Is The AI Debt Bubble Finally Popping? | Weekly Roundup

The sharp selloff in equities, particularly AI-related stocks, reflected a growing question: can the industry finance its enormous infrastructure buildout if revenue and returns fail to keep pace? Major technology companies are using up free cash flow and turning to debt markets, while Jensen was said to expect AI capital expenditure to reach $4 trillion by 2030. The concern is not just whether AI demand will grow, but whether companies can fund long-lived infrastructure when debt and equity financing may be unavailable or costly.

A Financial Times report reportedly put Anthropic’s ARR at 16.3, though the unit was not specified, and said the company was targeting a public valuation above $1 trillion. Its pricing roadshow had not begun and was expected in three to four weeks. The report coincided with declines in memory stocks and neoclouds as investors questioned growth, competition from open-weight models, and monetization. Those concerns do not establish that Anthropic is losing share: Gavin Baker argued that OpenAI and Anthropic are still growing collectively. Santi doubted Anthropic would reach a $100 billion run rate by year-end.

The economics and financing of AI infrastructure remain contested. Rob cited figures that Anthropic spent $1.60 on compute for each dollar of revenue, while noting that reports of positive free cash flow in Q3 relied on adjusted measures. A roughly 89% gross margin estimate for Anthropic’s enterprise API came from SemiAnalysis and was not confirmed; even high API margins do not answer whether customers will stay or switch to cheaper models. Rob said CoreWeave and related businesses depend on commitments from OpenAI and Anthropic, leaving them exposed if those customers slow spending.

Rob said borrowing costs for infrastructure companies had risen from roughly 4.5–5% a year earlier to low-teens rates on new issuances with warrant coverage, with yields around 15–16%. CoreWeave credit default swaps had at one point implied a 60% chance of default within the next few years. One speaker put CoreWeave’s debt and liabilities at roughly $40–50 billion but acknowledged limited knowledge of its finances. Rob expected painful restructurings among several businesses and said some cannot cover growth, debt obligations, and interest without additional capital; he did not expect CoreWeave to go bankrupt. The broader risk is a mismatch between large, often debt-funded capital costs and uncertain access to future financing.

The speakers expected a shakeout: tier-one AI companies may survive, while tier-four companies could be wiped out quickly and tier-three companies within a year or two. Frontier-model firms face pressure to keep spending even as advantages may last months rather than years. Open-weight models still require paid inference, and frontier providers may counter that better performance per chip lowers total inference costs. Speakers were skeptical that many venture-funded AI applications would become durable businesses, and favored established companies using AI to cut costs over businesses racing to build the most capable models. They anticipated continued volatility, including possible 10% single-day moves in AI-related stocks and crypto, and warned against leverage.

At Token 2049, Santi saw more institutional interest and fewer retail attendees than usual, with discussion centered on financial rails and institutional uses such as stablecoins, tokenization, Robinhood Chain, and prediction markets. Pearl attracted attention with a proposal to pay miners for AI inference on a Bitcoin-derived layer-one network, using matrix multiplication in a system it calls “proof of useful work.” The thesis is that miners shifting to better-paid inference could weaken Bitcoin’s hash rate while supporting Pearl’s network. A participant questioned how Pearl could verify useful computation and prevent fake or minimal work; another questioned the token’s role and whether miners would simply sell it.

The discussion also covered demand for dollar-based financial products and tokenized equities. Tokenized-equity wallet holders reportedly rose from 170,000 on January 1, 2025, to 4.7 million at the time of the discussion, while volumes increased 550-fold. Nubank’s CEO predicted that the world’s 190 currencies could shrink to 10 over 10–20 years, but participants disagreed, citing monetary sovereignty and governments’ interest in controlling money. They said U.S. stablecoins could widen access to the dollar; one speaker said stablecoins account for about 75% of volume in Nubank’s crypto wallet, with users flocking to USDC. The speakers expected blockchains to make cross-border money movement easier, while acknowledging that some governments may block it.