US Tightens AI Chip Checks for Nvidia and AMD Shipments

U.S. officials are tightening controls on access to advanced American AI chips and computing. New Commerce Department rules require overseas data centers serving companies in China, Russia, Iran and Cuba to identify their end users, and impose licensing requirements on certain chip shipments to Singapore, Malaysia, Thailand, Indonesia and Brazil. Chips for U.S. companies operating in those countries are exempt; existing operators have 90 days to obtain required licenses after the rule appears in the Federal Register. The measures aim to prevent remote computing and regional data centers from bypassing export restrictions.

The controls arrive amid scrutiny of Singapore-based Megaspeed, which U.S. agencies have investigated since 2024 over suspected diversion of Nvidia chips to China. Investigators sought Nvidia sales records in May 2025, and Nvidia put Megaspeed on its no-sell list the following month. U.S. and Singaporean officers inspected the company and interviewed its then-CEO Alice Huang; the Justice Department later took the lead in the criminal investigation. Megaspeed says it has begun its own inquiry. U.S. officials also say chip sales to regional operators, including buyers of Nvidia or AMD products, face added licensing checks.

A selloff in AI-linked stocks is prompting investors to examine exposure across portfolios. MSCI estimates at least $25.3 trillion is invested in major companies with direct AI exposure, while indirect effects are harder to measure. PGIM is screening nearly $1 trillion in fixed-income holdings, including bonds, loans and structured products, for AI-related business, credit and market risks. Equable Institute estimates 52 AI-related companies represented 38.5% of large pension funds’ public-equity holdings in June 2026, up from 20.7% in early 2010; the estimate excludes debt, venture capital and private equity.

Several companies are building AI tools around context and workflow. Second, described by a16z partners Joel de la Garza and Martin Casado, is designed to retain user context across apps; it runs on-device by default and lets people inspect and move its memory, select models and control app access. HPE says AI agents drove token use up five to ten times in a year. It plans to route 60% to 80% of tokens to on-premises systems for predictable costs, while retaining cloud for flexibility and advanced reasoning. HPE tracks use by person, organization and app, reviews experiments separately from production, and uses an MCP gateway to standardize agents’ access to company systems.

Codex tools described in another report aim to provide context without requiring users to paste it into prompts: screenshots can include an app’s accessibility tree, and plugins can bundle MCP servers, skills and other resources. Remote MCP connections can follow users across the web, Slack and development environments. Its Ultra Fast option runs eight times faster at six times the cost, though build, compile and testing time may reduce the practical gain. Separately, Google says Gemini can handle tasks such as appointments, bills and grocery shopping by connecting to email, calendars and payment methods. The feature uses the Gemini 4 Hydrogen model and is rolling out free to hundreds of millions of people in the U.S., Japan, South Korea, Canada, Australia and India; some advanced capabilities launch first in the U.S. Google says tasks run in isolated, hardware-protected environments.

AI adoption remains uneven in workplaces. Lenovo’s survey of 1,200 IT leaders at large organizations across 12 countries found 80% consider the digital workplace critical to AI transformation, but only 10% have fundamentally redesigned workplaces for AI. While 76% ranked AI transformation among their top IT priorities and 70% had started adoption programs, only 27% had scaled AI across business functions, 12% measured how well employees use the tools, and 10% continuously measured AI’s effects on workflows.

In government, the Pentagon plans to make AI a core training requirement for all cybersecurity roles, citing adversaries’ use of AI for reconnaissance, vulnerability discovery and social engineering. Gorak also described skills-based hiring that tests candidates in cyber ranges. He said the department is short 24,000 people in a cyber workforce of about 240,000, and is planning industry rotations and a workforce summit at National Defense University from April 7 to 9.

Other reports highlight risks and limits. Fireworks AI said an attacker reached its development machines after breaching a cloud provider hosting its research and development systems on October 23, 2026; Fireworks detected activity the next day. Uber is notifying nearly 150,000 drivers that personal details—including names, addresses, birth dates and driver’s license numbers—were stolen through the vendor. A benchmark of six large language models across 91 trading-agent runs found no statistically significant alpha across a rally, bear market and crash; market and style exposure explained a median 78% of returns, and agents did no better than randomly selected stocks in the same sector.

Similarweb reported third-quarter revenue of $83.0 million, up from $71.8 million a year earlier, and its first positive GAAP result, net income of $3.5 million. AI-related business made up almost one-fifth of sales and two-thirds of new annual recurring revenue. The company raised its full-year revenue outlook for the third time, to $318.0 million–$320.0 million. The quarter ended September 30, 2026; co-founder Or Offer’s final earnings call as CEO preceded Michael Akkerman taking over on November 2. Similarweb reported 1,790 accounts worth more than $25,000.

Key Takeaways

  • U.S. rules require end-user checks for overseas data centers serving companies in China, Russia, Iran and Cuba, and add licensing requirements for certain AI-chip shipments to five countries in Southeast Asia and Brazil.
  • Megaspeed has been under U.S. investigation since 2024 over suspected diversion of Nvidia chips to China; Nvidia placed the Singapore-based company on its no-sell list in June 2025.
  • MSCI estimates at least $25.3 trillion is invested in major companies with direct AI exposure; PGIM is examining nearly $1 trillion of fixed-income investments for AI-related risks.
  • HPE says AI agents increased its token use five- to tenfold in a year and plans to direct 60%–80% of tokens to on-premises systems.
  • Google is rolling out Gemini task handling for chores such as scheduling, bill management and grocery shopping, with access to email, calendars and payment methods.
  • Lenovo found only 10% of surveyed companies had fundamentally redesigned workplaces for AI, while 27% had scaled AI across business functions.
  • The Pentagon plans to require AI training across cybersecurity roles; Gorak cited a 24,000-person shortfall in a workforce of about 240,000.
  • Fireworks AI disclosed a breach affecting customer data; Uber says nearly 150,000 drivers are being notified after personal information was stolen through the vendor.
  • A benchmark covering 91 trading-agent runs found no statistically significant alpha from six LLMs; market and style exposure explained a median 78% of returns.
  • Similarweb recorded $83.0 million in third-quarter revenue, up from $71.8 million, and $3.5 million in GAAP net income; AI-related business generated almost one-fifth of sales.

US investigates Megaspeed over suspected Nvidia chip shipments to China

The US Department of Justice and Department of Commerce said they have investigated Singapore-based Megaspeed since 2024 over possible export control violations. Officials suspect the company helped send Nvidia’s advanced chips to China despite US restrictions. Commerce investigators requested Nvidia sales records in May 2025, and Nvidia put Megaspeed on its “no sell” list the next month. US and Singaporean officers later inspected the company and interviewed its then chief executive, Alice Huang, before the Justice Department took the lead in the criminal investigation. Megaspeed says it has started its own investigation.

US restricts remote access to advanced AI chips to curb China’s use

The Trump administration issued rules to restrict remote access to advanced US chips and servers, aiming to stop China from using US technology to train AI models. Overseas data centers serving companies in China, Russia, Iran and Cuba must identify their end users, and chip sales to Indonesia, Malaysia, Singapore, Thailand and Brazil will face extra licensing requirements. China’s Foreign Ministry spokesperson Guo Jiakun accused Washington of interfering with trade and technology cooperation, and said China would protect its interests. US lawmakers have also proposed tougher controls on cloud access and ways to track chips. The announcement came as officials discussed the investigation into Singapore-based Megaspeed, suspected of helping divert Nvidia chips to China.

US tightens AI chip sales rules for data centers in Southeast Asia

On November 16, the US Department of Commerce announced tighter rules for sales of advanced AI chips to data center operators in several Southeast Asian countries. The article says operators in Indonesia and Malaysia must get a special license to buy chips from US companies such as Nvidia or AMD. The rules respond to concerns that high-end chips bought in the region could be illegally redirected to Chinese companies. The Bureau of Industry and Security also requires existing overseas data center customers to obtain similar licenses if they provide services to others.

US requires licenses for AI chips sent to data centers serving China

The US Commerce Department issued an interim rule requiring licenses for advanced AI chips sent to overseas data centers that serve companies based in China. Exporters and data center operators must check who owns their customers. The rule also adds license requirements for chips sent to Singapore, Malaysia, Thailand, Indonesia and Brazil, which the department considers at higher risk of diversion to China. Chips for US companies operating in those countries are exempt. Existing operators have 90 days to obtain licenses once the rule appears in the Federal Register.

New US rules target China’s remote access to American AI technology

The Trump administration introduced rules to limit China’s remote access to advanced US chips and servers, including through data centers in other countries. The rules require overseas data centers serving companies in China, Russia, Iran and Cuba to identify their end users, and add license requirements for chip sales to Indonesia, Malaysia, Singapore, Thailand and Brazil. The Commerce Department said the measures aim to prevent remote computing arrangements from bypassing limits on access to sensitive US technology. The changes follow scrutiny of Singapore-based Megaspeed, suspected of diverting Nvidia chips to China, and reports that Singapore and Malaysia are investigating related chip shipments. The rule took effect immediately.

AI Selloff Drives Investors to Audit Hidden Portfolio Risks

After a sharp selloff in AI related stocks, investors are trying to measure how much AI risk their portfolios carry. The MSCI Global Intelligence Index estimates that at least $25.3 trillion is invested in major companies with direct AI exposure, while indirect effects are harder to track. PGIM is screening every bond, loan and structured security in nearly $1 trillion of fixed income investments for AI related risks. The Equable Institute estimates that 52 AI related companies made up 38.5% of large pension funds’ public equity holdings in June 2026, compared with 20.7% in early 2010.

Investors Audit AI Risks as $25 Trillion in Holdings Comes Under Scrutiny

A sharp selloff in AI related stocks has pushed fund managers to reassess their exposure to the technology. The MSCI Global Intelligence Index estimates that at least $25.3 trillion is invested in major companies with direct AI exposure, while the effects on other businesses are difficult to measure. PGIM is reviewing about $1 trillion in bonds, loans and securitized products for AI’s effects on business models, creditworthiness and market risks. Equable Institute estimates that 52 AI related listed companies made up 38.5% of major pension funds’ listed equity holdings, up from 20.7% in early 2010. The article also notes that this estimate leaves out debt, venture capital and private equity investments.

a16z Backs AI Memory That Users Can Take Between Apps

a16z partners Joel de la Garza and Martin Casado describe Second, a background engine that builds context over time and can share it across apps. The system runs on a user’s device by default, and users can inspect and move its memory, choose the AI model, and decide which apps can access it. The goal is to prevent users from having to teach each assistant the same preferences again. The article says Ken Liu researched AI and data privacy at Stanford, but its remaining details about the team and product are incomplete.

HPE’s AI Agents Push Token Use Up Five to Ten Times

HPE’s Dan Welble said AI agents increased the company’s token use by five to ten times in one year. HPE plans to direct 60% to 80% of tokens to on premises systems for more predictable costs, while keeping cloud services for flexibility, experiments and advanced reasoning. The company tracks usage by person, organization and app, and now updates its Chat HPE plans quarterly instead of annually. Welble said stronger open weight models have made on premises systems suitable for reasoning heavy coding, and described separate governance reviews for experiments and production. HPE also uses an MCP gateway to standardize how agents access company systems.

Codex Tools Aim to Add Context Without More Prompting

Codex features described in the article aim to give AI tools more context without requiring users to paste it into prompts. An app screenshot can include its accessibility tree, while plugins can bundle MCP servers, skills and other resources. Remote MCP connections follow users across the web, Slack and development environments. The article describes a design system plugin made with one prompt, including a button to add it to the composer and search that inserts component context. Ultra Fast runs eight times faster at six times the cost, but Kundel warned that build, compile and testing time can erase the benefit.

Fireworks AI Breach Exposes Data From Uber, Notion and Cursor

Fireworks AI confirmed that attackers accessed some of its systems and affected customer data. The company said a third party breached a cloud provider that hosted its research and development machines on October 23, 2026, then reached Fireworks development machines the next day. Fireworks detected the activity on October 24 and said it took steps to fix the problem. Uber is notifying nearly 150,000 drivers that attackers stole personal details, including names, addresses, dates of birth and driver’s license numbers, through the vendor. Uber confirmed that Fireworks.ai had access to some of its data.

Study Finds LLM Trading Agents Earn Market Returns but No Proven Alpha

Researchers introduced rta-bench-1, a benchmark that tests large language model trading agents across three U.S. markets: a rally, a bear market and a crash. It shows news only after publication and checks that agents do not use future information. Across 91 runs of six models from four vendors, no model produced statistically significant alpha, or returns beyond market factors. Market and style exposure explained a median 78% of returns, and agents did no better than randomly chosen stocks in the same sector. The study also found that agents could follow instructions reliably, while their returns were no higher without news headlines.

Lenovo Finds Only 10% of Companies Have Redesigned Workplaces for AI

Lenovo’s sixth Work Reborn report found that only 10% of surveyed companies have fundamentally redesigned their workplaces to support AI, even though 80% of IT leaders say the digital workplace is critical to AI transformation. The report is based on a survey of 1,200 IT leaders at organizations with at least 1,000 employees in 12 countries, conducted in August 2026. While 76% ranked AI transformation among their top IT priorities, only 27% had scaled AI across business functions. Just 10% had a continuous system for measuring how AI use affects workflows and guiding future investments. Lenovo also found that 70% had started AI adoption programs, but only 12% measured how well employees and teams used the tools.

Google Adds Task Handling to Gemini for Everyday Chores

Google says Gemini can now handle tasks such as scheduling appointments, managing bills and shopping for groceries. The assistant can connect to users’ emails, calendars and payment methods to complete tasks, and Google says it works in isolated, hardware-protected environments to protect private data. Users can try the features by switching the Gemini app from Chat to Task, texting Gemini on Android or iOS, or forwarding emails to [email protected]. The update is rolling out free to hundreds of millions of people in the U.S., Japan, South Korea, Canada, Australia and India, using the Gemini 4 Hydrogen model. Some advanced features are launching first in the U.S.

Pentagon Makes AI Training a Core Requirement for Cybersecurity Staff

The Pentagon will make AI a fundamental training requirement for all cybersecurity roles, according to Gorak, who spoke at the FedTalks summit on Tuesday. The Defense Department says the change responds to adversaries using AI to automate reconnaissance, find zero-day vulnerabilities and expand social engineering attacks. Gorak also described a move toward skills-based hiring, where candidates must show they can perform job tasks in a cyber range rather than relying mainly on education or work history. He said the Pentagon is short 24,000 people in a cyber workforce of about 240,000. The department is also planning industry rotations and a cyber workforce summit at National Defense University from April 7 to 9.

Similarweb posts first GAAP profit as AI sales near 20%

Similarweb reported third-quarter revenue of $83.0 million, up 15.6% from $71.8 million a year earlier, and GAAP net income of $3.5 million, its first positive GAAP result. The New York and Tel Aviv based company said AI-related business made up almost one-fifth of sales, while AI deals accounted for two-thirds of new annual recurring revenue. Similarweb raised its full-year revenue outlook for the third time this year, to $318.0 million to $320.0 million. The quarter ended September 30, 2026, and was co-founder Or Offer’s final earnings call as CEO; Michael Akkerman took over the role on November 2. The company also reported 1,790 accounts worth more than $25,000.

Sources

NOTE:

This news brief was generated using AI technology (including, but not limited to, Google Gemini API, Llama, Grok, and Mistral) from aggregated news articles, with minimal to no human editing/review. It is provided for informational purposes only and may contain inaccuracies or biases. This is not financial, investment, or professional advice. If you have any questions or concerns, please verify all information with the linked original articles in the Sources section below.

AI chips U.S. export controls AI chip licensing data center compliance Nvidia AI investment risk AI agents enterprise AI on-premises AI Model Context Protocol (MCP) Google Gemini AI workplace adoption AI cybersecurity training cyber workforce data breaches LLM trading agents AI benchmarks Similarweb AI business revenue

Comments

Loading...