CXL Isn't Killing HBM. It's Handing Hyperscalers a Weapon. There's an excellent explainer making the rounds arguing that CXL — pronounced dead in 2024 — is back. I agree with the frame. The key line is that CXL adds a floor to the memory pyramid; it doesn't replace one memory with another. HBM sits on top (bandwidth), DRAM in the middle, SSD at the bottom — and CXL wedges into the gap where "lukewarm" data lives. Three things revived it, and I'd sign my name to all three: 1. Training → inference. Training wanted bandwidth (NVLink's world). Inference wants capacity — the KV cache alone eats 80–120GB per GPU. Capacity is exactly what CXL does cheaply. 2. Crisis economics. Memory went from ~8% of hyperscaler capex to ~30% in two years. At that price, pooling idle "stranded" memory and reviving DDR4 you were about to scrap (Meta's Vistara paper) suddenly pencils out. 3. PCIe finally got fast enough. A PCIe 6.0 x16 link now carries ~2.5 DDR5 channels' worth of bandwidth. The road to the distribution center is finally wide enough to use. And the tell that matters most: in 2023 CXL was vendors pushing. In 2026 it's buyers deploying — Microsoft (Azure M-series), Meta (full-stack, own ASIC). That shift, from a seller's pitch to a buyer's order, is the whole story. So who benefits most? Read it as a value chain, not a ticker: The pick-and-shovel layer wins first and cleanest: CXL controllers and memory-interface IP (Montage, Marvell, Astera Labs, Rambus, and the EDA/IP houses). Every CXL module — at every deployment stage — needs a controller. That's the layer that gets paid regardless of who wins. The biggest dollar winner is the buyer itself — the hyperscaler. The entire point of CXL is to cut the memory tax: recycle what you already own, pool what sits idle, decouple capacity from the CPU. The value CXL creates is captured largely by the people deploying it. That's the counterintuitive part everyone misses. DRAM makers are mixed. Near term, CXL modules sell more standard DRAM and open a new outlet. Long term, pooling and recycling suppress incremental DRAM demand. Cut both ways. HBM is largely insulated. CXL solves capacity, not HBM's bandwidth job. "Does HBM demand take a hit" is the wrong question — CXL is complementary, not a substitute. Personal notes for informational purposes only. Not investment advice, not a trade signal.