Nanya Technology sees the global memory market in 2026 reaching over $800 billion, according to a July 21 speech by company executives reported by Economic Daily. The World Semiconductor Trade Statistics (WSTS) spring 2026 forecast also projects memory at over $800 billion and the overall semiconductor market at $1.51 trillion. That puts memory alone at roughly 53% of the total—more than half. As AI applications expand from model training to inference, pushing up memory content from servers to devices, the sharp rise in market value is also strongly driven by supply shortages and price increases.
$800 Billion Exceeds Last Year's Entire Global Semiconductor Market
WSTS expects memory revenue in 2026 to grow approximately 250% year-over-year, surpassing $800 billion. The forecast for the overall semiconductor market is $1.51 trillion, up 90% year-over-year, with memory serving as the largest growth driver. By comparison, logic is expected to grow 37%, microprocessors 20%, and analog just 10%—a stark divergence in growth rates across product categories.
Comparing this to the previous year helps put the scale in perspective. The Semiconductor Industry Association (SIA), in a report published in February 2026, put 2025 global semiconductor sales at $791.7 billion, of which memory accounted for $223.1 billion. Under the WSTS 2026 forecast, memory—a single product category—would exceed the entire semiconductor market's total from the previous year. Nanya Technology's explanation that memory's share will rise to more than half aligns with WSTS's figures.
However, the $800 billion figure represents sales value, not shipment bit volume. As high-priced High Bandwidth Memory (HBM) increases its share and DRAM and NAND prices rise, the market size expands beyond what volume growth alone would suggest. WSTS cites AI infrastructure, HBM, and accelerated computing as the primary growth drivers. For 2027, the forecast calls for overall semiconductor growth of 27% to approximately $1.9 trillion, with memory growing 32%—both of which are revenue forecasts incorporating both demand and price.
How Far Will Inference Push Memory Demand?
Wu Chih-hsiang, Senior Vice President of Strategic Planning and Management at Nanya Technology, explained the mechanism by which AI inference is expanding memory demand at a forum held by the Industrial Technology Research Institute (ITRI) in Taipei on July 21. According to Economic Daily, in 2026 DRAM bit demand is expected to be composed of roughly 10% from HBM and over 40% from conventional server DRAM. Combined, these two categories already exceed 50% and could rise to over 60% going forward.
In training, HBM bandwidth—which continuously feeds data to numerous compute units—determines performance. In inference, HBM demand for data center GPUs persists, but at the same time, large-capacity DRAM on the CPU side and KV cache for maintaining conversation history also expand. Wu Chih-hsiang explained, as configuration examples, that a GPU equipped with eight stacks of 12-layer HBM could require 96 high-density DRAM dies, while an inference-oriented CPU platform could require over 200. Regarding AI PCs, he noted the possibility of memory increasing from the conventional 16GB or 24GB to 128GB.
Here, interpreting "the spread of inference" as "the retreat of HBM" would miss the reality. Nanya Technology's own market outlook for Q3 2026 states that development of AI-oriented GPUs, CPUs, and ASICs is pushing up demand for HBM, LPDDR5, and DDR5 simultaneously. In cloud inference, HBM and server DRAM both increase, while on the device side, different configurations optimized for low power consumption and short response times are required. Rather than narrowing memory demand to a single type, inference branches demand across different use cases.
Prices and Product Mix Jumped Ahead of Shipment Volume
Nanya Technology's Q2 2026 earnings starkly illustrate the mechanism behind rising market value. DRAM bit shipments remained flat quarter-over-quarter, but average selling prices rose in the 60% range. As a result, revenue increased 68.2% to NT$82.549 billion, and gross margin rose to 79.5%. For this company at least, quarterly growth is driven far more by price than by volume.
What's supporting prices is the concentration of demand on AI-oriented products and the reallocation of supply. Nanya Technology explains that AI servers and general servers in cloud data centers are absorbing HBM and RDIMM, squeezing supply for memory used in smartphones, PCs, and automotive/consumer electronics. In the company's first-half revenue, products for AI infrastructure and servers accounted for over 20%. The rise in memory unit prices isn't contained within AI servers—it also ripples through to component costs for consumer devices.
Supply isn't catching up quickly. Nanya Technology expects the shortage to persist for several quarters or more, and plans to gradually increase production capacity starting in 2028, based on demand tied to customers' multi-year contracts. The 2026 capital expenditure ceiling is NT$52 billion, with production equipment accounting for roughly 30% of that. The high market growth rate projected for 2026 also factors in price increases that occur before new manufacturing capacity enters mass production.
UWIO and Supply Capacity Through 2028
What Nanya Technology is pursuing for inference isn't a path that simply chases the same products as HBM. According to Mirror Daily, the company is developing custom UWIO DRAM that directly connects DRAM with CPUs or ASICs using 3D IC technologies such as Wafer-on-Wafer (WoW). Wu Chih-hsiang stated a target of boosting bandwidth 5 to 10 times compared to HBM while reducing energy consumption per bit transferred to between one-third and one-tenth. This is a claim aimed at edge AI and on-device inference, where power and latency constraints are strict.
Nanya Technology announced in Q1 2026 that UWIO is already generating initial revenue. However, no independent benchmarks under matched configuration conditions, nor a mass production timeline, have been disclosed for the claimed 5-to-10x bandwidth and power consumption figures. Rather than being a blanket replacement for HBM, its success should be measured as a product designed in combination with customers' CPUs or ASICs.
The timeline for manufacturing capacity also requires attention. According to Nanya Technology's official materials, equipment installation at the new fab will begin in Q1 2027, but the first phase—30,000 wafers per month—is planned for 2028. At full capacity, output is expected to reach 45,000 wafers per month, with total capital expenditure projected at NT$480 billion. What underpins the 2026 forecast of over $800 billion is existing capacity and pricing; the new fab won't begin altering supply until after that.
WSTS also expects memory revenue to grow 32% in 2027. The key factors for confirming the forecast's durability will be whether shipment bit volumes—across Nanya Technology and other companies—begin rising alongside prices, and whether the new capacity planned for 2028 moves into mass production as scheduled. If UWIO is adopted in customer products, inference demand will expand from HBM-centric data centers to 3D-stacked memory in end devices. Until those conditions are in place, the $800 billion figure should be read as a market size simultaneously shaped by AI demand and supply constraints.