AI Bubble, Stablecoin Boom, and Runnin' Down a Dream | BG2 w/ Bill Gurley and Brad Gerstner

Summary

• Bill Gurley is stepping back from co-hosting BG2 after two years to promote his book “Running Down a Dream” (out late February) and work on U.S.–China relations, regulatory capture, U.S. healthcare, and nuclear — “life begins where your comfort zone ends.” Brad keeps the pod, name, and mission, with Gurley returning as an occasional guest.
• Gurley’s core AI-financing warning: he fed six “non-normal” transaction structures into ChatGPT and the model itself “would find its way toward company names like Enron and WorldCom.” The circularity started with the original Microsoft–OpenAI credits deal — a “cashless transaction” that becomes income-statement revenue — and is now competitive: “we’re fairly pregnant with it.” It raises the odds of over-provisioning while “hiding some of the signs that would tell you things are slowing down.”
• The single most peculiar deal per Gurley: Nvidia’s promise to buy any CoreWeave capacity it can’t sell — it helps CoreWeave’s debt financing but means investors “probably won’t be told” if real demand at a pure-play softens. Brad’s counter: these are disclosed public-company transactions and every analyst will now ask the question every quarter.
• Gurley’s continuum framework: judge each deal by “would this revenue have been purchased but for this investment?” Brad says Nvidia deploying a fraction of ~$450B in coming free cash flow into oversubscribed companies (OpenAI, xAI) doesn’t worry him; Gurley says he isn’t concerned as an Nvidia shareholder. A chip with one customer funded by $10B from its own manufacturer would raise red flags. Expect yellow flags “further out the risk curve” — neoclouds and startup chips desperate for capital.
• Neither calls it a bubble on multiples — Brad cites Howard Marks: “you can’t be on bubble watch if the multiples aren’t high enough” — but the scale is “remarkably unprecedented”: Mag 5 CapEx hits 66% of operating cash flow in 2025 ($379B vs $156B in 2023), consensus fading to 45–50%. Meta’s stock works this time, unlike Reality Labs, because investors see the earnings.
• Brad frames the Broadcom announcement as well over $1T of incremental CapEx; Gurley reads OpenAI’s broader deal blitz as manufacturing “escape velocity” — “daring people to follow them.” Brad’s math: OpenAI is on the hook for ~$150B of CapEx in 2030, needing at least $150B revenue — “more than plausible” for OpenAI, but it makes the game “very, very difficult for anybody else other than the hyperscalers.”
• Both want federal preemption of state AI laws — Colorado’s algorithmic-discrimination act and California’s SB 243 (private right of action for chatbot “emotional harm”) create “mud” with “zero chance” of not slowing U.S. players versus China, which is “run by engineers” while America is “run by lawyers.” Brad floats a moratorium and even AI companies blocking citizens of offending states.
• Stablecoin rails are ready now: Coinbase/Circle’s 4% “rewards” (interest by another name, post-GENIUS Act bank lobbying) plus instant, pennies-cost transfers; Gurley says, “The rails are there. They’re ready.” Brad’s call: hyperscalers — Amazon, Meta — will re-enter stables because “money is a network effects business”; Gurley: “I hope the incumbents aren’t able to strangle this thing in Washington.”

Deep dive

1. Gurley steps back

• The lead, not buried: after two years, Gurley steps back from co-hosting to create room for bigger projects — the book, plus regulatory capture, U.S.–China, and U.S. healthcare. His framing: “Life begins where your comfort zone ends.” Brad keeps the pod, the name, and the analyst-first mission, with upcoming episodes featuring Sam Altman and Sacha.
• Gurley’s send-off carries genuine weight — “I know there are going to be people that are upset with me… I’m sorry” — and credits Brad’s Invest America win as the inspiration: “You made it look easy.”

2. ChatGPT itself flags Enron — the circularity diagnosis

• Gurley’s exercise: he described six “non-normal” transactions to ChatGPT as accountant and investor, and “the AI itself would find its way toward company names like Enron and WorldCom… merely by describing the type of transaction.” His read on why AI multiples aren’t stretched: “there are red flags that people are looking at.”
• The origin story: the original Microsoft–OpenAI credits deal — credits go in as in-kind investment and come back against Azure. “That’s a cashless transaction… but it becomes an income statement revenue item for Microsoft.” The practice has since spread to Amazon and Google, and now “it’s become part of the competitive landscape” — boards saying “if we don’t do it, everyone else is doing it.”
• Via Paul Kedrosky’s argument (which Gurley finds credible): some players have loaded so much CapEx and debt that they’ve hit their limit — hence structures like Meta agreeing to cover debt risk on a facility it doesn’t own: “to me, that’s classic off-balance sheet financing. If they own the risk of it, just because they don’t own the paper… I don’t see the difference.”
• Bottom line, hedged exactly as delivered: the dynamic “increases the chance that we go over the top” — which Gurley “kind of felt was unavoidable anyway” — but worse, it “maybe pushes out when we find out,” because “you’ve created more virtual leverage on the whole system.”

3. The continuum: sham at one end, normal commerce at the other

• Gurley’s framework: at one pole, a true sham — “I send you a billion dollars, you send me the billion dollars back,” no underlying demand. At the other, massive real demand plus an incidental investment relationship. The middle is where quality of revenue gets questioned: “would this much product have been purchased but for this investment?”
• Applied to Nvidia, Brad notes the opportunity but not obligation to invest; OpenAI’s opportunity but not obligation to use the chips (evidenced by the Broadcom inference-chip announcement and the AMD deal); and Nvidia funding it from a fraction of ~$450B of free cash over 3 years. OpenAI was oversubscribed, Elon “could certainly raise it in other places” — and Gurley says he isn’t concerned as an Nvidia shareholder.
• Where Brad would worry: “imagine a chip that there’s only one customer for… that chip manufacturer gives a customer $10 billion, and that customer turns around and buys that chip.” Expect yellow flags further out the risk curve — startup neoclouds and startup chips without balance sheets or market leadership. Calling them out is how “we keep the wall of worry there.”
• Gurley’s rejoinders, worth keeping: we only know about many of these because “some auditor somewhere made them disclose them,” and equity investment is more risk-seeking than the customer loans that got Cisco in trouble — “when you switch from a loan to an equity, you no longer have to pay it back.”

4. CoreWeave’s demand backstop

• The deal Gurley calls one of the most peculiar (disclosed in a CoreWeave filing): Nvidia promised to buy any CoreWeave capacity it can’t sell elsewhere. It plausibly helps CoreWeave raise debt — but if you wanted an early warning of slowing demand, “you’d say, let’s look at one of the pure plays. And now that’s money” — investors “probably won’t be told” when offloading to Nvidia begins.
• Brad’s pushback: both are public companies, the transaction is disclosed, and “every analyst on every CoreWeave call for the next eight quarters” will ask the question. The supply chain is tracked obsessively — “look at Dylan Patel’s business at SemiAnalysis… the second they see something that smacks of any leakage in demand, boom.”

5. Not a bubble on multiples — but the CapEx is like nothing before

• On the glut question: Jensen told Brad the chance is zero for two-three years because build-out goes to hyperscalers running core businesses — “until we fully convert all general purpose computing to accelerated computing.” Gurley then said the chances are “extremely low.” The scale: ~$3T of build-out (~60 gigawatts, not all incremental) over 5 years; Nvidia consensus revenue $200B this year (~4–5 gigs) growing to ~$350B–$400B (~9 gigs) by 2029–2030.
• Brad, invoking Howard Marks: “you can’t be on bubble watch if the multiples aren’t high enough.” But “you’d have to be a fool not to notice that these numbers are so remarkably unprecedented.” His both-things-true stance: “it’s okay to recognize that the market’s great and still think that these transactions shouldn’t happen this way.”
• The chart: Mag 5 CapEx at ~66% of operating cash flow in 2025 — the peak per consensus, fading to 45–50% as cash flow grows 15–20%/year. Total CapEx: $156B in 2023 → $379B this year.
• Brad’s Meta contrast: in 2022 Reality Labs spend got the stock “obliterated” (“the CFO sent me a hat that says free cash flow”); today’s AI CapEx is tolerated because “we’re seeing the benefits in the earnings” — versus 2022’s “we don’t even know what we’re building.”

6. OpenAI’s dare: escape velocity and the $150B question

• Gurley’s read on the deal blitz: OpenAI “is trying to create escape velocity… it creates an interesting stress test for anyone else in the ecosystem: are you going to lay chase?… It feels like they’re daring people to follow them, and I suspect a bunch don’t. It may work.”
• Brad’s caveat — these are frameworks, not etched in stone (AMD and Broadcom still must deliver workable chips) — with Gurley’s needle: “wouldn’t have to be contractual for it to be RPO” at Oracle. Brad concedes the ancillary wins: recruiting (“all the best researchers want to work at the place with the most compute”) and locking supply.
• The arithmetic: summing the deals, OpenAI is on the hook for ~$150B of CapEx in 2030, requiring “at least $150 billion of revenue” — which Brad argues is “more than plausible.” The tradeable implication: scale compute “makes it very, very difficult for anybody else other than the hyperscalers” — Google, Meta, Amazon will be there; few others.

7. State AI laws: well-intentioned mud, federal preemption needed

• The trigger: Colorado’s AI Act defines “algorithmic discrimination” across 12 protected classes (including limited English proficiency, reproductive health) with liability flowing back to frontier models; Newsom just signed SB 243, giving consumers a private right of action for “emotional harm” from chatbot companions. Brad: “You can’t make this stuff up.”
• Gurley’s mechanism, honed from his regulatory-capture speech: “the intent of the policy is different from what happens once the policy is implemented.” With 50 state regimes that foreign competitors don’t face, “there is zero chance that’s not going to create mud and slow down the U.S. players.” Brad traces some of the legislative passion to Jonathan Haidt’s Anxious Generation — congressmen who feel they missed social media.
• Brad goes further: it’s “way worse for little tech,” existing law (Civil Rights Act, Fair Housing, ADA) already covers discrimination, and he wants a moratorium on all state AI laws — even wondering “whether OpenAI or some other company should consider blocking the citizens of those states.” Both support preemption; Brad notes it nearly passed in the big, beautiful bill.

8. Stablecoins: the rails are ready — watch the strangle move

• Gurley’s excitement: the Coinbase/Circle deal pays 4% on stablecoin balances — daily, from $10 to $1M — with instant, pennies-cost transfers, no savings-to-checking ACH shuffle. “The rails are there. They’re ready.” The catch Brad adds: GENIUS Act bank lobbying banned crypto firms from paying interest, so it’s “rewards” — “from a consumer perspective, a reward and interest at 4% is indistinguishable.”
• The absurdity Gurley reads aloud: the administration investigating Brazil’s PIX for “unfairly undercutting” Visa and Apple — when Visa/MasterCard “have like the top two operating incomes in the history of American business… there’s no one that needs less protection.” History’s warning: incumbents’ “go-to move” is to “run at it and strangle it,” as with debit vs credit cards.
• The bull case for openness: cheaper rails breed startups — WeChat Pay and Alipay “happened because of that government instant pay product, not in spite of it,” and Nubank’s CEO told Gurley PIX was huge for his business. “Probably bad for a laggard bank, but for a bank that embraces it, it becomes a better feature.”
• Brad’s out-on-a-limb call: hyperscalers re-enter stablecoins — “money is a network effects business,” and universality requires merchants, which Amazon and Meta have. Gurley suggests Meta “might just be kicking themselves” over Libra; Brad half-seriously proposes buying David Marcus’s Lightning startup. Gurley’s close: “I’m jumping on board the crypto train and I hope the incumbents aren’t able to strangle this thing in Washington.”

9. Trump accounts and the book

• Invest America update: now law as Trump accounts — every child under 18 qualifies (65 million kids), every child under 2 automatically gets $1,000 in a 401k-like account; website launch “maybe early December,” accounts funded and established by July 4, 2026. Bessent’s Treasury team (with Airbnb’s Joe Gebbia on front-end design) is “attacking this the way I would expect a Silicon Valley startup to.” Brad’s framing: 60% of people never own compounding assets, likely Mamdani is winning in New York — “the answer to this drift into socialism is more capitalism.”
• The book’s evidence base: Gallup finds only 23% thriving or engaged at work; Gurley’s own survey (validated with Wharton) found 6–7 in 10 would do their career differently if starting over. Daniel Pink’s “boldness regrets”: “over time we are much more likely to regret the chances we didn’t take than the chances we did… what haunts us is the inaction itself.” Brad’s confession: Bezos’s regret-minimization framework is “literally taped to my monitor.”
• Structure and audience: alternating profiles (stories of success) and principles (tools), eight years in the making, out late February — pre-order links coming, audio in Gurley’s own voice by popular demand. For the stuck, the overwhelmed teenager, and parents pushing the “lawyer doctor banker framework” — “I’m not sure that’s healthy, especially in this AI world where those jobs may be under risk as well.” His mantra: “life is a use it or lose it proposition.”
• What’s next after February: not more books but “meme flips” — like nuclear, where the Diablo Canyon episode was a small part of a movement that went “seemingly overnight from a very negative mindset to recognizing it’s very clean energy.” Target list: regulatory capture, healthcare, nuclear. “I’m kind of fired up and nervous at the same time.”