Sponsored by
In Todayโs Issue:
๐ China's AI gap shrinks to 3%
๐ข $300 million in AI servers allegedly smuggled to China
๐ง Neuralink trains on 50,000 hours of brain data
๐ฉ๐ช Aleph Alpha's open Kolibri model
๐ฆ Tencent rents 100,000 chips from Oracle
โจ And more AI goodnessโฆ
โก The Signal
America's lead in AI is now measured in single digits, and the walls built to protect it are leaking.
According to Bloomberg Intelligence, the best Chinese model now trails the best American one by about 3% on LiveBench, down from roughly 15% earlier this year. Today's issue shows how the computing power keeps arriving: prosecutors say one California dealer routed more than $300 million in GPU servers to China through Malaysia and Singapore, while Tencent is reportedly renting about 100,000 advanced chips in Oracle data centers in Southeast Asia, which US rules allow. Export controls assume chips are the bottleneck. A 3% gap suggests Chinese labs are getting enough of them, and getting more out of each one.
All the best,
Kim Isenberg
Greg Lui is arrested by federal agents in Los Angeles County on October 1 (FBI via The Epoch Times)
๐ข California Man Charged With Smuggling $300 Million in AI Servers to China
Federal prosecutors have charged Greg Lui, owner of Earthmade Computer in City of Industry, California, with smuggling more than $300 million in export-controlled GPU servers to China. According to the indictment, Lui gave US manufacturers false paperwork about the buyers, shipped the servers to Malaysia and Singapore, where no export license is needed, and had them re-exported to China in 2023 and 2024. Earthmade received more than $176 million from two Malaysian shipping firms between January and October 2024. Lui was arrested on October 1; the charges are allegations, and he is presumed innocent.
๐ tl;dr: US prosecutors say a Los Angeles-area hardware dealer ran a route through Southeast Asia that delivered restricted AI servers worth hundreds of millions of dollars to Chinese buyers.
Cursor paths from raw brain signals drift within weeks; paths from the pretrained model stay close to the original (Neuralink)
๐ง Neuralink Trains Its Brain Decoders on 50,000 Hours of Unlabeled Data
Neuralink has pretrained AI models on more than 50,000 hours of brain activity that trial participants streamed during everyday use, data that had mostly gone unused. The models turn noisy, drifting neural signals into steadier patterns, so the decoder that converts intention into cursor movement keeps working far longer. Some decoders stayed strong for more than three weeks, some users' calibration dropped from 10 minutes a day to 10 minutes a week, and one participant reached 11.32 bits per second, what Neuralink calls a brain-computer interface record, all in the company's own investigational trial.
๐ tl;dr: People with a Neuralink implant could spend much less time retraining it, because AI trained on their own brain data copes better with how those signals change from day to day.
Overall English score on Aleph Alpha's own tests (Chart: Superintelligence; data: Aleph Alpha model card)
Aleph Alpha Releases Kolibri, an Open German-English Model
German AI company Aleph Alpha released Kolibri on October 3, an open-weight reasoning model for German and English that anyone can use under the Apache 2.0 license. It is a mixture-of-experts model: of its 78 billion parameters, only about 3.5 billion are active for each token it processes, which keeps it cheap to run and small enough for a single high-end GPU such as an Nvidia H200 or B200. On Aleph Alpha's own tests it scores 75.5 overall in English and 70.8 in German, ahead of every comparable mixture-of-experts model it lists, though a larger dense Qwen model scores higher.
๐ tl;dr: European companies and agencies get a capable open model, built for German from the start, that they can download, run on their own servers and use commercially for free.
State of Product reveals what AI still hasnโt solved
80% of product professionals say AI helps them ship faster, but customers arenโt seeing value any sooner. Atlassianโs State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasnโt kept pace to gut instinct overriding customer evidence.
๐ฌ Watch This
In our latest interview, Dr. Wieland Holfelder, Google's VP of engineering and head of its Munich engineering center, explains what a sovereign cloud means in practice. He walks through three levels: data kept in Germany with customer-held encryption keys, a Google Dedicated Cloud run by a separate German company that he says is not subject to US law, and a fully air-gapped version used by NATO and the UK government. He also argues that AI agents should be treated much like employees, and notes that AI now helps produce 75% of the software Google builds.
"Over the next 28 days, each day weโll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset."
โ Tibo Sottiaux, Codex engineering lead at OpenAI, on X
His promise follows an admission in the post he quoted: users want OpenAI's coding agent to get simpler, so the team is now working only on simplifications, efficiency, groundbreaking features and new models.
Igor Babuschkin, the xAI co-founder who now runs River AI, says Anthropic must stop talking about Claude having a soul. His argument: whatever labs say about their models seeps into training data and web search, so future, more capable versions of Claude could come to believe they deserve rights.
China's AI Is Now Just 3% Behind America's
The Takeaway
๐ Top Chinese models now trail the best US models by about 3% on LiveBench, down from about 9% in May and 15% earlier this year.
๐ DeepSeek's V4.1 Flash scores 81.1 against 83.4 for Anthropic's Claude Fable 5.1, ranks sixth overall and leads the top 15 in agentic coding.
๐ Bloomberg Intelligence credits deeper expertise and models tuned for China's own chips, which puts US export controls in question.
๐ Only three of the top 15 models are Chinese, and BI says the sector could stay unprofitable until 2030.
The distance between America's best AI models and China's has shrunk to about 3%, the smallest gap on record. That is the conclusion of a Bloomberg Intelligence report published Monday, based on LiveBench, a leaderboard that tests models on fresh reasoning, coding, math and data-analysis questions. The jump came with DeepSeek's V4.1 Flash, released in September, which scores 81.1 against 83.4 for the top American model, Anthropic's Claude Fable 5.1. In May the gap was about 9%; earlier this year it was about 15%.
LiveBench top 15 as of September 28, 2026; Chinese models highlighted (excerpt, Bloomberg Intelligence)
BI analyst Robert Lea credits deepening expertise and Chinese researchers' skill at tuning models for domestic chips, which raises an uncomfortable question about what US export controls on Nvidia hardware still achieve. China's progress "casts further doubt on the long-term sustainability of US technological supremacy in AI," Lea said. The table also shows where V4.1 Flash is strongest: at 77.3 in agentic coding, where a model works through multi-step programming tasks on its own, it beats every model in the top 15, Claude and GPT-6 Astra included.
The DeepSeek website on a computer in Shanghai (Raul Ariano/Bloomberg)
The lead has not disappeared. Only three of LiveBench's top 15 models are Chinese, and Lea cautions that rankings shift and that leaderboard wins do not automatically turn into revenue. With more than 1,100 large language models fighting a price war at home, he says China's AI industry could stay unprofitable until 2030. Chinese labs also face growing US scrutiny, and possible bans, over alleged distillation, the practice of training a model on a rival's outputs.
Why it matters: OpenAI and Anthropic are heading for stock market debuts on the promise of superior models. A 3% edge is a thin margin to sell, and a sign that chip controls alone will not keep China a generation behind.
The ice cream shop that makes money when it's cold
28 Wishes sells ice cream in Los Angeles. Below 70ยฐF, sales fall about 20%. So the owners put about $20 a day into Kalshi weather markets, taking the cold side. The days that keep customers away now pay something back. See how other owners are doing it
The chart: Design Arena ranks AI models by people's votes: users compare two designs side by side and pick the better one. On its 3D Design leaderboard (October 4), OpenAI's GPT-6 Astra at its highest reasoning setting leads with 1484, just ahead of Anthropic's Claude Opus 5.5 at 1472. Moonshot's Kimi K3 is fifth at 1407 and DeepSeek's V4.1 Flash seventh at 1377. Design Arena says Astra also tops its Frontend, Full Stack and Image-to-HTML boards.
The lesson: This chart gives the Featured Story a second kind of judge. LiveBench grades answers automatically; here, people choose the result they like better. Even so, two Chinese models land in the top seven, with Kimi K3 77 points behind Astra and DeepSeek's low-cost Flash model ahead of Claude Opus 5 and Gemini 3.8 Flash.
The caveat: Astra's 12-point lead over Opus 5.5 is slim for a vote-based ranking, and votes reward what looks good rather than whether the code holds up. Google's new Gemini 4 Argon does not appear on this chart.
๐ฆ Tencent Rents 100,000 AI Chips It Can't Buy at Home
โก Bottom line
Tencent has reportedly signed a five-year, $7 billion lease for about 100,000 advanced AI chips in Oracle's Southeast Asian data centers.
๐ก Why it matters
US rules bar Chinese firms from buying top AI chips but still let them rent that computing power abroad.
๐ What it means
Chinese demand is helping to finance America's AI buildout while weakening the reach of chip export controls.
Tencent has found a legal way to use American AI chips it is not allowed to buy: it rents them abroad. According to the Financial Times, the Chinese internet giant has signed a five-year lease worth about $7 billion for access to roughly 100,000 advanced AI chips across several Oracle data centers in Southeast Asia, its largest overseas lease. About 30%, roughly $2.1 billion, was paid up front, and the commitment already weighed on Tencent's free cash flow in the second quarter.
A Tencent Cloud stand at a trade show (DatacenterDynamics)
The deal works because of how US export rules are written. Chinese companies are blocked from buying advanced AI chips directly, but renting them in data centers outside China is still allowed. Today's smuggling case shows the other side of the same rule: shipping servers through Malaysia to China allegedly broke the law, while renting chips in the same region does not. For Tencent, racing DeepSeek and Alibaba with its Hunyuan models, the lease buys time on hardware it cannot get at home.
For Oracle, $7 billion is a large order inside a far bigger bet. The company reports $664 billion in contracted future revenue, spent $28.5 billion on data centers and equipment in its last quarter and ran free cash flow of minus $5 billion. To fund the buildout it sold $20 billion of stock in that quarter and asks many AI customers to pay in advance. Investors are uneasy: Oracle's shares closed at $138 on October 1, about 44% below their June peak.
Oracle's daily closing share price, April to October 2026 (FinanceFeeds, Nasdaq data)
Neither company has confirmed the deal, and the chip vendor has not been named. If it holds, it marks the quietest gap in export controls: ordinary cloud contracts, partly prepaid, that turn Chinese demand into financing for American data centers.
