Newsletter · · Ashutosh Agarwal

The AI Bubble Debate Gets Its First Margin Call - The VC Read - Week of August 6, 2026

A synthesis of what venture and markets podcasts said from July 30 to August 6, 2026, as an AI-concentrated hedge fund blew up, chip stocks crashed, and Jim Chanos and Gavin Baker squared off over whether the AI capital-spending boom is a bubble, for the week of August 6, 2026.

The VC Read

Week of August 6, 2026: The AI Bubble Debate Gets Its First Margin Call

The week the AI-bubble argument stopped being theoretical: a chip-stock crash, a blown-up fund, and Wall Street's most famous short-seller squaring off against its most credible AI bull. Plus, the venture model quietly bifurcates, and two founders explain what VCs get wrong. Covering podcasts from July 30 to August 6, 2026.

The Big Debate: Is AI a Bubble, and Did It Just Get Its First Margin Call?

For two years the "is AI a bubble" conversation has been a parlor game. This week it got teeth.

The backdrop was a genuine market scare. Chip stocks cratered, the Nasdaq fell roughly 10%, Meta dropped about 10% in a single stretch after it lifted its capital-spending guidance by another ~$5 billion without explaining how it earns that money back, and, the detail everyone fixated on, a hedge fund blew up. Leopold Aschenbrenner's "Situational Awareness" fund, an AI-concentrated bet that had reportedly returned 400% in six months, got a margin call and was forced to dump assets to Citadel at roughly a 10% discount. On All-In (July 31) it was billed bluntly as a "$20B Fund Margin Called." The Journal walked through the wreckage in "How the 'Nostradamus of AI' Got It Wrong" (August 4). On TBPN (July 31), fellow quant-fund founder Richard Craib argued the lesson wasn't leverage itself but concentrated volatility, "150-squared units of volatility" masquerading as skill.

That was the spark. Here's the argument it lit.

The bear case, steelmanned.

The sharpest voice was Jim Chanos, the short-seller who called Enron, on Prof G Markets, "Jim Chanos: We're In The Golden Age Of Fraud" (July 31). His case isn't "AI is fake." It's that we are in "an unprecedented CapEx boom," and "CapEx booms tend to end badly. They tend to leave behind very productive assets... but investors often get burned along the way in financing that build-out." He put valuations "right up there with 1999-2000."

His most important point is an accounting one, and it's worth slowing down for. When a hyperscaler spends billions on chips, it doesn't count that as an expense right away, it spreads (depreciates) the cost over 5 to 10 years. But the sellers, "the NVIDIAs of the world, the Caterpillar tractors of the world, the utilities," book the revenue and profit immediately. So the same dollar shows up as a big profit for the seller now, while the buyer's cost is smeared out over a decade. Chanos argues this mismatch is inflating S&P 500 profits by "hundreds of billions of dollars a year." His tell: profits are "growing somewhere like 28 or 29%" when a healthy economy would justify "8 or 9."

He layered on the circular financing, the explosion in AI debt, and debt "increasingly going off of the balance sheets of the hyperscalers and stuffed into these SPVs." And he offered a sobering historical footnote: U.S. GDP grew about 6% a year in the decade before Netscape and about 6% a year in the decade after, "for all of the wonderment of the Internet... you wouldn't have really kind of noticed it in the aggregate." Asked whether we're in "1997 or 1999," he leaned to 1999, citing the one signal he trusts most, equity issuance. It "didn't start picking up until late 98, 99" last time; now, he said, "Wall Street's printing press... is now going full bore," on track for possibly record issuance in 2026.

The most viscerally alarming version came from tech critic Ed Zitron on Better Offline, "The AI Demand Bubble with Ed Elson" (August 5). His claim: the cloud growth investors are cheering is dangerously dependent on two money-losing customers. Citing analyst work from Wells Fargo, Barclays and UBS, he said roughly 73–74% of Microsoft's AI revenue traces back to OpenAI and Anthropic, "two of the most unprofitable companies in the history of companies," with Amazon and Google similarly exposed. Then the circular part: Amazon just handed OpenAI $35 billion (more than Amazon's own estimated AI revenue for the entire year), Amazon has given Anthropic $5 billion with ~$20 billion more coming, and Google has given Anthropic $10 billion with ~$30 billion more owed. His line: "It's just people handing money to each other and being like, yep, look how profitable we are." He also skewered "annual run rate" as "a bullshit metric" and noted Meta's free cash flow is "down 91%" while Mark Zuckerberg dodged analysts' direct questions about return on the spend.

Other bears filled in the edges. On Intelligent Machines (July 30), Henry Blodgett said it's "absolutely a bubble," recalling that 99% of dot-coms failed even though the internet was real, and only a handful (Amazon, eBay, Cisco) created lasting value. On TFTC, "Yields Must Rise, Fed Must Hike with Michael Howell" (July 30), Howell said we're "probably jointly in a valuation bubble and an earnings bubble," with 25% downside to the S&P if central banks tighten, while stressing AI is a genuine must-have tool.

The bull case, steelmanned.

The most credible counter came from Gavin Baker on Invest Like the Best, "Gavin Baker – AI Market Jitters" (August 4), and what makes it credible is how self-aware it is. "I look at what's happening in the stock market and I feel like a foolish optimist," he said. "And then when I talk to people... I'm like bearish relative to essentially everyone."

His argument is built on ground-level demand, not vibes. The tell everyone worries about, falling prices, is running in reverse. He described a hot startup renting thousands of Nvidia Blackwell chips today in the "mid $2 per GPU hour" range and expecting to pay "just under $4" seven months later, prices up 50–60% in half a year. An inference-cloud company, he said, publicly plans to pay "100% more" when its contract renews. If chip rental prices are rising, that means "all the hyperscalers are under-earning," the opposite of a demand cliff.

He turned that into a credit point. Consensus models the buildout as if it monetizes at the rate of "Ampere," a chip two generations old. If it monetizes even at a discount to today's Blackwell, he figures roughly $700 billion of credit demand simply disappears, and the whole thing can be funded out of operating cash flow, "maybe all of it." His one honest bearish data point: third-party data suggests Anthropic's growth curve "started to go off of its trajectory a little bit," but, he added, "you have OpenAI and open source massively accelerating," so the sum is "net accelerating." His real worry isn't demand; it's "the bullet you don't see," a market falling on nothing obviously new except credit anxiety.

On All-In (July 31), David Sacks argued the frontier labs' public "let's slow down AI" hand-wringing is "all performative," cover, regulatory capture, and "monopoly masking." His read: frontier AI is already "a commanding duopoly," with Anthropic pushing into "70 plus billion of ARR" (annual recurring revenue) on its way toward a 10x year, from $10 billion to $100 billion, likely "110, 120," at "80 plus percent gross margins," while OpenAI's Sarah Friar said July added more net-new ARR than all of Q2. Jason Calacanis took the other side: startups are "token-maxing" cheap open-source models like Kimi K2 and DeepSeek at "80, 90% cheaper," and big customers, he named ElevenLabs, Figma, Lovable, are quietly building their own models and will defect.

And the nuanced middle: on Down the Middle, "Breakthrough or Bubble?" (July 31), Peter Mallouk of Creative Planning said this looks like the early internet, huge winners and losers, unknowable champions, and pointed out software incumbents like Salesforce and Oracle have already fallen 40–70% as the market bets AI displaces them. On the Elon Musk Podcast, "AI is the new heavy industry" (August 3), the argument was that even if the financial bubble pops, the concrete, copper, and power stay in the ground, the "dark fiber" of this cycle. And on The Rest Is Money (August 2), the case was that it's not a bubble yet: AI revenue around $110 billion annualized, growing 3.5x, has finally crossed the depreciation-cost threshold.

Where it nets out. Both sides agree on the facts: enormous capex, circular deals, two dominant labs, real and accelerating usage. They disagree on one question: does the demand curve keep bending up fast enough to earn back the spend before the financing gets nervous? Chanos says the financing always gets nervous first. Baker says the ground-truth demand is stronger than the tape. This week, for the first time in a while, the market sided with Chanos.

Signals

  • The exit machine is still jammed, and secondaries are now the plan, not the backup. On The Private Equity Podcast, "Secondaries Are No Longer a Liquidity Tool. They're a Private Equity Strategy" (August 4), Rothschild & Co's Adrian Siew described four-to-five consecutive years of IPO and M&A slowdown breaking the LP "self-funding flywheel." His striking stat: roughly 15–16% of LP distributions in 2025 came through continuation vehicles rather than real exits. Echoed on Alt Goes Mainstream with Coller Capital's Jake Elmhirst, "secondaries coming in first" (August 4). Translation for founders and GPs: the cavalry (a booming IPO window) still isn't coming.

  • The venture model is quietly bifurcating, and "the free passes are out of the system." The sharpest VC-native take of the week came from Mark Peter Davis on The Full Ratchet, "AI's Effect on the Demand Curve, Investing in Durable Barriers..." (August 3). His argument: AI erased the old built-in moat, it used to simply be hard to start a company, and that difficulty protected startups. Not anymore. So venture-scale returns now live only in companies with genuine barriers (switching costs, network effects, proprietary data) and big markets. He described a founder with a $5 million business at 80% margins but only a $10 million market, "a great business," but not venture-fundable, and the founder "made the great decision of not raising venture capital." His verdict: investing now "requires a sharper pencil... It's less forgiving than it was in 2015."

  • The power law is alive and well. On The Distribution by Juniper Square (August 4), HighVista's Raphi Schorr reminded listeners that early-stage returns are almost entirely about the 10x/100x/1000x winners, a 50–55% loss ratio is "immaterial" next to the size of the home runs. A useful antidote to the doom: the math of venture was never about avoiding losers.

  • Capital keeps concentrating in a handful of giants. On Prof G Markets, "Apple's War On OpenAI Just Got Personal" (August 5), the eye-popping figure: Anthropic has raised roughly $132 billion in venture funding plus $35 billion in debt, and still needed a $36 billion loan from Blackstone to pay for compute. The daily 20VC roundup (July 30) captured the rest of the barbell: Travis Kalanick raising $1.7 billion for Atoms, Francisco Partners raising a $21 billion fund, Etched raising $300 million to take on Nvidia, and Google Cloud growing 82% even as the market tanked. Elsewhere (AI Update, August 4): Mistral valued at $23 billion, AMD putting $5 billion into Anthropic, and Anthropic signing $10 billion-plus compute deals.

  • "Is SaaS dead?": the debate has a body count now. On The Forward Slash Podcast, "/seats for everyone" (August 5), Chris Teeling argued seven-figure enterprise software deals "will slowly die" as companies realize they use only 20–30% of the tools and instead build custom agentic software (his example: an in-house agent processing a 30,000-row spreadsheet in 15 seconds versus 20–50 hours a week). The counterweights: Workday's Gerrit Kazmaier on Tech Disruptors (August 4) argued AI enhances enterprise software rather than replacing it; Kevin Tong on Innovation with Mark Peter Davis (July 30) reframed the shift as "vertical SaaS becoming vertical AI," software that acts autonomously instead of just showing you a dashboard; and the Intuit two-parter on The Investor's Podcast, "Intuit (INTU): The S&P 500's Biggest Loser" (August 2), debated whether trusted incumbents with proprietary data survive the agent era at all.

  • Founder-vs-VC tension got two unusually candid confessions. On Masters of Scale (July 30), ButcherBox founder Mike Salguero recounted raising ~$30 million from Google Ventures, First Round and others for a custom-goods marketplace, then, when the model clearly wasn't working and he wanted to pivot, being told: "if you guys don't do what we say, you probably going to be blackballed. You probably won't be able to raise money again." He said flatly, "I lost my integrity," and bootstrapped ButcherBox afterward with no outside money. And on TruthWorks, "Why Founder Mode Is A Lie" (August 4), Redfin's Glenn Kelman diagnosed a structural problem: "VCs are a little less likely to really invest in making you better and having a hard conversation... because they're worried about deal flow. So they talk bad stuff about the entrepreneur when that person isn't there." He also flagged the mega-seed trap, "the $50 million seed round at a $400 million pre-price. Every bit of growth for the next four years is priced into that round," which turns board meetings into theater instead of problem-solving.

Quote of the Week

"I look at what's happening in the stock market and I feel like a foolish optimist. And then when I talk to people... I'm like bearish relative to essentially everyone."

Gavin Baker, on Invest Like the Best, "Gavin Baker – AI Market Jitters" (August 4). One sentence that captures the entire week: the people closest to the demand can't square what they're seeing on the ground with what they're seeing on the screen.

Runner-up, from the other side, Ed Zitron on Better Offline (August 5), on the circular financing among the AI giants: "It's just people handing money to each other and being like, yep, look how profitable we are."