AI Policy · Daily

Top congressional leaders signaled likely support for new China chip export limits in the annual defense policy bill, but negotiators have not yet agreed to keep the provisions in the final version. Former Nvidia chief Jensen Huang agreed to brief House lawmakers privately before Thanksgiving, heading off subpoenas that two committee chairs had warned were likely. A forensic analysis found that six recent Chinese frontier models were likely initialized on export-controlled Nvidia Blackwell chips, though it cannot show where the chips were or whether export rules were broken. Top tax writing Democrats opened an inquiry into President Trump's Nvidia chip deal, which lets the company sell advanced AI chips to China in exchange for a government cut, and asked the Commerce Department to preserve its records.

I.Top Stories

Congressional leaders back China chip curbs in defense bill, but no deal yet

Senate Majority Leader John Thune told Semafor there is a "high probability" Congress adds limits on chip exports to China, and House Speaker Mike Johnson said there is "a developing consensus in Congress" for more restrictions. Three of the bills ride in the Senate version of the National Defense Authorization Act (NDAA), the annual defense policy bill. The Chip Security Act would require location checks on exported AI chips, the MATCH Act targets loopholes on chipmaking equipment, and the AI Overwatch Act would give Congress oversight of chip licenses to countries such as China. Thune and Senate Armed Services Chair Roger Wicker (R-Miss.) said negotiators have not agreed to keep them in the final bill. A predecessor measure was stripped from last year's House NDAA after pressure from Nvidia, the chip industry and the White House.

Read at Semafor ↗

Huang agrees to closed-door House briefing before Thanksgiving, averting subpoena fight

Former Nvidia CEO Jensen Huang agreed to brief House lawmakers behind closed doors on national security threats before Thanksgiving, Rep. Brian Mast (R-Fla.), chair of the House Foreign Affairs Committee, told The Hill. "I don't want to see anything hurt in Nvidia, but he has some very serious questions to answer," Mast said. The agreement follows a deadline from Mast and Rep. John Moolenaar, chair of the House Select Committee on China, who warned of likely subpoenas.

Read at The Hill ↗ • Read at TechCrunch ↗

Former OpenAI researcher finds export-controlled Blackwell fingerprints in six Chinese AI models

Six recent models from DeepSeek, Moonshot AI, MiniMax and Ant Group were likely initialized on Nvidia Blackwell GPUs, Will DePue of Unemployment Labs reported Nov. 12. A model's starting weights carry a pattern set by the chip that generated them. A rented B300 reproduced 445 rows of DeepSeek-V4-Flash's starting values bit for bit, and a GB200 matched Moonshot's Kimi K3, according to Tech Times. The same labs' earlier models match older Hopper chips. The analysis cannot show where the chips were, how the labs got them or whether export rules were broken. A senior U.S. official said in February that DeepSeek had trained on Blackwells in Inner Mongolia. DePue released his code so others can check the results.

Read at Unemployment Labs ↗ • Read at Tech Times ↗

Senior tax writing Democrats open inquiry into Trump's Nvidia H200 export deal

Sen. Ron Wyden (D-Ore.), ranking member of the Senate Finance Committee, and Rep. Richard Neal (D-Mass.), ranking member of the House Committee on Ways and Means, asked Commerce to preserve its records on President Trump's decision to let Nvidia sell H200 chips to China, per The Hill. In a letter to Commerce Secretary Howard Lutnick, they called the arrangement giving the government a cut of those sales "an unconstitutional export tax," according to Reuters. The two said they are examining approvals of chip sales to Chinese buyers tied to the Chinese government, military and sanctioned entities.

Read at The Hill ↗ • Read at The Business Standard ↗

White House officials weigh pushing out Commerce Secretary Lutnick over Nvidia scandal

Senior White House officials are debating Lutnick's future, including whether to push him out, two people familiar with the situation told Reuters. Hours earlier, a White House official had said President Trump maintains complete confidence in Lutnick. Lawmakers have criticized the administration for keeping dozens of Chinese firms off the Entity List, the Commerce Department's trade blacklist, which has had no new postings since October 2025. Sen. John Kennedy (R-La.) said on CNN he did not yet know whether Lutnick should lose his job. Sen. Elizabeth Warren (D-Mass.) said Lutnick and Commerce Under Secretary Jeffrey Kessler owe an explanation for leaving export control loopholes open.

Read at The Business Standard ↗ • Read at SCMP ↗

Singapore says U.S. gave little help on Megaspeed probe after it held Alice Huang's passport for months

Singapore's government said the United States offered little assistance in its investigation of Alice Huang and Megaspeed, the Singapore company she ran, The New York Times reported. Authorities held her passport from June 2025 until February over possible false declarations tied to Megaspeed, then released her for insufficient evidence, issued a warning and barred her from returning. That hold continued well after President Trump was briefed in fall 2025 on her relationship with Jensen Huang. Singapore said it flagged server maker Supermicro's poor cooperation to the Justice Department and the FBI in July 2025 and asked for their help, but has received little information or assistance since. Singapore police are still investigating Megaspeed.

Read at The New York Times ↗

II.China Watch

Baidu to release open weights of new Ernie model; Alibaba previews next Qwen flagship

Baidu released a preview of ERNIE 5.5 Flash on Nov. 11 and plans to publish its weights under the permissive Apache 2.0 license by Nov. 25, letting anyone download, modify and run the model, 36Kr reported. In scores Baidu ran itself, the company claims first place on two of eight benchmarks, both measuring workplace automation tasks, ahead of Anthropic's Claude Opus 5.5. The model trailed Opus 5.5 by 17 points on a command line coding test. Baidu called the preview an intermediate version that may still show gaps from incomplete training. Baidu founder Robin Li said in 2024 that open models would fall further and further behind. Alibaba opened a preview of Qwen4-Max the same day with no pricing or benchmark data, and its documentation retires the preview after Nov. 25.

Read at 36Kr ↗

ByteDance AI chip chief resigns after in-house accelerator falls short

Shi Yunfeng, who leads ByteDance's AI chip and data processing chip teams, submitted his resignation, per Chinese outlet LatePost. Shi joined the TikTok owner in 2018 and reports to Yang Zhenyuan, head of its cloud business, Volcano Engine. ByteDance set up its chip team in 2019 to build a custom AI chip, modeled on Google's Tensor Processing Units, for its own AI workloads. The chip's finished design went to manufacturing this year, but its performance fell below expectations. The team is now developing a successor chip.

Read at TechNode ↗

Robot chip venture backed by Chinese GPU maker and humanoid robot company raises about $70 million

Xilect, a robot chip startup co-founded by Chinese GPU maker MetaX and humanoid robot maker UBTech Robotics, raised 500 million yuan ($70 million) in an early angel funding round, TechNode reported. Auto industry outlet Gasgoo put the startup's post-money valuation at 4 billion yuan ($560 million). Xilect is designing a single chip that handles perception, decision-making and motion control on the robot itself, plus software that lets robot makers run existing AI models on it. The money will go to developing that chip, hiring engineers and testing the platform with robot manufacturers. Xilect plans to give partners a development and testing platform in 2027 and to release its first chip in 2028.

Read at TechNode ↗ • Read at Gasgoo ↗

Chinese AI stocks slide on fears Huang's exit will bring tighter U.S. chip curbs

Investors sold shares of Chinese AI model developers on bets that Washington will further restrict China's access to advanced U.S. chips, with Z.ai down 12% and MiniMax down 8.4%, per Reuters. Data center suppliers Zhongji Innolight and Eoptolink Technology, which rely heavily on U.S. customers, dropped more than 8%. Tencent and Alibaba each fell more than 2% as investors weighed the risk of losing access to Nvidia powered processors in overseas data centers, Investing.com reported. Domestic chipmakers moved the opposite way, with Cambricon up more than 11% and Moore Threads up about 7% in the prior session on bets that tighter controls would speed China's search for substitutes.

Read at Financial Times ↗ • Read at Euronext ↗ • Read at Investing.com ↗

Alice Huang's layered stakes link her to Shenzhen data center serving Huawei Cloud

Alice Huang co-owned Shanghai Hexi, a technology and data center investor that in 2021 bought 27% of developer Shanghai Century Longteng, according to a CNBC investigation built on research from private intelligence firm Tesari. Century Longteng held a controlling 50.1% stake in Shenzhen Yifeng Technology, a data center whose key customer was Huawei Cloud, the cloud arm of Nvidia's main Chinese rival. Hexi also guaranteed loans that financed the Shenzhen facility. CNBC could not determine whether that data center obtained Nvidia chips. Brian Andersen, a Tesari executive and former supervisory special agent at Homeland Security Investigations, the Department of Homeland Security's investigative arm, said authorities should open an investigation. Any charges would turn on evidence of unauthorized transfers of controlled equipment, he said.

Read at CNBC ↗ • Read at Yahoo Finance ↗

III.Policy Tracker

Rep. Foster says Nvidia privately called Chip Security Act tracking feasible

Nvidia representatives told Rep. Bill Foster (D-Ill.), a cosponsor of the Chip Security Act, that the bill's location verification requirements for advanced AI chips could be met, Foster told MIT Technology Review. "They never put anything in writing," Foster said, but in technical talks their reaction was "yeah, of course we can do this." In March, the Semiconductor Industry Association, whose members include Nvidia, opposed the bill and called its tracking mandates untested and possibly unworkable. Nvidia's Fleet Intelligence software, announced in December 2025, can indicate which country a customer's chips are in, the company confirmed to Reuters at the time. Foster plans to introduce a stricter bill next year that would require a verified license before an advanced AI chip can connect.

Read at MIT Technology Review ↗

Sen. Cotton asks Commerce to blacklist Chinese memory chipmaker CXMT

Sen. Tom Cotton (R-Ark.), chair of the Senate Intelligence Committee, urged Commerce Secretary Howard Lutnick to put ChangXin Memory Technologies (CXMT) on the Entity List, the department's trade blacklist, according to a letter obtained by CNBC. "CXMT, one of Communist China's national memory chip champions, poses a growing risk to our economic prosperity and military superiority," Cotton wrote. A listing would stop U.S. firms from doing business with CXMT and deny it American technology. Cotton also pointed to CXMT's push into high-bandwidth memory, a component of the AI processors built by Nvidia, Google and Amazon Web Services. Counterpoint Research estimates CXMT's slice of worldwide revenue for DRAMDRAMThe working memory that phones, PCs, and AI servers use to hold data that a processor is actively handling., the main working memory in computers and phones, at 10% in the second quarter of 2026, up from 4% a year earlier.

Read at CNBC ↗ • Read at Quartz ↗

Georgia tells local governments to set AI rules before buying AI tools

The Georgia Technology Authority's Office of Artificial Intelligence released a 53-page Local Government AI Playbook aimed at cities and counties, especially those without dedicated AI teams, StateScoop reported. Written by Nikhil Deshpande, the state's chief digital and AI officer, the guide advises local governments to adopt policies, catalog the AI systems they already use, form oversight committees and train staff before purchasing AI products. It says AI should support employees' judgment, not replace it.

Read at StateScoop ↗ • Read at Georgia.gov ↗

IV.Capability & Research Watch

CSET maps where AI kill switches could sit and what each would disrupt

Georgetown University's Center for Security and Emerging Technology (CSET) published a guide for policymakers on AI kill switches, mechanisms that quickly halt an AI system during a harmful incident. The authors say some policy proposals would let government officials order their use. Shutdowns can happen at levels ranging from software and network access to hardware, such as cutting power to one chip or to an entire data center. The broader the intervention, the more collateral damage it causes, CSET said. A malicious user of a commercial model could be stopped by blocking an IP address or restricting its access to the model's programming interface, while an AI agentAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. that escaped a contained environment may require broader, coordinated action.

Read at CSET ↗

Google makes Gemini 4 Hydrogen the default model in Search's AI Mode

Google DeepMind released Gemini 4 Hydrogen, a faster and cheaper model in its Gemini 4 family, across Google Search, the Gemini app, the Gemini API and all tiers of Google AI Studio, according to the company's blog. Hydrogen now powers Search's AI Mode by default, Google Search product vice president Robby Stein said, per The Verge. Google said Hydrogen scores 51 at its default setting on the Artificial Analysis Intelligence Index, an independent composite of several tests, edging the larger Gemini 4 Argon at 50, and reaches 54 at its highest setting. It costs $1.50 for every million tokens, or chunks of text, of input and $7.50 for every million of output. Google said it more than tripled its predecessor's score on a command line coding test, to 57% from about 18%.

Read at Google DeepMind ↗ • Read at The Verge ↗

V.Industry & Market Watch

SpaceX to put over $74 billion into AI supercomputing sites at Starbase and in Arizona

SpaceX said it will spend more than $24 billion on a supercomputing campus at its Starbase site in Texas and roughly $50 billion on another in Cochise County, Ariz., per Reuters. The company's AI segment produced more than half its revenue in the third quarter. The Arizona project needs county approval, and 2,400 people have petitioned officials to reject the permit ahead of a December county hearing. The Starbase campus would generate its own power: the city's posted plans call for a gas-fired power plant and a supercomputing facility on about 444 acres. Starbase city commissioners are scheduled to take up a property tax break for the project Dec. 9.

Read at The Wall Street Journal ↗ • Read at CNA ↗

Huang was also involved with a second, unnamed Nvidia employee, Reuters reports

Jensen Huang had also been intimately involved with a second Nvidia employee, two people familiar with the matter told Reuters, which did not name her. The report brings to three the women Huang is reported to have been involved with, alongside Alice Huang and Kimberly Powell, who leads Nvidia's health care business. Nvidia declined to comment, while Huang, Powell and the company's board did not respond.

Read at Reuters ↗

VI.Global & Geopolitics

Malaysia checks whether Megaspeed unit re-exported AI chips, says it never cleared the firms

Malaysia's Investment, Trade and Industry Ministry said it is working with customs to determine whether AI chips imported by Speedmatrix, the Malaysian affiliate of Singapore's Megaspeed, were shipped on to other countries and whether that broke Malaysian law, per Malay Mail. The ministry said it never cleared either company and that its visits to Speedmatrix's rented data center space were routine monitoring of the sector. Since July 2025, re-exporting or transshipping high performance U.S. AI chips through Malaysia has required a ministry permit. "Malaysia welcomes investment in data centres and technology, but will not allow itself to be used to get around export controls," the ministry said. Prime Minister Anwar Ibrahim said Malaysia has received no requests for help from Singapore or the U.S., according to Reuters.

Read at Bloomberg ↗ • Read at Malay Mail ↗ • Read at Yahoo News Malaysia ↗

Chinese national becomes fifth person charged in Singapore's Nvidia server fraud case

Singapore charged Shi Bingzhe, a 34-year-old Chinese national, with 41 counts of consenting to money laundering by two companies where he was a key officer, in a fraud case linked to the movement of Nvidia chips, CNA reported. Police said the two firms, Superphi and NJ Global Financial, bought 917 servers from Aperia International through 41 purchases from May 2024 to February 2025, and charge sheets value the servers at more than $100 million. Four people charged earlier, including Aperia's former chief executive, have cases pending. Shi faces up to 10 years in jail and a fine of up to 500,000 Singapore dollars (about $390,000) on each charge. His case returns to court Jan. 4.

Read at Bloomberg ↗ • Read at CNA ↗