1/16 the ai industry is making an unusual bet: buy compute for...
sridhar@RamaswmySridhar
91 views
Oct 29, 2026
~3 min read
Advertisement
1
1/16 the ai industry is making an unusual bet: buy compute for years, sell intelligence by the token. banks, commodity merchants, airlines, utilities and cloud providers each explain a different part of the risk.
2
2/16 banks borrow short and lend long. deposits can leave tomorrow; mortgages repay over decades. a bank earns a spread by transforming duration. its nightmare is a run.
3
3/16 ai labs almost invert this: lock in years of compute purchases, sell usage that customers can cut or reroute tomorrow. it's long-term cost versus short-term revenue, not runnable deposits versus loans.
4
4/16 the banking parallel is still real: fixed obligations amplify shocks, and access to fresh capital matters. a funding pullback can cascade. but customers can't demand their money back like bank depositors.
5
5/16 commodity merchant is closer. sign a 10-year take-or-pay lng contract, then sell cargoes at market prices. if prices or volumes collapse, the supplier still gets paid. that's the lab's compute book.
6
6/16 the math is brutal. commit $100 to capacity expecting $150 in revenue: $50 of contribution before other costs. a 30% revenue miss cuts that to $5. a small demand or pricing error can erase nearly all the spread.
7
7/16 airlines lease planes for years and sell seats one flight at a time. profit depends on load factor and yield. an empty seat at takeoff is gone forever. so is yesterday's unused gpu-hour.
8
8/16 ai adds faster obsolescence. aircraft can fly for decades; chip generations and model economics change much faster. a better model or cheaper algorithm can make a capacity plan obsolete before the contract ends.
9
9/16 utilities add another lens. power plants get financed against long-term offtake agreements. ai infrastructure is similar: secure promised payments, then finance chips, buildings and power. risk is pushed toward the offtaker.
10
10/16 gpu neo-clouds are the project-finance layer. coreweave said over 98% of its 2025 revenue came from committed contracts; lenders financed assets against pledged customer cash flows. less spot risk, more credit and hardware risk.
11
11/16 cloud providers aren't exempt. aws, azure and google cloud also commit to infrastructure years ahead of demand and sell some of it on demand. they too can overbuild.
12
12/16 their advantages: many customers and workloads, one- to three-year committed-use plans, some ability to repurpose infrastructure, and large cash-generating businesses. a frontier lab must keep winning the next model cycle.
13
13/16 but watch circularity. a csp can invest in a lab, sell it compute, distribute its models and book its spending as backlog. one ecosystem promise can look like several independent revenue streams. it isn't.
14
14/16 anthropic's new s-1 shows the scale: $265.6b in future fixed/minimum purchases and $321.4b of leases not yet commenced as of june 2026. separate long-dated buckets, some conditional—not cash due today.
15
15/16 also, don't confuse exposure with failure. anthropic had $16.3b of first-half revenue; compute expense was 55% of sales versus 201% a year earlier. the core question is whether demand keeps outgrowing committed cost.
16
16/16 the synthesis: banks teach funding contagion; commodity merchants teach spread risk; airlines teach utilization; utilities teach offtake finance; csps teach pooling. watch paid gpu use, price per unit and firm obligations. a capacity bust is more likely than a bank run.
