While AI stocks slide on funding fears, Nomura is calling the opposite direction. According to ITHome citing Korea's Chosunbiz, Nomura's October 20 forecast argues that despite slowing AI revenue growth and fundraising jitters, global memory chip demand will keep growing fast through 2028, the shortage will last into 2028, and storage stocks are severely undervalued.
1. A Game of Chicken Nobody Dares Leave
Nomura's global memory research team describes big tech's relentless AI investment as a game of chicken: even under free-cash-flow pressure, no giant wants to be first to slow down, since easing off means falling behind in the compute arms race. Meanwhile, long-term supply agreements have already locked in 2027 volumes and prices, underpinning industry momentum. Unless capex contracts materially, the tight supply-demand balance in memory is unlikely to reverse.
Key Facts at a Glance
· 2028: Nomura sees the memory shortage lasting through that year
· $1.043T: global semiconductor revenue in the first 8 months of 2026, up 119% YoY (Bernstein)
· ~80%: memory's share of semiconductor growth (of $443B added, $379B from price)
· Near 20%: Q3 memory contract price increase quarter-over-quarter
2. The Data Behind a Price-Led Supercycle
Bernstein's latest report footnotes Nomura's case: semiconductor revenue hit $1.043 trillion in the first eight months of 2026, up 119% year-on-year, with memory contributing roughly 80% of growth — of about $443 billion in added revenue, $379 billion came from price increases and only $64 billion from volume, showing pricing power concentrated in the hands of memory makers. Q3 contract prices rose nearly 20% quarter-over-quarter, and SK Hynix expects possibly the worst memory shortage on record next year. During the recent panic, the Philadelphia Semiconductor Index logged its longest losing streak since the 2008 financial crisis.
3. The Bottleneck Is Shifting from GPU to Memory
The underlying logic of this memory supercycle is bandwidth hunger: HBM sits next to GPU and ASIC silicon as a “second compute” that scales with model size, and OpenAI's deepening HBM cooperation with Samsung shows head players locking supply early. For China's compute supply chain, HBM adjacencies, processing-in-memory and near-memory architectures are becoming the next strategic slot after advanced process nodes — the first half of the compute race was about GPUs; the second half may be about memory. (Compiled from public reports by Wallstreetcn, ITHome, Gelonghui and C114.)