📌 AI Server Expansion Tests Storage Pricing Power
The profit distribution of AI infrastructure may increasingly depend on memory, not just GPUs. Micron demonstrated a 512GB DDR5 server memory module that can configure a 24-slot server with up to 12TB. This indicates that individual servers are raising memory-capacity requirements to support AI workloads. If such high-capacity configurations enter actual deployment, storage manufacturers will be selling not only more chips, but also higher-value modules.
The cost structure of next-generation systems also makes this shift more important. According to a breakdown of Nvidia Rubin systems, DRAM accounts for 13% of bill-of-materials costs and HBM accounts for 9%, for a combined 24%. The GPU itself still makes up the largest share, but changes in memory prices or supply are already enough to materially affect total system costs, server manufacturers’ purchasing pace, and cash spending on AI infrastructure projects.
Suppliers’ bargaining power becomes more convincing only if demand-side price increases spill over from AI servers to consumer electronics. Apple reportedly accepted Samsung’s DRAM and NAND quotes for the first quarter of 2027 that were 30% to 40% higher than before. This means device manufacturers may begin bearing higher procurement costs, while memory suppliers have an opportunity to increase revenue and gross profit. However, whether end-market demand can absorb the rising costs will still determine how long the price increases can be sustained.
All three pieces of information are currently retellings from the same X account and are not accompanied by original manufacturer announcements, procurement documents, or cost lists. Memory-capacity upgrades, system cost shares, and Apple’s quotes cannot yet directly prove an industry-wide price increase. Subsequent verification should examine the actual shipment mix, contract prices, and whether memory manufacturers’ gross margins improve in tandem.
▌ Sources