# HBM demand — X 热门讨论 (2026-09-24 20:00 UTC)

## @OrhanErgunCCDE (Orhan Ergun) · 09-24 14:45 · ♥30 ↻5 💬1 If you are a Network Engineer and want to move into AI Networking, I would follow this learning path:

AI/ML Fundamentals – Training, inference, models, transformers GPU & Compute Architecture – GPU, HBM, PCIe, NVLink/NVSwitch Distributed AI – DP, TP, PP, EP and collective operations AI Network Fabrics – Clos, rail-optimized designs, Ethernet vs. InfiniBand RDMA & RoCEv2 – RDMA, GPUDirect RDMA and transport fundamentals Congestion Control – PFC, ECN, DCQCN, buffers and incast AI Traffic Engineering – Adaptive routing, packet spraying, ECMP and UEC NCCL & GPU Communication – How GPU communication becomes network traffic AI Data Center Architecture – Scale-up/scale-out, sscale-accross, storage, multi-rail and failure domains

As a Network Engineer, you don't need to become an ML Engineer.

You need to understand **how AI workloads behave, what they demand from the network, and how to design the infrastructure around them.**

These are exactly the areas we cover in my new **AI Infrastructure & Networking Bootcamp**.

Our November session filled up very quickly, and registration is now open for the upcoming sessions. If you want to build these skills with us, reserve your place in one of them:

https://t.co/315sJtQnFb https://x.com/OrhanErgunCCDE/status/2103133640368464298

## @itsmichaelluu (Michael | Hypermarkets) · 09-23 21:43 · ♥33 ↻2 💬3 I like $DRAM calls for Jan 2028 $100 for $9 before $MU $SKHY $SNDK breakouts in the next 3 months.

$DRAM can hit $20 then $50+ at some point in the next 3 years.

Here's how it can explode if the AI trade keeps running for 3 more years:

Every AI chip needs memory. Each GPU Nvidia ships comes paired with stacks of HBM, and only three companies make it at scale: Samsung, SK Hynix and $MU.

$DRAM holds all three, roughly 75% of the fund. Why it can blow up:

• Memory is the bottleneck of the AI buildout, and new fabs take years to come online

• Every hyperscaler capex raise flows straight into memory orders

• Inference is memory hungry. More AI users = more memory demand

• Memory is cyclical. When pricing turns up, earnings don't just grow, they explode

• 2 more years of buildout = 2 more years of pricing power

The risk: this is the most cyclical corner of semis. It dropped 30%+ this summer before bouncing.

If AI spending slows, it falls hard.But if the AI trade keeps running, memory is where the leverage is. https://x.com/itsmichaelluu/status/2102876523661586735