The data wall, not power, is AI's real constraint: ScaleFlux's CXL bet
Key points
Key takeaways from a theCUBE interview at the CASPA CEO Summit 2026 with Greg Ernst, chief business officer and former Intel revenue leader, on ScaleFlux (October 2026):
The data wall. Power dominates the AI-infrastructure debate, but the defining constraint is memory: growing context windows and multi-user token retention outstrip what HBM can hold on a GPU, and the memory wall may not be solved for five years.
Why ScaleFlux. Ernst joined the 12-year-old company for its third-generation CXL memory controller, NAND controllers, proprietary compression algorithms and full firmware stack — assets he believes make it a disruptor in solving AI's memory problem.
CXL's HBM moment. Memory pooling across compute nodes has existed for five years without at-scale adoption; Ernst expects it to take off as AI did for HBM, with hyperscalers pooling DDR4 as they move to DDR5.
ScaleFlux inside drives. Its "ScaleFlux inside" model lets hyperscalers and OEMs build their own NAND drives — controller and firmware from ScaleFlux, NAND from any supplier — a business Ernst says is already booming.
Demand, not bubble. He pushes back on bubble narratives: Amazon is sold out through 2026-27, Google sold out for next year, and neoclouds are converting old data centres into computers — a demand curve he calls undeniable.
Latency is everything. With GPUs and TPUs getting faster, the latency between memory tiers (HBM, DRAM, NAND, hard drives) keeps growing as the constraint, pushing disaggregated architectures.
Industry as team sport. Ernst cites Intel-Nvidia custom products as proof the semiconductor industry now co-designs across rivals — the collaborative posture of the CASPA ecosystem.
Roadmap. Near term is hyperscaler/OEM engagement on the controller-firmware model; the larger opportunity is CXL memory pooling with memory, GPU, TPU and CPU vendors, plus confidential innovation projects.
Read more: theCUBE on YouTube
