Moore Threads posts 147% H1 growth, plans Hong Kong listing
China's AI economy is consolidating fast — its flagship GPU maker is nearly break-even and heading for Hong Kong, a startup is open-sourcing a small vision model that out-perceives far bigger rivals, and Alibaba is testing a new way to charge for open weights.
Moore Threads, China's answer to Nvidia, is planning a Hong Kong listing after first-half revenue jumped 147% — and its losses nearly vanished. The Beijing chipmaker, founded by former Nvidia China chief Zhang Jianzhong, reported revenue of 1.74 billion yuan (about $257 million) for the first half of 2026 — already more than its entire 2025 figure — with the net loss narrowing 95.7% to roughly 11.6 million yuan (about $1.7 million), putting it within sight of break-even. On Sunday it said it plans to issue H-shares on the Hong Kong Stock Exchange's main board, opening the door for international investors to a stock that has surged more than 420% since its December debut on Shanghai's STAR Market, where its 8-billion-yuan IPO produced the biggest first-day pop for a major listing since China's 2019 market reforms.
The numbers show Beijing's Nvidia-replacement push turning into real revenue: sales were driven by its KUAE intelligent-computing clusters built on the MTT S5000, R&D spending hit 769 million yuan (44% of revenue), and the next-generation "Huagang" architecture promises native FP8/FP4 support with a tenfold gain in compute efficiency. The timing is deliberate — the results land the same weekend China switched on its first fully domestic 100,000-card supercluster, and a Hong Kong listing would hand the country's flagship GPU champion a fresh war chest and international credibility for the export-control fight. The open question is whether Moore Threads can hold this growth curve long enough to turn near-break-even into actual profit.
Om AI, a Chinese physical-AI startup, has open-sourced VLX-Seek 1.5, which it bills as the first on-device-native streaming multimodal model — and its 3B version is beating far bigger rivals at seeing the real world. Released on August 6 alongside a funding round of several hundred million yuan led by Qianhai FOF, the model family (0.6B to 10B parameters, with 10B weights already on GitHub) is designed from the architecture up to run on robots, drones and edge devices, processing video streams locally instead of shipping frames to the cloud. In Om AI's own benchmarks, the 3B model beat Nvidia's same-size LocateAnything by 13% on open-vocabulary detection (LVIS Mean 57.5 vs 50.7) and by 40% on drone-perspective detection (RefDrone F1 73.2 vs 52.3), while cutting object hallucination roughly fourfold; the earlier VLX-Seek-3B also edged Gemini 3 Pro on referring-expression tasks. The claim that matters isn't the scoreboard, though — it's the philosophy: "on-device native" rather than train-big-then-compress, trading raw scale for the latency, privacy and cost that make physical AI deployable at all.
Alibaba reportedly plans to take a cut of revenue from large commercial users of its next open-weight Qwen model — a first for a family that has long been free to self-host. Reuters, citing two people familiar with the plans, says the revenue-sharing terms would roll out alongside the open-weight release of Qwen3.8-Max, expected as soon as this week, with the rate still under negotiation. The move follows the template Moonshot set with Kimi K3, whose license requires companies making more than $20 million a year from the model to sign a commercial deal — including revenue shares of up to 30%. It signals the end of the free lunch for frontier Chinese models, and it blurs "open weight" and "open source": pay-or-negotiate terms aren't OSI-approved open source, which is exactly how Chinese labs can monetize adoption without closing their weights.
What to watch: whether Moore Threads' H-share filing draws international demand — and whether Alibaba confirms a revenue-share rate when Qwen3.8-Max's weights land.
Chinese labs are finding ways to charge for "open" models — does revenue-sharing kill the open-weight advantage? Tell us in the comments.
Sources: VLX-Seek (GitHub)
