# HBM demand — X 热门讨论 (2026-09-25 02:32 UTC)

## @SemiconductorsX (Semiconductor Insider) · 09-24 05:17 · ♥31 ↻4 💬4 HBM isn’t easing. JPMorgan’s CoWoS-adjusted model still shows shortage every year through 2028E.

• 2023: –32% • 2024: +1% (the only surplus year) • 2025: ~–17% • 2026E: –20% • 2027E: –19% • 2028E: –16%

Demand rises from 1.2B GB-eq in 2024 to 10.8B in 2028. Supply trails it the whole way.

The constraint isn’t just memory wafers. Extra HBM dies don’t ship without CoWoS packaging slots. That pairing is the ceiling for GPUs and ASICs. Spec cuts and more 8Hi stacks change the mix, not the deficit.

JPM still sees ~63% HBM bit demand CAGR over 2026–28 and HBM taking a much larger share of DRAM capacity.

Why it matters: shortages this long usually mean pricing power stays with suppliers and conventional DRAM gets crowded out. Watch Micron, SK Hynix, and Korea exposure.

Not a guarantee. Forecasts move. But the latest JPM cut still leaves the market tight, not loose.

$MU $SKHY $DRAM $EWY > 引用 @MojoTricks: Even after lowering assumptions, JPM still sees HBM in shortage every year through 2028E. 2024 was the only surplus year (+1%). Then it flips back and stays tight: –20% / –19% / –16%. CoWoS is the real ceiling - extra dies don’t ship without packaging slots. $MU $SKHY $DRAM $EWY https://t.co/pGL0dXrkY9 https://x.com/SemiconductorsX/status/2102990697867657421

## @jiahanjimliu (Jim Liu) · 09-24 18:45 · ♥30 ↻0 💬5 AI Infrastructure Spend Concentration

On X, there's an overwhelming amount of statistics and research on AI Infrastructure. @tryramp has some of the cleaner datapoints.

Ramp, @tryramp, is an AI startup that helps companies including a large proportion of AI Natives with AI automated accounting, accounts payable, expense management. Thus they have good visibility in what AI Natives are spending on. Ramp has 1B ARR, real SaaS ARR.

Inference Spend Concentration @arakharazian points out a legitimate concern that AI spend is heavily concentrated with 99.5% of Inference is from top 10% of companies.

There is a strong reason for this. As GPU hour prices rise so does inference token prices. What is happening is that GPU compute supply is so tight, inference prices are bid up so that only top 10% of companies can afford it.

This is capitalism: when a good becomes scarce, price increases until fewer are able to afford it so that demand can be reduced to meet supply.

Thus if the supply of GPUs increases, the price of GPU hours and thus tokens would drop and we would see the bottom 90% of firms buy more tokens.

Spend for Bare Metal is 80% of AI Applications From the long article, "Despite its much smaller customer base, spend on GPU vendors reached nearly 80% of spend on AI applications over the same period, emphasizing just how costly compute infrastructure is."

The market for raw GPU compute is extremely high and reaching the total spend on AI applications because live GPUs and behind it the HBM, Power, Datacenters are the constraints. Many AI Natives can drive revenue from bare metal GPUs. > 引用 @arakharazian: Over the last few weeks, the AI bull discourse has shifted over to AI infrastructure and away from AI models.

Why? AI models are seeing threats to their growth, due to greater competition driving price cuts and model efficiencies. Meanwhile, GPUs and infrastructure are becoming more scarce and rental rates are rising.

I'd argue AI infrastructure is not the safe bet investors think it is. As a spend category, it's even *more* concentrated than AI spend, or any other spend category we track.

The entire bet is based around a narrow group of highly correlated buyers. And simple growth in rental rates or demand does not account for outsized concentration risk.

See for example Ramp data today in Torsten Slok's daily spark https://x.com/jiahanjimliu/status/2103194176745820163