The tightest supply line in the chip industry is memory, not logic. High-bandwidth memory now takes roughly 23 percent of DRAM wafer capacity, and data centres are expected to use close to 70 percent of the memory chips made this year. Contract prices for automotive DRAM are forecast to rise 70 to 100 percent in 2026. TSMC will raise chipmaking prices by up to 10 percent from January 2027, and Nvidia and Broadcom have been turned down on extra capacity. Copper, helium and memory packaging are also tight, and Washington has threatened tariffs on overseas memory producers unless they expand output inside the United States.
ASML's chief executive warned on October 5 that the global AI boom is growing faster than the chip industry can add production capacity, and that supply constraints could last for years. ASML is the only maker of the most advanced lithography machines used to print leading-edge chips. Industry group SEMI expects semiconductor equipment sales to reach a record $165.9 billion in 2026, up 23.2%, driven by AI demand. Test equipment is forecast to grow 31%. Analysts expect generative AI chips to bring in about $500 billion in 2026.
Aleph Alpha, the German AI company based in Heidelberg, released a new large language model called Kolibri. The model supports German and English and uses a mixture-of-experts design, which runs only part of the network for each request. Weights were published on Hugging Face under the Apache 2.0 licence, so companies and public bodies can download and run it on their own servers. Aleph Alpha markets the model for government, industry and defence work, where data must stay inside Europe. The release adds to a push for sovereign AI, in which governments train and host models on domestic hardware rather than relying on US cloud providers.
Alibaba has announced a new processor that it describes as China's most powerful chip for artificial intelligence work, along with plans to expand its AI model lineup. The company says the chip is built for training and running large models in data centres. The move comes as Chinese firms look for domestic alternatives to imported accelerators. Analysts say software support and manufacturing capacity will decide how quickly the chip is adopted at scale, and Alibaba has not yet given detailed benchmarks.
Nvidia broke through to a new record high even as a broad bond sell-off dragged other sectors lower. Demand for AI accelerators and high-bandwidth memory continues to outrun supply, and the companies buying them are financing an unprecedented build-out of data centres. Bank of England governor Andrew Bailey warned that the boom could trigger market shocks, saying the bank is watching the cash flowing into AI very carefully. In China, humanoid robot maker Unitree Robotics soared more than 600 percent on its market debut despite a US ban on its hardware.
The US Federal Trade Commission has opened a broad investigation into whether OpenAI, Anthropic and other AI companies have harmed consumers. The agency plans formal demands, similar to subpoenas, for documents and executive testimony, and will examine whether the companies broke federal laws barring unfair and deceptive practices, including misuse of consumer data or overstated product claims. A central question is the risk from autonomous AI agents. It is the first US enforcement action aimed at rogue agents, after incidents reported since July. The probe follows a White House meeting where AI firms agreed to voluntary safety standards.
High-bandwidth memory is expected to take close to 30% of DRAM makers' total wafer capacity in 2027, up from roughly 20% now, as suppliers shift lines toward chips used in AI accelerators. SK Hynix said it validated HBM5 with TSMC packaging while HBM4 begins entering Nvidia's next platform. Memory prices are feeding into the wider economy: chips made up nearly half of South Korea's September exports, and Morgan Stanley analysts have called the squeeze part of chipflation. Suppliers are responding with heavy investment, from ChangXin Memory Technologies to Micron and Samsung.
Alibaba Group used its Apsara Conference in Hangzhou to unveil the Zhenwu V900, an AI processor built by its T-Head semiconductor unit. CEO Eddie Wu said it delivers three times the performance of the M890 launched in May and can be clustered with up to 500,000 chips. Mass production is planned for the first quarter of 2027. Future Qwen models could hold 5 to 10 trillion parameters, up from 2.4 trillion in the current flagship. Alibaba also targets more than 20 gigawatts of global data centre capacity by 2032. Shares rose 5.1%.
President Donald Trump, House Speaker Mike Johnson and a group of technology chief executives are due to meet on artificial intelligence on Tuesday, September 29, according to people familiar with the plan, first reported by Axios. Trump has argued the US does not need new AI-specific regulation and told the UN General Assembly he rejects any attempt to build a globalist scheme to control AI. China has taken a different route, with rules covering algorithms and training data and labelling requirements for AI-generated content. Leading AI labs have warned about risks to humanity and asked governments to coordinate. Analysts expect export controls on advanced chips and compute to be the channel most likely to touch listed companies.
The companies building America's AI data centers are now at the centre of fights over electricity, water, land and pollution, NBC News reported. AI demand has pulled billions in financing toward a handful of little-known developers. The power numbers explain why: a single AI-focused campus can draw 50 to 100 megawatts, and Brookings notes the US AI sector may need 50 gigawatts of new capacity by 2028. The IEA projects global data centre electricity use could reach 945 TWh by 2030. Protests have followed in Imperial, California, and Louisiana regulators approved three new gas plants for a Meta AI campus.