The More You Buy, The More You Lose
Source: https://www.wheresyoured.at/the-more-you-buy-the-more-you-lose/
Platform: blog
Author: Ed Zitron
Original date: 2026-07-28
Summary (zh)
Zitron 把 AI 基建从叙事拉回资产负债表：超大规模厂商 2022–2026 大举堆 PP&E 与债务，却几乎不披露可验证 AI 收入；未来支出承诺暴涨，部分债券利差走阔。更关键的是买更多 AI 服务器推高 HBM/DRAM 价格，抬高下一批服务器成本，形成越买越贵、越贵越举债的循环。
Summary (en)
Ed Zitron argues hyperscaler AI CapEx and debt are compounding while verifiable AI revenue remains opaque, and that buying more AI servers raises HBM/DRAM prices and future server costs.
One-liner
算力军备竞赛的硬账：CapEx、债务与内存涨价互相放大，宣传倍数对不上利润表。
Fetched / evidence body
The More You Buy, The More You Lose：买得越多，亏得越狠
- url: https://www.wheresyoured.at/the-more-you-buy-the-more-you-lose/
- score: 8.5
- feed: Ed Zitron
- archived_at: 2026-07-29T20:14:57.277238+08:00
The More You Buy, The More You Lose
发布时间: 2026-07-28T16:29:37.000Z
原文链接: https://www.wheresyoured.at/the-more-you-buy-the-more-you-lose/
If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic and OpenAI’s finances, and the AI bubble writ large. My Hater's Guides To the SaaSpocalypse, Private Credit and Private Equity are essential to understanding our current financial system, and my guide to how OpenAI Kills Oracle pairs nicely with my Hater's Guide To Oracle, as well as the Hater’s Guide To Oracle (Part 2).
Subscribing to premium is both great value and makes it possible to write these large, deeply-researched free pieces every week.
- * *
_Soundtrack:_ _Queens of the Stone Age — Infinity_
- * *
Two years ago, NVIDIA CEO Jensen Huang said that “the more you buy, the more you save,” referring to its new (at the time) Blackwell GPUs that would “reduce LLM inference operating cost and energy by up to 25x.” Two years later, those supposed gains have been pared back to 10x, based on case studies with private inference providers that do not share their margins and are most-decidedly not profitable, and absolutely nobody seems to mind that NVIDIA overstated the gains on Blackwell (in a vacuum, in specific circumstances) by 150%, partly because these numbers are utterly meaningless, and partly because the media in most cases ardently refuses to criticize this company.
Blackwell being “10x better” than Hopper does not appear to have made any AI startups profitable (or even _more_ profitable), it does not appear to have lowered anyone’s costs in a way that we can measure using dollars and cents, and as a result, I feel very little when I’m told that Vera Rubin provides “_up to_ 10x more tokens per megawatt,” especially as that was with DeepSeek R-1, a year-and-a-half-old open source model.
Nevertheless, all of this is immaterial to the larger problem that none of this appears to have resulted in anything tangible other than horrendously-overstuffed balance sheets and spuriously-puffed stock prices.
Hyperscalers will have sunk over $1.3 trillion dollars into generative AI by the end of 2026, and have plans to spend a trillion dollars more next year. On a very rational level, nothing that large language models (LLMs) have done, do or will do in the future can or will ever bring in the more than $2 trillion (or more) in _brand new revenue_ that will be required to make any of this worth it.
To be more specific, between March 2022 and July 2026, Meta, Google, Amazon, and Microsoft added over $850 billion in property, plant, and equipment (PP&E), nearly tripling their PP&E from $498 billion or so, and in a period where they spent over $1 trillion in capital expenditures.
In that same four year period, none of them have disclosed their actual revenues from AI or AI-related services, and, as of their latest quarters, capital expenditures now represent 24.4% of Amazon’s, 33.7% of Meta’s, 37.3% of Microsoft’s, and an astonishing 43.4% of Google’s revenue, a number that’s steadily increased over the last three years.
They’ve also added over $307 billion in _on-balance sheet_ debt, leaving them with a total of $557 billion, doubled from $250 billion or so in March 2022. I mention _on-_balance sheet because Nikkei reports that Meta, Google, Amazon and Microsoft have over $1.35 trillion in _off-_balance sheet debt — either data centers/GPUs yet to be delivered, or debt raised via SPVs that shift the actual “ownership” of them over to another party as a means of making them look less-indebted than they really are.
To be clear, it’s totally _fine_ accountancy-wise to not include leases or commitments yet-to-commence, but it’s very important to know how big an anvil hyperscalers are conjuring above their heads. Google, for example, has $811 billion in contracted future spending commitments as of its latest quarter, increasing by a dramatic $661 billion ($478 billion or so in the latest quarter) in the last 6 months, and Meta has over $237 billion in non-cancellable contractual commitments.
Over $167 billion of that on-balance sheet debt has been raised in bonds across Google, Meta, and Amazon, with its $25 billion bond sale from July receiving (per Bloomberg) a cool reception, with “demand \[settling\] at 1.6 times the deal’s size…\[and to\] put that in perspective, US high-grade corporate deals have seen orders average around four times their size this year.”
For some _further_ perspective, per Freedom Broker’s Saken Ismailov, there was around $100 billion of demand for $20bn of Google’s three to fourty-year-long bonds (5x) and around £9.5 billion of demand for its £1 billion 100-year bond sale (9.5x).
As of last week, Google’s century bond has already lost 10% of its value.
This is a problem, as all four are certain to become repeat visitors to the bond markets. Herman Chan of Bloomberg Intelligence estimates that hyperscalers will need to raise $1.5 trillion in investment-grade debt in the next five years just to keep up with their trillions in estimated capital expenditures.
To make matters worse, hyperscaler bonds are, to quote Bloomberg, “...underperforming on almost every metric,” and are “in the red on average,” though that includes Oracle, whose credit just got downgraded to a single rung above junk by S&P Global.
Sidenote: As an aside, whenever you hear bonds are measured in something called “spreads,” it’s how much more an investor would expect to get paid above the current US treasury bond rate in “basis points,” with 100bps referring to 1%.
The thing is, when treasury rates go up, bond prices go down, even though you’re still getting paid on the coupon (the yield, IE: the money it pays regularly) and the payoff at the end of the bond’s life. As a result, spreads exist to tell you how much more or less it pays than an equivalent US treasury bond, or a similar ultra-low-risk government bond.
Bonds are also generally raised in tranches, in different currencies, at different lengths, which makes their prices less useful than you’d think. Furthermore, bonds can be resold to third-parties, and often for cheaper than the original purchase price — which is something that would happen if people started to get worried about the company not being able to pay back its debts.
Let’s give you a (hypothetical) example. US Treasuries are 5%, and a hyperscaler raises $10 billion in bonds at 6.5%.
The “spread” here would be 150bps — which suggests that investors think they’re mostly safe. In general, 150-300bps is worth keeping an eye on, 300bps+ is worrying, and 700bps+ is a company that the market is concerned about repayment. For example, a high-credit company's bonds would have an average OAS of 20-80bps, or an investment grade (like Amazon at 118bps) would have between 80-150bps.
And, well, then there's the rest.
For example, CoreWeave’s $1.25 billion in bonds raised in June 2026 currently sit at an option-adjusted spread of 756bps, despite being issued somewhere around 540bps on the day of issuance. Put another way, bondholders have dumped the shit out of them in the last month and have material concerns that CoreWeave won’t pay. To make matters worse, the average for its bond/credit rating (Ba3) is 200-400bps, meaning the market is really, really concerned about whether CoreWeave pays its bondholders.
All of this is a way of telling, in realtime, how much riskier a bond might be than the rock-solid guarantee of the US government (or whatever currency it was raised in). When I say “option-adjusted spread,” that’s a forward-looking model that strips out things like if a company can recall (IE: buy back a bond early) a bond to give you a generalized spread that tells you how the market feels about a particular bond.
Things can muddle a little bit depending on the company, or should I say one company in particular.
EDITOR'S NOTE: To be clear, this is an edit to the piece, because I wanted to clarify some stuff and the article I previously linked got things wrong.
Also, thank you to eagle-eyed premium subscriber Clayton for noticing!
So, SpaceX's "investment-grade" bonds (issued in late June) are rated BBB - the lowest level possible - and have an average OAS of around 270bps, with the widest spread sitting at 346 (for its 30-year dated bond). I previously linked to a piece that said it was "trading like junk," and I was completely right in that assessment, but didn't give you, the reader, the juice to understand why.
As of writing, the average spread of Bloomberg's Corporate High Yield Index - which, to be clear, is made up of junk bonds in most cases many rating levels lower than SpaceX \- is around 282bps. In other words, Musk's "investment-grade" bonds are trading like junk!
As complex as all of this sounds, it’s all pretty simple: hyperscalers have
… truncated for storage …
Obsidian evidence
OpenClaw定时任务/AK-RSS-Digest（89源精选）/2026-07-29-AK-RSS-Digest.md- Run date: 2026-07-29