Chips & semiconductors · Trend watch (posts originally published July 20–21)
From the Aug 5, 2026 daily brief
The most structurally valuable post in this Teortaxes batch. On the software side, Chinese model architecture has converged on mixture-of-experts with wide expert parallelism — mixture-of-experts (MoE) meaning each token wakes only a few "expert" sub-networks, which saves per-chip compute and shifts the load onto memory and chip-to-chip communication — with the experts spread across large numbers of chips, and the motive stated outright: to "survive on weaker NPUs" — meaning China's own domestically produced accelerators. On the hardware side, the system vendors now design to that shape in return, building supernodes — dozens to thousands of chips fused into one big pool over high-bandwidth interconnect: "Everyone's got to have a 'supernode' now. It's not just Huawei" (Teortaxes, July 20). The mechanism's premises each carry an independent anchor: Moonshot, the developer of Kimi, officially recommends at least 64 accelerators as the floor for deploying its flagship (relay, July 19); and Susan Zhang, the training-side practitioner, worked through a 1,024-chip Chinese cluster — total compute roughly equal to 1,024 NVIDIA A100s, total memory exceeding 1,024 of the new-generation B200s (Zhang, July 17) — weak per card, thick as a system: exactly this shape. But "not just Huawei" names no second company and comes with no specs — the sentence most in need of verification is the one with the least evidence behind it. This brief logs the whole line as a framework awaiting a ruling: if it holds, the metric that matters for export controls moves from "per-card compute ceiling" to "how large a compute domain can be fused." And settlement doesn't wait on the technical debate — count two numbers: how many Chinese system vendors ship supernode-class products, and how many Chinese fabs run 7nm-class volume production. Today both counts sit at one: only Huawei, only SMIC.
Just an email address, unsubscribe anytime. This is the only thing we ask of you.
The same commentator argues the real story in Chinese advanced process is diffusion — not …
Starting from a proof of concept he himself calls toy-grade, Teortaxes offers a structural…
In late April, Gooaye aired a piece of information that cut against the host's own holding…
Microsoft unveiled its next-generation in-house AI chip, Maia 200, this January; the same …