PHYSIX FRONTIER EASTVOICE
2026.10.19 · Monday
Issue 062
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
423 tokens per second per user
Kimi K3 serving throughput on GB300 NVL72
130 million
DeepSeek free user base to convert
100,000-GPU
China's first all-domestic AI cluster in Shenzhen
¥1.42 billion
Chinese rehab exoskeleton market in 2025
Top Stories 3 · dual commentary ｜ Section Picks 25 · one-line takes
China's first all-domestic 100,000-GPU cluster, DeepSeek's product division, and Xiaohongshu's AI reshuffle.
SemiAnalysis reports that an upstream vLLM optimisation lifted serving throughput on the 2.8-trillion-parameter Kimi K3 to 423 tokens per second per user on GB300 NVL72 hardware, a 1.68× gain over the same silicon without the patch.
Editor's take｜A 68% throughput gain on unchanged hardware is a price cut nobody had to announce. Serving cost, not leaderboard position, decides which Chinese models get used outside China.
On October 16 DeepSeek merged its Harness team with the Web, App and Platform groups into a single product division. The company now has to convert a 130 million-strong free user base into something a prospectus can explain, and a ChatGPT-style subscription is the most obvious answer on the table.
What it means for you｜DeepSeek now has to convert a 130 million-strong free user base into something a prospectus can explain, and a ChatGPT-style subscription is the most obvious answer on the table.
Editor's take｜Charging Chinese consumers directly combines the widest reach with the thinnest margin. The revenue will show up on the API line long before it shows up in the app.
Chinese startup StartLux, or Yuandian Xinghui, reached first place on Hugging Face's Decision Index 0.3.2 with StartLux-Decision-v1.1-27B Instant at 69.1 points, nearly three points clear of the runner-up across 173 reproducible models.
Editor's take｜Decision-making is the benchmark category Chinese labs have been contesting quietly. A 27B model at the top also undercuts the assumption that reasoning over tools requires trillion-parameter scale.
25 AI stories worth your time, each with a one-line take.
The first fully domestic 100,000-GPU AI cluster built in China went on public display at the National High Performance Computing Conference in Shenzhen, where organisers also announced ten flagship application results, including work aimed at building digital twins of biological systems.
One-line take｜Cluster counts are easy to print and hard to audit. What decides whether this pile of accelerators is useful is domestic interconnect bandwidth, and that figure is rarely published.
A report circulating in Chinese tech media says Google offered roughly $70 million to acquire and distil data from Spirit Airlines, the discount carrier that filed for bankruptcy this spring after years of losses.
One-line take｜A dead company leaves behind process knowledge with exactly one buyer worth calling. Purchases like this will get more common, and the litigation around them will follow.
Chinese outlet AI Era reports that NVIDIA has begun selling RTX Spark, pitched with a 74 billion parameter model small enough to run on a notebook.
What it means for you｜A laptop that holds a 74B model changes what developers are willing to ship without a cloud bill attached.
One-line take｜Local inference is drifting from demonstration to default. A laptop that holds a 74B model changes what developers are willing to ship without a cloud bill attached.
Jack Clark, co-founder of Anthropic, told an audience during the IMF and World Bank annual meetings that internal projections show AI adding close to 50% to US economic output by 2030, qualified by specific conditions the company attached to the forecast.
One-line take｜The words carrying the weight here are "under specific conditions". That clause is where capability forecasts hide their assumptions.
A TMTPost analysis argues that through 2026 the deciding variables have moved away from model rankings and GPU counts toward inference cost, developer ecosystems, cloud distribution, enterprise deployment and, increasingly, who writes the rules.
One-line take｜Open weights against closed ones stopped being a technical dispute the moment procurement teams started scoring them.
An assistant called "Xiaowei" appeared as a green dot in the top-left of WeChat for some users in June and is still labelled a test build. Swiping right on the home screen opens it, with no separate app and no menu hunting, the first time in years WeChat has surrendered such a prominent position to a new feature.
What it means for you｜A mediocre assistant inside WeChat reaches more people in a week than a good one reaches in a year.
One-line take｜Distribution does work that model quality cannot. A mediocre assistant inside WeChat reaches more people in a week than a good one reaches in a year.
Beijing, Shanghai, Jiangsu and Zhejiang have folded medical rehab exoskeleton training into medical insurance or long-term care pilots, after the national insurance administration said new devices should be covered by commercial health plans. The Chinese market reached ¥1.42 billion (about $200 million) in 2025.
What it means for you｜Insurance coverage is what turns a robotics demonstration into a purchase order.
One-line take｜Insurance coverage is what turns a robotics demonstration into a purchase order. The rehab segment will fill with hardware vendors before it fills with hospitals.
Xiaohongshu has rearranged its AI responsibilities twice since April, when the Dots team was elevated to a first-level AI unit covering models, infrastructure, engineering and product. Product and part of engineering moved to the community organisation in September, and on October 16 the company announced another change.
One-line take｜Two structural changes in six months usually means the first one never produced a number worth reporting. Consumer AI inside a content platform either shows up in engagement or it gets reorganised again.
The Financial Times reports Anthropic has been in talks with law firms over a fresh problem, how an AI agent should negotiate on its user's behalf when the stakes are haggling over a price, settling a dispute or discussing a salary.
One-line take｜An agent that haggles needs a mandate, a record and a hard limit. None of those exist as standards yet, which is why the lawyers were called first.
Meng Pu told an AI industry conference in Shenyang that on-device AI is already doing material work in deployed systems, and that the spread of agents accelerates the shift toward running inference on the handset.
One-line take｜Every agent vendor eventually hits the same wall, which is that memory and permissions live on the device. That is a chip argument, and Qualcomm is making it early.
John Felton told Axios that spending on AI infrastructure and services is sustainable because it funds genuine innovation, and put the payback period on AI server investment at under three years.
One-line take｜A CFO answering a debt question with a payback number is answering the only question investors are asking this quarter.
A Huxiu piece built on a16z's "Agents Denied" chart pack from October 16 asks what happens when the buying decision is delegated to agents that compare, negotiate and check out on their own, and where the retail margin ends up when that happens.
One-line take｜Whoever controls an agent's default choice controls the shelf. Retailers spent twenty years buying that position and are about to learn it can be leased to somebody else.
A Huxiu report on companies that neither lay off nor hire points to a 2026 labour case in Hangzhou involving a worker whose job was checking whether model answers to user questions were accurate, one of the first roles being hollowed out.
What it means for you｜People whose job is evaluating model output are the first to be told their output is no longer needed.
One-line take｜Freezing headcount conceals the transition without stopping it. People whose job is evaluating model output are the first to be told their output is no longer needed.
A year after launching robot after-sales service, JD.com says it has service agreements with close to 100 robotics companies, including AgiBot, DEEP Robotics, EngineAI and Songyan Power, and projects more than 100,000 robot maintenance engineering roles over five years.
One-line take｜A repair network is the unglamorous half of a robotics boom and the clearest sign that one is really happening.
A TMTPost report from a 5,000-square-metre training ground in Zhangjiang, Shanghai describes more than 100 robots from a dozen brands worked by young operators who repeat a single motion more than 500 times a day, with each scene requiring roughly a million accurate data points and about 200 days of collection per operator.
One-line take｜Data is the least scalable input in embodied AI, and it is currently being produced by hand, five hundred repetitions at a time.
Bloomberg's Mark Gurman reports that Apple plans a HomeView home hub fitted with a robotic arm that can move the screen, with the earliest launch in 2027.
One-line take｜Apple's first robot is a screen that follows you around a room. That is a far cheaper bet than a humanoid and it reuses the same sensor stack.
Agility Robotics will use Nyobolt's battery technology in the next-generation Digit 5, targeting nine minutes of charging for roughly 90 minutes of work and the goal of keeping the humanoid on shift instead of tethered to a cable.
One-line take｜Uptime is the only humanoid metric that converts into an hourly rate. Charging speed is what a warehouse manager actually notices.
In a podcast appearance with biographer Walter Isaacson, Musk said his attempt to cut federal spending from inside government did not fix the deficit, and argued that AI and robotics are the only remaining route out.
One-line take｜Productivity-driven growth is the standard answer from anyone whose product is meant to supply it.
Musk said Roadster owners will be able to instruct the car to perform the stunt sequence shown at the Texas launch event on October 16, with the vehicle driving itself throughout, and promised a ring track at the Texas headquarters.
One-line take｜A car that can be told to jump a ramp is a marketing artefact with a spectacular bill of materials. The autonomy claim underneath it is the part Tesla needs to actually ship.
Tesla stretched its Austin robotaxi window to run from 6 a.m. to midnight, the latest in a series of schedule extensions, with FSD v15 expected to support round-the-clock operation.
One-line take｜Service hours are the honest autonomous driving metric. Every extension is a published claim about performance at night, in weather and at the end of a shift.
In a MotorTrend walkthrough, Tesla chief designer Franz von Holzhausen discussed the Roadster's camera arrangement, and the report concluded the car cannot run the complete FSD stack, with its extreme design the main limiting factor.
One-line take｜Sensor placement is a design constraint rather than a software gap. The car carrying the loudest autonomy claims is the one that ships with the partial build.
XPeng said its GX has begun large-scale shipments to Middle East and African markets, with more than 3,000 regional orders taken before overseas pricing was even published.
One-line take｜Taking orders before publishing a price tells you which side of the market is short. Chinese exporters are finding demand easier to secure than supply.
A Korean semiconductor engineer claims Tesla pushed for generous terms in its foundry agreement with Samsung, seeking sensitive process information and audit rights alongside lower pricing, and says the Terafab project has been pulling talent out of the chipmaker.
One-line take｜Whoever writes the chip order also writes the terms. Vertical integration is a negotiating position long before it becomes a factory.
VAST released and open-sourced EdenLive, a real-time world model that runs at 720P and above 24 FPS on a single consumer-grade graphics card, aimed at interactive and creatable environments.
One-line take｜Real-time on a retail GPU is the threshold that decides whether world models become a product or stay a benchmark entry. Open-sourcing it puts the cost argument on the table.
The Elec reports that Micron is developing Edge Inference Memory, a memory chip aimed at physical AI, with the company outlining the work at a semiconductor forum in Seoul on October 16 and targeting lower data-processing latency and higher bandwidth.
One-line take｜Robots and cars cannot wait on a round trip to a data centre, which turns memory bandwidth into a robotics bill of materials line. Micron is positioning for the part of physical AI that is not a model.
📖 The full EastVoice issue on the Chinese forum