Tech Express | October 19, 2026 — Jensen Huang: Nvidia 'Absolutely Must' Cooperate With Chinese AI Firms

1. Jensen Huang: Nvidia 'Absolutely Must' Work With Chinese AI Firms

Nvidia CEO Jensen Huang said at a media briefing in London that artificial intelligence is a potentially dangerous technology and that Nvidia "absolutely must" cooperate with Chinese AI companies to jointly monitor and address related risks, according to a report published on October 19. He stressed that the way to cooperate is dialogue, and that both the United States and Britain should do so.

The remarks come as Nvidia's share of China's AI chip market shrinks rapidly. IDC data shows that in the first half of 2026, Nvidia's share of China's data center AI chip market fell to 34%, with shipments of 678,000 units, down 44% year on year. Tightening U.S. export restrictions are the root cause: its most advanced Blackwell chips are completely barred from China, while revenue from the export-approved H200 accounts for less than 1% of its total data center revenue.

Nvidia CFO Colette Kress said in the company's second-quarter 2026 earnings report that its forward guidance assumes no compute revenue from Chinese data centers, meaning the company has temporarily excluded the China market from its financial forecasts. The shift in wording, from being forced to give up the market to saying cooperation is absolutely necessary, reflects the dilemma Nvidia faces in China.

Source

2. Huawei's Ascend Ships 572,000 Chips, Closing Gap With Nvidia to 6 Points

China's data center AI chip market changed markedly in the first half of 2026, according to a new IDC report. Nvidia remained first with a 34% share, but its shipments were only 678,000 units, down 44% year on year, while Huawei's Ascend took 28% with 572,000 units, narrowing the gap to just 6 percentage points. Nvidia's most advanced Blackwell chips cannot enter the Chinese market, and revenue from the export-approved H200 accounts for less than 1% of its data center business. Huang Renxun has repeatedly criticized the policy, saying a $50 billion Chinese AI chip market is being forced to Huawei.

The wave of domestic substitution is benefiting the whole supply chain. Besides Huawei, Alibaba's T-Head ranked third with 7%, while Cambricon, Kunlunxin and Hygon held 5%, 4% and 3% respectively. Nine of the 11 Chinese chipmakers tracked by IDC grew shipments in the first half.

In a more symbolic shift, shipments of non-GPU accelerators reached 1.016 million units, up nearly 80% year on year, surpassing GPUs at 991,000 units for the first time. Huawei, T-Head, Kunlunxin and Cambricon together accounted for 95% of the non-GPU market, with government agencies, internet companies and telecom operators the main adopters. IDC said competition is shifting from single-chip performance to overall system capability. Morgan Stanley expects China's AI chip self-sufficiency rate to rise from 35% in 2025 to 45% in 2026 and to 68% in 2027.

Source

3. Chinese GPU Unicorn Turns Profitable: MetaX Swings to Gain as C600 Ramps

MetaX said in its preliminary results for the first three quarters of 2026 that it expects revenue of 2.541 billion to 3.106 billion yuan, up 105.60% to 151.29% year on year, and net profit attributable to shareholders of 587 million to 717 million yuan, swinging from a loss of 346 million yuan a year earlier.

The main driver is commercial-scale deployment of MetaX's Xiyun C-series GPUs. The C500, benchmarked against Nvidia's A100, entered mass production in February 2024. The C600 opened up growth room: it entered mass production in May 2026 with 144GB of HBM3e memory, native FP8 support and tensor transpose instructions, sharply improving memory bandwidth and compute efficiency. Industry estimates put its memory specification above Nvidia's H20 at 96GB. The C600 is built on a fully domestic supply chain, relying on no overseas advanced process from wafer fabrication to packaging. The next flagship, the C700, is expected to begin small-volume production in 2027, targeting Nvidia's H100.

Public-market performance contrasts sharply with the results. MetaX listed on December 17, 2025 at an issue price of 104.66 yuan; it opened at 700 yuan and its market value once reached 280 billion yuan. It has since fallen to 166.3 billion yuan, a drop of 113.7 billion yuan from its post-listing peak.

Source

4. Another Chinese Model Lands Tonight, Tipped to Beat DeepSeek V4.1 on Price

Chinese large models have entered a new, dense release cycle. DeepSeek V4.1 Flash-1017 has already arrived, and industry leaks suggest another domestic model that is both stronger and cheaper will be released tonight, Mydrivers reported on October 19.

AI influencer @MaxForAI said he was told a closely watched model would launch today, with agentic capability better than DeepSeek's V4.1-Flash-1017 and a lower price, though that price excludes caching, where DeepSeek's advantage is hard to beat. The account has previously previewed and tested several Chinese models with a good hit rate, but is under an NDA and cannot say which company is behind it.

On performance alone, Zhipu's GLM-5.5 or Kimi K3.1 could surpass V4.1 Flash, but as flagship models they are unlikely to undercut it on price, ruling them out. Combined with leaks earlier in the day, the launch may be MiniMax's M3.1 Flash: clues have appeared on its website, including a 1M context window, multimodal support and thinking depth from low to max that cannot be switched off, a lightweight model optimized for coding and agents.

Source

5. Google Open-Sources Harness²: Models Rewrite Their Own Harness Mid-Run

Google researchers have open-sourced a framework called Harness² that lets large models rewrite their own add-on code, system prompts, tool chains and skill libraries while running, QbitAI reported on October 19. The project is fully open source and already supports OpenCode and Codex.

Earlier attempts at recursive self-improvement (RSI) largely failed in the open world: lab benchmarks have unit tests and ground truth, but in real office work nobody grades the model, so it cannot tell whether a prompt rewrite made it smarter or just more inventive. A harness modified for one task becomes a burden on the next and the system slides into a noise spiral. The "squared" in Harness² refers to two nested recursion loops, task-level and harnessing-level, paired with a Propose-Probe-Compose mechanism: one track fixes cross-domain defects such as chaotic planning or runaway tool calls, another targets domain-specific constraints, and behavioral contrast checks whether a change actually resolved the sticking point.

With Harness², GLM-5 rose from 64.2 to 75.7 on the office benchmark WorkBuddy-Bench under OpenCode, a gain of 11.5 points, and from 64.4 to 75.3 under Codex. Gemini 3.5 Flash climbed from 72.4 and 74.5 to 77.5 on both. On the end-to-end JobBench, the two gained 4.5 and 3.6 points; on the legal benchmark LAB, GLM-5's partial pass rate rose 11.8 percentage points. With a parallel budget of five candidates, WorkBuddy-Bench gained up to 3.9 more points and JobBench 4.2 more.

Source

6. Study: AI-Altered Images Can Plant False Memories, Held With Confidence

MIT scientist Pattie Maes told the Washington Post about her team's research: participants viewed a series of photographs and, after a break, some were shown the identical images while others saw AI-altered versions. In a later memory test, those who saw the altered images were not only more likely to misremember the originals, but were also overly confident in their false recollections, Gizmodo reported.

Human memory is fickle, unreliable and highly open to suggestion, as psychological experiments dating to the 19th century have documented. That is what makes misinformation dangerous: some studies show it can keep shaping decisions, including political ones, even after corrections are issued. A study last year found that both deepfakes and AI-generated text effectively planted false memories about entirely fabricated political scandals, going so far as to reduce voting intention.

The problem is that an average person's ability to identify AI-generated content is basically a coin toss. An internal U.S. Department of Defense investigation reported by Bloomberg last month concluded that a military strike on an elementary school in Iran earlier this year, which killed more than 150 people including at least 123 children, was partly due to overreliance on a Palantir-developed AI system that recommended the school as a day-one target based on outdated data.

Source

7. iPhone 18 Pro Series Tops 4.08 Million Units in China, a Million in One Week

Sales of the iPhone 18 Pro series have kept climbing in China since launch, far exceeding industry expectations, Mydrivers reported on October 19. A digital-industry blogger's latest tally puts cumulative sales at about 4.0809 million units as of week 41 of 2026, the seven days from October 5 to October 11.

The same blogger had reported cumulative sales of 2.8909 million units at the end of week 40, meaning the series added more than a million units in a single week, the fastest pace for an Apple flagship after launch in recent years. The top three best-selling models were all iPhones, with the iPhone 18 Pro far ahead. Another blogger said the iPhone 18 Pro Max set a record for premium large-screen models, with nearly 900,000 activations in week 41 alone.

Memory chip prices are surging globally, pushing up smartphone pricing across the industry. Even as Apple raised its own prices, the iPhone's overall value for money looks stronger by comparison against Android flagships that have repeatedly raised prices. The industry consensus, however, is that prices will not fall noticeably any time soon, which will suppress replacement demand and further lengthen upgrade cycles.

Source

8. PlayStation 6 Price Seen Above $1,000 With RAM and SSD Alone Over $680

Sony is reportedly still targeting a late 2027 release for the PlayStation 6, even though current market conditions make launching a new console extremely difficult, Wccftech reported. Fewer gamers will be able to afford it at launch: RAM and SSD costs alone are approaching $700, making a price above $1,000 almost certain.

Well-known leaker KeplerL2 said on the NeoGAF forums that "RAM + NAND alone is over $680," before other components are counted, suggesting the bill of materials has likely pushed past the $1,000 threshold. He noted that when Sony and Microsoft began designing the console SoCs, GDDR7 was projected to cost $3 per GB; it now runs about $20 to $23 per GB.

Based on current prices, the most probable configuration is 32GB of RAM with a 2TB SSD, especially if Sony launches a digital-only console, which is considered highly likely given plans to phase out disc production from early 2028. A widely rumored alternative is 40GB of RAM with 1TB of storage. Because delaying would cost far more than launching at a high price, the PS6 is almost certain to arrive in late 2027 or early 2028, and the high price will likely produce the longest cross-generation overlap ever seen.

Source

9. SpaceX Launch Slips to Tuesday, Another 'Space Jellyfish' Possibly Visible

SpaceX had planned to launch its Bandwagon-5 mission from Cape Canaveral Space Force Station in Florida on Monday morning but pushed it back because of unfavorable weather, WRAL reported. The launch is now targeted for 6:30 a.m. Tuesday, October 20, with a window from 6:30 to 7:30 a.m. Eastern.

The Falcon 9 rideshare mission will carry multiple satellites to orbit, and the timing could produce an illuminated exhaust plume visible hundreds of miles away. The display happens because the rocket climbs into sunlight about three minutes after launch while the ground below is still dark, lighting up exhaust that spreads at high altitude into a shape resembling a jellyfish with long, wispy tentacles. An earlier liftoff in the window is better: sunrise in Raleigh is at 7:24 a.m., and by 7 a.m. the sky will be too bright.

Last week's launch was visible from Alabama to Maryland and Tennessee because the rocket flew northeast, carrying the plume up the Atlantic coast. Tuesday's Bandwagon-5 will head southeast instead, giving the best show to the Caribbean and central and southern Florida. Observers in North Carolina may still see it, but lower in the sky and less prominently. Look toward the southeast, avoid bright streetlights and give your eyes a few minutes to adjust.

Source

10. Huawei's Smaller Foldable PC May Arrive in November at Around 10,000 Yuan

Huawei's smaller foldable PC now has a clearer launch window and price, Mydrivers reported on October 19, after earlier rumors that such a device was coming. Digital blogger Dingjiao Shuma said the machine is expected in the last two months of this year, most likely November, and may debut alongside a new n-series foldable.

Compared with the MateBook Fold on sale now, the biggest changes are size and price. The blogger said the new model will be half the size of the current MateBook Fold, with pricing dropping to 1x,999 yuan, meaning just over 10,000 yuan. For comparison, the MateBook Fold Ultimate Design uses an 18-inch folding display that becomes 13 inches when folded, and starts at 24,999 yuan.

If the leak is accurate, the price barrier for a foldable PC will fall noticeably, letting more consumers try the category. A smaller body should also be easier to carry for mobile work and study. The device may add an external screen so users can check information and handle simple tasks while it is closed. Huawei already uses its Xuanwu waterdrop hinge and UTG ultra-thin glass on the MateBook Fold, and whether those carry over to the smaller model is worth watching.

Source