Astera says customers are rethinking HBM needs for AI accelerators

The shift in HBM demand observed by Astera Labs signals a fundamental reconfiguration in AI accelerator architecture. As external storage interfaces like NVLink and CXL demonstrate superior efficiency in KV cache scenarios, the traditional reliance on HBM for high-bandwidth applications is under pressure. This trend directly impacts memory vendors such as SK Hynix and Micron, while forcing NVIDIA and AMD to reassess their product strategies. From a geopolitical standpoint, supply chain fragmentation—particularly involving Taiwan, China and Hong Kong, China—intensifies operational risks for semiconductor firms. In the short term, NVIDIA may accelerate hybrid memory solutions, while AMD could strengthen partnerships with HBM suppliers. Over the next 12–24 months, AI infrastructure will likely adopt more flexible memory hierarchies, reducing HBM’s share, though its role in high-performance computing remains entrenched.