Is The AI Debt Bubble Finally Popping? | Weekly Roundup
Summary
Public AI has become more volatile than crypto because markets are repricing the financing chain beneath frontier-model growth. AI-linked equities have fallen 20–30%, with 10% daily moves becoming routine, while Bitcoin held near $82,000. Anthropic’s apparent slowdown, an ARR leak around $16.3 billion, and market-share losses to open-weight models are forcing investors to ask whether today’s roughly $1 trillion of AI capex can credibly become the roughly $4 trillion that Jensen Huang appears to envision by 2030.
CoreWeave’s credit is the clearest warning that even correct long-term AI demand forecasts can produce painful near-term failures. Bonds that comparable borrowers placed near 4.5% a year ago now require low-double-digit yields and warrant coverage; CoreWeave debt traded around 15–16%, while its CDS at one point implied roughly 60% default probability over the next few years. Rob does not expect bankruptcy, but he sees “a very painful restructuring” if neoclouds cannot fund growth and service existing obligations without continually raising capital.
The unresolved Anthropic question is whether attractive enterprise API margins outweigh the enormous cost of staying at the frontier. Rob cited leaked figures suggesting $160 of compute expense per $1 of revenue and questioned whether reported third-quarter free-cash-flow positivity survives numerous adjustments; Character AI was allegedly still spending $9 per revenue dollar. Santiago’s pushback was that enterprise APIs may carry 80–90% gross margins, with consumer subscriptions functioning like Slack’s land-and-expand funnel—but open-model substitution and weak customer stickiness remain the threat.
The frontier-model race resembles the Black Queen paradox: labs must spend furiously to preserve an advantage measured in months, even when incremental capability barely matters to 98% of businesses. That shifts the more attractive economics toward trusted incumbents that can embed cheaper intelligence into existing distribution, cutting costs and lifting productivity. Open weights do not eliminate monetization because inference “still has to be run somewhere,” although the timing mismatch between eventual demand and present debt can still break providers.
The panel thinks the era of indiscriminate AI capital is ending, with survival determined by business quality rather than exposure to the narrative. Jason expects tier-four companies to disappear almost immediately, tier three within one or two years, and tier-two SaaS businesses that cannot adapt to be vulnerable; Anthropic and OpenAI may be tier one without being attractive at trillion-dollar valuations. Santiago compared the frenzy to OpenSea’s former $200 million in monthly fees: hypergrowth makes durability easy to ignore while “we’re evaporating money right now.”
Pearl emerged from Token2049 as the speculative bridge between AI inference and Bitcoin’s long-term security-budget problem. Jason said Rob had a position in the project; its miners perform matrix multiplication useful for inference, receive dollars from service providers plus Pearl issuance, and simultaneously secure the blockchain. Rob framed the thesis as Bitcoin losing hashrate to AI while Pearl gains it. Jason challenged whether miners would manufacture easy, fake workloads, but Rob argued that a Together AI inference endpoint creates demand for genuine work rather than token farming alone.
Crypto adoption is globalizing faster than regulation can unify liquidity or defend local currencies. US rules may force Hyperliquid and similar venues into isolated, KYC-gated pools, even as 65% of tokenized-stock volume on Solana occurs outside US market hours and wallet holders have risen from 170,000 to 4.7 million since January 1, 2025. Nubank’s David Vélez predicted today’s roughly 190 currencies could shrink to ten within 10–20 years as dollar stablecoins “save” the dollar, setting up a direct conflict between user choice and sovereign control.
Deep dive
1. AI equities are repricing growth before crypto does
Jason opened the market discussion with a striking cross-asset divergence: AI-linked stocks had fallen 20–30%, and 10% down days were becoming normal, while Bitcoin and broader crypto were holding up unexpectedly well. Santiago’s practical conclusion was simple: “Don’t use 100% leverage” when public AI can now move more violently than crypto.
Anthropic had reportedly begun pre-marketing ahead of a roadshow expected within three or four weeks, potentially above a $1 trillion valuation. An after-hours leak placed ARR around $16.3 billion and pressured all AI stocks and unprofitable companies; the market read it as evidence that growth was slowing.
Santiago conceded that closed models are losing share to open weights, though he preserved Gavin Baker’s counterpoint: OpenAI and Anthropic may be losing relative share while still growing rapidly in aggregate. The difficult leap is from roughly $1 trillion of infrastructure spending today to Jensen Huang’s apparent $4 trillion view for 2030.
2. CoreWeave credit is signaling a financing wall
Rob’s chain of risk runs downstream from OpenAI and Anthropic: CoreWeave is underwritten partly against commitments from those buyers, so slower frontier-lab growth weakens the contracts supporting neocloud debt. The same question then reaches hyperscalers and every capital-intensive layer built around continued capacity expansion.
Credit markets have already moved dramatically. Deals that cleared around 4.5% a year earlier now require low-double-digit rates plus warrant coverage; CoreWeave bonds traded near 15–16%, and its CDS at one point suggested roughly a 60% probability of default over the next few years.
Rob stopped short of predicting CoreWeave bankruptcy, but the financing loop looked unstable: providers must choose between funding growth and servicing debt unless they can repeatedly raise more capital. With the market almost completely closed to new debt issues and definitely closed to new share issuance, he expects debt-to-equity conversions or a similarly painful restructuring.
Santiago’s pushback was that scarce computing access and harder US data-center approvals could keep capacity valuable, while customers beyond OpenAI and Anthropic—including five or six plausible substitutes such as Meta—might assume leases. He nevertheless agreed that constrained supply cannot rescue every neocloud: “It all comes down to structuring.”
3. Anthropic’s unit economics remain the decisive unknown
Rob cited leaked Anthropic materials suggesting that, for every $1 of revenue, the company had spent $160 on compute. The materials reportedly said the company would become free-cash-flow positive in the third quarter, but the “corrected this, corrected that” wording left him unsure whether that represented genuine free-cash-flow positivity or profitability after numerous adjustments while training costs remained high.
Character AI provided the harsher specimen: even after shifting toward open-source and simpler models, its bankruptcy filings allegedly showed $9 of spending for every revenue dollar. Rob’s inference was carefully scoped but bleak—many heavily funded AI applications may have similarly weak economics despite declining model prices.
Santiago’s rebuttal centered on enterprise APIs rather than subscriptions. Citing SemiAnalysis estimates of 80–90% gross margins, he described consumer plans as a Slack-like go-to-market funnel: seed usage inside organizations, then convert that adoption into much larger enterprise API accounts.
His concession was the crucial one: high margins invite customers to switch toward Llama or other open models, making stickiness and durability uncertain. Until Anthropic’s S-1 reveals what “free cash flow positive” actually means, he expects the entire public AI complex to remain unusually nervous.
4. Falling intelligence costs favor distributors over frontier labs
The frontier race fits the Black Queen framing: “You have to run as hard as you can just to stay in place.” Labs spend extraordinary sums to preserve an advantage lasting months, while Santiago argued that each incremental capability increase makes little difference to 98% of mainstream business use cases.
That asymmetry led the panel toward incumbents with trust, distribution, and durable economics. Healthcare and established enterprises can embed existing intelligence, reduce production costs, raise employee productivity, and benefit as model costs fall—without financing the competition to build the most productive frontier model.
Open weights do not make the infrastructure layer worthless because inference “still has to be run somewhere.” Santiago noted that DeepSeek and other open models can remain expensive to serve; frontier developers may plausibly argue that superior performance per chip makes their total inference cost lower despite a higher model price.
Santiago also questioned the durability of the GPU market. Venture capitalists say older GPUs retain significant residual value, but he found that some companies mainly make money brokering old GPUs; he heard of one allegedly making $2 million a week. That reinforced his concern that money is chasing activities that may not be durable.
Jason’s hierarchy captured the coming shakeout: tier four could disappear almost instantly, tier three within one or two years, and tier-two SaaS companies could fail if they cannot adapt. Anthropic and OpenAI may remain tier one, but he would not buy either at one- or two-trillion-dollar valuations today.
5. Pearl turns AI inference into a Bitcoin-security critique
Token2049 felt more institutional and infrastructure-heavy than prior editions, with large asset managers, banks, stablecoins, tokenization, Robinhood’s network, and prediction markets dominating fewer retail attendees. The genuinely new obsession was Pearl, alongside renewed discussion of Zcash and alternatives to Bitcoin as stores of value.
Jason said Rob had a position in the project, and Rob framed Bitcoin’s declining block subsidy and miners’ migration toward higher-paying AI data-center work as a future security-budget problem. Pearl replaces part of conventional hashing with matrix multiplication used for inference—crypto’s “proof of useful work”—so the same machines can earn inference revenue while securing the blockchain.
Rob cited a Gemma 4 endpoint that he said involved Pearl and Together AI: service-provider dollars pay for actual inference, while Pearl issuance finances a discount and adds a second miner reward. Unlike Bitcoin miners, Pearl miners receiving cash for operating expenses may face less pressure to sell every token they earn.
Jason’s objection was that miners might fabricate the easiest AI-looking calculations merely to harvest issuance. Rob answered that genuine inference demand comes through paying providers, but he remained measured on displacement: Bitcoin is “in a league of its own,” while Pearl is a distinct way to access inference growth and improve proof-of-work economics. He also framed the security-budget critique as bullish for Zcash, while saying it did not imply Bitcoin would fall to zero.
6. Crypto’s global-liquidity prize collides with US market structure
Beyond Pearl, Hyperliquid’s Jeff discussed HIP-4, options, and their complementarity with perpetuals, reviving a category whose earlier onchain iterations never became effective. Backpack’s proposition similarly extends beyond an exchange toward licensed, globally shared liquidity infrastructure, with capital efficiency and cross-margining as the prize.
Jason argued that US plumbing prevents truly shared liquidity. Hyperliquid retail users coming through NinjaTrader would face KYC, FCM-provided margin, a separate HIP-3 pool, and clearing through the Bitnomial DCO; market makers would need separate onboarding, while firms such as Millennium, Citadel, and Tower would avoid offshore arbitrage.
Santiago’s counterargument was not that those giants would participate: smaller firms value cross-margining and common liquidity more, and the structure is designed to solve their capital-efficiency problem. Jason later emphasized Backpack’s day-one focus on licensing across Europe and the US, while Santiago remained doubtful that even innovation-friendly US regulators would permit a genuinely global liquidity pool.
The demand may primarily sit outside America anyway: 65% of tokenized-stock volume on Solana occurs after US market hours. Since January 1, 2025, volume reportedly increased 550-fold and wallet holders climbed from 170,000 to 4.7 million—evidence that tokenized equities currently solve foreign access more clearly than domestic trading.
7. Stablecoins are turning monetary sovereignty into a user-choice fight
Jason highlighted Nubank founder David Vélez’s prediction that roughly 190 currencies could shrink to ten within 10–20 years. His mechanism was fiscal decay: billions of people will seek offshore dollars, Bitcoin, and other assets, making dollar stablecoins “a great savior of the US dollar” by democratizing access.
Nubank supplies the working example. Across its roughly 140 million users in Brazil, Colombia, and Mexico, its initially modest crypto wallet evolved until stablecoins represented about 75% of volume; users were not primarily choosing Bitcoin, but moving “en masse to USDC.”
Jason called the broader contest a “currency cold war”: weaker countries may eventually align with the dollar, China, or the EU as citizens and economies respond to more stable external units. Blockchains add what the internet did not—freer movement of money—potentially breaking capital barriers much as the internet weakened barriers to information.
Santiago resisted the forecast of mass currency extinction. Sovereigns retain local money because it lets them support the financial pyramid; he reduced the policy choices to raising taxes, lowering taxes, or supporting the pyramid, while state power can enforce the system. China’s repeated Bitcoin restrictions and Argentina’s shadow-dollar rate show why free capital movement is “a direct threat to sovereignty.”