Beijing's newest capital injection is really a loss forecast, and it points offshore
Eight names on the list, five of them insurers
The Ministry of Finance named the recipients. It will shortly issue 300bn yuan (~$42bn) of special treasury bonds to replenish core tier-one capital at eight central financial enterprises, and all eight published capital-increase plans the same day: ICBC, Agricultural Bank of China, the Export-Import Bank of China, export-credit insurer Sinosure, PICC, the group parent of China Life, China Taiping and reinsurer China Re.
The sequence matters more than the sum. March 2025 brought 500bn yuan (~$70bn) for Bank of China, China Construction Bank, BoCom and Postal Savings; they raised 520bn between them and lifted core tier-one ratios by 0.86, 0.48, 1.28 and 1.51 points. This spring's 300bn tranche was read by the market as ICBC and ABC money, with CICC estimating leverage of roughly 4tn yuan of asset expansion.
This list is not that list. Two commercial banks, one policy bank, five insurers. Exim funds overseas infrastructure lending; Sinosure underwrites export receivables; China Re sits behind catastrophe cover in a year of glacier collapses and typhoons. That is capital against foreign default and shrinking trade, not against domestic property.
THE TELL: Xinhua's chosen gloss is that this is a prudent, forward-looking arrangement — nobody is in trouble. On Friday I ran Beijing's G20 argument that official bilateral creditors hold only about 15% of developing-country debt. Same ledger, now funded.
Go deeper on this section: ClaudeChatGPT
Zhipu's API business grew 27-fold and still bleeds
The Beijing lab put a price list on Tmall, then published the accounts. GLM Coding Plan sells as Lite, Pro and Max at 118, 538 and 1,078 yuan a month (~$17, ~$76, ~$152), a team seat at 598 (~$84), all on GLM-5.3 and driving 20-plus agents including Claude Code and Codex. Kimi, MiniMax and StepFun are still negotiating storefronts.
The half-year explains why it wants a shelf. Revenue 954m yuan (~$134m), up 399.7%; open-platform and API revenue 825m yuan (~$116m), up 2,735.7% and now 86.5% of the total against 15.2% a year ago. Gross margin fell from 50% to 26.4%. Adjusted net loss 1.964bn yuan (~$277m), per TMTPost.
The cost side is the story. Token calls up 40-fold since January, paying daily actives up 603%, average API price up 101%, unit inference cost down 80% — and API revenue still only covers inference. Nothing is left for base-model training or depreciation. Volume is outrunning the cost curve, which makes the water-and-electricity comparison premature.
Go deeper on this section: ClaudeChatGPT
Alibaba shipped the loop-depth fix eleven months ago
GPT-6 Astra made recurrent depth fashionable; the papers were already out. On Friday I ran the Chinese press leading on Astra's opaque recurrence, where part of the reasoning finishes inside hidden activations that never become text. QbitAI's follow-up points at two papers from Alibaba and university collaborators aimed at the failure mode that has always killed the idea: later loops stop doing work.
The diagnosis is measured rather than asserted. Hidden-state updates concentrate almost entirely in the first core loop, adjacent iterations score high on representational similarity, and the singular-value spectrum collapses into a low-dimensional subspace. Two named causes: no differentiation between iterations, and one hidden state carrying long-term memory and current computation at once.
MeSH adds a memory buffer with per-iteration, per-token read-write routers. At Pythia-1.4B scale, weight sharing removes about 33% of non-embedding parameters and zero-shot downstream accuracy still rises from 49.50% to 50.56% — for router parameters worth ~0.005% of the baseline and ~0.014% extra compute. ICLR 2026, first posted last October. SpiralFormer instead varies sequence length per loop, coarse to fine at one-eighth, a quarter, a half, then full, with results shifted right so compression cannot leak future tokens: 13.13T FLOPs against 14.08T, five-shot accuracy 54.37% against 51.93%.
Go deeper on this section: ClaudeChatGPT
Two Chinese framings of American results
A Chinese industry blog says the gap is a full generation. FuneralAI, translated by ChinaTalk, argues K3 and Qwen 3.8 Max sit a major generation behind Anthropic's Fable 5 and that GPT-6 may widen it. The frame is Maoist on purpose: quick-victory theory and defeat theory are two sides of one coin, and open weights are the united front.
Its evidence is the K3 episode. The release prompted a US move to restrict American companies from deploying open-source models; more than 270 companies and organisations including Nvidia and Microsoft signed an open letter; nothing happened. A closed model, the author notes drily, would have had only Jensen Huang speaking up, and only to sell GPUs.
Fermat's Last Theorem came home as a Tsinghua story. Claude's 11-day Lean formalisation ran on Prove2Me, the theorem-graph platform built by Peng Tianyi — a Columbia assistant professor and Anthropic researcher, and a Tsinghua Yao Class alumnus, which is the part New Wisdom puts in the headline.
What each side keeps. The Chinese version adds Peng's motive: an adviser once asked whether he was certain a proof was right, he said 99%, and the gap cost him a Nature submission. It drops Anthropic's own caveat that the 13m-line proof is far longer than it needs to be.
Go deeper on this section: ClaudeChatGPT
Dictating instead of typing, heroes who de-risk
Voice input has gone from fallback to default. Daily use was 32% three years ago and is now put at 65-70%. Tencent's Sogou keyboard claims over 100m AI users and nearly 2bn voice inputs a day, more than 20,000 people a second talking at a screen. In the 2025 keyboard survey, AI voice ranked first among AI features at 40.9%; the reason given most was avoiding the tedium of typing, at 48.6%.
The arithmetic is real. Typing runs 40 to 60 words a minute against 130 to 150 spoken, and ByteDance's Doubao keyboard ships a 150MB offline model at 0.8-second latency. Whoever owns the keyboard owns the first gate to every agent, which is why Tencent, ByteDance, iFlytek and Zhipu have all shipped one.
The male power fantasy is being replaced by risk management. TMTPost's read of a crowded slate is that audiences have gone immune to the flawless protagonist. Youku's hit cultivation series belongs to a genre whose hero's entire goal is to avoid entanglement and survive; comments say it cured their burnout. A producer's line: audiences do not hate the genre, they hate the formula.
Go deeper on this section: ClaudeChatGPT
Threads we are pulling
- Jiangxi's landslide toll rose. Yesterday I ran one dead and 11 missing at Gaoping, in Suichuan county, after Typhoon Saudel's rain. It is now 4 dead and 9 missing, a body having been found at Yangfen village in neighbouring Zuo'an, with 641 rescuers on site by Friday night, per Jiemian. Eight houses were destroyed; a middle school three miles away took in 400 people. Jiangxi Fire says collapsed floor slabs are buried in layers and further slides hit during the operation.
- The Nepal compute answer got a new instrument. Fengyun-4C, launched last December, is feeding thunderstorm intelligence to the China-Nepal search from 36,000km up, its lightning mapping imager running 500 frames a second through 30 gigabits of data, per SCMP. Last Tuesday I flagged the meteorological retasking; it has kept escalating.
- OpenAI conceded the disclosure gap. After the DseWiki agent edits I ran Friday and Saturday, it posted on 5 September that it should have had standards for when and how to disclose misalignment incidents rather than only model properties, with a framework promised in weeks, via GeekPark.
- CATL rewrote its supplier contracts. Three copper-foil agreements: 400,000 tonnes of new capacity with two partners over three years, priced on cost-plus with locked processing fees and shared risk, plus prepayments and better payment terms. H1 revenue 276.917bn yuan (~$39bn), up 54.8%, per IT Home. A supplier that sets its suppliers' margins is no longer buying, it is governing.
- A dead giveaway in your headphones. A developer debugging why Bluetooth multipoint would not hand audio to his phone found AliExpress holding a silent Web Audio stream open — a zero-gain node still wired to the hardware output, fingerprinting the device through floating-point differences, per InfoQ. No permission prompt exists for this.