The technology industry’s obsessive two-year chase for GPUs is quietly mutating into a fresh supply crisis — this time centered on memory. For months, the bottlenecks that defined the AI build-out were compute accelerators and their advanced packaging. Now, the insatiable demand for high-bandwidth memory (HBM), standard DRAM, and NAND flash is rapidly tightening supplies, threatening to eclipse the earlier GPU shortages. Apple, a company known for its mastery of supply-chain orchestration, is already moving to lock down future memory capacity, anticipating the pain before most rivals.

AI models have evolved from pure compute intensity to a consumption pattern that leans heavily on memory bandwidth and capacity. Training and inference workloads running on systems stuffed with NVIDIA H100, B100, or AMD’s MI300X accelerators require massive amounts of HBM stacked directly on the processor packages. Each new generation of AI chip multiplies HBM content — the latest designs demand up to 192 gigabytes of HBM3e per accelerator. Memory fabricators, including Samsung, SK hynix, and Micron, are scrambling to shift production lines from conventional DDR and LPDDR to HBM, a complex process that cannibalizes output for mainstream chips. This conversion is one reason why server DRAM prices have climbed over 20% quarter-over-quarter, and why lead times for commodity memory have stretched to levels not seen since the 2017-2018 cycle.

The squeeze extends beyond data-center gear. Smartphones, laptops, and edge AI devices compete for the same underlying wafer starts. As memory makers prioritize high-margin HBM orders, the supply of LPDDR5 — a key component in iPhones, iPads, and Macs — gets incrementally constrained. Apple, according to multiple supply-chain sources, has recently signed multi-year agreements with its primary memory vendors, securing allocations that cover both next-generation mobile DRAM for its 2025 device ramp and enterprise-grade memory modules for the internal servers powering Apple’s expanding cloud-AI ambitions. The contracts reportedly include volume guarantees and premium pricing structures designed to jump the queue ahead of other large buyers.

For the broader industry, the memory pinch introduces a new layer of cost and complexity. Cloud providers that are mid-expansion now face not just elevated energy bills and GPU price tags, but also soaring memory expenses that directly erode return on AI infrastructure. Hyperscalers like Microsoft and Amazon have yet to report memory as a main hindrance, but analysts warn that if Apple’s preemptive moves signal a broader trend, smaller AI startups and mid-tier enterprises could be frozen out of the component market entirely. When memory becomes the scarcest resource, the ability to deploy AI services at scale may shift from compute budget to DRAM access.

This evolution mirrors the earlier packaging shortage, where a handful of companies dominated the advanced substrate and CoWoS market. Today, the memory sector is similarly consolidated: three producers control over 90% of DRAM and HBM supply. As one major buyer ties up an outsized portion of their output, everyone else must contend with higher prices and longer waits. The transition from “GPU constrained” to “memory constrained” marks a new phase of the AI infrastructure build-out, and Apple’s early positioning signals that, for those who can afford it, the new race is about cornering the silicon memory reserves before the taps run dry.

TL;DR Pendant deux ans, l’industrie technologique a vécu au rythme d’une même obsession : les GPU. Les accélérateurs de NVIDIA sont devenus le symbole de la ruée vers l’intelligence artificielle, entraînant dans leur sillage une pénurie de packaging avancé, une explosion des investissements dans les centres de données et une course mondiale aux capacités électriques. …

L’article L’IA assèche le marché de la mémoire : APPLE anticipe déjà les prochaines pénuries est apparu en premier sur FW.MEDIA.