Alex Feng

JPMorgan, MUFG, SMBC, and Morgan Stanley are actively shopping ways to offload data center construction debt. They have spent more than six months trying to distribute $38 billion of construction loans tied to a single Oracle leased project across Texas and Wisconsin. Some banks have already sold portions to nonbank lenders — at a discount. JPMorgan does not spend six months shopping $38 billion in construction debt unless internal risk limits are getting hit. The only way to keep originating new loans into the AI buildout is to make room first. Selling at a discount is not a negotiating posture. It is a signal that the balance sheet constraint is real and the queue of new deals behind this one is long. The structures being explored are what should be getting attention. A traditional significant risk transfer spreads exposure across dozens of loans so no single default sinks the trade. What is being shopped now is a modified single-borrower SRT — investors taking the riskiest tranche of one concentrated loan, to one operator, backed by one or two anchor tenants, carrying full construction risk. Deals in the $500M range, backed by a single name, already in market. That structure is not the diversified portfolio transfer Europeans have used for years. It is closer to the bespoke single name credit instruments that were being layered through the system in 2005 and 2006 — before anyone had stress-tested what concentration looked like when the underlying cycle turned. The cycle risk here is specific. These projects are underwritten on the assumption that frontier model spending continues indefinitely, that hyperscaler tenants honor long term leases, and that construction timelines in secondary markets hold. Maine passed a statewide data center moratorium in April, adding regulatory risk to projects that already carry construction and concentration risk. The question is not whether any of these projects default today. It is what the recovery looks like on a half built data center in a secondary market when the AI capex cycle decelerates and the anchor tenant's own model economics have changed. The banks are doing exactly what rational risk managers should do. They are identifying concentration, seeking to distribute it, and repricing where necessary.  The fact that they need to is the signal. When the institutions leading the financing of the largest infrastructure buildout in a generation start choking on deal size nine months in, that is not a footnote in the credit markets. It is the leading indicator actually. Is this the beginning of AI credit stress — or just normal risk management at the edge of a supercycle??

Some rough math! (All napkin math...) Assume Colossus 1 has 220k GPUs Assume 150k H100s, 50k H200s, 20k GB200s Pricing Assumptions: - $2.30 / hour for H100s - $2.60 / hour for H200s - $5 / hour for GB200s - blended rental rate across the entire fleet of $2.60 / hour Assume it's all take-or-pay style deals (you pay for 24x365 usage) This translates to ~$5b of annual rev to Xai. We have a new neocloud! On top of that - on recent Dwarkesh podcast, Dario ran through some napkin math on unit economics (he framed it all as industry math vs Anthropic specific - which is important, he wasn't disclosing anything Anthropic specific). What he mentioned was take $100b of compute spend (he just picked a round number). There will be a mix shift of that spend between training and inference. Skew too much on training and you don't generate enough revenue. Skew too heavy on inference and you kneecap future R&D progress. He thought the industry is currently 50/50 on training / inference of compute spend. He said as in industry, could turn that $50b inference spend into $150b of revenue (called out these are most likely the unit economics of the industry in 1-2 years) So taking this back to the Xai deal. Under above assumptions, Anthropic paying $5b / year. Let's say they turn that into $15b / year in rev (60-70% gross margin) Win win!!

Boom - nailed the N+1 Lane theory