ByteDance (Doubao)

Beijing issued the Action Plan for Accelerating the Development of the Token Economy (2026-2028) jointly through the Municipal Bureau of Economy and Information Technology and the Development and Reform Commission, dated 15 September, published on the 18th and running to the end of 2028. It descends from State Council Document Guo Fa [2025] No. 11, the "AI Plus" directive, and is the first policy at provincial level in China written around the token as an economic object rather than a technical one. Ten numbered measures follow. Beijing will write a graded evaluation standard for what it calls token factories, a compute centre bolted to a model-tuning operation, grading them on models adapted, token throughput, time to first token, cache hit rate and power usage effectiveness, and back the leading ones with land, energy quotas and network guarantees. It will build a token distribution platform with a unified interface, model aggregation and token metering and settlement, then fold that platform into the city new-infrastructure plan. It will have financial institutions develop products for token consumption scenarios in order to extend medium and long-term credit. And it will retrain working software engineers as token engineers. At the press briefing the next day, Chen Guangming, deputy director of the bureau digital industry division, said standards and evaluation systems for token factories are not yet unified and that the problem of token idling still exists. The Chinese term is 空转, the word for an engine running in neutral, and it carries the weight of 资金空转, capital idling inside the banking system without reaching a factory, which regulators spent most of the last decade on. His colleague Zhang Jinrui said Beijing will schedule compute across Beijing, Tianjin, Hebei and Inner Mongolia and build a chain running from electricity through compute invocation to token production, distribution and application, which the bureau expects to pass a trillion yuan during the current five-year plan. The lending is already happening: Beijing E-Town arranged close to 2 billion yuan of credit lines partly on the basis of token consumption, with the Agricultural Bank of China, CITIC, Industrial Bank, Bank of Beijing, Minsheng and Huaxia participating, and Bank of China Guangzhou ran the first token loan in August. Lou Feipeng of Postal Savings Bank named exaggerated token-use data as the first obstacle to scaling it. Leiphone reported from the cloud sales floors at the big three platforms, where model-as-a-service quotas run above 10 million yuan per salesperson against a few hundred thousand to a million last year, one regional team resigned together in an episode classified internally as a management accident, and vendors resell Zhipu models at 60 percent of list and hand over cloud-service vouchers when the discount runs out of room, so the customer cost falls while the model revenue line does not move. At Volcano Engine token call volume is a single-veto item. The token count Beijing wants banks to lend against is, at three of the largest producers of it, a quota with a subsidy attached. In the Briefing: Huawei set commercial dates for the Ascend 950 cluster cloud service, 30 September in China and 30 November globally; the Ascend 960 SuperPoD takes 4,096 NPUs on near-packaged optics; CXMT posted an 82 percent operating margin for April to June, the highest of any memory maker in the world; and the securities regulator gave banks informal window guidance raising the bar for humanoid robot listings.

Renting out a humanoid robot in China was a 3.3-day payback trade a year ago and is now a way to lose money, and the register has not noticed. tmtpost published a piece by 硅基降临 on 16 September reporting that a 100,000 yuan Unitree G1 could command up to 30,000 yuan a day at the peak, while JD’s own self-operated listing now rents a Unitree U1 for 1,571 yuan a day, about 234 dollars. Most machines have fallen from 10,000 or 20,000 yuan a day into the hundreds, and the rental platform Qingtianzu ran a ten-city promotion offering two hours for one yuan that drew over 500,000 visits with the fastest slots gone in seven seconds. Qichacha data cited in the piece puts 153,000 robot-rental companies on the Chinese register, 38,200 of them added during 2025 and another 9,114 in a bit over three months of 2026, close to a hundred new firms a day, against a sector worth about 1 billion yuan in 2025 and forecast at 10 billion for 2026. What the collapse exposes is what the orders were for: Qingtianzu’s own figures put 65 percent of them at entertainment performance and commercial marketing, where the novelty was the product and there is no second order, while light industrial work that renews year after year is about 10 percent of platform orders. The asset then depreciates at 20 to 30 percent a year, an engineering prototype bought in 2024 for 300,000 to 800,000 yuan now carries a second-hand residual of 30,000 to 60,000, and no machine works unattended, so a 2,500 yuan bare-machine day becomes 5,000 once the operator’s time and travel are added. The three businesses surviving are not renting robots: Hangzhou’s Xingshu Intelligence rents hardware near cost and sells choreography at 500 to 2,000 yuan per custom routine, Qingtianzu owns no machines and brokers idle inventory, and light industrial patrol and material handling renews. This matters beyond the rental trade because rental companies were named, by a founder who had just taken his own robotics company public in Hong Kong, as one of two vehicles the sector uses to manufacture revenue, the other being the data-collection centres covered on 11 September. The Information reported on 9 September that regulators had given banks informal window guidance raising the bar for humanoid-robot listings to recurring revenue, narrowing losses or real technical innovation, and bankers confirmed the tightening to Yicai on 14 September, against a queue of about 45 robot companies at some stage of capitalisation. Unitree, the one with the revenue, closed on 16 September at 474.28 yuan for a market capitalisation of 191.83 billion yuan, about 28.6 billion dollars, 3.1 times its 150.80 yuan issue price and 56.9 percent below the 1,100 yuan it opened at on 19 August.

Enflame’s allocation result, filed to the exchange on 7 September, priced 43,035,173 new shares at 142.18 yuan and records the online retail tranche subscribed 6,109.34 times. That tripped the clawback rule, which moves 10 percent of the public offering out of the institutional book once retail demand passes 100 times, growing the retail tranche from 6,885,500 shares to 10,328,500 at a final allotment ratio of 0.02455315 percent. The offline book was barely cooler: 291 institutions across 11,272 placement accounts bid 72.795 billion shares against a tranche of about 27.5 million, or 2,642.99 times. Enflame has never made a profit, so listing rules bar it from quoting a price-to-earnings multiple and its own investment-risk notice prints price-to-sales instead, 55.62 times pre-issue and 61.80 post-issue, valuing the company at 61.187 billion yuan. The same notice’s peer table, dated 28 August, averages 93.51 times sales, so the deal was priced under its comparables on purpose. Moore Threads went the other way the same day: 25,774,510 shares, 5.48 percent of the company, came off a nine-month lockup and the stock closed limit-down at 415.49 yuan against Friday’s 519.35, taking 48.8 billion yuan off its market value, with institutions net sellers of 1.462 billion. Its free float roughly doubled, from 6.43 percent to 11.91 percent, and the unlock is the only input that moved. Enflame’s own day-one free float is 17,900,325 shares of 430,351,728, or 4.16 percent, with 88.09 percent still locked and Caixin reporting close to 40 percent of the share capital coming free on 7 December.

The Cyberspace Administration published a phase-two note on its Qinglang campaign against chaos in AI applications on 2 September, and it places obligations at three points in the pipeline rather than one. Platforms including Douyin, Weibo, Tencent and Xiaohongshu are named for expanding face, voiceprint and violation sample libraries, with 46 governance announcements published between them. Four model products, Doubao, Yuanbao, Qwen and ERNIE, are told to guard the content review pass on original training corpora and to label AI-generated synthetic content so as to prevent illegal information at source, which is an obligation on the pre-training set rather than on a feed. The third layer is distribution: Huawei, Xiaomi, OPPO and vivo are named as setting developer admission standards for AI applications and spot re-testing apps already listed, so four handset makers become the admission gate. The note landed one year and one day after the labelling Measures took effect on 1 September 2025. No published standard, audit or methodology sits behind the corpus clause, so what changed this week is scope rather than enforcement capability.

Harvey, the OpenAI-backed legal AI company, published its first post-trained model on 20 August. Harvey Tenet is a Kimi K3 base, post-trained with a rank-64 LoRA over the full network on a dataset of about 1,750 tasks. It completes almost twice as many held-out tasks on the Legal Agent Benchmark as base Kimi K3, and places second on that benchmark overall. Harvey ran ablations on candidate judge models against heavier frontier models and settled on Kimi 2.6 as the grader. Its M&A diligence work uses a GLM-5.2 orchestrator, taking rubric criteria pass rate from a 43.8 percent best baseline to 60.1 percent after self-distillation post-training.

Chinese cloud vendors are reselling competitors’ models at roughly 30 percent of list price, below their own cost, because in AI coding the customer names the model. Leiphone reported the discount ladder on 21 August from named cloud sales sources: 60 to 80 percent of list six months ago, 48 percent now common, roughly 30 percent at the low end. The same day, Approaching.AI vice-president Guan Jiawei told the South China Morning Post that domestic chips can only serve the low-quality inference tier, the one with weak demand and weak monetisation, and that high-quality tokens, especially coding, still depend on Nvidia. iFlytek said the same week that a flagship model trained entirely on domestic compute would reach the domestic first tier on code capability.