Musk Says AI's Biggest Bottleneck Has Shifted to Memory
Musk stated that the main constraint on AI expansion has shifted from computing power to memory, particularly memory capacity and bandwidth. He pointed to the training and inference efficiency of xAI rather than the chip supply itself.
This means the industry's competitive focus is shifting from "buying more GPUs" to "how to feed models into memory and read/write faster." Once memory becomes the bottleneck, the value of HBM, DRAM, packaging, and memory interconnects will be re-evaluated, and the expansion of computing power will have to yield to storage layer optimization.
Buyers will be more inclined to invest in the memory supply chain and advanced packaging, while selling pressure may shift from narratives focused solely on GPU shipments to those highlighting memory shortages. Event-driven funds typically first chase high memory and related equipment, then trade on who can alleviate bandwidth constraints first.
Supplementary data: Musk has previously mentioned that xAI's biggest headaches are memory usage and memory bandwidth, and this statement aligns with his consistent judgment.
Source: Public Information
ABAB AI Insight
Musk has repeatedly shifted the AI bottleneck from chips to infrastructure: first citing insufficient power, then focusing on specific links in the supply chain. This approach is consistent with how he handles rockets, factories, and charging networks, always identifying the "slowest link" first.
In terms of capital flow, money will continue to move from general GPU procurement to memory, packaging, interconnects, and system-level optimization. For xAI, the real scarcity is not in model ideas, but in the engineering capability to load, read, and rearrange larger models in shorter time frames.
Analogously, this resembles the transition from "capacity shortage" to "bandwidth shortage" in the DRAM cycle, rather than a simple boom in AI hardware. The industry's position is shifting from "expansion buying cards" to "architectural reconstruction," where the winners are no longer just those selling the most chips, but those who can enhance memory layer efficiency.
Structurally, this represents a technological substitution. Once the bottleneck shifts from computing power to memory, industry competition will transition from training scale to data handling efficiency, and the hardware value chain will migrate from a GPU-centric model to one centered on memory and packaging.
ABAB News · Cognitive Laws
Once the bottleneck changes, profits will shift tracks.
The true moat often lies in data handling.
Computing power is the beginning, but memory is the watershed.