Gary Huang, vice president and head of Asia-Pacific research and business development at Yole Group, said the imbalance is already showing signs of what he described as supply-chain "decoupling."
The global DRAM market continued its explosive expansion in the second quarter of 2026, as AI infrastructure spending drove demand for both conventional memory and high-bandwidth memory (HBM), while Chinese supplier ChangXin Memory Technologies (CXMT) rapidly increased its presence.
A surge of multibillion-dollar industrial expansions is quietly transforming California's capital region into a premier domestic semiconductor and advanced technology powerhouse, positioning it as a highly competitive gateway for global tech firms looking to scale beyond an overflowing Silicon Valley.
Sumitomo Chemical has started mass production and sales of indium phosphide (InP) epitaxial wafers, a core material for optical semiconductors used in next-generation data centers and high-speed networks. The move could help ease power and data bottlenecks facing AI infrastructure worldwide, as demand rises for faster, more efficient computing and communications.
As silicon photonics (SiPh) and co-packaged optics (CPO) move toward mass production, suppliers say testing is shifting earlier in the process to catch defects at the wafer stage, protect yields, and preserve data through final assembly. That change is also pushing equipment makers toward team-based bids, with a single integrator likely to coordinate the system.
The rapid expansion of artificial intelligence is forcing a radical shift from individual chip design to holistic system-level engineering, requiring competitors and supply chain partners to dismantle organizational silos or face crippling inefficiencies, top industry executives said at SEMICON Taiwan 2026.
Kioxia's plan to push NAND flash into work that DRAM currently does is less a product launch than an answer to a structural problem. Alone among the leading memory makers, it does not produce DRAM, and in the one market where it does compete, it is being squeezed from both above and below — by a resurgent Micron and by China's Yangtze Memory Technologies (YMTC), which has closed much of the gap in a single year.
Nvidia used IFA 2026 to argue that serious AI work can run locally as well as in the cloud, pairing an October launch date for its RTX Spark Windows PCs with free software that distributes independent inference requests across compatible machines already on a local network.
Nvidia's agreement to buy Hugging Face for roughly US$13 billion is being framed by analysts less as a move into software than as insurance against three ways its core business could deteriorate: cheaper open-weight models eroding the frontier labs that buy its chips, those same labs diversifying their silicon, and enterprises pulling workloads out of hyperscaler bundles. On each of those paths, owning the industry's main model repository pays off.
The mechanics of Nvidia's US$12.93 billion purchase of Hugging Face say as much about the deal's purpose as the strategic case does: a tenth of the consideration is set aside to stop the team from walking, part of the payout goes to Intel and AMD, and the whole transaction is framed against an open-weight model market in which the fastest-moving suppliers are Chinese.
Nvidia's agreement to acquire Hugging Face for US$12.93 billion pushes the world's dominant AI chip supplier past silicon and into the layer where developers actually pick their models — a position that gives it early sight of demand shifts, a structural hedge against customers designing their own accelerators, and a neutrality problem it has already had to answer for in writing.
More coverage
