# AI capex — X 热门讨论 (2026-09-25 20:35 UTC)
## @themarketear (The Market Ear) · 09-25 18:49 · ♥32 ↻8 💬7 AI capex looks insane. Until you see the backlog. https://t.co/Db9QNgWPUI https://x.com/themarketear/status/2103557515539517542
## @TheNavroopSingh (Navroop Singh) · 09-25 07:47 · ♥30 ↻10 💬4 "Hyperscalers' Debt Is Bloating the AI & Data Centers Buildout in America"
Oracle's 5-year CDS has been sitting near 200 bps since the spring, breaking above its 2008 financial crisis peak.
AI infrastructure debt is moving into retail retirement accounts through target-date and core bond funds, and most participants have no idea they're holding any of it.
Meta's $30 billion Hyperion financing through Blue Owl is the cleanest example of how hyperscalers are funding the buildout off-balance-sheet, with the rating dependent on a parent-company guarantee.
The bull case still works, but it requires demand, power delivery, and refinancing to line up on the same timeline. If one slips, credit markets reprice before equity does.
In October 2025, Meta (META) closed the largest private debt deal in history to fund a data center campus called Hyperion in Richland Parish, Louisiana. The total cost reached around $30 billion, yet Meta does not own most of it on its books.
Meta partnered with Blue Owl Capital to create a separate company called Beignet Investor LLC. Blue Owl owns 80%, Meta owns 20%, and Beignet is the legal entity that actually owns the campus. Meta then leases it back on a long-term lease.
Beignet issued $27.3 billion in senior secured bonds due 2049 at a 6.58% coupon. PIMCO anchored $18 billion of it, while BlackRock took $3 billion, and Apollo, JPMorgan, and a syndicate of insurers absorbed the rest, with Morgan Stanley arranging the package.
The bonds earned an investment-grade rating thanks in large part to Meta’s residual value guarantee and long-term lease. If the campus loses value and Meta walks at the end of the lease, Meta covers part of the loss.
This structure lets Meta capture the economic benefit and absorb most of the risk while keeping the debt off its consolidated balance sheet. It preserves credit metrics, free cash flow, and share price.
The lease even renews every four years to avoid triggering long-term liability concerns for equity investors.
This structure now seems to be appearing across the industry, including at Microsoft, Amazon, and Alphabet. Hyperion is the largest and cleanest case study for this article because the bond documents are public.
Why Are Hyperscalers Turning To Debt Markets
Hyperscalers are now projected to spend around $700 billion in 2026, with Morgan Stanley forecasting a path to roughly $1.1 trillion by 2027. That is more than double 2025 levels, with about 75% tied to AI. Capex is now consuming the vast majority of operating cash flow at the Big Five.
Estimates put 2026 capex at roughly $700 billion against operating cash flow in the $720-$760 billion range, which leaves little room after dividends and buybacks. Amazon and Oracle are both projected to post negative free cash flow in 2026.
If the cash isn't there, the only options left are equity dilution or debt. Hyperscalers raised over $121 billion in new debt in 2025, with $90 billion of that in Q4. Bank of America forecasts $175 billion of hyperscaler issuance in 2026, more than 6x the prior 5-year average.
Wall Street is now projecting the technology sector may need to issue $1.5 trillion of new debt over the next several years to finish the AI build. That's roughly the size of the entire U.S. high-yield bond market.
Investment-grade tech is trying to issue high-yield-sized supply through investment-grade-rated structures, and creative SPVs like Hyperion make that possible.
JP Morgan estimates AI-linked companies will be 14% of its investment-grade index in 2026, a top-three sector weight inside the bond benchmark that defines what "balanced" means in American retirement accounts.
The U.S. investment-grade corporate bond market is roughly $9-$10 trillion in size. So 14% represents around $1.4 trillion of AI-linked credit exposure inside the benchmark that index-tracking and benchmark-aware funds are mandated to hold in some proportion. https://x.com/TheNavroopSingh/status/2103390977386004552
## @phl43 (Philippe Lemoine) · 09-25 18:05 · ♥34 ↻3 💬4 I think AGI-pilled people constantly make fallacious arguments, where the conclusion is only plausible if you make a lot of implicit and highly non-obvious assumptions they clearly don't realize are required (because otherwise they would point out they are needed and offer a defense of them), but at the same time I think it's a good thing, because they may not commit as much money and effort to AI if they weren't convinced that amazing things like explosive economic growth or curing most diseases within a few years, so I think that's a case where bad reasoning is conducive to progress.
There were probably many similar cases in history where people had unrealistic beliefs or beliefs that happened to be true but were based on flawed arguments, but it led them to do things that were conducive to scientific, technological or economic progress. My only concern is that capex is truly insane at this point and, if revenue doesn't increase fast enough, the damage to the economy could be significant if/when the bubble bursts. However, even in that case, it will have led to faster progress on AI that would have happened otherwise. https://x.com/phl43/status/2103546463640863190
## @stark0xbt (stark0xbt) · 09-25 17:46 · ♥32 ↻1 💬6 Gavin Baker called Micron's 2024 run. His next call is the memory trade you're about to be too late for — and Elon Musk just pivoted Tesla's fab to prove it.
Baker's frame:
"DRAM is going to be 30-40% of all hyperscaler capex. Every hundreds of billions spent goes straight to memory."
He calls it the single biggest AI bottleneck. Above lasers, power chips, NAND, HDDs.
The setup:
- Only 3 firms on Earth make the memory AI servers need - Micron locked significant supply at prices well above historical cycle floors - xAI + Tesla FSD + Optimus + every hyperscaler competing for the same 3 vendors
Baker: "$MU used to trade at a commodity discount to $ASML and $LRCX. That discount is no longer earned."
Elon Musk is stockpiling memory. Wall Street is still pricing Micron like it's 2019.
Not financial advice. > 引用 @stark0xbt: Gavin Baker owned 15% of Nvidia and 10% of Tesla sub-$2 billion. His biggest regret won't be either. It'll be the price you'll pay for SpaceX after the IPO.
"A company like this comes once in your career. Thank God I took it. Plenty of people didn't."
His analog: British East India Company. Empire scale, not tech scale.
The setup:
- 10,000 SpaceX employees could have sold every 6 months for a decade - Almost nobody did - Secondary pricing is a decade underpriced
Baker on the Elon-Anthropic partnership: "Four months ago it didn't exist. Then it made sense. Everyone has a price."
Incumbents realign around SpaceX gravity when it moves.
The IPO doesn't price the company. It prices the last window. https://x.com/stark0xbt/status/2103541701289910517