Newsletter · 2026-08-27 · Ashutosh Agarwal

Venture Traces Who Is Actually Financing the AI Buildout - Startups and Venture - Week of August 27, 2026

Startups and venture newsletter for the week of August 27, 2026. The debate moved past whether AI is a bubble and onto who is financing it, pension funds, Nvidia and Wall Street's off-balance-sheet plumbing, right as Anthropic lines up what bankers say could be the largest IPO in history.

Startups and Venture

Week of August 27, 2026: Venture Traces Who Is Actually Financing the AI Buildout

The week venture stopped arguing about whether AI is a bubble and started tracing who is actually financing it, pension funds, Nvidia, and Wall Street's off-balance-sheet plumbing, right as Anthropic lines up a record IPO to hand the whole thing to the public.

A quick note on where this week's material came from, because the newsletter is only as honest as what fed it. What the week had, in unusual volume and quality, was a single argument told from a dozen angles: venture GPs, a former Fed president, a data-center CEO, a chip-financing desk, and the bankers running the Anthropic deal all circling the same question. Who is paying for the AI buildout, and what happens to venture if that money stops? So that is the issue.

The Big Debate: the AI buildout has become a debt machine, and Anthropic is about to ask the public to refinance it

For most of the past two months the debate on these podcasts was "is AI a bubble." This week the smarter voices moved past that and onto a more specific, more useful question: the buildout is no longer being paid for out of the hyperscalers' own cash, it is increasingly being paid for with other people's money, routed through structures most investors cannot see. And the single biggest private bet in that machine, Anthropic, just told bankers to line up the largest IPO in history. So the real question for a GP is no longer "is it a bubble." It is "if the financing hiccups, who is holding the bag, and does the exit window stay open long enough for the paper marks to turn into real cash?"

The bear case, steelmanned

Start with the mechanics, because the mechanics are the story. On the Big Technology Podcast, the hosts walked through the Wall Street Journal's dissection of Meta's Hyperion data center in Louisiana, "the size of about 1,700 football fields." Meta is building it, but "neither Hyperion nor the $27 billion in debt that's financing its construction show up on Meta's balance sheet." A Blue Owl Capital vehicle owns the campus; Meta is the minority partner and the tenant and the backstop, "guaranteeing that it would make bondholders whole if it doesn't stay the entire two decades." The kicker: "The company doesn't think payments under the guarantee are probable, so it hasn't recorded any liability on its balance sheet." As of June, Meta disclosed $347 billion in total lease obligations that have not kicked in yet, against a market cap of about $1.39 trillion. One host's summary: "the term 'bet the company' is used way too often. But this is bet-the-company type activity here." Why do it this way instead of just spending the cash they have? "If there's a buyer, why would you draw down your own cash? If you can find someone else's money… you would rather do that." And, more cynically: "Isn't part of the reason to do this to obscure it from Wall Street… so you don't take a hit? Of course. That's why you do it" (Big Technology Podcast, "Big Tech's Insane Hidden AI Spending, Ranking Anthropic vs. OpenAI," Aug 21, 2026).

Then there is the chip vendor turning into a bank. Multiple podcasts converged on the same event: Nvidia is now financing its own customers at scale. On Eurodollar University, Jeff Snider called Nvidia "a gray swan hiding in plain sight," arguing that Wall Street shadow banks are using special-purpose vehicles to lend to Nvidia's customers through vendor financing and residual-value guarantees that never touch Nvidia's balance sheet, structures he compared directly to the 2008 SIVs and to the Nortel and Lucent vendor-financing blowups of the dot-com era (Eurodollar University, "Nvidia's Hidden Debt Risk Could Crash the ENTIRE AI Bubble," Aug 21, 2026).

The most precise breakdown came from Andy Constan on Excess Returns, who took apart Nvidia's roughly $500 billion arrangement with Wall Street firms. Previously, he explained, Nvidia did these deals one data center at a time: take an equity stake, arrange a third party to lease capacity, agree to buy back the chips as they depreciate. This new structure bundles it: the data-center assets drop into a trust ("I call it a CDO"), Nvidia provides a credit guarantee, "a stop-loss for the assets," and then "senior, secured, and super-senior traditional corporate buyers," names like BlackRock and Apollo, "step up and lend money to this structure." His read on why Nvidia switched from direct deals to this: Nvidia "is out of capacity to do direct deals… but still very, very in need of supporting its customer, and just had to come up with another way to do it." He paired it with an announced $240 billion credit backstop for OpenAI's chip investment. And his real fear is not fraud, it is timing: "everybody is incented to get these deals done… for a period of time the capital markets roar. And then investors look at their portfolios and say, Jesus, what do I own? And the capital markets close. They just close." The question he is watching: "Do they close before the trillion is through the door? If they close at $500 billion and there's $500 billion of new funding that's needed… what happens?" (Excess Returns, "We Asked Andy Constan What Happens If AI Funding Breaks Before the Thesis," Aug 20, 2026).

Former New York Fed president Bill Dudley put valuation numbers on the same anxiety twice this week. The Shiller cyclically-adjusted P/E is at 41, and "the all-time high was 44 in December 1999," versus a 25 to 30-year average around 17, while the Buffett indicator (market cap to GDP) is near 240%, where Buffett called 100% "risky." His mechanical worry: it is "not the level of investment, it's the change in investment that matters," and 2027's increase in capex will be smaller than 2026's, which squeezes supplier earnings just as "the financing… is getting pretty incestuous, as the suppliers are lending to the hyperscalers and private credit is doing their own piece." His bottom line on whether the revenue ever shows up: can the hyperscalers "generate the $2 trillion of revenue they need to justify their investment?" And he expects the industry to go "from seven or eight [AI firms] to two or three… pretty painful" (Bloomberg Talks, "Former Fed President Bill Dudley Talks AI Bubble, Bond Yields," Aug 20, 2026; and The Financial Exchange Show, "The AI Bubble May Be Running Out of Time," Aug 20, 2026, where Dudley pegged the burst to end-2027).

Even a 25-year venture veteran leaned bearish on the debt specifically. Jerry Murdock, co-founder of Insight Partners (which manages over $90 billion), told 20VC that the thing making this cycle different is leverage: "So much debt. All the hyperscalers have taken on much more debt than they ever have before." His fear is a credit-market disruption, potentially triggered by the Iran war, sometime between October 2026 and March 2027, and he invoked the dot-com fiber overbuild: "the companies that laid the fiber… all went bankrupt," even though the fiber was valuable. When you are "heavily dependent on debt and there's a dislocation, the underlying value of the asset declines… and that's when you have margin calls." His most quotable prediction: of the crop of "neoclouds" (the GPU-rental startups renting out AI compute), "at least half of them go away within 36 months," and in a real dislocation "a lot of them will go away right away" (The Twenty Minute VC, "The AI Bubble Will Burst: Half the Neoclouds Will Die…," Aug 22, 2026).

The bull case, steelmanned

It is genuinely strong, and it came from people who actually run these businesses.

The best rebuttal came from a CoreWeave co-founder on The Rundown, who did something the bears mostly did not: explained why the debt is structured rather than reckless. CoreWeave carries over $30 billion of debt against $30 billion-plus of capex, and the co-founder argued investors misread it. Most of the debt sits at the asset level, backed by "multi-year take-or-pay agreements," where customers "must pay over that entire duration… they pay the same amount if they use 100% of the infrastructure or 0%." The math per dollar of revenue: "75 cents goes to paying the debt, the interest and the data-center operations, and then 25 cents, a contribution margin of 25%, goes up to the parent." Because it is self-amortizing against roughly $104 billion of backlog on 4.5 to 5-year contracts, "if we didn't build another GPU cluster… all of that debt gets paid off by the existing take-or-pay contracts." His proof the market is warming, not cooling: CoreWeave's cost of capital has collapsed "from SOFR plus 850" two and a half years ago "to SOFR plus 225" today, and its customer base has diversified from one hyperscaler to Meta, Caterpillar, and "10 of the 10 top AI labs on the planet." His sharpest point cut right at the circular-financing fear: the credit markets do the diligence equity never sees, since "the debt world is not allowed to lose money," so lenders comb every contract, every SLA penalty, every cancellation clause, which "just benefits the equity investor." He even pushed back on the GPU-obsolescence worry: CoreWeave depreciates chips over six years, and just signed a three-year lease on a 2020-vintage A100 at 2025 prices ("nine years of life… no pricing degradation"), because "not everyone needs a Ferrari," and plenty of workloads want the older, cheaper chip (The Rundown, "CoreWeave Co-Founder on Sold-Out Compute and What the Market Gets Wrong about GPUs," Aug 23, 2026).

a16z's Martin Casado made the deeper structural case for why the capital is being deployed rationally, not sprayed. His core insight: for the first time in the history of the industry, small teams can convert enormous capital directly into capability. "What would have happened 10 years ago if I gave you a billion dollars? You'd hire a ton of people… and you would blow up." Now, he said, one of the most popular multimodal models in the world "was built with a team of about 20 people," at a cost "probably $2 billion-plus." "In the history of humanity… we've never been able to have 20 people being able to productively use $2 billion." The old venture rule that too much money too early kills a company no longer holds, because "you put in $10, and then some amount out pretty directly," whereas historically "we've never been able to put in $10 and get anything back." He flipped the usual causality: it is not that AI happens to absorb capital, it is that "the more money goes to private markets, the larger the markets are going to grow," which is why companies stay private longer and accrue value privately. And he warned against judging these businesses on the balance sheet, on churn, margins and "revenue quality," when the real value is the "strategic control point." The frontier labs, he noted, have raised "$220 billion… more than the entire downstream ecosystem combined," own 95% of the market on three years of data, and can hold pricing power "just by being an epsilon higher." His forecast: supply constraints ease around 2028, after which the big labs keep roughly 80% of the market dollar-weighted but only ~60% token-weighted, with open source and the app layer taking the rest (The a16z Show, "Martin Casado on Where the Value Is Going in AI," Aug 22, 2026).

And then the demand data, which is the bulls' trump card. Wellington's head of late-stage growth, Matt Witheiler, an Anthropic and Databricks investor whose firm runs ~$10 billion private and ~$1.3 trillion public, laid out the number that makes bears nervous: Anthropic hit "$65 billion of annualized revenue run rate in July," up from "$9 billion of ARR" at the start of the year. That is "$56 billion of annualized revenue added in 7 months, the amount of revenue that Intel does in a year… Anthropic adding an Intel-sized business in 7 months." Even if that growth cools, he argued, it is "slowing from absolutely phenomenal to maybe just phenomenal," and Anthropic may already have been cash-flow positive in Q2, "a pot of gold at the end of the revenue-growth rainbow." On the fear that collapsing token prices (down "like 100x" in a year) signal weakness: "a token at an Anthropic frontier model does not produce the same amount or quality of output as a token at an open-source model… Jevons paradox wins" (Squawk on the Street, "10AM Hour: Jefferies Chief Market Strategist, Investing in AI," Aug 21, 2026).

The swing factor: Anthropic's IPO

This is why it is a venture story and not just a macro one. Anthropic's IPO is the moment the biggest private AI bet tries to convert paper into cash, and it is using the exact machine the bears are worried about to do it.

Bloomberg reported Anthropic expects an IPO that could "match or even top" SpaceX's record $75 billion raise, with a confidential S-1 already on file and a public flip possible by month-end; CFO Krishna Rao is leading early investor meetings, and 2025 net losses were "nearly $42 billion, a five-fold increase from 2024." Last private valuation: $965 billion, with bankers now debating "$1 trillion? $1.5 trillion?" And, tellingly, the same report noted Broadcom is in talks to raise more than $60 billion of debt to finance AI chips "that will benefit Anthropic and others" (Bloomberg Tech, "Anthropic Preps for Blockbuster Public Listing," Aug 21, 2026).

The Elon Musk Podcast added the specifics: a $2 trillion valuation floated (up from under $1 trillion "a remarkably short period prior"), a $25 to 35 billion target raise, Morgan Stanley, Goldman Sachs and JP Morgan running it, and the crucial circular detail: Anthropic's recent $65 billion round included "$15 billion committed directly from hyperscalers," with "Amazon alone" putting in $5 billion, money Anthropic then "turns around and spends right back on Amazon Web Services." Their analogy is the one to remember: "a railroad company funding a coal mine solely so the coal mine will pay to use its trains." On valuation, the podcast conceded the bull framing is real: at $965 billion Anthropic trades around 5x projected revenue, versus Palantir near 53x and Cloudflare around 42x, "so by the strict measure of revenue multiples, the most staggering number in AI might actually be a highly conservative bet… if the growth holds" (Elon Musk Podcast, "Anthropic's two trillion dollar IPO will defeat SpaceX," Aug 21, 2026).

So the two sides are not really disagreeing about demand. Everyone agrees AI demand is real and enormous. They are disagreeing about fragility: whether financing a permanent-seeming boom with off-balance-sheet leverage, vendor guarantees, and hyperscaler-funded customers is rational engineering (Casado, CoreWeave) or a daisy chain that snaps the moment credit tightens (Snider, Constan, Dudley, Murdock). The thing that settles it is liquidity, whether the capital markets stay open long enough for the biggest bets to cash out. Which is exactly why Anthropic filing its S-1 "as soon as next week" is the single most important venture event of the quarter. If it prices well, the bulls were right and the exit machine works. If it stalls, Constan's question, do the markets close before the trillion is through the door, stops being hypothetical.

Signals

• "Half the neoclouds die," and the winners are about capital efficiency, not scale. Jerry Murdock's neocloud cull comes with a stock-picker's logic: what separates survivors is "who is running it" and how capital-efficient it is. He would "bet on Fireworks over Base10, 10 times better business," purely on capital efficiency and willingness to earn a profit, and needled the app-layer land grab, with companies taking "zero margins just to get the real estate." He also rejected Gavin Baker's "a token is a token" line: "the more you customize the model, the more the token changes its value." On pricing discipline, he is blunt that we are "crazier than 2021," with founders now saying "we're raising 100" as a seed, though he concedes the OpenAI and Anthropic mega-rounds at $100 to 150 billion "might actually have been cheap." (The Twenty Minute VC, "The AI Bubble Will Burst: Half the Neoclouds Will Die…," Aug 22, 2026)
• Venture's liquidity crisis is spawning a whole new wrapper. On How I Invest, a secondary-market veteran of 16 years explained why he abandoned deep-value discipline to raise a $408 million closed-end fund buying "the highest-priced, high-growth companies" at full price, Stripe included, losing traditional LPs who said "we didn't invest in you to go buy Stripe at full price." The logic: a closed-end, permanently-traded wrapper is "a better way to exit" an illiquid asset class, with fees on NAV and vehicles that "last forever." His warning about how frothy the late-stage secondary market has gotten: "every Tom, Dick and Harry, every sovereign wealth fund, every family office" now buys common stock "at the last-round price or sometimes even higher," so real returns only survive in "diamonds in the rough" smaller companies where you can still get a 40% discount. He also flagged the ethics gap: "It's not illegal to have insider trading" in private secondaries, since "one party can have all the information and be selling to another that doesn't." (How I Invest with David Weisburd, "E419: Venture Capital Has a Liquidity Crisis," Aug 21, 2026)
• The application layer is quietly declaring independence from the frontier labs. On This Week in Startups, Jason Calacanis seized on Harvey, the legal-AI company that raised $200 million at an $11 billion valuation in March 2026 (led by GIC and Sequoia, originally seeded with $5 million by the OpenAI Startup Fund), releasing its own in-house model, "Harvey Tennant," post-trained on the open-weight Kimi K3. His read: big app-layer customers spending "$10, $20, $30 million a month" with OpenAI and Anthropic will move "99% of their spend… off the frontier models," because they do not want to hand their proprietary data to a vendor that might compete with them. "Open source is going to win it all… the overwhelming majority of tokens in corporate America will not be on the frontier models." His message to founders taking free tokens from a platform: "There's no free beer… nobody who went to bed with Microsoft in the 80s, Facebook in the 2000s, or Sam Altman now woke up without their throat slit." (This Week in Startups, "Open source is going to win it all: Harvey proves it | E2328," Aug 21, 2026)
• How big is "real" AI spend, and how much is just the industry paying itself? On TechStuff, Azeem Azhar shared his research team's triangulation: AI revenue across the industry hit $126 billion in the year to July 2026, "growing roughly three times per year," the fastest revenue ramp of any technology wave in history. But the composition matters: only ~20 to 25% now comes from the AI companies spending on themselves, "down from closer to 60% a year and a half ago," with just ~$4 billion from venture-backed startups, meaning demand is increasingly real enterprise money, ~80% of it enterprise rather than consumer. His measured take on Nvidia-as-financier: it is what Ford and GM did for car dealers a century ago, since "when a core supplier has the strongest balance sheet and the sector is growing very quickly, it makes sense they become the bank," but "it does introduce a new class of risk that doesn't exist if customers are always paying for everything." (TechStuff, "Everyone Is Saying Something Different About the AI Economy," Aug 21, 2026)
• The macro backdrop the venture debate keeps bumping into. On Monetary Matters, Luke Gromen tied the AI-financing surge to the government's own predicament: the SEC just changed the rules "around securitizations for anything related to AI," and Jensen Huang is "bragging about creating a compute security derivative," all of it, in Gromen's read, "making it easier to get more credit" to keep the boom going, so "the bubble's not over yet." He would take profits, not short it: "every CapEx boom smaller than this one in U.S. history, going back 200 years, has burst," and once you are two to three years into one, you would historically have done better selling and buying gold. He also flagged the employment sting hidden in the productivity story: healthcare administration is the biggest employer in 39 states and "uniquely suited to being disintermediated by AI," which "undermines the tax base." (Monetary Matters, "Why Bessent Blinked | Luke Gromen on Doubling of Treasury Buyback Plan," Aug 20, 2026)
• A regulator is now looking at venture itself. The Equity podcast covered the DOJ opening an investigation into a16z, set against a backdrop of AI-startup consolidation and the acqui-hire wave (Groq, Relay, Zipline, Wispr and others were name-checked). Worth watching as a governance and antitrust overhang on the whole "labs and their backers absorb everything" thesis. (Equity, "The DOJ is investigating a16z. What does this mean for venture capital?," Aug 21, 2026)
• The IPO tape outside AI is running hot too, and it is Chinese. Unitree Robotics surged 460% on its Shanghai debut after raising $905 million, landing at a $51 billion valuation, with Morgan Stanley pegging China's humanoid-robot market growing from $2 billion to $15 billion by 2030, a reminder that the "IPO window is open" story is not confined to Silicon Valley, and that the robotics narrative is increasingly being priced in Asia. (The Rundown, "Unitree Surges 460% In Wild IPO Debut, Anthropic Looks to One-Up SpaceX," Aug 21, 2026)
• One structural detail founders and LPs should file away. Anthropic is reportedly considering supervoting shares to keep founder control through the IPO, the same dual-class control mechanism that let founders "raise billions in the public markets while maintaining total operational authority." As the OpenAI and Anthropic S-1s land, expect governance and disclosure, not just valuation, to become a live venture debate. (Elon Musk Podcast, "Anthropic's two trillion dollar IPO will defeat SpaceX," Aug 21, 2026)

Quote of the Week

"In the history of humanity, in the history of engineering efforts, we've never been able to have 20 people being able to productively use $2 billion. What does that even mean?"

Martin Casado, a16z, on why this venture cycle breaks the old rules (The a16z Show, "Martin Casado on Where the Value Is Going in AI," Aug 22, 2026).

Runner-up, for the bears:

"For a period of time the capital markets roar. And then investors look at their portfolios and say, 'Jesus, what do I own?' And the capital markets close. They just close… Do they close before the trillion is through the door?"

Andy Constan, Excess Returns (Aug 20, 2026).