Chip suppliers are giving sharply different accounts of the memory market. Industry trackers reported on September 26 that Acer and Phison Electronics disagree about where DRAM prices are heading, as supply improves for mainstream PC memory while AI server demand keeps the wider market constrained. Foundry capacity remains the bottleneck: Broadcom has warned that advanced-node capacity will stay tight through 2026 as AI accelerators take priority. SK hynix has committed about $38 billion to new memory output but says relief is roughly three years away. Samsung's foundry recovery is gaining pace, though 2-nanometre yields and the ramp of its Taylor fab in Texas remain hurdles.
Chinese AI lab DeepSeek is reportedly developing its own AI chip to cut its reliance on Nvidia and Huawei, and has been expanding its use of Huawei's Ascend processors for model training. Huawei said at its Connect conference on September 17 that it is moving its next data-centre accelerator, the Ascend 960DT, up to the first quarter of 2027. Supply is the bottleneck. US export controls keep Chinese designers away from advanced overseas foundries, SMIC's 7nm and 6nm capacity limits domestic output, and high-bandwidth memory is tight. Analysts say the programme is still early and will depend on how fast domestic foundry and memory capacity can grow.
Memory prices are still rising, just more slowly. TrendForce expects conventional DRAM contract prices to gain 13% to 18% quarter on quarter in Q3 2026, with NAND up 10% to 15%, as long-term supply deals cap server prices. The damage lands on consumer hardware: notebook makers are passing higher memory costs into retail prices, and smartphone vendors are lifting prices and trimming production plans. Gartner had forecast a 130% surge in combined DRAM and SSD prices by the end of 2026, pushing PC prices up 17% and phone prices up 13%.
About 20 million AI accelerator chips are already running in data centres worldwide, and the count is doubling roughly every nine months, according to Epoch AI research reported by The New York Times. At that pace the fleet would reach around 200 million chips by the end of 2028. Hundreds of data centres are under construction. The constraint is shifting from chip supply to power, materials and construction: data centres already draw roughly as much electricity as Germany, and that figure is expected to quadruple by 2030. New York State has imposed a one-year moratorium on new data centres.
OpenAI's Project Camellia, announced in July, puts a 1,441-acre data centre campus at the Savannah Gateway Industrial Hub in Effingham County, Georgia, near Rincon. The plan calls for four buildings and 3.2 gigawatts of power under a 25-year agreement with Georgia Power, delivered in phases between 2028 and 2032. The disclosed initial investment is at least $20 billion, and the company's vice president of compute strategy has said the full build-out will exceed $30 billion. It is the first campus OpenAI designs, builds and operates itself. The schedule highlights the real bottleneck: model releases keep arriving while power, land and chip supply do not arrive until 2028 at the earliest.
Samsung Electronics plans to double output of HBM4 and HBM4E memory next year, according to Seoul Economic Daily, one of the largest capacity commitments the memory industry has made this decade. The near-term supply of HBM3E is sold out and manufacturers have shifted wafer capacity away from commodity DRAM, pushing DRAM prices up about 171 per cent year over year and roughly quadrupling DDR5 spot prices since late 2025. CSIS analysis suggests a single next-generation Nvidia processor could carry as much as 384GB of HBM by 2027. PC makers including Dell, Lenovo, Apple and Microsoft have tied price increases to memory costs.
The competition in AI chips is shifting from processor design to memory and packaging. Huawei said it will launch new AI chips in 2027 aimed at Nvidia's data center lineup, while Qualcomm signed a reported 60 billion dollar AI chip agreement with Amazon Web Services. ASML extended its lead in lithography as customers adopt High NA EUV tools, and Nexperia formed a partnership with India's Tata Electronics after falling out with its Chinese parent. TrendForce expects TSMC's 3nm monthly wafer starts to reach 180,000 by early in the fourth quarter. Industry group SIA says semiconductors account for 95 percent of the value of an AI server rack, from logic to memory and power devices.
Huawei's Ascend processors now account for roughly half of the AI chips sold in China, according to market estimates, as export controls keep the most advanced Nvidia parts out of the country. Estimates put Huawei's share at 50% to 60% this year, with revenue near $12 billion. Chinese cloud providers ordered about 800,000 Ascend 950PR chips for 2026. China has also set out a five-year plan for national AI data centres worth roughly $295 billion, requiring at least 80% domestic technology. The limit is high bandwidth memory: domestic HBM supply holds finished Ascend output to well under 300,000 units a year.
High-bandwidth memory used in AI accelerators now takes about 23% of all DRAM wafer capacity, and the shift is tightening supply for phones, laptops and game consoles. Micron told the Hot Chips 2026 conference that HBM uses roughly three times the silicon of standard DDR5 and that the gap widens with every generation. Memory prices have roughly quadrupled over the past year, and analysts do not expect relief before new fab capacity arrives. Nvidia has raised the price of its DGX Spark desktop from $3,999 to $4,699 and warned large customers that AI server prices could rise by more than 15%. Micron reported $41.5 billion in quarterly revenue, up 345% year over year.
Huawei will launch two new artificial intelligence chips in 2027, rotating chairman David Wang said on Thursday at a company event in Shanghai. The 960DT is planned for the first quarter of 2027, followed by the Ascend 960PR in the third quarter. Wang said Huawei's UnifiedBus interconnect technology will be a key link in its next generation of large AI systems, letting many chips work together as one machine. Huawei says it has shipped more than 1,000 of its large AI systems to more than 370 customers. U.S. export controls have pushed Chinese firms to build domestic alternatives to Nvidia hardware.