Alex Altmann says investors remain in denial about an AI bubble, while financing risk quietly builds beneath bullish consensus.
Altmann says a Barclays survey found around 60% of investors expect AI financing to be a problem, and that neoclouds like CoreWeave and Cipher carry the riskiest funding profiles.
Barclays forecasts about $1.1tn of net investment-grade and high-yield issuance this year, a break from the usual pattern where issuance rises as rates fall.
Why AI's financing cracks may start with debt-heavy neoclouds
Alex Altmann argues Wall Street remains bullish on AI while overlooking a financing problem that has already arrived. Debt-reliant neoclouds play the role subprime played in 2007, and record bond issuance despite higher rates shows capital turning selective. The trade holds only while hyperscalers keep spending and until a consumer application layer emerges to pay for the build-out.
Takeaways
AI financing stress is starting with neoclouds. Altmann sees CoreWeave and Cipher as the riskiest end of AI financing because they have little operating cash flow and rely on debt. CoreWeave's bonds sold off when Cipher tried to raise money, not because of anything CoreWeave said.
The market is in denial about financing. Altmann says consensus remains bullish and investors are still in the denial phase of a bubble. He compares it to the financial crisis, which hit subprime vehicles 12 to 24 months before banks. About 60% of investors Barclays surveyed expect AI financing problems.
AI is not immune to interest rates. Barclays data back to about 1995 shows issuance rises when rates fall. 2026 has broken that pattern, with about $1.1tn of net investment-grade and high-yield issuance forecast. The hosts agree funding is shifting from seemingly unlimited money toward selectivity.
Cheap open models squeeze frontier economics. Altmann says open-weight models have grown from about 20% to as much as 80% share and token costs have fallen to about a hundredth. Meta and NVIDIA's Nemotron offer capable local models. Weisenthal notes GPU rental rates are still rising.
Consumer applications, not frontier models, decide payoff. Altmann argues enterprise revenue will not close the revenue gap and frontier models will not change most lives. He will turn constructive once the consumer layer is solved, and predicts a 100 IQ agentic workflow will be almost ubiquitous within 12 to 18 months.
A hyperscaler capex cut is the trigger. Altmann expects doubts to surface if a hyperscaler reins in spending, which this week's results did not show. Alphabet fell 7.9% and Microsoft rose 6.4% as Azure grew 48% against a 44.8% estimate. Oracle capped extra fundraising at about $20bn through May.
The spending dwarfs anything in American history. Altmann puts AI capex at about 12% of GDP and expects about 75% of next year's S&P 500 earnings growth from AI-linked sectors. SpaceX's roughly $30tn addressable market is triple Microsoft and Apple's cumulative inflation-adjusted revenue since 1990.
Why calm indexes hide violent AI rotations
Multi-manager fund assets grew from about $300bn in 2020 to more than $1tn, according to Altmann.
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
These funds use Barra or Axioma risk models to neutralise factors such as quality, value, size and growth, so rotations show up in stocks but not in index volatility.
Weisenthal notes the big Alphabet and Microsoft moves washed out on net for indexed investors, and argues recent winning fund managers were tied to the AI factor.
Altmann concedes being short AI has probably not been lucrative over the past three years.