China Desk

China News, Summarized 20 Oct 2026, 12:07 UTC 34 stories 22/23 sources

Today's brief

Washington hit China's AI funders and buyers in one day, and DeepSeek shipped regardless

Washington sanctions the investor, bans the model

The target is the funder. On Monday the State Department sanctioned five Chinese entities under the Protecting American Intellectual Property Act. One is Shanghai Zhisuan Technology, founded by Ding Linwei, the ex-Google engineer convicted in January of stealing AI trade secrets. The other four are MiraclePlus, the Beijing venture firm that seeded Zhisuan, and three affiliates.

Investor liability is the precedent. The theory is that MiraclePlus took equity in Zhisuan and gave Ding a stage to pitch replicating Google's infrastructure. Its founder, Lu Qi, is a former Microsoft AI executive who developed Y Combinator's China arm, then kept it going under his own brand. Every Chinese VC now has a reason to vet founders' former employers. That inference is mine.

The DeepSeek order sat for four months. Homeland Security Secretary Markwayne Mullin signed it on 26 June; it surfaced on 19 October. Civilian agencies must stop buying from DeepSeek and must remove its products, including third-party tools "developed using DeepSeek implementations". Private use remains legal. I can't tell whether surfacing it a month before the Shenzhen summit was deliberate.

Both sides are stacking leverage. Negotiators are racing to close investment deals before the leaders meet at APEC in Shenzhen next month. Congress has the MATCH Act, which would restrict immersion lithography exports. ChinaTalk's guest says China bought about 90 such systems a year in 2024–25, nearly two-thirds of ASML's immersion exports.

The coverage gap. Bloomberg's headline sells a brand: "China's Y Combinator". State's own text describes a funding chain. No mainland outlet in today's feed mentions either action. They covered DeepSeek's code release, a fired intern's funding round and Singles' Day. I found no response yet from Beijing or MiraclePlus.

WHAT WE'RE WATCHING: whether Beijing answers before Shenzhen, and whether it names the VC firm.

Go deeper on this section: ClaudeChatGPT

DeepSeek opened the plumbing anyway

DeepInfer is the serving stack. DeepSeek open-sourced DeepInfer under Apache 2.0. It runs on Nvidia's SM100 GPUs and Huawei's Ascend 950 chips. Zhidx relays DeepSeek's claims: 155 trillion tokens a day from one cluster, and $38 of revenue per card-hour at off-peak prices. Both are self-reported, and I have seen no reproduction.

The design is worth reading. Decode runs wide expert parallelism across 64 to 128 cards. KV-cache blocks live on a roughly 100 PiB SSD cluster, so any prefill node can hit any prefix, at the cost of prefix-affinity routing. "Bounded replay" recomputes only the last 128 tokens of the sliding-window layers, nearly halving prefill compute.

The catch is what's missing. The KV-cache system and load balancers are not in the release, though a complete service needs both. My read: DeepSeek published the engine and kept the cache tier behind its unit economics. A procurement ban reaches buyers, not a repository, and this landed a day after the order.

MiniMax matched it within 48 hours. MiniMax's M3.1 Flash launched Monday at 1 yuan (~$0.14) per million input tokens and 4 yuan (~$0.56) output, with a 1M-token context and 200-plus tokens a second. By MiniMax's own numbers it ties DeepSeek's V4.1-Flash-1017 on agent benchmarks, at half DeepSeek's peak price.

Chinese labs can now raise prices. TMTPost's tally: Chinese models consumed 86.74 trillion tokens on OpenRouter in the week of 5 October, 4.5 times US models, a lead 24 weeks long. Zhipu, the lab behind GLM, raised average API prices 101% without losing customers. MiniMax tried a hike, fell 15.71% in a day and halved prices. OpenRouter excludes most direct traffic to OpenAI and Anthropic, so it measures only contestable demand.

Go deeper on this section: ClaudeChatGPT

Who the AI economy actually rewards

A fired intern raised $30 million. Tian Keyu was a Peking University doctoral student and ByteDance research intern in 2024. ByteDance says he repeatedly killed colleagues' training jobs over compute allocation. It fired him, sued for 8 million yuan (~$1.1m) and won 500,000 yuan (~$70,000). This month his unnamed world-model lab, about ten people, raised nearly $30m at a $200m valuation from 5Y Capital and IDG Capital.

The anger is about the price gap. After a 12 October podcast where he called it "plain justice", five ex-colleagues published a rebuttal. They allege worse damage and an open model passed off as in-house. Those are allegations. TMTPost's commentary reads the fury as arithmetic: a penalty worth about three years of intern pay, a reward of $200m. The investors have said nothing.

Same question, two enforcers. Washington sanctions the investor who backed a founder carrying someone else's secrets. China's internet interrogates its investors by shame. The pairing is mine; neither source draws it.

The drama factory broke. Per TMTPost, Soy Sauce Culture, an AI-drama studio founded by Huang Haonan, grew past 1,200 staff in late 2025. This year it halved targets of 1bn yuan (~$141m) revenue and 200m yuan (~$28m) profit. About 150 finished titles sit unreleased, costing 12m yuan (~$1.7m), and staff are down to 600-odd.

The unit economics are brutal. TMTPost's film-industry piece says an AI-made drama costs 10,000 to 20,000 yuan (~$1,400 to $2,800) against about 300,000 (~$42,000) for live action. But revenue per ten thousand plays has fallen from 30–100 yuan (~$4–$14) to 5–10 yuan (~$0.70–$1.40).

Yuewen opened its vault. The web-novel publisher reported first-half short-drama and animated-drama revenue of 4.3bn yuan (~$606m), triple last year. It says it has adapted under one in a thousand of its stories. Huang told reporters no AI-drama studios will remain after 2026, only one-person companies. My read: when a few people can make a drama in days, margin moves to whoever owns the rights and the shelf.

Go deeper on this section: ClaudeChatGPT

Threads we are pulling

  • The export draft got a rulebook. On 16 October I called Beijing's draft a gatekeeper, and on 18 October I couldn't say why labs weren't flinching. Sinocism's summary of Friday's proposal covers 23 AI technologies, from models to compute clusters. Public technology is excluded, new open releases could get automatic licensing, and API and SaaS delivery generally sit outside. Registration applies immediately, including to transfers underway. I now read it as a ledger more than a gate. DeepInfer on Ascend tests the cluster category.
  • Unitree warned on itself. After the slide from 1,100 yuan to about 420 that I tracked on 15 October, the robot maker's shares rose enough over three sessions to trip the Shanghai exchange's abnormal-trading rule. Its notice cites a trailing price-to-earnings ratio of 388.25 against 39.51 for its industry, and warns of valuation reversion.
  • The home car market kept shrinking. On 19 October I argued that Chen Deming, the former commerce minister, had diagnosed weak demand yet advised going abroad. 36Kr now cites industry data showing domestic car retail down 28% in September. Li Auto, Wenjie and Xiaomi swung to first-half losses. Wenjie will enter Europe in autumn 2027 with Bosch supplying its driver assistance.
  • CATL's profit lagged its sales. The battery maker, reportedly adding about 10bn yuan to DeepSeek's round, posted third-quarter revenue up 58.58% to 165.2bn yuan (~$23.3bn). Net profit rose only 14.54%, to 21.2bn yuan (~$3.0bn).

Go deeper on this section: ClaudeChatGPT

The river · 34

DeepSeek open-sources its inference engine and production tokenizer

# AI Zhidx 1h

DeepSeek released DeepInfer and deepseek-recipe, exposing substantial production inference engineering while retaining some infrastructure-coupled components privately.

Also: Hacker News

More
  • DeepInfer supports distributed inference, speculative decoding, structured output, and operation on both Nvidia and Huawei Ascend hardware; deepseek-recipe aims to match the official API's tokenization behavior.
  • The report describes separate prefill and decode layouts, with wide expert parallelism for decode and asynchronous CPU-side scheduling to keep GPUs busy between steps.
  • A central systems choice is long-lived KV caching on a separate storage cluster, allowing any prefill machine to reuse a prefix but giving up prefix-affinity routing.
  • For sliding-window attention, DeepSeek says its newer bounded replay recomputes only the latest 128 tokens instead of persisting a large cache, nearly halving total prefill computation in the described setup.
  • The company says its V4.1 Flash service reaches 1.55 quadrillion tokens per day per cluster and reports revenue equivalent to $38 per card-hour at its stated pricing basis; these are company disclosures, not audited unit economics.
  • KV-cache management and load balancing remain closed because they are tightly coupled to internal infrastructure, so this is a meaningful but incomplete release.
Open at Zhidx Ask Claude ChatGPT

Beijing proposes export controls on AI, biotech, and photovoltaic technology

# Policy Sinocism 13h

Chinese coverage foregrounds a draft export-control catalog spanning AI, biotech, and solar technologies, alongside a coming Party plenum and an official 4.5% third-quarter growth figure.

More
  • The Ministry of Commerce and Ministry of Science and Technology proposed adding 23 AI technologies, including large models, agents, chip design, computing clusters, and selected cyber and information-processing capabilities.
  • The proposal would exclude technology already public, and the newsletter says API and SaaS delivery generally fall outside large-model export management; the final scope remains uncertain.
  • Registration requirements for covered exports, including transfers already underway, reportedly took effect immediately even as comments on the wider catalog remain open.
  • The proposed rules create a conditional pathway for international cooperation, with Beijing promising more facilitation for technology exports supporting members of the World Artificial Intelligence Cooperation Organization.
  • The move treats model architecture, compute organization, and related technical know-how as exportable strategic assets, not merely commercial products.
  • A Qiushi editorial cited in the newsletter says the upcoming Fifth Plenum will focus on strict Party governance; the excerpt does not provide the plenum's eventual decisions.
Open at Sinocism Ask Claude ChatGPT

Chinese chipmaker Zhongxing Micro enters the high-speed interconnect market

# Analysis QbitAI 1d

Zhongxing Micro says it has shipped 800,000 interconnect products and now offers PCIe 6.0 switching, 800G networking, and storage controllers for AI clusters.

More
  • The company's product range addresses a bottleneck that grows with cluster size: networking and storage interconnect can limit effective compute even when accelerator capacity increases.
  • Its PCIe 6.0 switch supports up to 144 lanes and CXL memory pooling, while the 800G adapter targets high-bandwidth AI fabric.
  • The company reports PCIe 5.0 bandwidth comparable to an international incumbent and higher RAID write bandwidth in a specified test, but independent test results are not supplied.
  • A domestic alternative matters because interconnect is deeply coupled to system software, reliability, and supply assurance; product breadth and shipment claims still need customer-scale validation.
  • The report also cites a national industrial plan that names high-speed interconnect as a priority, linking a private vendor's roadmap to China's broader compute self-reliance agenda.
Open at QbitAI Ask Claude ChatGPT

Calls grow for China to release detained US scholar Min Zin

# Policy China Digital Times 9h

A coalition of democracy groups is calling for Min Zin's release after China detained the US citizen on suspicion of espionage; Chinese authorities have offered little public information, the groups say.

More
  • The source says Min Zin was detained in Kunming on June 3 after traveling at a Chinese university's invitation for an academic workshop.
  • China's Foreign Ministry reportedly said he was suspected of activities endangering national security through espionage; the supplied account does not include evidence supporting the accusation.
  • The groups say his location, condition, and access to counsel remain unclear, while the US government has designated the case as wrongful detention.
  • The case raises risks for foreign scholars and conference participants working in China, particularly where academic exchange intersects with national-security investigations.
  • The coalition's characterization is advocacy, not a court finding; the distinction matters because the underlying Chinese case has not been publicly explained in the excerpt.
Open at China Digital Times Ask Claude ChatGPT

MiniMax M3.1 Flash targets coding and agent workloads at high speed

# AI Zhidx 21h

MiniMax released a one-million-token model that it says reaches more than 200 tokens per second and approaches DeepSeek V4.1 Flash on several agent benchmarks.

Also: Hacker News

More
  • The report says M3.1 Flash scored 62.2 on an automation benchmark, slightly above DeepSeek's 60.5, while DeepSeek scored higher on the DeepSWE coding test.
  • MiniMax lists input pricing at 1 yuan and output pricing at 4 yuan per million tokens, with cached reads at 0.1 yuan; it says agent-task costs fell 70% versus its prior model.
  • The article describes a sparse-attention design and reports median output speed above 170 tokens per second; the measurements are vendor claims and benchmark comparisons may not transfer to production workloads.
  • The release, arriving shortly after DeepSeek's Flash model, signals an emerging contest among Chinese labs for inexpensive, fast agent inference rather than a single obvious domestic leader.
  • The company says anonymous versions already attracted heavy use on model-routing platforms, but the article supplies no audited adoption or revenue numbers.
Open at Zhidx Ask Claude ChatGPT

HIL-UMI trains robot policies without putting a robot in the loop

# AI Leiphone 8h

A Peking University and Qiyuan Robotics team proposes a human-in-the-loop method that uses model disagreement to collect targeted correction data without a physical robot.

More
  • A human demonstrates a task with a handheld UMI gripper while a policy predicts actions from the same camera view; large differences trigger additional demonstrations in likely failure cases.
  • The system samples multiple action chunks and uses an energy-based disagreement signal to detect out-of-distribution situations rather than recording every routine success.
  • An advantage estimator filters for demonstrations that advance the task, intended to avoid teaching meaningless movement as well as correct actions.
  • After training on the targeted data, the model is used to identify new blind spots in the next collection round, creating an iterative data-selection loop.
  • The design addresses a major robotics scaling cost: conventional human-gated DAgger requires a robot, operator, and repeated scene resets to capture corrections.
  • The supplied report describes the method and a paper, but the excerpt ends before giving quantitative results or evidence of transfer to deployed hardware.
Open at Leiphone Ask Claude ChatGPT

DeepSeek's Flash model tops an open coding leaderboard

# AI Zhidx 1d

DeepSeek V4.1 Flash-1017 led Arena's Code Arena WebDev ranking among open models, with the report highlighting low API prices and strong browser-based prototype generation.

More
  • The report lists a score of 1,670, just 12 points below GPT-6 Sol, and says the model's average task cost fell about 60% from its prior version.
  • In outlet-run demos, it generated a 3D Kaifeng scene and a browser racing game for about 6 yuan combined; these examples show prototyping, not general software reliability.
  • The model's listed output price is 4 yuan per million tokens off peak and 8 yuan at peak, with much cheaper cached input.
  • The release strengthens DeepSeek's position in low-cost coding, but leaderboard rank and two demos do not establish performance on large production codebases.
Open at Zhidx Ask Claude ChatGPT

US lawmakers weigh restricting DUV lithography exports to China

# Policy ChinaTalk 1h

ChinaTalk argues that immersion lithography is a key remaining US-allied leverage point over Chinese chipmaking, while noting China's large recent equipment purchases.

More
  • The podcast summary says China imported roughly 90 immersion lithography systems a year in 2024 and 2025, nearly two-thirds of ASML's immersion exports; that is the author's cited estimate.
  • The proposed MATCH Act would restrict DUV immersion exports and servicing, an effort that would depend on allied coordination rather than US action alone.
  • The strategic claim is that equipment access, maintenance, and upgrades could constrain China's ability to scale advanced AI chip production over the next decade.
  • The argument also emphasizes uncertainty: broad restrictions could provoke retaliation, and the long-term effect depends on China's ability to indigenize the manufacturing stack.
Open at ChinaTalk Ask Claude ChatGPT

US sanctions five Chinese entities over alleged AI trade-secret theft

# Policy SCMP 13h

Chinese coverage reports sanctions on five entities linked by Washington to an AI trade-secrets case; US coverage singles out MiraclePlus and its alleged connection to Zhisuan.

Also: Techmeme

More
  • The State Department imposed measures under the Protecting American Intellectual Property Act, targeting Shanghai Zhisuan Technology, MiraclePlus, and three related entities, according to the report.
  • Zhisuan was founded by former Google engineer Ding Linwei, whom the article says was convicted of seven trade-secret theft counts in January.
  • The sanctions are a US government response to alleged theft and ties; they do not establish that each named entity committed the underlying offense.
Open at SCMP Ask Claude ChatGPT

China Telecom's 29B intent model extracts structured action parameters

# AI Zhidx 23h

The open-source Xing4.0-29B-Intent aims to combine intent classification with extraction of entities and values into constrained JSON for agent workflows.

Also: New Wisdom

More
  • In a lighting example, the model assigns separate brightness and color-temperature values to two rooms while leaving a third light untouched, according to the article's test.
  • The report says end-to-end latency is below 150 milliseconds and describes the model, training data, and inference SDK as open source.
  • The article compares it with Jev-style decision models, arguing that bounded classification does not by itself solve free-form parameter extraction.
  • The tests and performance claims come from the outlet's own scenarios and the release announcement; no independent benchmark is supplied in the excerpt.
Open at Zhidx Ask Claude ChatGPT

Chinese aircraft and ships tracked US-Japan drills near a Taiwan contingency

# Policy SCMP 7m

Japan says PLA aircraft, drones, and ships monitored drills focused on a possible Taiwan crisis, including a vessel that could support the Liaoning carrier.

More
  • The reported activity placed reconnaissance aircraft and drones in international airspace between Japanese islands before they returned toward mainland China.
  • The presence of a supply ship potentially associated with carrier operations points to the logistics dimension of PLA activity, though the excerpt does not confirm a direct carrier link.
  • The drills and tracking took place as Tokyo and Washington rehearsed for a Taiwan contingency, sharpening the signaling value of the encounter.
  • Japan's account is the source for the sightings and scramble; the excerpt does not include a detailed Chinese military response.
Open at SCMP Ask Claude ChatGPT

Yuewen opens its IP library to AI-assisted one-person studios

# Culture 36Kr 6h

Chinese publisher Yuewen is connecting its fiction catalog to AI animation tools and independent creators, while restricting IP access and limiting AI-generated fiction on its own writing platform.

More
  • Yuewen says short-drama and animation revenue reached 430 million yuan (about $61 million) in the first half, more than triple the year-earlier period; it also reports 5,000 creator or studio partners.
  • Its DramaBuddy platform is intended to connect licensed stories, AI production, creator matching, and distribution rather than earn mainly from selling model tokens.
  • The company says fewer than one-thousandth of its IP catalog has been adapted, while AI has shortened sample production from years of studio work to roughly a month in some cases.
  • Access is not fully open: Yuewen says creators must demonstrate production ability because weak adaptations could damage valuable rights.
  • The platform is also drawing a boundary around authorship: it says literary works themselves should remain human-written, is limiting daily output, and is cleaning up low-quality AI-written material.
  • This is a strategic shift from controlling only the catalog to controlling the workflow and creator marketplace that converts IP into derivative content.
Open at 36Kr Ask Claude ChatGPT

China says its navy followed a US ship near disputed South China Sea waters

# Policy SCMP 1h

Beijing accused a US Navy vessel of entering waters near the Spratly Islands without approval; Chinese coverage foregrounds China's sovereignty claim and the PLA's monitoring response.

More
  • The Southern Theatre Command Navy said it monitored and followed the USS Charleston near Sabina Shoal; the US position is not included in the supplied excerpt.
  • The report uses Beijing's description of the entry as an illegal intrusion, while the central fact is a recurring freedom-of-navigation encounter in contested waters.
  • Such incidents make competing legal claims visible without changing who exercises day-to-day control; the excerpt provides no escalation beyond tracking.
Open at SCMP Ask Claude ChatGPT

China protests UK searches and phone seizures at its border

# Policy SCMP 6h

Beijing formally protested after Chinese travelers and trade delegates were questioned for extended periods and had phones or other devices seized by UK border authorities.

More
  • The Chinese embassy says the affected travelers included tourists, trade visitors, and people visiting relatives; the supplied account does not include the UK's explanation.
  • The dispute comes amid broader espionage tensions and raises a practical concern for business travel: personal devices may be treated as security evidence at the border.
  • The episode shows how reciprocal scrutiny can spill from intelligence disputes into ordinary travel and commercial exchange.
Open at SCMP Ask Claude ChatGPT

Neolix raises nearly $400 million for driverless delivery expansion

# Analysis Leiphone 7h

Chinese autonomous-delivery company Neolix says it raised nearly $400 million to scale operations abroad and invest in its end-to-end driving models.

More
  • The company describes a cloud-edge-vehicle model architecture and a vision-led driving stack, with data from its delivery network feeding model development and operations.
  • Neolix says it has deployed in the Middle East, Europe, East Asia, Southeast Asia, and Australia; the report supplies no audited vehicle counts or revenue by region.
  • The financing arrives as Chinese transport authorities are discussing pilot applications for unmanned delivery vehicles, tying commercialization to national logistics and industrial policy.
  • A key proof point will be paid utilization and safe operation across different local rules, not the size of the funding round or claimed fleet leadership.
Open at Leiphone Ask Claude ChatGPT

Chinese AI labs are testing whether high usage can translate into pricing power

# Analysis TMTPost 2h

Chinese models lead one third-party usage ranking, and several labs have raised prices without an immediate reported collapse in demand, challenging the industry's discount-first story.

More
  • The article cites OpenRouter data showing Chinese models accounted for 47.9% of measured global calls in one week; OpenRouter does not capture all direct usage by US labs, so this is not a market-wide revenue comparison.
  • It says Zhipu raised average API prices while usage on its MaaS platform grew, and DeepSeek introduced peak pricing while maintaining demand; the article attributes these figures to company and press sources.
  • DeepSeek is reported to have exceeded $1 billion in annualized revenue, with a 44.6% overall gross margin in the first seven months; these figures are not audited in the supplied text.
  • Kimi reportedly reached a $1 billion annual revenue run rate after its K3 release, while the article says MiniMax reversed a price increase after developer reaction and a stock decline.
  • The strategic milestone is not raw token volume but evidence that some users may pay for capability and availability; a single ranking or price change does not prove durable economics.
  • The piece distinguishes open-model API sales, Alibaba's open-source ecosystem, and ByteDance's internally generated product demand as different routes to scale.
Open at TMTPost Ask Claude ChatGPT

China shows a Y-20B tanker-transport aircraft in an official image

# Policy SCMP 2h

A PLA Daily image appears to show a Y-20B heavy transport aircraft fitted with aerial-refueling pods, suggesting a tanker-transport configuration may have entered service.

More
  • The aircraft is shown with refueling pods under its wings and a cargo ramp open, indicating a design intended to combine transport and tanker roles.
  • The excerpt describes this as an apparent first official photograph and says the image suggests service entry; it is not an explicit confirmation by the military.
  • A tanker variant would extend the reach and endurance of Chinese aircraft, but the report gives no fleet size or operational deployment evidence.
Open at SCMP Ask Claude ChatGPT

Chinese maker prepares a 0.1-nanometer-resolution electron microscope for sale

# Analysis SCMP 8h

Hefei-based CIQTEK says its 200 kV transmission electron microscope reaches a 0.1 nm information limit, bringing a Chinese product into a field led by US and Japanese suppliers.

More
  • The instrument's reported point resolution is 0.25 nm, while the 0.1 nm information limit is the more demanding figure highlighted in the report.
  • The story marks a move from China's growing microscopy research base toward commercial instruments that could compete with established workhorse models.
  • The report does not provide independent benchmark comparisons, customer deployments, or production volumes, so the commercial challenge remains to be demonstrated.
Open at SCMP Ask Claude ChatGPT

Chinese buyers say privacy and reliability drive demand for on-device AI

# Analysis GeekPark 1d

Five Chinese industry executives argue that local inference is attractive when AI handles sensitive records, must stay available, or needs low latency in homes and workplaces.

More
  • Examples include an elder-care assistant that reminds residents about medication and fraud, and enterprise tools handling invoices, financial statements, and orders that customers do not want sent to public cloud services.
  • The panelists say buyers judge experience through latency, noise, uptime, and integration, not simply token savings or model size.
  • Local inference is not a free security guarantee: the article does not discuss device hardening, updates, or the risks of compromised endpoints.
  • The business model described is closer to deploying a role-specific digital specialist than selling a generic model, with hardware, software, and service delivery bundled together.
  • The claim that Chinese AI adoption will be edge-first is a panelist's prediction, not established market data; the examples suggest a real demand driver but not yet a universal architecture.
Open at GeekPark Ask Claude ChatGPT

US and China seek investment deals ahead of a leaders' meeting in Shenzhen

# Policy SCMP 14h

Negotiators are trying to finalize investment and economic arrangements before next month's leaders' meeting, while both governments bargain over leverage in extending a wider trade agreement.

More
  • The report, citing unnamed sources, describes ongoing talks over a complex package of institutional and economic frameworks.
  • The timing suggests the meeting is being used to produce concrete commercial outcomes, not only political optics.
  • The excerpt does not identify the likely deals or resolve the obstacles, so the scope of any agreement remains unclear.
Open at SCMP Ask Claude ChatGPT

China's AI short-drama boom exposes the limits of volume-first production

# Culture TMTPost 1h

A fast-growing AI animation studio cut its staff after a production-at-scale strategy stopped paying, while platforms and regulators tightened content controls.

More
  • The article describes Jiangyou Culture growing to more than 1,200 staff and producing about 100 AI dramas in one month, then shrinking to roughly 600 as many titles remained unreleased.
  • Its reporting says around 150 unreleased productions represented about 12 million yuan (roughly $1.7 million) in production costs, while the company's revenue and profit targets were halved.
  • The underlying economics changed as AI lowered production barriers: output volume became less scarce, while scripts, rights, quality, and audience demand remained constraints.
  • Red Fruit platform figures cited in the article show thousands of low-quality or noncompliant dramas blocked or removed over several months; the piece cautions that the studio's losses cannot simply be attributed to enforcement.
  • China's short-drama rules, in force from September according to the report, require stronger content governance and AI-generated-content labeling, narrowing the space for a low-cost content lottery.
  • The story connects automation to labor as well as margins: a model that once scaled through large teams can become a fixed-cost burden when a few-person crew can produce comparable material.
Open at TMTPost Ask Claude ChatGPT

iQIYI's AI push reshapes both content production and internal work

# Analysis Leiphone 5h

A long profile traces iQIYI's multi-year effort to apply generative AI to script analysis, production, and viewing personalization while training staff across departments.

More
  • The platform's 'skip-view' feature uses viewing behavior and AI to identify plot highlights, offering a shorter route through long-form dramas.
  • For script analysis, content specialists and engineers annotated scenes and narrative structure in works including The Bad Kids, creating structured material for model development.
  • The company trained employees broadly and ran cross-functional AI projects, showing how domain expertise can become part of an internal model-development pipeline.
  • The article emphasizes the risk of simple model wrappers: iQIYI reportedly rejected consumer AI products that merely called general-purpose model APIs without creating a defensible capability.
  • The profile is strongly company-framed and provides no independent measurement of whether these tools improved retention, production economics, or creative quality.
Open at Leiphone Ask Claude ChatGPT

A Chinese medical-imaging model filters noisy MRI inputs before fusion

# AI Leiphone 23h

A Chinese university team proposes CoReFuse-Med, which tries to prevent low-quality MRI modalities from contaminating multimodal lesion segmentation.

More
  • The paper reports that adding T2 and FLAIR data reduced performance versus high-resolution T1 alone on one dataset, despite spatial alignment.
  • Its method separates shallow feature cleanup from deeper modality balancing: anisotropic encoders and residual filtering address resampling artifacts, while a shared-query cross-attention module recalibrates modalities.
  • The authors report a Dice score of 0.7326 on the EPVS task, above the listed MedNeXt baseline of 0.7085; the supplied excerpt does not give enough detail to judge external generalization.
  • The central technical point is that geometric alignment does not guarantee equal information quality; a weak modality can still introduce strong inference-time noise.
Open at Leiphone Ask Claude ChatGPT

Aito turns to Bosch for overseas driver assistance and cockpit systems

# Analysis 36Kr 2h

36Kr reports that Aito's European models will use Bosch's full intelligent-vehicle package, while Chinese coverage frames the deal as part of a shift in overseas control from Huawei to Seres.

Also: IT Home

More
  • The reported package spans advanced driver assistance, the digital cockpit, and chassis systems for overseas M6 through M9 models; Bosch is said to be developing it now.
  • The report says Aito's domestic identity is built around Huawei-Seres cooperation, but Huawei has handed overseas vehicle leadership to the automaker.
  • Bosch's role may help with local regulatory adaptation and brand recognition, but the report does not establish how much of Huawei's software or hardware will remain in the exported cars.
  • The same partnership is covered in a shorter item; this account adds the wider export strategy and reported organizational context.
Open at 36Kr Ask Claude ChatGPT

A founder's past workplace sabotage collides with China's AI startup culture

# Culture TMTPost 2h

Chinese online debate erupted after a researcher accused of disrupting colleagues' model training received fresh startup funding, with many questioning investors' standards and accountability.

More
  • The article says a court ordered the founder to pay ByteDance 500,000 yuan after he repeatedly shut down colleagues' training jobs; two years later, his new lab reportedly raised nearly $30 million at a $200 million valuation.
  • Five former colleagues publicly disputed his account of the earlier incident and alleged that he misrepresented open-source work as original research; these are allegations in the report, not established findings here.
  • The controversy shifted from what happened at ByteDance to whether early-stage investors should treat founder character as relevant diligence or prioritize technical potential.
  • The article compares US stories of rebellious technical founders, but notes that serious workplace system interference can also bring criminal liability there; its cultural contrast is more complex than a simple China-versus-Silicon Valley framing.
  • The deeper public grievance described is the perceived mismatch between a relatively small legal penalty and the large financial upside of a new fundraising round.
Open at TMTPost Ask Claude ChatGPT

Cell Press investigates allegations that a journal editor sought bribes

# Culture SCMP 11h

Cell Press opened an investigation after a prominent Chinese biologist accused a medical-journal editor of soliciting payments from mainland authors.

More
  • The allegation comes from Rao Yi, a Peking University professor and former university president; the excerpt does not identify the journal or provide the editor's response.
  • If substantiated, the claim would point to a vulnerability in international publishing incentives where authors face pressure to secure visibility and acceptance.
  • The publisher's investigation is the reported action; the allegation should not be treated as established fact.
Open at SCMP Ask Claude ChatGPT

Chinese quantum-computing firm shifts from chips toward callable cloud services

# Analysis QbitAI 21h

A Chinese quantum-computing company is positioning its hardware as a service through a planned thousand-qubit cloud platform, while acknowledging that engineering maturity remains the bottleneck.

More
  • The report describes a hybrid-computing vision in which quantum processors complement classical systems for selected optimization and molecular-simulation problems rather than replace GPUs.
  • The company says moving from lab prototypes to usable systems requires improvements across qubit fidelity, control electronics, cryogenic operation, calibration, and software tooling.
  • The article notes that current applications in finance and molecular simulation are mostly validation exercises; qubit count alone does not demonstrate useful computational advantage.
  • A cloud-access model could lower the capital barrier for users, but programming still requires specialized quantum circuits and the article offers no production performance data.
Open at QbitAI Ask Claude ChatGPT

Alibaba's AliExpress plans to invest in early-stage AI hardware brands

# Analysis Leiphone 1d

AliExpress and Tmall say they will combine startup investment with domestic brand support and overseas commerce services, aiming to incubate more than 100 AI hardware brands.

More
  • The plan moves the marketplace upstream: instead of waiting for finished products, the platform wants to finance companies before their brands and products are established.
  • The proposed package combines capital with Tmall product launches and AliExpress marketing, logistics, and compliance support, aligning investor returns with future sales through the platform.
  • That can ease founders' distribution burden, but it also gives Alibaba an early influence over which brands reach its channels and where they sell.
  • The article uses industry anecdotes about rapid product copying and overseas demand, but offers no investment terms or evidence that the program has produced successful brands.
Open at Leiphone Ask Claude ChatGPT

A Peking University scientist is building an AI foundation model of the brain

# AI 36Kr 11h

LattixOrigin, led by a Peking University neuroscientist, raised nearly $10 million to build a multiscale brain model from fragmented human-cell datasets.

More
  • The team says it integrated more than 200 studies containing nearly 100 million human brain cells, addressing a field where samples are scarce and datasets lack consistent standards.
  • Its DigitalBrain-M1 is reported to align with imaging-based analysis in some brain regions; the article supplies no external validation or prospective drug-discovery result.
  • The research ambition is to connect molecular, cellular, regional, and whole-brain signals so models can help identify disease mechanisms and intervention windows.
  • The scientist argues that AI should reorganize biological research around standardized data rather than merely automate steps in researcher-defined workflows.
  • The next technical test is whether the dataset and model can predict outcomes in independent human samples and wet-lab experiments.
Open at 36Kr Ask Claude ChatGPT

Virtual power-plant startup raises tens of millions of yuan

# Analysis 36Kr 5h

Hefei-based Yuanli Energy raised several hundred million yuan (roughly $30-60 million) to expand its software and trading services for distributed energy assets.

More
  • The company says its platform links solar, storage, charging infrastructure, and industrial loads, but only about 2 GW of the more than 10 GW it has aggregated actively participates in market trading.
  • In a cited 2.5 MW/5 MWh storage project, the company says it changed dispatch after local pricing shifted from fixed time-of-use rates to spot-price pass-through, lifting the asset return from negative to about 4%.
  • This points to the commercial pivot for virtual power plants: connecting devices is not enough; operators need market access, forecasting, dispatch, and settlement capability.
  • Revenue is currently in the tens of millions of yuan and the company is not yet profitable; its 2027 profitability target remains a forecast.
Open at 36Kr Ask Claude ChatGPT

Huiyuan's trademark fight ends in a 97 million yuan damages ruling

# Policy TMTPost 1h

A Chinese court ordered related companies to stop using the Huiyuan name and pay more than 97 million yuan (about $14 million), prompting one side to launch a replacement brand.

More
  • The ruling concerns use of the Huiyuan trademark and company name by two firms; the article says the court ordered them to stop production and sales of the disputed products.
  • The damages calculation described by Beijing Huiyuan totals 97.3 million yuan (about $14 million), including lost profits, punitive damages, and legal costs.
  • The dispute sits within a long-running restructuring conflict: the article reports a disagreement over an unpaid 850 million yuan (about $120 million) investment balance and control of the revived business.
  • Huiyuan Group's announced move to a new name and brand is framed as both a response to the judgment and an effort to separate consumers from products it says no longer come from its operating system.
  • The underlying financing claims remain contested; the article presents competing accounts from the parties and notes related litigation is unresolved.
Open at TMTPost Ask Claude ChatGPT

Ant Group leads a funding round in financial-AI startup Shengjian

# Analysis Zhidx 1d

Ant Group led a new round in Shengjian Technology and plans to collaborate on financial agents and domain models for investment research.

More
  • Shengjian's AlphaEngine combines financial research documents with retrieval and analysis tools; the company says it serves more than 100,000 institutional investors across 9,000 organizations.
  • The startup previously distilled DeepSeek reasoning into its FinGPT model and has since launched an agent that can produce spreadsheets and research reports.
  • The investment extends Ant's AI activity toward institutional finance, beyond its better-known consumer-facing products.
  • Financial hallucinations, data rights, explainability, and customer willingness to pay remain substantial barriers; the article gives no new validation on these issues.
Open at Zhidx Ask Claude ChatGPT

GoGoX sells its mainland China business for one yuan

# Analysis Jiemian 17m

Hong Kong-listed GoGoX agreed to sell its mainland operating subsidiary for one yuan and says it will concentrate on Hong Kong and overseas markets.

More
  • The reported deal removes the mainland business from the company rather than merely changing its consumer-facing brand.
  • GoGoX listed in Hong Kong in 2022 as a leading urban freight platform; the article says its market value had fallen by more than 99% from its peak by mid-October 2026.
  • The nominal price underscores the weak value attributed to the divested operation, but the excerpt does not include the subsidiary's liabilities or full transaction rationale.
Open at Jiemian Ask Claude ChatGPT

Chinese film workers describe AI adoption as management theater

# Culture TMTPost 3h

A profile of film workers says some companies have shifted attention from projects to AI demonstrations, while generative tools cut production costs and intensify pressure on labor.

More
  • One producer describes weekly AI meetings and mandatory tool demonstrations displacing discussion of scripts, budgets, and project delivery; this is an interview account, not an industry-wide survey.
  • The article estimates AI short-drama production can cost 10,000 to 20,000 yuan versus about 300,000 yuan for a comparable live-action production, but says much of the savings is spent on traffic acquisition.
  • It reports that revenue per 10,000 views fell from 30-100 yuan to 5-10 yuan, illustrating how lower production costs can be offset by platform economics and competition.
  • The piece documents a labor shift from experienced producers toward AI tool experimentation, with employees describing projects that appear busy but fail to reach usable production.
  • The article's strongest insight is organizational: adopting AI as a visible ritual can satisfy anxious leadership without fixing the underlying shortage of viable projects.
Open at TMTPost Ask Claude ChatGPT
Source ledger
ChinAI 403 0/0 HTTPError: 403 Client Error: Forbidden for url: https://chinai.substack.com/feed
ChinaTalk 200 1/20
Pekingnology 200 0/20
Ginger River Review 200 0/20
Sinocism 200 1/20
QbitAI 200 5/10
36Kr 200 8/30
Solidot 200 12/20
Leiphone 200 15/20
GeekPark 200 4/30
IT Home 200 12/60 +34 over cap
SCMP 200 10/50 +11 over cap
Initium 200 3/15
Jiemian 200 8/30 +14 over cap
China Digital Times 200 1/15
The Wire China 200 0/50
Techmeme 200 6/15
Hacker News 200 3/20
V2EX 200 8/50 +40 over cap
Zhidx 200 10/20
InfoQ China 200 8/20 +12 over cap
New Wisdom 200 8/15 +2 over cap
TMTPost 200 8/17 +9 over cap