For the past two years, China's AI companies have competed on a single axis: whose model performs best. That contest is starting to give way to a harder one — whether those models can actually run, at scale, on chips made in China, as US restrictions on advanced AI chip exports raise the stakes of that question.
China's second humanoid robot games, held in Beijing from August 22 to August 26, 2026, sent videos and clips racing across social media and underscored how fast humanoid robotics is advancing. For Taiwan, the spectacle also highlighted both the promise and the risks of robots in policing, firefighting, and even warfare.
Nvidia's data center business now splits almost evenly between hyperscalers and everyone else — US$48.7 billion against US$40.3 billion in the July quarter. The obvious reading is that the customer base is broadening. The sequence says something more specific.
Nexperia China is accelerating its separation from the Dutch chipmaker's traditional European manufacturing network, raising the prospect that one of the world's largest suppliers of automotive and industrial semiconductors could operate through increasingly distinct China- and Europe-based supply chains.
Nvidia has reportedly notified major customers that server systems using its AI chips will rise by more than 15%, with the new pricing set to apply to models shipping from early 2027, including Grace Blackwell and the next-generation Vera Rubin platform. The move is expected to ripple through the chip and system supply chain, while industry figures say it could also lift the appeal of inference ASICs.
The memory industry is undergoing a structural shift as AI demand surges, and Etron Technology chairman Nicky Lu said the sector's boom will stretch beyond 2027, with shortages likely lasting through 2028 and possibly into 2030. He said supply is tightening further as AI and edge computing create rigid demand, pushing the supply chain to focus on higher-value applications.
SK Hynix is moving on several fronts as it pushes deeper into system-level memory competition. From high-bandwidth memory (HBM) and high-bandwidth flash (HBF) to a recently published roadmap proposing to extend co-packaged optics (CPO) to the memory interface, the company increasingly appears to be competing not just over the next generation of HBM, but over how future artificial intelligence (AI) systems organize and access memory.
The 2026 World Robot Conference (WRC) opened in Beijing on August 19, and the biggest change on this year's show floor was not the larger number of humanoid robots — it was that robots started working on site. From coffee making and retail service to printed circuit board (PCB) handling, logistics sorting, auto parts loading and unloading, and even home organizing, laundry, and table clearing, robot makers are no longer just showing robots walking, dancing, and boxing.
Rapid changes in the power and thermal requirements of AI accelerators are forcing data-center designers to incorporate chip-level information earlier in the planning process, according to Cadence Design Systems. This is narrowing a gap that once separated semiconductor development from facility engineering.
Foxconn is repositioning its AI data center business around system integration, modular deployment, and sovereign computing, as the rapid growth of GPU infrastructure changes how customers design and procure computing capacity.
ASIC and GPU shipments will reach a "golden cross" in 2027, with ASIC shipments expected to surpass GPU shipments for the first time at 15.3 million units, DIGITIMES Intelligence analyst Stella Weng said. She also said AI is driving a "qualitative change" in memory architecture inside AI servers.
Autonomous driving technology is maturing, and a future in which cars entirely drive themselves is quickly approaching in the rearview mirror. Yet safety and trust remain crucial barriers to overcome before fully autonomous driving becomes a reality, and a key part of this is the "eyes" of the car.
As AI develops rapidly, data centers are driving equally rapid growth in electricity demand. The stability of future energy supplies will directly determine whether AI computing capacity can continue to expand.
As AI chips continue to grow in size, the shift in advanced packaging from 300 mm wafers to square panels is becoming clearer, driving parallel development of new packaging technologies including FOPLP, CoPoS, and Glass Core. Manz Asia general manager Robert Lin said yield will determine whether Glass Core can move into mass production.
The global memory market is entering a sharp upcycle as cloud providers boost capital spending and demand for AI infrastructure lifts DRAM, NAND, and HBM prices. According to DIGITIMES, the three largest upstream memory chip makers are headed for a more than threefold jump in combined DRAM and NAND revenue in 2026, with supply constraints likely to keep pricing firm until new capacity arrives in 2027.
