What TradingKey 美股財經 said in this video:
- AI Capexpositive for stockslong-termAfter NVIDIA's HBM downgrade, the AI memory bottleneck shifts from buying HBM to high-speed cross-device connectivity; capital and compute demand will be squeezed into the undervalued CXL memory connectivity technology. Morgan Stanley estimates that by around 2030 the CXL and related connectivity chip market could approach USD 6 billion.Source
- NVDAneutralNVIDIA's next-generation Rubin GPU in some versions will cut HBM capacity from 288GB to 192GB. Customers accept the lower-spec version because they are forced by capacity and cost constraints; insisting on top specs could cause material shortages preventing shipment, while installing less memory per unit allows limited capacity to ship more machines.Source
- MUneutralMicron (MU) needs to see whether more system shipments can fill the gap in memory demand caused by the lower HBM configuration.Source
- Otherpositive for stocksSanDisk may conveniently pick up some NAND flash memory demand, but this was only a hedged passing mention, and SanDisk is not a US-listed instrument.Source
- ALABneutrallong-termALAB, being much smaller than MRVL, will see its earnings elasticity greatly amplified by its small revenue base once CXL volumes ramp and the Leo product enters mass production in 2027; CXL is a 2027 super growth option for it. However, the current revenue driver is the Scorpio switch chip, and CXL still has to wait for Leo to take over.Source
- MRVLneutrallong-termMRVL has a more complete platform, able to scale from single-server expansion to rack-level shared memory; but because the company is large and diversified, CXL is more like an attractive additional upside rather than the sole lifeblood. Management expects its CXL business to contribute over USD 1 billion in revenue around 2028.Source