Damnang
· Damnang’s Substack
· July 06, 2026 at 02:55
· ⏱ 17 min read
| Read on Substack ↗
Summary
CXL (Compute Express Link) is experiencing a revival driven by the shift from AI training to inference, a severe memory capacity crisis pushing hyperscaler memory spend from 8% to 30% of capex, and PCIe 6.0 bandwidth now matching memory-channel speeds. This creates a new memory tier between DRAM and SSD, expanding total DRAM demand without cannibalizing HBM, but the article provides no explicit trade recommendations or disclosed positions.
•Hyperscaler memory spend rose from ~8% of total capex in 2023–2024 to ~30% in 2026, a nearly fourfold jump driven by AI demand.
•CXL was declared 'dead in the AI era' in early 2024 because training workloads prioritized GPU-to-GPU bandwidth over memory capacity.
•Three factors revived CXL: inference workloads require large KV caches (80–120GB per GPU) that fit poorly in expensive HBM; memory crisis makes pooling and reusing old DRAM economical (Meta's Vistara paper); PCIe 6.0 x16 delivers ~128GB/s, matching 2.5 DDR5 memory channels.
•CXL memory adds a tier between local DRAM (latency ~100ns) and SSDs (latency ~100μs), with application-level latency penalty of only 5–10μs at 95–100% throughput in Montage Technology demos.
•Deployment evolves in three stages: direct attach (current volume), pooling (shared 100TiB+ pools), and fabric (rack-scale shared memory from CXL 3.1 onward).
•Microsoft previewed CXL-attached memory in Azure M-series VMs in Nov 2025; Meta deployed CXL with revived DDR4 at hyperscale production; CXL 4.0 spec arrived Nov 2025 with PCIe 7.0 base at 128GT/s.
The article states DRAM prices are 'soaring' and hyperscalers cannot secure enough volume, and CXL modules use 'ordinary DRAM' — so every CXL-attached server adds DRAM content. Micron is a major DRAM
The article states DRAM prices are 'soaring' and hyperscalers cannot secure enough volume, and CXL modules use 'ordinary DRAM' — so every CXL-attached server adds DRAM content. Micron is a major DRAM supplier and benefits from the structural increase in total DRAM demand per server, even as CXL creates a new tier rather than replacing HBM.
Risk: CXL adoption may ramp slower than expected if PCIe 7.0 integration faces delays or if hyperscalers prioritize HBM3E capacity expansion over CXL.
The article outlines a CXL ecosystem layer including 'switches' for memory pooling, and Broadcom is the dominant supplier of PCIe switches and retimers. The shift to PCIe 6.0/7.0 and CXL 3.1/4.0 neces
The article outlines a CXL ecosystem layer including 'switches' for memory pooling, and Broadcom is the dominant supplier of PCIe switches and retimers. The shift to PCIe 6.0/7.0 and CXL 3.1/4.0 necessitates advanced switch silicon, and Broadcom's Tomahawk and Jericho families are direct beneficiaries of data center memory disaggregation.
Risk: Competition from Marvell or custom hyperscaler switch ASICs could pressure Broadcom's CXL switch market share.