# AI capex — X 热门讨论 (2026-09-23 15:34 UTC)
## @conorsen (Conor Sen) · 09-23 14:04 · ♥33 ↻5 💬6 The rate hikes need to hurt the AI capex. > 引用 @DeItaone: U.S. PMI SURGES, SIGNALING STRONG GROWTH AND RISING COST PRESSURES
U.S. business activity accelerated sharply in September, with manufacturing PMI rising to 57.0, services to 58.7, and the composite index to 58.4, all well above expectations.
The data points to the strongest growth in more than five years, supported by robust hiring and demand.
However, input costs jumped to their highest level since 2022, driven by higher energy and transport costs, adding to inflation concerns even as selling-price increases remained relatively contained. https://x.com/conorsen/status/2102761098613297302
## @SouthernValue95 (SouthernValue) · 09-23 14:31 · ♥32 ↻2 💬3 Among AI bear cases, of which there are many valid ones, I find this one the least compelling. It is very obviously not correct that 50% of chips are sitting idle in warehouses. I’d like for someone to find one of these warehouses filled with idle chips.
There has to be some timing lag from when a chip is purchased to when it is plugged in and monetized, it cannot happen in a day. It involves real setup work. But suggesting these chips are “warehoused” is both inaccurate and framed specifically to support the idea there are too many chips relative to demand.
There is ~1-2 quarters of time from when NVDA books revenue to when a hyperscaler is monetizing the chip, and hyperscalers have worked to bring these times down over time. The lag is driven by: 1) freight; 2) rack integration and testing; 3) installation, cooling, hookup, cabling, burn-in; 4) more testing.
On as AI chip revenues grow in triple digits, 1-2qtrs of “in process” chips can be a very large number! But that doesn’t mean those are warehoused chips. It means it is capex spent by hyperscalers that customers are clamoring for them to bring online as quickly as possible, and it’s why Hyperscale revenue growth took longer to accelerate than front-end supply chain picks and shovels.
Basically every server chip deployed in the last 6+ years is being used, customers are desperate for more, willing to pay ever higher prices for them, begging suppliers to produce more. The details of Ed’s analysis and assumptions can be debated (CIP isn’t the same as “in a warehouse”, some of the CIP is from DC shell builders only who don’t buy chips, DC shells stay CIP for longer than servers as a general matter so the 50-50 split may be wrong), but the basic premise of the article doesn’t pass the common sense test. > 引用 @edzitron: Free newsletter: I estimate that ~50% of AI chip sales - $200bn to $300bn+ - are sitting in warehouses, with NVIDIA selling hundreds of billions of GPUs years in advance. Data centers take way longer to build than hyperscalers are leading us to believe. https://t.co/q52aNNmPg8 https://x.com/SouthernValue95/status/2102767899429536005
## @SemiconductorsX (Semiconductor Insider) · 09-23 11:01 · ♥32 ↻2 💬3 Hyperscaler capex is exploding. Memory is taking an even bigger slice.
UBS now puts memory at 14% of AI capex in 2025, 37% in 2026, and 64% by 2027. That is $71 billion, then $367 billion, then $923 billion, while total AI spend heads toward $1.45 trillion.
This is not just more servers. Next-gen GPUs and TPUs need stacked HBM that eats about 3x the wafer area of regular DRAM. Capacity is being pulled from DDR and NAND at the same time inference, agent workloads, and storage demand rise.
Result: shortages and higher ASPs through 2027. Hyperscalers already capture most of the relevant HBM, DDR, and NAND supply. Memory has gone from a line item to the line item.
The boom is real. The mix is shifting even faster. Watch the memory names as closely as the GPU names.
If memory is 64% of 2027 AI capex, what part of that $923 billion do you think is HBM versus DDR and NAND? $MU $SKHY $NVDA $GOOGL https://t.co/G0zE3FnjT2 https://x.com/SemiconductorsX/status/2102714874120081698