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

2025-10-14 - 62 min - source - Read full transcript

Brad Gerstner (host)Bill Gurley (host)

Key insights

Feeding descriptions of six unusual AI vendor-financing transactions to ChatGPT causes it to independently pattern-match them to Enron- and WorldCom-style structures.

Gurley says he described the transaction types (without naming companies) to an AI and asked for analysis as both an accountant and an investor; the AI converged on historical fraud comparisons on its own, which he takes as evidence these structures carry real red flags even though none are illegal.

ai-capex-bubble

Gurley frames AI financing deals on a continuum from sham round-tripping to legitimate co-investment, with a gray-zone test in the middle.

One end is a true sham (money sent back and forth with no underlying demand); the other end is normal commercial investing plus a real purchase. The useful test for the murky middle is: would this much revenue have been purchased but for the investment? If not, the quality of that revenue is in question even if nothing illegal occurred.

ai-capex-bubble

The clearest red-flag pattern is a single-customer chipmaker funding the only buyer who could not otherwise afford the chip.

Gurley describes a hypothetical where a chip has essentially one customer, the manufacturer gives that customer $10 billion, and the customer uses it to buy the chip it could not have purchased otherwise. He says this is where real concern belongs, more than in Nvidia's current, better-capitalized customer investments.

ai-capex-bubble

Gerstner isn't worried about Nvidia's own investing behavior, only about smaller, less capitalized players further out the risk curve.

Nvidia is projected to generate about $450B of free cash flow and is investing a small fraction of that in customers like OpenAI and xAI, most of whom could raise capital elsewhere. The concern shifts to neoclouds and startup chipmakers with weaker balance sheets and no market leadership, where more desperate financing structures are likelier to appear.

ai-capex-bubble

The Mag 7's CapEx is consuming an unprecedented share of their operating cash flow, projected to peak around 66% in 2025.

Total Mag 7 CapEx rose from $156B in 2023 to $379B in 2025. Consensus has the CapEx-to-operating-cash-flow ratio falling to roughly 45-50% as free cash flow keeps growing 15-20% a year, but Gerstner flags the ratio itself as a key thing to watch, distinguishing today's AI CapEx (which investors believe in) from Meta's earlier Reality Labs spend (which tanked the stock because investors didn't know what they were buying).

ai-capex-bubble

Gerstner estimates OpenAI's aggregated deal commitments put it on the hook for roughly $150B of CapEx by 2030, which sets a revenue bar so high it effectively locks out everyone but the hyperscalers.

Summing OpenAI's announced compute and chip deals, Gerstner arrives at about $150B of CapEx by 2030, meaning OpenAI would need comparable revenue to justify it. He thinks that revenue level is plausible for OpenAI specifically, given consumer and enterprise ChatGPT usage, but the sheer scale required makes it very difficult for anyone besides OpenAI, Google, Meta, and Amazon to compete at this scale of compute investment.

ai-capex-bubble

CoreWeave's disclosed deal where Nvidia agrees to buy any unsold CoreWeave capacity is one of the more unusual structures because it could mask a real demand slowdown from investors.

This isn't a normal investment; it functions like a demand backstop that could help CoreWeave get more debt financing while obscuring whether it is actually starting to offload unsold capacity to Nvidia. Gurley notes this is exactly the kind of pure-play signal analysts would normally watch for early evidence of an AI compute glut.

ai-capex-bubble

A state-by-state patchwork of AI laws (Colorado's algorithmic discrimination law, California's SB 243) risks damaging US AI competitiveness against China and disproportionately burdens startups over incumbents.

Colorado's AI Act creates liability at the frontier-model level for discriminatory chatbot outputs across 12 protected classes; California's SB 243 gives consumers a private right of action to sue over chatbot-caused emotional harm. Both hosts argue 50 different state compliance regimes create friction that competitors operating without such rules won't face, and that smaller companies without legal staff are hit hardest; they call for federal preemption or a moratorium on state AI laws.

ai-regulation

Coinbase and Circle's stablecoin 'rewards' program is functionally identical to interest for consumers, a workaround born from Genius Act lobbying by banks.

The Genius Act banned stablecoins from paying interest after bank lobbying, so the deal is structured to pay 'rewards' instead. From a consumer's perspective, a 4% reward and 4% interest are indistinguishable, and unlike bank interest it doesn't require a direct-deposit relationship and settles instantly and cheaply.

stablecoin-payments

Payments are a network-effects business, so the hyperscalers with existing universal merchant reach (Amazon, Meta) are better positioned to build winning stablecoin rails than crypto-native issuers alone.

Visa and Mastercard's moat is that they're accepted everywhere; stablecoin issuers like Circle lack that universal merchant acceptance. Gerstner predicts Amazon and Meta will get more directly involved in stablecoin infrastructure, partly because Meta already tried this with Libra and has WhatsApp distribution to leverage.

stablecoin-payments

Research replicated with a Wharton sample found roughly six to seven in ten people would restart their career differently, and the dominant regret is inaction rather than the risks they actually took.

Gurley cites a survey he ran (70% said yes) and a follow-up Wharton academic survey (60% said yes) asking people if they'd redo their careers differently. He connects this to Daniel Pink's 'boldness regrets' research, which finds people regret the chances they didn't take far more than the ones they did, regardless of domain.

career-and-life-purpose

Invest America ('Trump accounts') will auto-seed every US child under 2 with $1,000 in a portfolio account, positioned as capitalism's answer to rising anti-capitalist political sentiment.

Every child under 18 already qualifies, and starting around a launch in early December, kids under 2 get an account automatically seeded with $1,000 they can roll into a brokerage of their choice; accounts must be funded and established by July 4, 2026. Gerstner frames this against the backdrop that 60% of people never own compounding assets, and cites rising anti-capitalist political wins as the problem it's meant to counter.

capitalism-and-invest-america

Books referenced

Running Down a Dream: How to Thrive in a Career You Actually Love - Bill Gurley - Gurley's upcoming book (out late February), the main topic of the episode's second half; built around alternating profiles and principles.
The Anxious Generation - Jonathan Haidt - Gerstner cites it as background for why state legislators are rushing to regulate AI chatbots the way they wish they had regulated social media.
The Coddling of the American Mind - Jonathan Haidt and Greg Lukianoff - Cited for the 'resume arms race' framing of how kids are pushed toward grinding rather than passion.
The Power of Regret - Daniel Pink - Source of the 'boldness regrets' research Gurley leans on to argue people regret inaction more than failed risks.
Atomic Habits - James Clear - Mentioned because Clear retweeted and republished a transcript of Gurley's original 'Running Down a Dream' talk, helping it go viral.

Media referenced

Plain English - podcast - Paul Kedrosky's episode is cited for the theory that circular AI financing deals happen because some players have already taken on so much CapEx and debt they can't take the next funding step directly.
Founders - podcast - David Senra's podcast is cited as a channel that has amplified Gurley's original career presentation to a wider audience.
BG2 Pod: Jensen Huang episode - podcast - Referenced repeatedly for Huang's answer that the odds of an AI compute glut in the next 2-3 years are close to zero.
BG2 Pod: Diablo Canyon episode - podcast - Gurley names it one of his most rewarding episodes, crediting it with helping shift public sentiment toward nuclear energy.
BG2 Pod: China trip episode - podcast - Gurley's other favorite episode, framed as a personal learning expedition rather than a normal recording.

Companies

OpenAI - Center of the circular-revenue debate; its Microsoft credit deal, new Broadcom chip, and AMD deal are all discussed as evidence of both risk and non-binding deal structure.
Broadcom - New OpenAI partnership to build a proprietary inference accelerator, adding well over a trillion dollars of incremental CapEx to the AI build-out.
Nvidia - Its customer investments (OpenAI, xAI, CoreWeave) are scrutinized for circularity, but Gerstner says he isn't concerned given Nvidia's ~$450B free cash flow and low leverage.
Microsoft - The original OpenAI deal, where in-kind Azure credits were booked as OpenAI revenue, is identified as the origin point of circular-revenue practice in AI.
CoreWeave - Disclosed in an SEC filing that Nvidia will buy any of CoreWeave's unsold service availability, flagged as an unusually opaque structure that could hide real demand weakness.
Meta - Discussed for its CapEx-to-free-cash-flow ratio, its earlier Libra stablecoin project, and as a candidate to re-enter stablecoin payments given its merchant and consumer reach.
Amazon - Named alongside Meta as best positioned to build competing stablecoin rails because it already has universal merchant acceptance.
Coinbase - Pays users 4% in stablecoin 'rewards' (structured to avoid the legal ban on stablecoin interest) with instant settlement.
Circle - Issuer behind the stablecoin used in the Coinbase rewards deal.
Visa - Cited as the payments incumbent facing tokenization disruption, and named in Brazil's antitrust probe into PIX for allegedly undercutting it.
SemiAnalysis - Dylan Patel's firm is cited as tracking the AI chip supply chain closely enough to catch early signs of demand slowdown.
Oracle - Raised as an example in the discussion of whether AI deal announcements are contractually binding or just RPO frameworks.
AMD - OpenAI's new chip deal with AMD is cited as evidence that these big AI announcements are frameworks, not guaranteed commitments.
Cisco - Historical comparison point: Cisco's customer-loan practices got it in trouble in a prior cycle, contrasted with today's equity-style AI vendor financing.

Techniques and frameworks

Circular-revenue continuum - Gurley's framework for judging AI vendor-financing deals, ranging from sham round-tripping with no underlying demand, to legitimate co-investment, with a gray-zone test of whether the revenue would exist but for the investment.
Regret-minimization framework - Bezos's decision tool for leaving D.E. Shaw to start Amazon, imagining himself at 80 looking back; Gerstner says he has it taped to his monitor.
CapEx as a percent of operating free cash flow - Gerstner's metric for tracking how much risk the Mag 7 are taking on with AI spend, currently peaking near 66% in 2025.

Summary

Bill Gurley opens by announcing he's stepping back from co-hosting BG2 Pod to focus on his upcoming book and other passion projects, then the two spend the bulk of the episode on the AI CapEx debate that has followed their prior Jensen Huang episode. Gurley lays out a framework for judging the wave of circular-sounding AI financing deals: a continuum running from outright sham round-tripping (no underlying demand) to ordinary co-investment alongside real purchases, with a gray-zone test in between - would this revenue exist but for the investment? He traces the practice back to Microsoft's original OpenAI deal, where in-kind Azure credits were booked as revenue, and flags CoreWeave's disclosed arrangement where Nvidia will buy any unsold capacity as a genuinely unusual structure because it could hide a real demand slowdown from investors.

Gerstner pushes back gently on the panic framing, distinguishing Nvidia's own investing behavior (low leverage, huge free cash flow, mostly well-capitalized counterparties) from riskier, less capitalized players further out the curve like startup neoclouds and chipmakers. They walk through hard numbers: Mag 7 CapEx has grown from $156B in 2023 to $379B in 2025, consuming a peak 66% of operating cash flow this year before an expected decline to 45-50%. Gerstner estimates OpenAI's aggregated deal commitments put it on the hook for roughly $150B of CapEx by 2030, a bar he thinks OpenAI specifically can clear given its usage growth, but one that locks nearly everyone else out of the compute-scale race except Google, Meta, and Amazon.

The conversation pivots to AI regulation, where both hosts are alarmed by a growing patchwork of state laws - Colorado's algorithmic discrimination statute and California's SB 243 chatbot liability law - that they argue burdens startups more than incumbents and damages US competitiveness against China. They call explicitly for federal preemption or a moratorium on state AI legislation.

A shorter stablecoin segment covers Coinbase and Circle's 4% "rewards" program, which the hosts note is functionally identical to interest but structured to route around a Genius Act ban on stablecoin interest that resulted from bank lobbying. They frame payments as a network-effects business and predict Amazon and Meta, who already have universal merchant reach, are best positioned to build competing rails, unlike crypto-native issuers alone.

The back half turns personal: Gerstner gives an update on Invest America ("Trump accounts"), which will auto-seed every child under 2 with $1,000 in a brokerage-style account by mid-2026, framed as capitalism's answer to rising anti-capitalist political sentiment. Gurley then discusses his book Running Down a Dream, which grew out of a talk he gave at UT Austin's MBA program that went viral through James Clear and David Senra. He cites research (replicated with a Wharton sample) that roughly 60-70% of people would restart their careers differently, and leans on Daniel Pink's "boldness regrets" work and Bezos's regret-minimization framework to make the case that people should take more career risk, not less.

Notable Quotes

"There's this continuum. On one end of the continuum is a true sham transaction. There's no underlying demand for the product. I send you a billion dollars, you send me the billion dollars back." - Bill Gurley

"Imagine there's a chip that there's only one customer for... and that chip manufacturer gives a customer $10 billion, and that customer turns around and buys that chip. That to me raises big red flags." - Bill Gurley

"If we implement 50 different state rules that these companies have to jump through, and companies that are competing in the broader world don't have any of them, there is zero chance that's not going to create mud and slow down the US players." - Brad Gerstner

"A reward and interest if it's 4% is indistinguishable [to a consumer]." - Brad Gerstner

"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." - Bill Gurley, citing Daniel Pink's The Power of Regret