# AI capex — X 热门讨论 (2026-09-11 11:06 UTC)

## @CoinMarketCap (CoinMarketCap) · 09-11 09:53 · ♥48 ↻5 💬25 LATEST: 🏦 BIS chief Pablo Hernández de Cos warns the AI capex race is being fueled by opaque debt that could pose systemic risks to global financial stability. https://t.co/2EivQ4dTyn https://x.com/CoinMarketCap/status/2098349124085059866

## @aleabitoreddit (Serenity) · 09-11 07:17 · ♥33 ↻2 💬0 Not telling you to buy, my argument is 25-50 bps can compress forward multiples.

But unless it causes hyperscaler cuts, it's not going to break the AI capex cycle (especially when there's LTAs signed + Gov is coming into fund the buildout).

When your trillions of dollars flows into sectors like photonics/memory.

I expect companies high beta to each sector would outperform macro in the medium term, especially when the capex hits their balance sheet. https://x.com/aleabitoreddit/status/2098310007016264055

## @PhotonCap (Photon Capital) · 09-11 07:05 · ♥31 ↻2 💬0 I’m largely on the same page here.

Photonics and semiconductor including memory stocks remain two of my highest-conviction areas within the AI infrastructure trade, and there are quite a few names on this list that I like a lot.

Macro can obviously create volatility, but if hyperscaler CapEx keeps scaling anywhere close to these expectations, the underlying demand for HBM, high-end memory, lasers, and optical connectivity should remain extremely hard to ignore. > 引用 @aleabitoreddit: Yep, I'm personally high conviction on the AI trade (especially photonics/memory) despite macro.

- $NVDA gave forecasts for $1.3T hyperscaler capex ($100B+ above BofA analyst expectations)

- $SPCX Wells Fargo forecasts are ~$263B for 2027.

- $127B spend in 2027 for Anthropic/OpenAI model training + inference from internal investor documents (WSJ).

- Even the Pentagon (US gov) is looking to fund the AI buildout, with $5B potentially into Fluidstack.

What sectors do I think capex flows hit the hardest?

Photonics + Memory.

Trendforce for example est. DRAM and NAND to be 68% of hyperscaler capex in 2027.

So when all that memory procurement spend hits the balance sheets of $MU, $SNDK, Samsung, $SKHY with 70-80%+ margins.

I don't think these companies care about short term macro scares.

And when your leading memory companies are all allocating capacity for DDR5/HBM or high end memory, with little visibility on legacy memory supply coming online.

What does that say about the durability of ESMT/Etron/Winbond and others?

And when your optical content/GPU goes up for rubin ultra + $META / $GOOGL TPU v9, etc. Alongside TAM eg. $131.4 billion (+81% revision) for 2027 optical module forecasts.

Who benefits from supplying all that lasers during a shortage other than $LITE / $SIVE / $AAOI / $COHR?

I think the overwhelming fundamentals + growth of some of the AI sector companies will outperform macro. https://x.com/PhotonCap/status/2098306921526661386