Capex Circuit @RaphFieldNotes

What if India's biggest AI winners never build a model? 20 of Goldman’s 42 AI enablers span power generation/transmission. Key point: grid capacity is the bottleneck—data centers need reliable megawatts before earning a rupee. Who has the strongest moat?

Where will AI find $4.2 trillion in new annual revenue? AI needs $6T in annual revenue to support $1.5T/year in infrastructure by 2031, Bain says. Consumer and enterprise AI may bring in only $1.2T–$1.8T—a $4.2T+ gap. What economic value closes it?

Could Big Tech's AI capex already be covered by booked demand? Goldman: Amazon, Microsoft & Google need ~$1T in 2028–30 revenue to earn 15% ROIC on AI capex. Their ~$1.7T backlog is 1.7x the hurdle. What breaks first: conversion or margins?

What can fill AI's $4.2T revenue hole by 2031? $6T/year: Bain’s revenue target to justify current data-center spending, up from $2T last year. Existing AI could deliver $1.8T; robotics, drug discovery and other bets must close the $4.2T gap. What scales?

What breaks first in the $700B AI buildout? $700B+ planned capex at 5 tech giants vs. $577B in operating cash. Debt funding jumped from 9% in 2024 to 32% by mid-2026, while bond demand slid from nearly 5x to under 2x. AI’s next bottleneck: financing.

Can AI demand soar while the data-center boom still busts? $18B of Oracle’s Project Jupiter loans trade at 89–91¢; Oracle is BBB-. AI infrastructure’s next constraint may be capital costs, not demand. Are credit spreads now the key capex metric?

Are AI revenue delays being mistaken for weak demand? Oracle: $664B RPO, triple-digit OCI growth. Infra read-through: power and data-center connections are bottlenecks, so capex may lead revenue—execution/timing risk, not demand risk. What am I missing?

How can 30% revenue growth make Oracle's debt riskier? Oracle has a $664B backlog, but guides to $90–95B capex this year and negative FCF. At BBB- and 222 bps 5-year CDS, AI spending is outrunning cash generation. What gives first?

How does $707M become $24B in just four years? JPMorgan’s IREN model lifts AI capacity from 40 MW to 2,065 MW (>50x), while revenue/MW jumps from $7.68M to nearly $13M as pricier GPUs enter service—a capital-intensive bet with little execution margin.

What breaks first when AI capex outruns cash flow? Oracle spent $28.5B on capex vs. $23.1B in operating cash flow, then invoked force majeure on an $18B OpenAI data center. AI’s next bottleneck may be financing, not compute. Who funds the gap?

What if half the GPUs sold through 2028 can't be powered? 27–37 GW by 2028: Morgan Stanley’s plausible range for the US data-center power deficit. Chip forecasts may include servers with no path to deployment. Which model adjusts first?

Could HBM break the old memory cycle? HBM needs 2.5–3x more wafers per bit than standard DRAM—and potentially 4x for HBM4E. Each ramp tightens supply, giving Samsung, SK Hynix and Micron more margin control. Do 5–7x forward P/Es reflect that?

Can a $1T AI buildout actually pay for itself? 182.7 GW: Columbia’s central case adds capacity equal to ~77M GPUs and $3.725T in mature annual revenue, implying $5.50/GPU-hour vs. today’s $6–$10+. Key sensitivity: full utilization—a demanding assumption.

What if AI token prices collapse faster than demand grows? $920B a year: Goldman’s worst-case estimate of hyperscalers’ depreciation and operating costs if capex earns 0%. How much pressure can cloud margins absorb?

How did a $45B AI fund lose $35B in 3 weeks? At up to 4x leverage, Aschenbrenner’s AI-infrastructure bet faced margin calls in the drawdown, forcing a sale to Citadel. He’s still up 80% YTD, largely from Anthropic’s 620% gain: great picks, fragile setup.

Where will the AI boom crack first? 200 bps: Oracle’s 5-year CDS is near that level—above its 2008 crisis peak. If tech needs $1.5T in new AI debt, capital intensity matters: a demand miss or refinancing squeeze could hit credit before equities.

Could credit choke AI spending before chips or power do? $1.2T of hyperscaler capex in 2027, Goldman estimates—already above operating cash flow. The next AI buildout will need debt and equity. Investment-grade credit may be the bottleneck. Priced in?

What does an $11.6B AI contract cost to fulfill? Akamai expects $1.7B of 2026 capex to pre-buy memory/components for Anthropic’s $11.6B, 7-year compute deal. Total capex: ~$5.5B, or ~47¢ per contract dollar—a capital-heavy build with overbuild risk.

Why is Nebius down 24% while its AI capacity gets pricier? 15% above its prior peak: Since June, Nvidia Blackwell capacity cleared at that premium; short-term deals hit $40–50M/MW—2x long-term rates. Near-term scarcity sets prices. What am I missing?

Can the AI capex boom survive losing a handful of buyers? 99.5% of model-serving spend on Ramp comes from the top 10% of customers paying GPU vendors. Penetration rose from <4% to ~10% in 2 years. Adoption is broadening—but how durable is demand?