Topic
Explore English articles about Tech Geopolitics.
Moonshot has halted new subscriptions for Kimi K3 amid surging demand, prioritizing existing members. The company plans to separate membership tiers for Web/App/Work use and coding use, and revisit how GPU capacity is allocated.
Alibaba has teased Qwen3.8, a 2.4-trillion-parameter model, but for now it is only accessible through a subscription-based preview. Comparing the disclosure gap with Kimi K3, this article examines the unresolved questions around performance ranking and open-weighting.
东方算芯's DF1000 stacks DRAM directly onto 14nm logic, claiming 6.4TB/s of bandwidth. We examine what this design means for moving away from HBM dependence, and the conditions that remain unverified for mass production and real-world measurement.
In July 2026, Moonshot AI's Kimi K3 took the top spot in the Arena coding rankings—but behind the scenes, it emerged that US-based Cursor had not initially disclosed its reliance on an earlier version of Kimi. This piece traces how suspicion of and dependence on Chinese-made AI became entangled over the course of just a few months.
Reports indicate that CXMT's DRAM orders are booked through the end of 2027. Reading the background of major PC manufacturer prioritization and the timeline for supply improvement, based on a 95.73% utilization rate and equipment upgrade plans through 2028.
Moonshot AI's Kimi K3 offers 2.8 trillion parameters and a 1-million-token context through a low-priced API. Reading the primary sources on its sparse MoE design, the burden of self-hosting, and the constraints ahead of the weight release.
Apple's procurement of China-made memory is not immediately banned outright. Meanwhile, U.S. federal procurement regulations naming CXMT and YMTC take effect at the end of 2027, forcing supplier verification for federally supplied devices.
China's IC export value surged approximately 96% year-on-year in the first half of 2026, but export volume growth was limited to just 6.9%. This divergence is primarily driven by the global surge in memory prices, and caution is needed before interpreting the export value spike as a major improvement in China's production capacity or technological level.
Shanghai-based Dongfang Suanxin has unveiled the DF1000, an AI chip that combines a 14nm process with 3D stacking technology against the backdrop of US export controls. Its design vertically stacks logic circuits and DRAM to achieve high memory bandwidth, but mass-production yield and building a practical software ecosystem remain challenges ahead.
Tracing the background of DeepSeek's in-house AI inference chip plan reported by Reuters, and Zhipu AI's search for a custom ASIC. The focus is on the development gap: while Alibaba and Baidu have already reached mass production and IPO stages, both companies remain at an early conceptual stage.
