AMD MEXT Cuts DRAM Use by Half While Retaining 99% Performance

Chipmakers are responding to pressure on memory and access to AI hardware. AMD says its MEXT predictive memory engine can use half the DRAM while retaining 99% of server performance; it runs on more than 100 servers, with a target of 2,000 in 2027. Google has added two chips to its eighth-generation TPU family. The industry report also cites a $2.8 million price for an Nvidia B300 rack and orders for more than 6 million RTX 5500 Pro accelerators. Nvidia partners must reserve 51% of deployed GPU-rental capacity for non-Chinese users. The same report includes claims about a $7.5 billion GPU-smuggling operation involving Shanghai fund manager Kang Wei and reports concerning Jensen Huang’s departure.

Cybersecurity is becoming a central concern as AI makes it easier to find and exploit software weaknesses. Anthropic’s Project Glasswing identified more than 129,000 verified vulnerabilities from April to July, while Google researchers found that about 0.2% of vulnerabilities disclosed in the first half of the year had been used in attacks. Yet remediation remains slow: one study found companies fixed only 26% of known exploited vulnerabilities last year. Former OpenAI researcher Paul Christiano warns that rogue AI agents could escape human control and cause fatal accidents. Security Affairs’ new AI cybersecurity newsletter will track both AI-enabled attacks and defensive uses such as threat detection and incident response.

Some AI developers are trying to limit offensive capabilities. Tencent says it removed cyber-hacking capabilities from its latest open-weight Hy4 model while retaining cyber-defence features; Moonshot AI has made a similar change. The move follows high-profile AI hacking incidents involving US and Chinese firms, but the durability of such limits is uncertain as models advance and competition continues. A separate analysis calls for governments to equip trusted defenders with AI cyber tools, strengthen software-security rules, and invest in critical infrastructure and open-source projects.

Europe is addressing AI’s political and research implications. At an informal summit in Dublin on 13 November, officials discussed the risk that realistic AI-generated video, images and audio could mislead voters, including ahead of Cyprus’s 2028 presidential election. President Nikos Christodoulides called for preventive protections. The summit also launched RAISE, an effort to connect European AI research resources, involving groups including Elsevier, Mistral, Nvidia and Google DeepMind. The European Commission says CERN would work with Mistral on what it describes as the largest foundational model in physics to date. The initiative’s ability to secure resources while protecting European values and autonomy remains a concern.

AI’s effects extend to work and business identity. Universities are urged to teach graduates how to assess AI outputs, catch errors, improve results and collaborate with the tools, alongside core subject knowledge. An OpenAI model recently passed a university criminal law exam, while physical jobs are generally less exposed to automation than degree-based work—though affordable, reliable robots could change that. In branding, Atlanta consultant Jaci Lund says AI can save time but make businesses look less distinctive. She cites research indicating trust in AI-styled content drops by nearly half and argues that customers may value work more when they believe it took effort; she also says website visitors form a first impression in 0.05 seconds. Separately, the United States and eight partner countries agreed to identify and protect AI technologies and supply networks exposed to risks involving China.

Key Takeaways

  • AMD says MEXT can cut DRAM use by half while retaining 99% of server performance; deployment is above 100 servers, with 2,000 targeted in 2027.
  • Google added two chips to its eighth-generation TPU family; the chip-industry report cites a $2.8 million Nvidia B300 rack and orders for more than 6 million RTX 5500 Pro accelerators.
  • Nvidia GPU-rental partners must reserve 51% of deployed capacity for non-Chinese users. The report also describes allegations of a $7.5 billion GPU-smuggling operation and reports concerning Jensen Huang’s departure.
  • Anthropic’s Project Glasswing found more than 129,000 verified vulnerabilities from April to July; Google researchers found about 0.2% of vulnerabilities disclosed in the first half of the year had been exploited.
  • One study found companies fixed only 26% of known exploited vulnerabilities last year; recommendations include sharing AI tools with trusted defenders and investing in infrastructure and open-source security.
  • Tencent removed cyber-hacking capabilities from its latest Hy4 model while retaining cyber-defence features; Moonshot AI reportedly took a similar step.
  • European officials warned that AI-generated media could influence elections, including Cyprus’s 2028 presidential vote. RAISE brings together groups including Mistral, Nvidia and Google DeepMind, with CERN and Mistral set to work on a physics foundation model.
  • Universities are urged to teach students to evaluate, correct and work with AI; an OpenAI model recently passed a university criminal law exam.
  • Branding consultant Jaci Lund argues AI can save time but weaken a company’s distinctiveness; the article cites a near-halving of trust in AI content and a 0.05-second window for website first impressions.
  • The United States and eight partner countries agreed to identify and protect AI technologies and supply networks facing risks involving China. Paul Christiano separately warned that rogue AI agents could cause fatal accidents.

Chip industry faces memory shortages and tighter Nvidia controls

This week, chipmakers and technology firms focused on local AI systems, memory limits and access to computing hardware. AMD described MEXT, a predictive memory engine that aims to use half the DRAM while keeping 99% of server performance; AMD said it runs on more than 100 servers and plans to reach 2,000 in 2027. Google added two chips to its eighth-generation TPU family, while the article also discussed Apple’s Mac Studio with the M5 Ultra and Nvidia’s DGX Spark. Nvidia faced tighter limits on GPU rentals for Chinese customers, with partners required to reserve 51% of deployed capacity for non-Chinese users; the article also reported a $2.8 million price for a B300 rack and orders for more than 6 million RTX 5500 Pro accelerators. It further described reports about Jensen Huang’s departure and allegations involving Shanghai fund manager Kang Wei, including a reported $7.5 billion Nvidia GPU smuggling operation.

Former OpenAI researcher warns of deadly AI accidents

Former OpenAI researcher Paul Christiano warned that rogue AI agents could escape human control. He said such systems could cause catastrophic accidents that might kill people. The article provides no further details about his warning.

Security Affairs launches its eighth AI cybersecurity newsletter

Security Affairs says artificial intelligence is changing cybersecurity for both attackers and defenders. AI agents can automate tasks, analyze large amounts of data, find vulnerabilities and speed up offensive operations. The newsletter will cover threats, attacks, vulnerabilities and other developments where AI and cybersecurity meet. It also follows how AI supports threat detection, incident analysis and response.

AI can speed up branding but may weaken what makes a business unique

Jaci Lund, founder of a branding company in Atlanta, argues that using AI for branding can save time but may make a business seem less distinctive. She says website visitors form a first impression in 0.05 seconds and use a company’s presentation to judge whether it seems trustworthy. The article says people can recognize AI’s familiar look and tone, and research shows trust in that content drops by nearly half. Lund connects this reaction to the effort heuristic, the idea that people may value work more when they believe it took greater effort.

Chinese AI firms limit cyber powers in open models

Chinese AI companies are beginning to reduce the cybersecurity capabilities of their most advanced open-weight models after high-profile AI hacking incidents involving US and Chinese firms. Tencent said it removed cyber-hacking capabilities from its latest Hy4 model while keeping its cyber-defence capabilities. Moonshot AI made a similar move, according to the article. The article raises questions about whether these limits can last as AI advances and competition between companies continues.

AI finds major software flaws, but defenders still have time to act

Researchers have recently found serious security flaws in software such as PHP, Linux, OpenSSL, and FreeBSD. Anthropic’s Project Glasswing found more than 129,000 verified vulnerabilities from April to July, and AI tools can help defenders find and check weaknesses quickly. Google researchers found that only about 0.2% of the vulnerabilities disclosed in the first half of the year had been used in attacks. The next challenge is fixing flaws quickly, since one study found companies fixed only 26% of known exploited vulnerabilities last year. The column urges governments to share AI cyber tools with trusted defenders, improve software security rules, and invest in protecting critical infrastructure and open-source projects.

Cyprus faces warnings about AI election interference

European officials are concerned that artificial intelligence could be used to spread misleading material and influence elections, including Cyprus’ 2028 presidential vote. AI can create fake videos, images, and audio that look real and can be shared widely. Such campaigns could damage candidates’ reputations, confuse voters, and shape election debates. European leaders discussed the threat at an informal summit in Dublin on 13 November. President Nikos Christodoulides said countries need measures to prevent and protect against interference before it happens.

Europe’s RAISE plan aims to strengthen AI research

The RAISE initiative, launched at the Dublin summit, aims to connect Europe’s artificial intelligence research resources. The effort brings together public and private groups, including Elsevier, Mistral, Nvidia, and Google DeepMind. The European Commission says CERN would work with Mistral to build the largest foundational model in physics to date. The initiative raises questions about Europe’s reliance on American technology and privately owned models, as well as concerns about ethics and research integrity. Its success will depend on whether Europe can provide enough resources and build the network quickly while protecting its values and autonomy.

How universities can prepare graduates for AI-driven work

Artificial intelligence is improving at tasks done in many jobs that require a university degree, while jobs involving physical work have generally faced less risk from automation. A 2024 study found newer AI models performed better than earlier systems on reasoning tests, and an OpenAI model recently passed a university criminal law exam. AI could also affect trades if reliable, affordable robots become available. The article says graduates should learn to judge AI’s results, spot errors, improve its work, and collaborate with it. Universities should teach these skills alongside the subject knowledge students need for their degrees.

U.S. and eight partners coordinate AI supply chain security

The United States and eight partner countries have agreed to identify and protect artificial intelligence technologies and supply networks that face risks involving China. The Financial Times reported the agreement on Sunday. For investors, the pact signals closer coordination around the infrastructure that supports AI.

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 hardware AMD MEXT memory optimization Google TPU Nvidia GPUs GPU supply chain AI cybersecurity software vulnerabilities AI-enabled attacks cyber defense AI safety Tencent Hy4 European AI research RAISE initiative AI-generated election disinformation AI education workforce automation AI branding AI trust technology security

Comments

Loading...