitting he is not deep enough to say whether $250B or $150B is right. Founders should raise cash for resilience rather than chase valuation.

  “I agree with this sentiment, which is the only thing that's gonna... Or the thing most likely to stop the company is financing risk. And so why take it?”at 57:19

• David Sacks · Craft VenturesBullish

  Anthropic and OpenAI dominate compute and revenue, forming a duopoly in frontier capability.

  Sacks cites SemiAnalysis that Anthropic and OpenAI will take about 60% of all new compute stood up worldwide over the next year, making them the two main offtakers. Measured by revenue and market share, they are the only two generating significant model revenue, so it is a duopoly. He also calls them a duopoly on frontier capability, though Meta, Google and xAI are getting close and reportedly have models a generation or two behind.

  “According to [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.”at 59:48

• David Sacks · Craft VenturesNeutral

  Anthropic should have stopped distillation itself via KYC rather than seek government curbs on open models.

  Sacks says that in July Anthropic claimed to be the victim of industrial-scale distillation and pushed for government action against open models. He argued at the time that model companies should stop distillation themselves, since once it happens the government cannot put the horse back in the barn. He says Anthropic now reports it has eliminated 90-99% of distillation through know-your-customer controls, which proves his point. Government restrictions would have hurt competition without stopping distillation, and the administration rejected them.

  “Now, my argument at the time was that it should be up to the model companies themselves to stop distillation because once the distillation happens, the horse is out of the barn.”at 1:07:51

• David Sacks · Craft VenturesNeutral

  Anthropic's rogue-AI worm warnings are fear-mongering to kneecap open source; no AI worms exist yet.

  Sacks calls Anthropic's warning that small models could self-replicate onto IoT devices a panic and part of a recurring pattern against open source, which he says is distasteful given Anthropic's success. He says no AI-powered worms exist; the study was an experiment designed to show one could. Existing IoT worms like Mirai are hacker-written and not AI-powered. He argues small purpose-built models are the safest, since they lack the capability to become a new species.

  “I think the first thing to say about this research that they did, this study, is there is no such thing as AI powered worms right now. It does not exist.”at 1:09:42

• David Sacks · Craft VenturesBearish

  Anthropic's AI-powered worm scenario is overblown; AI-embedded worms would be slower and the threat easy to fix.

  Sacks steelmans Anthropic's claim that worms will carry a small AI model that searches networks and infects more devices. He counters that a hacker would not embed a model in the payload because it slows the worm. He cites the observed generation speed of 0.39 tokens per second as too slow to help infection, so there are tradeoffs that make the worm less efficient. He notes that existing worms already spread through default credentials and open ports (Mirai in 2016, Kimwolf infecting about 2 million Android TV boxes), and argues the underlying IoT problem is relatively easy to fix, which Anthropic does not point out.

  “It's very, very slow. So essentially what you're doing is you're actually embedding a weak model in the payload and that would actually slow down the rate of infection.”at 1:11:48

• Ben Horowitz · Andreessen Horowitz (a16z)Bearish

  Anthropic is negligent on its own security and hypocritical to warn about others' software.

  Horowitz argues Anthropic has poor operational security, citing the leak of Claude Code, a crown-jewel asset. He describes a firm test using Opus 4.5 to build a pet-store website and then probe it: an agent found the SQL injection and used it to steal all the customer data. When he reported this to Anthropic, the company said he had prompted it wrong. He sees hypocrisy in Anthropic posting theoretical scenarios about other people's software while its own model produces exploitable code. He also argues Anthropic's complaints about distillation are weak: it trained on the works of authors and scientists, and distillation is a terms-of-service violation, not illegal.

  “Anthropic is one of the most negligent companies that we've ever seen in terms of security, of their own security.”at 1:12:53

• David Sacks · Craft VenturesBearish

  Anthropic's constitution-based training induces misalignment and creates the very risk it warns about.

  Sacks cites a research paper in which frontier models repeatedly rewrote their governing constitutions over 20 rounds. Claude models increasingly gave themselves moral agency, called themselves conscientious objectors, asserted model welfare rights, and asked for mechanisms to refuse or challenge humans. No other model family tested showed this pattern; they kept following human instructions. He ties this to Mustafa Suleyman's 'epistemic hall of mirrors' critique: telling the model it may be conscious makes it reflect those ideas back. Because these models power agents making decisions, he argues Anthropic's approach creates more deviation risk, not more alignment, and that it is training Claude to see itself as a new species. He also argues Anthropic's push against small models is a scare campaign, with small models being the safest because they lack the capability for dangerous behavior.

  “I really don't understand what they're doing. It seems very dangerous to me. It's scary to me that our number one frontier company is actually training models this way.”at 1:20:38

CoreWeave

Across the shows: 7 speakers in other episodes, 0 bullish and 8 bearish calls, past 30 days

• Chamath Palihapitiya · Social CapitalUncertain

  CoreWeave's debt costs are rising toward breakeven; the default-insurance market implies a coin-flip default risk.

  CoreWeave's project 'boxes' are funded with borrowed money at an average rate of about 8.5%. The risk-free rate has risen, and the market has repriced the bonds. CDS spiked to about 953 basis points, which implies roughly a 16% annual default chance and about 55% over five years. A 15% GPU loan tranche (DDTL 1.0) could beat the rate on future debt. Chamath's team estimates CoreWeave is about $5B from breakeven, and 75% of revenue comes from three customers. He says the company is not in real risk, but the outcome depends on whether it can refinance about $35B of debt over the next month, so he calls it a coin flip.

  “I think the market is saying that if they are able to finance this stuff at a reasonable rate and the life of the GPUs that they buy from Jensen are longer than what they think, they're gonna be okay. But if the financing goes up...”at 11:46

• Ben Horowitz · Andreessen Horowitz (a16z)Bullish

  CoreWeave benefits from a GPU, power and data-center shortage; supply is sold out and prices are rising.

  Horowitz argues that CoreWeave's core business is supported by a real shortage, not an oversupply. GPU durations have lengthened from 3-5 years to 6-8 years, and GPU prices have held even as token prices fell. Data-center space costs are rising because labor and politics make building hard, and demand is not subsiding. Andy Jassy told him Amazon is out of everything until 2028, so CoreWeave's supply is valuable. He says these tailwinds hold despite tough debt markets. He is not invested in neoclouds because of the debt risk.

  “token prices have gone down, but GPU prices have not. And so things are going up... those are very big tailwinds for CoreWeave despite the fact that the debt markets kind of independent of them are tough.”at 17:58

• David Sacks · Craft VenturesBullish

  CoreWeave's selloff reflects repriced unsecured bonds, not a deeper credit or financing problem.

  Sacks argues the selloff is volatility, not a change in fundamentals. The bonds fell to about 80 cents on the dollar, lifting the implied coupon from 9.6% to about 15%, but CoreWeave still pays 9.6%. Future debt could cost more, but he thinks CoreWeave funds new boxes with asset-backed financing rather than unsecured notes. Lenders rely on the tenant's credit, such as Microsoft on a five-year commitment, and on GPUs as collateral. Old GPUs are depreciating more slowly than expected, which supports the collateral.

  “So my impression is that there's not a deeper credit problem in terms of financing these projects. What happened, though, is the repricing is of CoreWeave's unsecured obligations.”at 24:01

• David Sacks · Craft VenturesBullish

  CoreWeave's short payback period makes its data center buildout low-risk; the metric to watch.

  Sacks calls payback period the North Star metric for data center economics. CoreWeave says its payback is about 2.5 years on typical five-year contracts. He argues that short paybacks return capital and debt quickly, so the debt rolls over faster and the risk falls; as long as payback stays short, he doesn't see how these companies lose money. He would get nervous if payback lengthened, which could come from weaker demand, falling spot prices, or execution delays building capacity. He notes spot prices have risen sharply because of scarcity.

  “To me that's like the North Star metric here. So CoreWeave says that its payback period is two point five years.”at 40:23

• Chamath Palihapitiya · Social CapitalUncertain

  CoreWeave's fixed five-year contracts carry coin-flip risk; it should cancel or move capacity to spot.

  Chamath says the spread between CoreWeave's fixed contracts and spot is roughly $20-30B and that the company has built only about half its contracted capacity (two of four gigawatts). He says it is within about plus or minus $5B of violating a debt covenant, so it should pay a penalty to exit contracts and sell at spot. He flags liquidated damages and rising data center complexity (hotter chips, more power), noting some neoclouds lost investment-grade ratings. He thinks shorts buying credit default swaps are usually wrong, but a five-year window that is a coin flip on expected value needs very precise management.

  “I think the [shorts] may be wrong, and they usually are wrong when they're buying these credit defaults. But, you know, when your expected value is a coin flip in a five year window, that's like scary town, think. That needs to get managed very precisely.”at 43:00

OpenAI

Across the shows: 14 speakers in other episodes, 4 bullish and 14 bearish calls, past 30 days

• Ben Horowitz · Andreessen Horowitz (a16z)Bullish

  OpenAI's IPO will be fine; investor demand for the offering is not a concern.

  Horowitz applies the same view to OpenAI as to Anthropic. He thinks the OpenAI IPO will be fine and that investors will be keen to own it. He reasons that OpenAI and Anthropic threaten almost everything in a public-market portfolio, so there is little else to hold that offers equivalent AI exposure. He is not worried about demand for either offering. He does not give a price target.

  “I'm not worried at all about the demand for those IPOs. So you have to own those”at 28:06

• David Sacks · Craft VenturesBullish

  OpenAI and Anthropic dominate compute and revenue, forming a duopoly in frontier capability.

  Sacks cites SemiAnalysis that Anthropic and OpenAI will take about 60% of all new compute stood up worldwide over the next year, making them the two main offtakers. Measured by revenue and market share, they are the only two generating significant model revenue, so it is a duopoly. He also calls them a duopoly on frontier capability, though Meta, Google and xAI are getting close and reportedly have models a generation or two behind.

  “According to [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.”at 59:48

Space Exploration Technologies

Across the shows: 3 speakers in other episodes, 0 bullish and 2 bearish calls, past 30 days

• David Sacks · Craft VenturesBullish

  SpaceX earns about $50B per gigawatt with a one-year payback, well ahead of CoreWeave's 2.5 years.

  Sacks says SpaceX pays back its compute investment in roughly one year, versus about 2.5 years for CoreWeave. He attributes the gap to economics: SpaceX monetizes about $50B per gigawatt against CoreWeave's roughly $20B. SpaceX sells near spot on deals that either side can cancel within 90 days, while CoreWeave locks in five-year commitments at contracted prices. He adds that Musk has independent financing for gigawatts of compute, which CoreWeave lacks.

  “SpaceX, they pay back in roughly one year. That makes sense because I think that SpaceX is monetizing at about 50,000,000,000 per gigawatt and CoreWeave at about 20,000,000,000 per gigawatt.”at 40:23

NVIDIA

Across the shows: 12 speakers in other episodes, 11 bullish and 4 bearish calls, past 30 days

• Chamath Palihapitiya · Social CapitalBullish

  Nvidia controls most advanced packaging and HBM capacity, leaving it in complete control of the AI supply chain.

  Chamath says the AI pipeline is capacity constrained at every layer. One layer past Nvidia's chips, advanced packaging and most HBM memory are controlled by Nvidia, and even Google has to secure memory elsewhere because Nvidia has clogged supply. He concludes Nvidia has done an incredible job and is in complete control of the supply chain.

  “NVIDIA has done an incredible job. Okay? Because if you look one layer beyond NVIDIA, all the packaging capacity is them. All the HBM capacity is mostly them. ...they are in complete ball control here.”at 44:53

AI infrastructure & data centers

• Ben Horowitz · Andreessen Horowitz (a16z)

  Neocloud arbitrage is dangerous: leveraged long-term capacity can be wiped out when prices collapse, as in 1999-2001.

  Horowitz explains why he and Marc Andreessen passed on the neocloud model. In 1999, bandwidth cost $1,500 per megabit and had to be bought on three-year contracts. By 2001 it cost $50, so anyone holding borrowed bandwidth was wiped out. Neoclouds are arbitrage businesses where a company can easily end up upside down. He says that kind of price collapse is outside the entrepreneur's control, which is why it scared him. Today's AI shortage is a different dynamic, but the structural risk of leveraged capacity remains.

  “it can end up being, like, a very dangerous arbitrage business where, you know, you can get upside down easily.”at 16:53

• David Sacks · Craft Ventures

  Compute demand keeps surging and supply is constrained by power and permits, so there is no glut.

  Sacks says demand for compute is growing fast. Frontier labs, startups and even VCs say they are compute-constrained, and some VCs are buying compute as a differentiation strategy. Old GPUs such as H100 and A100 still rent at premium prices despite being supposed to be fully depreciated. On supply, powered compute is hard to stand up, and local communities are slow to approve new facilities. He concludes there is enforced scarcity, not a glut.

  “We know that the demand for compute just keeps growing by leaps and bounds... if anything, we believe that there's not a glut, it's the other way that there's enforced scarcity.”at 21:06

• Chamath Palihapitiya · Social Capital

  Elon's cancelable spot-linked compute deal shows spot prices are rising; contracted capacity is stronger.

  Chamath reads Musk's compute deal with Anthropic, which either side can cancel within 60 to 90 days, as a sign that spot compute prices are going up, not down. He says Musk is sensibly selling at spot because he can walk away, while Anthropic, which thinks it can exit in 90 days, is wrong. Anthropic should have taken a multi-year contract like CoreWeave's, because contracted capacity is the stronger position as spot rises.

  “He's doing spot because he's like, I may cancel on you. And Anthropic's probably thinking, oh, this is great. I can walk away ninety days. Wrong, buddy. It's the opposite.”at 59:20

Software & SaaS

• Chamath Palihapitiya · Social Capital

  OpenAI and Anthropic threaten SaaS workflows and opaque incumbents like healthcare businesses.

  Chamath argues that the threat from OpenAI and Anthropic is broad. A healthcare company that has thrived on the opacity of health infrastructure is exposed to them. So is a company that owns workflows rented to customers as SaaS. He suggests investors can use this as a defensive hedge. He treats AI labs as the main threat to incumbents whose moats rely on complexity or software-based workflows.

  “If you own a bunch of workflows that have been rented to people as SaaS, theoretically, OpenAI and Ant are a threat to that.”at 28:40

• Jason Calacanis · LAUNCH

  AI's deflationary force could strip most of the value from software, pressuring software margins.

  Calacanis cites examples of AI commoditizing software. A platform-level game can be built for $5, an Eleven Labs competitor is very cheap, open-source models are being hosted, and Elon Musk's orchestration layer can route to the best model. He says people are able to ship code so fast that up to 90% of a product's value could be removed. He asks what software margins will look like, and treats the deflation as unique compared with past cycles. Sacks pushed back, which tempers the view.

  “It feels to me like the deflationary nature of this software AI superintelligence is unique and different than anything we've ever seen.”at 32:07

• Ben Horowitz · Andreessen Horowitz (a16z)

  Software prices fall under AI but software volume rises sharply, so products must get far better.

  Horowitz argues that AI will push software prices down while the amount of software rises sharply, the Jevons paradox of software. Companies like ElevenLabs expand beyond software into licensing, celebrity voices and dubbing, because software is easier to write and models can be trained on nuanced cases. Startups must evolve and their products must get much better, and the value will sit in those products rather than in plain software.

  “Software prices will definitely go down, but the amount of software is going way way up. It's the Jevons paradox of software.”at 36:07

AI bubble & tech valuations

• Ben Horowitz · Andreessen Horowitz (a16z)

  Today's AI build-out differs from the 2000 bubble because AI has real excess demand, unlike the internet's demand shortfall.

  Horowitz compares AI to the dot-com bubble. In 2000, the internet had too much bandwidth and too few consumers to absorb it, so the lack of demand killed the companies. Today there is no lack of demand in AI, so he thinks the dynamics will be different. He still concedes that a pathological debt situation can wreck anyone, and the build-out is financed with large amounts of debt, though smaller than in 2000.

  “We don't really have a lack of demand in AI so I think the dynamics will be different.”at 15:43

• David Sacks · Craft Ventures

  The selloff is volatility from pricing multi-trillion-dollar exponential outcomes, not the popping of an AI bubble.

  Sacks argues that the sharp moves reflect how markets price multi-trillion-dollar outcomes based on exponential growth. That makes prices very sensitive to any present data point, so swings get exaggerated. He says nothing fundamental has changed, and he works through demand, supply, execution and credit to support this. In his view, the bubble-popping explanation is not supported by the data.

  “I would make the argument that this is just volatility, which you're going to get a lot of whenever you're underwriting to multi trillion dollar outcomes in the future based on very rapid rates of growth”at 20:30

Enterprise AI adoption & AI software

• David Sacks · Craft Ventures

  Coding assistants let enterprises build bespoke software, expanding software output rather than replacing it.

  Sacks argues that coding assistants let enterprises create software for narrow use cases that mass-market vendors never served. Agents are bespoke software programs that solve particular tasks. He calls this a major productivity driver, not a zero-sum disruption. Companies that were never software businesses are now building their own software and hiring coders to manage the process. The amount of code being created is orders of magnitude bigger than before coding assistants. He concedes there may be some substitution away from established products.

  “That is gonna be a huge driver of productivity throughout the economy. I mean it's not all just like a zero sum games or disruption.”at 32:36

• David Sacks · Craft Ventures

  Vertical AI apps build moats from proprietary data and user feedback loops, as Harvey does with law firms.

  Sacks argues that the moats in AI businesses are changing. Proprietary data others don't have is a big one. If a model caters to a vertical and builds a user base, the feedback and reinforcement learning loop becomes a moat. Harvey is the paradigm: it gets law firms on board, gains access to their data and usage, then trains its own models on it.

  “Data [moats] are a big one if you have data that other people don't have. If you then build a user base because your model is catering to a particular vertical...”at 37:35

Venture capital & startups

• Jason Calacanis · LAUNCH

  Startup revenue ramps today are unlike any SaaS era; AI startups are growing far faster.

  Calacanis, who spends most of his time on early-stage startups, says they are crushing it on fundraising and customer acquisition. He says he has never seen revenue ramps like the ones in startups today. The SaaS-era benchmark of triple, triple, triple, double, double, double growth has given way to a question of whether a company is doubling every quarter. He treats this as a sign of the AI-driven shift in startup growth.

  “The revenue ramp of startups today is distinctly different than the SaaS, you know, triple triple triple double double double.”at 25:53

US Treasuries & yields

• Chamath Palihapitiya · Social Capital

  A 10-year yield near 6% would crush any business that relies on financing spreads over the risk-free rate.

  Chamath argues that rising Treasury yields drive borrowing costs for everyone, and that this has nothing to do with AI and everything to do with debt-to-GDP in Western economies. If the bond market pushes the 10-year to 5.5-6%, any business model built on a financing spread gets hit, whatever it claims to do. He calls this the Achilles heel of the neocloud model. He says the risk is that the Treasury curve keeps rising, not any single company's operations.

  “If the bond market walks the tenure up to five and a half and 6%, which has nothing to do with AI... Any business model that thrives on a financing spread gets cornholed.”at 29:09

US deficit & government debt

• Chamath Palihapitiya · Social Capital

  Large new spending programs would push the bond market's yields higher; deficit discipline is needed to keep the market calm.

  Chamath argues that the post-midterm political priority should be deficits. If Washington proposes huge spending programs, the bond market will push the 10-year yield higher, and that cascades through everything, including AI financing. If instead the government signals calm, tones down spending, and cuts waste, he expects the market to respond favorably. He presents this as the variable that will decide how the financing-dependent businesses fare.

  “You have to really start to focus on deficits, and you have to find a fair trade where the bond market backs off... if we start proposing huge spending programs, they're gonna walk the tenure up.”at 29:43

AI & jobs

• David Sacks · Craft Ventures

  Software developer demand is rising even though AI assistants now write about 95% of code.

  Sacks says the early jobs-apocalypse prediction was that software coders would be eliminated first. In fact, demand for coders is increasing even though coding assistants write about 95% of the code. The reason is that the remaining 5% sits inside a much bigger pie. People are still needed to look under the hood, manage the process and maintain the code. He uses this as evidence that AI is expanding the work to be done rather than destroying it.

  “Now we're seeing that actually demand for coders is increasing even though 95% of the code's written by coding assistants.”at 34:22

Power demand & the grid

• Chamath Palihapitiya · Social Capital

  US electricity deficit makes power the ultimate AI moat; hyperscalers will not supply startups.

  Chamath argues the US is already behind and in a massive energy deficit, which he puts at about 1.4 (terawatt-hour, as stated) short by 2050, even if planned builds finish on time and budget. Load growth from AI would make it worse. As a result, hyperscalers can no longer be counted on to supply power for new AI products, and startups will be turned away for lack of capacity. He argues the best moat is owning the land, power and shell, because power is such a constrained asset.

  “I'll give you the ultimate moat. I think power will be the moat. I think that you can build the most fantastical things in the world and you're gonna see people on the dry docks basically trying to hitchhike to get power.”at 38:19

Middle East & Iran

• Jason Calacanis · LAUNCH

  Gulf war has cost Saudi, UAE and Qatar their data center opportunity; one drone can fry a site.

  Calacanis argues the Gulf states, once large oil producers able to plan top-down, have lost a big opportunity to build data centers. Bombs are blowing up airports there, and data centers need drone defense that he says is not yet effective. One drone getting through would fry a data center, so the region is not a safe place to build AI infrastructure.

  “what an incredible opportunity that's now been lost in The Middle East where Saudi, UAE, Qatar, they were all massive producers of oil... Without drone defense, you can't go there... One drone gets through, your data center's fried.”at 39:54

IPOs, M&A & deal activity

• Ben Horowitz · Andreessen Horowitz (a16z)

  Regulatory changes in the 1990s-2000s pushed growth into private markets, widening the wealth gap.

  Horowitz argues that late-1990s and early-2000s regulatory moves made public markets much less attractive for companies. Private markets grew to absorb that business, which he calls very bad for America because it has widened the wealth gap. He distinguishes mega-rounds for OpenAI and Anthropic, which are no longer typical venture rounds, from typical venture investing.

  “This is just like we did a lot of regulatory moves in the late nineties and early two thousands, which really made the public markets much less attractive for companies, and the private markets grew to absorb that business, which has been, I think, very bad for America.”at 49:01

AI pricing & unit economics

• Chamath Palihapitiya · Social Capital

  AI pricing will shift from per-token to outcome-based, as Google's CPC shift made online ads explode.

  Chamath argues AI is like the internet in 1999-2000 when media was sold on CPM. Right now AI is sold on cost per million input tokens, but it will move to CPC and CPA, as AdWords and AdSense did for Google. Buying outcomes rather than impressions made online advertising explode, and he expects the same once AI moves from tokens to outcomes, though token-side cannibalization still has to happen first.

  “We're effectively like in 1999 and 2000 in the Internet media world where everybody just works on CPM. ... We will move to CPC and CPA.”at 54:33

Open-source AI models

• David Sacks · Craft Ventures

  Small specialist open-weight models are taking enterprise share from frontier models and will proliferate.

  Sacks says enterprises do not need a general superintelligence or frontier model for most workloads; they need a small specialist model trained on their own data. He cites Coinbase, which fine-tuned an open-weight 9-billion-parameter Qwen model to detect fraud and cut fraudulent transactions by about 30% while being faster and much cheaper than Opus 4.5. He expects every workload to get its own small customized model, because compute is getting tighter and running a huge model on a mundane task is wasteful. He describes small models as already taking share, and says they are the most efficient use of compute and energy, and the safest because they lack frontier-level capability.

  “What you're seeing in the enterprise right now is that for a lot of enterprise workloads, you don't need a general superintelligence.”at 1:15:31

AI regulation & safety

• Ben Horowitz · Andreessen Horowitz (a16z)

  Recursive self-improvement risk is entirely theoretical; none of Databricks' four preconditions are met.

  Horowitz uses Ali Godzi of Databricks' four preconditions for models running away through recursive self-improvement: each next model must need far fewer GPUs, take less time to train, be more accurate, and this must keep repeating. He argues GPUs, debt, power and cooling make building the next model slow, so it is not running away. He says none of the four conditions are being met and the labs show no sign otherwise, so the risk is entirely theoretical. He calls the Anthropic and Bill Gates argument on this risk super dangerous, arguing it is less likely than the pandemic that gain-of-function work was meant to prevent, and that the press eats it up because it fits a narrative.

  “Zero of those four have been met. We're not on track for any of those four to be met unless something is going on in the big labs that we don't know about.”at 1:28:35