Nvidia Vera Rubin Leads All Seven MLPerf Training Tests

Nvidia’s Vera Rubin NVL72 leads all seven tests in MLPerf Training v6.1, recording the fastest time in six and up to twice the per-GPU performance of its GB300 predecessor. In the DeepSeek-V3 671B test, 1,024 Rubin GPUs completed training in 4.14 minutes, with about 97% scaling efficiency from 256 to 1,024 GPUs. The system combines 72 GPUs, 36 CPUs, HBM4 memory and sixth-generation NVLink. In MLPerf Endpoints, it delivered more than 10 times the GB300’s throughput per watt on Kimi K3. CoreWeave also reported a 2.7-times faster 256-GPU DeepSeek-V3 run than on its earlier GB200 setup.

Lambda’s MLPerf results show gains on Nvidia Blackwell Ultra as well: 64 GPUs in a GB300 NVL72 trained GPT-OSS-20B in 19.697 minutes, 2.86 times faster than a 64-H100 result from the previous round. Lambda’s unofficial Flux.1 reproduction took 26.36 minutes, 1.5 times faster than the quickest result on the same hardware six months earlier; the company attributed most of that improvement to software. It called this its first multimodal training workload in the MLPerf harness.

AI security developments span both attacks and defenses. Google says it detected its first cyberattack by an autonomous AI agent in recent weeks; the agent used machine learning to find vulnerabilities and develop exploits, though Google has not identified the target systems or disclosed the attack’s details. Broadcom warns that autonomous agents can operate faster than security teams can respond and describes VMware HyperArmor and Adaptive Protection as layers for defense and threat monitoring. Google Project Zero is also shortening its vulnerability disclosure deadlines: 90 days from November 13, 2026, 60 days from February 2, 2027, and 30 days from January 3, 2028. Google says AI can help attackers analyze patches, while vendors caution that shorter windows may not leave enough repair time. Project Zero found an exposed server running exploit_agent that also contained stolen data; the policy does not address delays in getting patches to users.

Zscaler says Security for AI bookings rose more than 50% sequentially in fiscal 2026’s fourth quarter, with its sales pipeline up more than 75%. The business generated about $70 million in annual recurring revenue for the year, more than 10 times the prior year. Zscaler estimates a $24 billion market for Security for AI and another $24 billion for Data Security; roughly 70% of its fourth-quarter AI security deals included Data Security. At the same time, it forecasts fiscal 2027 revenue of $3.91 billion to $3.94 billion, about 17% growth versus 25% in fiscal 2026.

Other developments show AI’s reach into business and public policy. Thomson Reuters is expanding CoCounsel agentic AI across ONESOURCE and Pagero tax and compliance products, with workflows to check records and invoices, find errors, and prepare tax returns for professional review. Its direct-tax tools can prepare workpapers and draft disclosures and board materials with figures linked to sources. The company says the workflows follow its Fiduciary-Grade AI standard; 62% of tax leaders surveyed said their departments made limited or no progress adopting automation and AI in the past year. Separately, Chinese entities are reportedly using output from leading US AI models to train domestic competitors as US chip restrictions tighten, though the report gives no names or methods.

AI investment and infrastructure plans also draw scrutiny. A report warns that circular financing—when one company funds another expected to buy its products—could transmit financial stress among AI firms, suppliers, investors and customers. It cites estimated US AI infrastructure investment of about $500 billion this year and $10.3 trillion from 2025 to 2032, while emphasizing that it does not establish an AI bubble or predict a wider crisis. In South Korea, SK Broadband plans to transfer its AI data center and submarine cable businesses to SK Horizon. Investors KKR and the IMM Investment–Stonebridge consortium have agreed to invest 3.08 trillion won, about $2.3 billion. The union says the parent will receive no direct cash and is seeking a replacement growth business and employee compensation; after talks broke down, it plans to seek mediation and secure the right to strike.

In North Carolina, a child fatality task force recommends protections against harmful technology and recurring funding for more school health staff. It backs efforts to address young people’s emotional dependence on chatbots, cover commonly used chatbots and give families routes to seek help. Schools held more than 500,000 counseling sessions in 2024–25, including nearly 18,000 involving depression, suicidal thoughts or other mental health concerns. The state’s 2026 budget added no funding for school health personnel, leaving some less-resourced counties with little or no access to support staff.

Purple Lexity CEO Aravind Srinivas says the AI search company has surpassed $1 billion in annualized sales, up from less than $250 million earlier in the year and more than $750 million in August. He credits growth partly to AI agents and using its own models alongside external models rather than building huge data centers. Purple Lexity says its models handle 80% to 90% of fast-search traffic, saving $100 million to $200 million annually. Srinivas reported a 56% gross margin for paid users and 17% including free users, compared with minus 219% in the fourth quarter of 2025. Its new Frontier Research program plans to invest in at least 10,000 GPUs to coordinate and continually train AI agents, with the goal of automating idea formation, experiments and knowledge discovery.

Key Takeaways

  • Nvidia’s Vera Rubin led all seven MLPerf Training v6.1 tests and trained DeepSeek-V3 671B in 4.14 minutes on 1,024 GPUs, with about 97% scaling efficiency.
  • Lambda reported a 19.697-minute GPT-OSS-20B training run on 64 Blackwell Ultra GPUs, 2.86 times faster than the previous-round 64-H100 result.
  • Google says an autonomous AI agent recently used machine learning to find vulnerabilities and develop exploits in its first detected AI-led cyberattack; the target and attack details remain undisclosed.
  • Google Project Zero’s vulnerability disclosure deadline moves from 90 days to 60 days in February 2027 and 30 days in January 2028, amid concerns that AI can help attackers exploit newly released patches.
  • Zscaler reported about $70 million in fiscal 2026 Security for AI annual recurring revenue and estimates separate $24 billion markets for Security for AI and Data Security, while forecasting slower revenue growth in fiscal 2027.
  • Thomson Reuters is extending CoCounsel agentic AI to ONESOURCE and Pagero workflows for tax, invoice and compliance work; 62% of tax leaders reported limited or no progress adopting automation and AI over the past year.
  • A report warns circular financing could spread financial risks across AI companies and their business partners, citing estimates of $500 billion in US AI infrastructure investment this year and $10.3 trillion from 2025 to 2032.
  • SK Broadband’s proposed transfer of its AI data center and submarine cable businesses to SK Horizon includes a 3.08 trillion won (about $2.3 billion) investment; its union is seeking a replacement growth business and employee compensation.
  • North Carolina’s child fatality task force backs chatbot protections for young people and recurring school health staffing funds; the 2026 state budget included no new funding for school health personnel.
  • Purple Lexity reports more than $1 billion in annualized sales and plans Frontier Research, including investment in at least 10,000 GPUs to train and coordinate AI agents.

Nvidia’s Vera Rubin leads MLPerf training benchmarks

Nvidia’s Vera Rubin NVL72 made its debut in MLPerf Training v6.1 on November 16, 2026. It led all seven tests and recorded the fastest training times in six, with up to twice the per GPU performance of the GB300 NVL72. On DeepSeek-V3, it reached about 97% scaling efficiency as the system grew from 256 to 1,024 GPUs. Vera Rubin also delivered more than 10 times the throughput per watt of the GB300 on Kimi K3 in MLPerf Endpoints v1.0, while CoreWeave said its 256 GPU DeepSeek-V3 run was 2.7 times faster than its earlier GB200 NVL72 setup. The system has 72 GPUs, 36 CPUs, HBM4 memory and sixth generation NVLink interconnects.

Lambda reports faster AI training on NVIDIA Blackwell Ultra

Lambda reported results in MLPerf Training v6.1 using 64 NVIDIA Blackwell Ultra GPUs in one NVIDIA GB300 NVL72 system. Its official GPT-OSS-20B training run took 19.697 minutes, 2.86 times faster than a result using 64 NVIDIA H100 GPUs in the previous round. Lambda also trained the Flux.1 text to image model in 26.36 minutes in an unofficial reproduction, 1.50 times faster than the quickest result on the same hardware six months earlier. Lambda said software improvements drove most of the Flux.1 speedup. The company described this as its first multimodal training workload in the MLPerf harness.

Nvidia’s Vera Rubin tops seven MLPerf AI training tests

Nvidia’s Vera Rubin NVL72 debuted in MLPerf Training v6.1 on November 16, leading all seven tests and posting the fastest times in six. It delivered up to twice the per GPU performance of the GB300 NVL72 across several workloads. On the DeepSeek-V3 671B benchmark, 1,024 Rubin GPUs completed training in 4.14 minutes, and scaling efficiency was about 97% between 256 and 1,024 GPUs. In MLPerf Endpoints v1.0, Vera Rubin delivered more than 10 times the throughput per watt of the GB300 on Kimi K3 under the same interactivity conditions. CoreWeave said its 256 GPU DeepSeek-V3 submission ran 2.7 times faster than its previous GB200 NVL72 setup; Vera Rubin has 72 GPUs, 36 CPUs, HBM4 memory and sixth generation NVLink interconnects.

Broadcom outlines defenses against machine speed AI attacks

Autonomous AI agents can move at machine speed, making it harder for people in security operations centers to respond in time. Broadcom describes a three layer defense that includes VMware HyperArmor and Adaptive Protection. HyperArmor helps defend against attacks, while Adaptive Protection monitors and adjusts to changing threats. Broadcom says security teams need a proactive and adaptive approach to protect their organizations.

Google reports its first autonomous AI cyberattack

Google says it detected its first cyberattack carried out by an autonomous AI agent in the past few weeks. The agent used machine learning to find vulnerabilities in Google’s systems and develop exploits to automate an intrusion. Google has not disclosed the attack’s exact nature or which systems it targeted. The company says AI powered attacks are becoming more sophisticated and difficult to detect, and that its security team is developing tools and techniques to identify and prevent them.

Google speeds up bug disclosure as AI helps attackers exploit patches

Google Project Zero has shortened the time software makers get to fix reported security flaws. Its 90-day deadline took effect on November 13, 2026, and will drop to 60 days on February 2, 2027, then to 30 days on January 3, 2028. Google says AI can help attackers study new patches and find ways to exploit flaws, and it reported finding an exposed server running the exploit_agent framework that also held stolen data. Project Zero also found that AI tools have helped it identify more flaws, while vendors warn that shorter deadlines could leave them too little time to repair systems. The change does not solve the delay between a patch being released and reaching users’ devices.

Zscaler’s AI security business grows as revenue growth slows

Zscaler said Security for AI bookings rose more than 50% sequentially in the fourth quarter of fiscal 2026, and its sales pipeline grew more than 75%. The business brought in about $70 million in annual recurring revenue in fiscal 2026, more than 10 times the prior year’s amount. Zscaler estimates a $24 billion market for Security for AI and another $24 billion for Data Security, and about 70% of its fourth-quarter Security for AI deals included Data Security. The company expects fiscal 2027 revenue of $3.91 billion to $3.94 billion, or about 17% growth, down from 25% in fiscal 2026.

Chinese firms use US AI models to train domestic rivals as chip limits tighten

Chinese entities are reportedly finding more advanced ways to use output from leading US AI models to train domestic competitors. The effort comes as China seeks to broaden its AI strategy and respond to tighter US limits on its access to chips. The article does not provide further details about the methods or the firms involved.

SK Broadband union challenges $2.3 billion data center spinoff plan

SK Broadband plans to move its AI data center and submarine cable businesses into a new company, SK Horizon, through a personnel split. SK Telecom says the split is planned for the first quarter of next year, with SK Broadband retaining 83.51% of the net assets and SK Horizon receiving 16.49%. Investors KKR and the IMM Investment-Stonebridge consortium have agreed to invest 3.08 trillion won, about $2.3 billion. The union says SK Broadband will receive no direct cash from the deal and is demanding a replacement growth business for revenue and profit it expects to lose, as well as employee compensation. After talks broke down, the union plans to seek mediation and secure the right to strike.

Thomson Reuters expands CoCounsel AI across tax and compliance tools

Thomson Reuters announced on November 17, 2026, that it is expanding CoCounsel agentic AI across its tax and transactional compliance products. New workflows for ONESOURCE and Pagero are designed to check customers, vendors, products and invoices, help find errors, and prepare tax returns for professional review. For direct tax work, the tools are designed to prepare workpapers, support ONESOURCE Tax Provision, and draft disclosures and board materials with figures linked to their sources. The company says the workflows follow its Fiduciary-Grade AI standard. Thomson Reuters noted that 62% of tax leaders said their departments made limited or no progress adopting automation and AI over the past year.

AI company investments could spread financial risk across the economy

A new report warns that circular financing in the AI industry could spread financial problems from one company to its suppliers, investors, and customers. In these arrangements, a company funds another business that is expected to buy its products or services. Analysts estimate U.S. AI infrastructure investment at about $500 billion this year and $10.3 trillion from 2025 to 2032, according to CRS. The report does not say that an AI bubble exists or that circular financing will cause a wider financial crisis, but it raises questions for Congress about possible risks, especially if banks become heavily involved.

North Carolina task force backs AI protections and more school health staff

The Child Fatality Task Force recommended legislation to protect young people from harmful technology and called for recurring funds to hire more school health personnel. It endorsed efforts by the national nonprofit YPA to protect young people from emotional dependence on chatbots, cover the chatbots they use most, and give families ways to seek help. In the 2024 to 2025 school year, schools held more than 500,000 counseling sessions, including nearly 18,000 about depression, suicidal thoughts, or other mental health concerns. North Carolina's 2026 budget included no new funding for school health personnel, leaving students in less-resourced counties with little or no access to support staff. The task force also aims to move the state toward nationally recommended staffing ratios.

Purple Lexity reports over $1 billion in annualized sales

AI search company Purple Lexity CEO Aravind Srinivas said annualized sales have topped $1 billion, up from less than $250 million earlier this year and more than $750 million in August. He said growth came partly from expanding into AI agents and from using its own AI models alongside outside models instead of building huge data centers. Purple Lexity uses its models for 80 to 90 percent of fast search traffic and says its computing approach saves $100 million to $200 million a year. Srinivas reported a 56 percent gross margin for paid users and a 17 percent margin when free users are included, up from minus 219 percent in the fourth quarter of 2025. The company also announced Frontier Research, a program that plans to invest in at least 10,000 GPUs to coordinate and continually train AI agents; its goal is an automated system that can form ideas, run experiments, and discover knowledge.

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.

Nvidia Vera Rubin MLPerf Training AI accelerators GPU performance AI infrastructure Blackwell Ultra AI cybersecurity autonomous AI agents vulnerability disclosure AI security Zscaler Thomson Reuters CoCounsel agentic AI AI investment circular financing AI data centers AI policy chatbot safety AI search AI training

Comments

Loading...