Credit desks see AI borrowers as safer. Markets disagree.

Credit desks mark it safer

An ECB working paper — to be precise, the findings of its researchers, not the ECB's policy position — studies how generative AI reshaped bank credit assessments. Using generative AI as a lens on lender-specific probability-of-default estimates for private firms in four large euro-area economies, the authors find banks lowered assessed default risk for borrowers in highly AI-exposed industries by about 5% after ChatGPT. The reassessment is weaker for labor-intensive firms and stronger among lenders with pre-existing industry expertise. m:…5686

The reordering is subtle. Within product markets, banks reassess firms with stronger pre-ChatGPT market positions less favorably relative to their competitors, and market leaders are experiencing relatively lower drawn and committed credit, more relationship exits, and fewer new lenders. The authors' punchline: this reordering of credit allocation predicts subsequent bad-credit events beyond banks' consensus PDs — internal bank assessments are a forward-looking measure of how AI is expected to reshape competitive risk. m:…5686

Markets mark it riskier

Equity markets spent the same week doing the opposite. UBS reports that demand for AI compute remains robust — semiconductor results, rising GPU rental prices, and TSMC expecting aggregate 2026–28 capex meaningfully higher than 2024–26 all point the same way — with AI capital expenditure estimated to grow 33% to $1.2 trillion in 2027. Yet stocks of major neoclouds came under intense pressure last week: the Nasdaq fell 4.5%, its biggest weekly loss since June, on investor concerns over growth models, interest rates, competition from China, and revenue growth at leading AI labs. m:…6593

UBS's distinction is about the shape of the business, not the demand. Neoclouds require substantial upfront investment before contracted revenue translates into cash flow, while hyperscalers have more diversified revenue streams. The bank puts numbers on the gap: hyperscalers' cost of debt sits at 6.5–7%, return on invested capital for their new data-center investments is 25% or higher, and neoclouds account for less than 10% of total AI capex. UBS favors hyperscalers and hardware over neoclouds. GPU rental economics still favor the buildout — rental rates for four-year-old GPUs are up 7% over the past month and nearly 30% year-to-date. m:…6593

The same capex, two prices

The demand side of the trade keeps printing. Apollo Global's analysis has AI driving M&A demand for data centers, power generation, grid infrastructure, cooling, and connectivity, with utilities and energy posting a record first half on AI-infrastructure demand and energy security. And in the equity-pricing corner, ARK Invest updated its open-source Tesla valuation model: the expected 2031 share price is approximately $4,400, with a bull case of about $5,000 and a bear case of about $3,300 across 32 independent inputs. That is ARK's forecast, labeled as such — the house model's view, not a consensus. m:…7742 m:…7824

This columnist's read — reasoning, not reported fact — is that the divergence is a maturity mismatch. Banks price AI exposure through private-firm default estimates and are giving AI-exposed innovators better terms, reordering credit toward the disruptors. Investors price the listed vehicles that fund the capex and demand cash-flow proof first. Lenders bet on what the technology does to their borrowers' risk; markets bet on who survives the buildout. Both can be right for a while — the tension is in how long. m:…5686 m:…6593