# AI capex — X 热门讨论 (2026-09-23 16:39 UTC)
## @amitisinvesting (amit) · 09-23 16:11 · ♥266 ↻10 💬27 Well, today was a weird day for macro data.
PMI came in at the HIGHEST level since…2022. Expectations were 53 and we came in at 57.
What does that mean? Well, people are spending, they have jobs, companies are investing…everything good for a functioning economy.
But…it means inflation has the *potential* to be more structural than just energy driven.
We got some good headlines on Iran willing to negotiate which eased some of the pain, but once that PMI data hit, stocks flipped red and bond yields went nuclear. The 10-yr yield went to a 20-year high at 5.085%.
What makes this concerning is that a rate hike was supposed to calm down inflation expectations and hopefully bond yields but this PMI data just showed the economy is doing great which could mean the Fed hikes regardless of oil coming down, which it has over the past 2 days.
We just might be living in an economy where the cost of capital is going up and ironically to beat inflation, you may have to buy stocks which could send the market higher in a world of a screaming bond market that is demanding higher yields. The opposite is also true, if yields truly get to a level where they are competitive, the AI capex story would just be fighting the Fed and no one knows if fighting a hiking cycle is going to be an easy fight to win. https://x.com/amitisinvesting/status/2102792957417976221
## @conorsen (Conor Sen) · 09-23 14:27 · ♥30 ↻3 💬10 Deep down I don’t think there’s anyone who really believes we can get inflation back to 2% without hurting some combination of (tech) stocks, AI capex, or the labor market, the only question is which piece is most important to them. https://x.com/conorsen/status/2102766685849886774
## @demian_ai (dylan ツ) · 09-23 15:18 · ♥32 ↻1 💬4 boy this is fascinating 👀
-> what if the buyers of compute become agents?
i read blackrock’s new AI paper so you don’t have to (it's secretly an inference paper)
i like how they frame this idea: 1. AI as machine native intelligence 2. digital assets as machine native money 3. agents as the glue
the tell they lean on is Stripe recently agreeing to buy Openrouter. Openrouter sits in front of 400+ models across 80+ providers (including @nebiustf) and routes by cost and fit. Stripe handles the payments. So easy to put them together and thinking that shopping for tokens could become billing infrastructure.
but let's talk about the compute part
Capex still gets the headlines. Some estimates they cite put AI build north of $5T through 2030. They have another point: watch opex. The hyperscalers are on a consensus path toward $1.1T rev by 2030. Right now training is concentrated to a few buyers, long contracts, with humans on both sides of the table.
inference is the mass market. The mckinsey chart they reprint is the whole essay in one figure: by 2030 inference is the largest slice of AI data center power, something like 43% in their cut.
-> It won 't only be chatbots, it will be millions of buyers asking for a unit of useful work: tokens, a job, reserved capacity, a region, a latency, a price.
so quite an interesting power-market logic when u think about it.
When the buyer is an agent, a normal card payment starts to feel weird (e.g the ticket might be tiny).
Blackrock points at always-on rails (stablecoins, x402-style "pay before the API answers" type) as the plumbing that might fit.
once that works, inference starts looking like something you can shop and settle without a person in the loop.
machine native intelligence needs machine native money. The quiet third term is machine native inference? > 引用 @BlackRock: Our latest research paper explores the growing connection between AI and digital assets and explains why broad AI adoption may drive new demand, utility and applications across the digital asset economy. https://t.co/z5T88Orble https://t.co/GMxQqTVmUN https://x.com/demian_ai/status/2102779672496865305