NVIDIA’s Vera Rubin systems post strong results across MLPerf Training v6.1 and Endpoints v1.0. In training, Vera Rubin NVL72 records the fastest time on six of seven benchmarks and delivers up to twice the performance per GPU of an optimized GB300 NVL72. NVIDIA says its platforms hold the fastest time on all seven benchmarks overall. On DeepSeek V3, Vera Rubin scales from 256 to 1,024 GPUs at more than 97% efficiency, while NVIDIA also reports a 10,240-GPU GB300 result; Microsoft Azure matches that scale with GB200 systems.
For agentic inference, NVIDIA’s provisional results report up to 10 times more throughput per provisioned kilowatt on Vera Rubin than GB300 for Kimi K3, nine times for DeepSeek-V4.1-Flash and five times for Qwen3.6-35B-A3B. At matched query concurrency, end-to-end interactivity is three, five and three times higher, respectively. A separate Kimi K3 comparison reports 41,414.2 tokens per second from 32 Vera Rubin accelerators at 100.89 kilowatts, versus 2,210.4 tokens per second from 48 GB300 accelerators at 60 kilowatts, with both systems above 50 tokens per second per user. MLCommons is reviewing the provisional result. NVIDIA cites prefix caching, state reuse, NVIDIA Dynamo routing and model-specific GPU optimizations. Its DeepSeek V3 scaling test cuts training time from 16 minutes to four.
Google Cloud reports that 640 Ironwood TPUs reached 0.765 accuracy on Qwen3.5-397B in 18.02 minutes, above a 0.69 target, with chips idle less than 3% of the time. The test uses pre-agentic reinforcement learning, training coding agents as they generate code and learn from execution; it measures convergence from an official reference checkpoint, not pretraining from scratch. NVIDIA says Vera Rubin NVL72 can deliver up to twice Blackwell Ultra’s performance in the same round.
Markets turn lower ahead of NVIDIA’s results and Anthropic’s planned IPO. In Monday, Nov. 16 trading, the Dow falls 1.26% to 1.41% across the two reports, the S&P 500 drops 1.78% to 1.92%, and the Nasdaq loses 2.30% to 2.49%; the semiconductor index is down 4.6% to 5.3%. One report says the Nasdaq is more than 10% below its Oct. 6 record close, a correction, and NVIDIA shares fall 4.2%. Analysts expect NVIDIA earnings to rise 90.4% year over year, with $109.10 billion in revenue. One article says markets price in a 95.7% chance of a 25-basis-point Federal Reserve rate hike in December. Anthropic’s anticipated IPO is also a test of confidence in AI stocks; a report describes a planned $1.4 trillion IPO and an expected $10 billion NVIDIA investment immediately afterward, while noting Anthropic has diversified hardware suppliers beyond NVIDIA.
Blackstone president John Gray says doubts about AI weigh on Korean stock valuations, although demand to build computing infrastructure could benefit the country. He says strong company results, steadier interest rates and geopolitical conditions could help revive IPO and merger activity, and points to a possible Anthropic IPO as a signal of AI-sector strength. Blackstone portfolio companies’ global sales rise more than 10% in the third quarter, alongside cash-flow growth.
Companies are also applying AI to operations and confronting risks around data. In sports, Aramark uses arrival information to improve venue entry, navigation, food service and restocking; the Detroit Lions use AI to recommend gates and offer food and drink discounts. The New York Islanders work with Viam to assess bathroom and concession lines. Industry speakers discuss agents that surface insights and learn from human guidance, while AMBSE’s Tameka Rish identifies connecting data for real-time use as a challenge. Separately, an article warns that old, duplicate and forgotten enterprise data can waste network resources and create security and compliance risks. About 80% of the world has data-protection laws, yet organizations struggle to enforce retention rules; AI agents may inherit permissions and act quickly without recognizing sensitive, outdated or restricted information.
An Efekta Education survey of 850 teachers in the UK and US finds regular AI use is reported by 86% of teachers aged 55–64, compared with 60% of those aged 18–24. The survey points to a need for teacher training and wins an ETIH Innovation Award. In another company update, AMD completes its $8.2 billion acquisition of World Labs and appoints founder Fei-Fei Li executive vice president and chief scientist. World Labs develops models that simulate 3D environments for training robots, and CEO Lisa Su says the research will inform AMD’s hardware, software and systems.
Agora reports third-quarter 2026 revenue of $42.1 million, up 19%, and $3.0 million in net income, its eighth consecutive GAAP-profitable quarter. Active customers edge up to 4,024, while its AI products handle more than one million agent calls per day. The company forecasts fourth-quarter revenue of $44 million to $45 million. In a separate sports story, Luis Gurule, a +700 underdog, beats Bilal Hasan by split decision at UFC 334, despite Hasan landing more significant strikes. Gurule records more total strikes and completes two of five takedown attempts; afterward, he mocks ChatGPT for predicting Hasan would win by first-round TKO.
Key Takeaways
- NVIDIA says Vera Rubin NVL72 leads six of seven MLPerf Training v6.1 benchmarks and delivers up to twice the per-GPU performance of optimized GB300 NVL72.
- On DeepSeek V3, Vera Rubin scales from 256 to 1,024 GPUs at more than 97% efficiency, cutting the reported training time from 16 minutes to four; NVIDIA and Microsoft also report 10,240-GPU-scale systems using GB300 and GB200, respectively.
- Provisional MLPerf Endpoints results show Vera Rubin throughput per kilowatt up to 10 times GB300 on Kimi K3, nine times on DeepSeek-V4.1-Flash and five times on Qwen3.6-35B-A3B. MLCommons is reviewing the Kimi K3 result.
- Google Cloud says 640 Ironwood TPUs reach 0.765 accuracy on Qwen3.5-397B in 18.02 minutes, exceeding a 0.69 target; the test measures convergence from a reference checkpoint, not full pretraining.
- Investors sell technology shares ahead of NVIDIA earnings and Anthropic’s planned IPO. Analysts expect NVIDIA revenue of $109.10 billion and 90.4% year-over-year earnings growth; reports differ on the precise Monday market declines.
- Reports describe Anthropic’s planned IPO at a $1.4 trillion valuation and an expected $10 billion NVIDIA investment after the IPO; Anthropic has also diversified its hardware suppliers beyond NVIDIA.
- Blackstone’s John Gray says AI-related investor doubts weigh on Korean stock values, while computing-infrastructure demand could benefit the country; Blackstone portfolio companies’ third-quarter global sales rise more than 10%.
- Sports organizations use AI for venue entry, gate recommendations, discounts, queue monitoring and restocking; an industry leader says connecting data for real-time use remains a challenge.
- An Efekta survey finds regular AI use among 86% of teachers aged 55–64 versus 60% of teachers aged 18–24, highlighting the need for training. A separate report warns that poorly governed enterprise data can expose AI agents to security and compliance risks.
- AMD completes its $8.2 billion World Labs acquisition and names Fei-Fei Li executive vice president and chief scientist. Agora reports $42.1 million Q3 2026 revenue, up 19%, and more than one million daily AI agent calls; UFC 334 underdog Luis Gurule defeats Bilal Hasan by split decision.
Vera Rubin delivers far more Kimi K3 tokens per watt
NVIDIA reported that 32 Vera Rubin accelerators produced 41,414.2 tokens per second at 100.89 kilowatts on Kimi K3. In comparison, 48 GB300 accelerators produced 2,210.4 tokens per second at 60 kilowatts, with both systems keeping end to end interactivity above 50 tokens per second per user. The Kimi K3 result is provisional while MLCommons reviews it. NVIDIA also said Vera Rubin scaled from 256 to 1,024 GPUs with about 97 percent efficiency on DeepSeek V3, cutting training time from 16 minutes to four.
Google reports fast Ironwood TPU training results
Google Cloud said 640 Ironwood TPUs reached 0.765 accuracy on Qwen3.5-397B in 18.02 minutes, exceeding a target of 0.69. The test used pre agentic reinforcement learning, which trains coding agents as they generate code and learn from running it. Google said the chips waited less than 3 percent of the time. The results measure training from an official reference checkpoint to convergence, not full pretraining from scratch. NVIDIA said its Vera Rubin NVL72 can deliver up to twice the performance of Blackwell Ultra in the same round.
Vera Rubin reaches up to twice the performance of GB300 in MLPerf tests
In its first MLPerf Training v6.1 appearance, NVIDIA’s Vera Rubin NVL72 had the fastest training time on six of seven benchmarks and delivered up to twice the performance per GPU of an optimized GB300 NVL72 system. NVIDIA used a 10,240 GPU GB300 NVL72 cluster to record the fastest DeepSeek V3 671B result, while Vera Rubin scaled from 256 to 1,024 GPUs with more than 97 percent efficiency. NVIDIA platforms submitted results for all seven benchmarks and held the fastest time on each. The tests covered large language models, reinforcement learning, image generation, and recommendation systems.
NVIDIA reports leading MLPerf results for Vera Rubin training and agentic inference
NVIDIA said Vera Rubin led across MLPerf Training v6.1 and MLPerf Endpoints v1.0, which tests agentic coding and enterprise tasks. Vera Rubin NVL72 delivered up to twice the performance per GPU of an optimized GB300 NVL72 system in training. On DeepSeek V3, Vera Rubin scaled from 256 to 1,024 GPUs with about 97 percent efficiency, while NVIDIA also ran GB300 systems at 10,240 GPUs. For the provisional Kimi K3 result, NVIDIA reported more than 10 times the throughput per watt of GB300 NVL72 at the same response speed. Microsoft Azure matched NVIDIA’s 10,240 GPU scale using GB200 systems.
Vera Rubin delivers up to 10 times more agentic throughput per kilowatt
NVIDIA’s provisional MLPerf Endpoints v1.0 results show Vera Rubin NVL72 delivering higher agentic throughput per provisioned kilowatt than GB300 NVL72 across three models. The reported throughput ratios were 10 times for Kimi K3, nine times for DeepSeek-V4.1-Flash, and five times for Qwen3.6-35B-A3B. At matched query concurrency, Vera Rubin’s end to end interactivity ratios were three times, five times, and three times, respectively. NVIDIA described techniques including prefix caching, state reuse, routing with NVIDIA Dynamo, and model specific GPU optimizations to reduce repeated work and improve serving.
Wall Street falls as AI worries grow ahead of Nvidia results
US stocks fell sharply on Monday, Nov. 16, as investors sold technology shares ahead of Nvidia’s results and Anthropic’s planned IPO. The Dow Jones Industrial Average fell 739.38 points, or 1.41 percent, to 51,608.17, while the S&P 500 dropped 1.92 percent and the Nasdaq fell 2.49 percent. The Philadelphia Semiconductor Index lost 5.3 percent, and the Nasdaq moved more than 10 percent below its Oct. 6 record close, putting it in a correction. Analysts expect Nvidia’s earnings to rise 90.4 percent from a year earlier, with revenue of US$109.10 billion. Markets priced in a 95.7 percent chance of a 25-basis-point Federal Reserve rate hike in December.
Wall Street slides as investors await Nvidia results and Anthropic IPO
Wall Street fell sharply on Monday as investors grew uneasy about AI stocks ahead of Nvidia’s earnings report and Anthropic’s planned IPO. The Dow Jones Industrial Average fell 659.41 points, or 1.26 percent, while the S&P 500 lost 1.78 percent and the Nasdaq dropped 2.30 percent. Nvidia shares fell 4.2 percent, leading the Magnificent Seven stocks lower. The Philadelphia Semiconductor Index dropped 4.6 percent. Analysts expect Nvidia to report a 90.4 percent year-on-year earnings increase and revenue of $109.10 billion.
AI tools reshape business intelligence in sports
Sports industry leaders say AI is changing business intelligence and operations, from helping fans enter venues to managing concessions. Aramark uses arrival data to improve entry, navigation and food services, while the Detroit Lions have begun using AI to recommend arrival gates and offer food and drink discounts. The New York Islanders use AI partner Viam to assess bathroom and concession lines, and Aramark uses real-time data to guide restocking across its sports venues. Panelists also discussed AI agents that can surface insights and learn from human guidance. AMBSE’s Tameka Rish said a key challenge is connecting data so organizations can use it in real time.
Efekta survey finds older teachers use AI more often
A survey by Efekta Education of 850 teachers in the UK and US found that 86 percent of teachers aged 55 to 64 use AI regularly. Among teachers aged 18 to 24, the figure was 60 percent. The findings point to a need for teacher training to support AI use in education. The survey won the ETIH Innovation Awards.
Anthropic IPO set to test investor confidence in AI
Anthropic’s planned $1.4 trillion IPO on Thursday will test the AI trade amid market turbulence, political backlash and a crisis at Nvidia. A Morningstar analyst told The New York Times that Anthropic has diversified its hardware suppliers beyond Nvidia. The company is also expecting a $10 billion investment from Nvidia immediately after its IPO.
Blackstone president says AI doubts weigh on Korean stocks
Blackstone president John Gray said investor doubts about artificial intelligence are keeping Korean companies’ stock values low. He said Korea could benefit from the need to build more computing infrastructure, and strong results from Korean companies could improve how investors value them. Gray said a possible Anthropic initial public offering could signal continued strength in the AI sector. He added that steadier interest rates and geopolitical conditions, along with solid company results, could help revive IPO and merger activity; Blackstone portfolio companies’ sales rose more than 10% worldwide in the third quarter, and their cash flow also grew.
AI agents need awareness of enterprise data risks
Many companies keep large amounts of old, duplicate, or forgotten data, which can waste network resources and create security and compliance risks. About 80% of the world has data protection laws, but companies often struggle to enforce their own rules for how long records should be kept. Exact-match tools can miss related files, such as different versions of the same contract. The article compares this problem to a loss of proprioception, the body’s sense of its position and movement. It warns that AI agents may act at machine speed while inheriting data access permissions without knowing whether data is sensitive, outdated, or restricted.
AMD completes $8.2 billion World Labs deal and names Fei-Fei Li chief scientist
AMD finalized its $8.2 billion acquisition of World Labs and appointed Fei-Fei Li as an executive vice president and chief scientist. World Labs builds AI models that simulate 3D environments to help train robots to move in virtual and real worlds. AMD CEO Lisa Su said the company aims to develop computing platforms and models together for future AI systems. She welcomed Li and the World Labs team, saying their research would help shape AMD’s hardware, software, and systems.
Agora revenue rises 19% as AI products handle over one million calls daily
Agora, Inc. reported Q3 2026 revenue of $42.1 million, up 19%, driven by growth in real-time engagement and conversational AI products. The company recorded $3.0 million in net income and reported GAAP profitability for the eighth straight quarter. Active customers rose slightly to 4,024, while its AI products handled more than one million agent calls each day. Agora expects Q4 revenue of $44.0 million to $45.0 million. As of Nov. 17, 2026, its stock traded at $3.98, up 0.51%, and its market value was $333.55 million.
Luis Gurule upsets Bilal Hasan and taunts ChatGPT after UFC 334 win
Luis Gurule, a +700 underdog, defeated Bilal Hasan by split decision at UFC 334, with judges scoring the fight 29-28, 28-29, and 29-28. Gurule mocked ChatGPT after the fight because it had predicted Hasan would win by first-round TKO. Hasan landed more significant strikes, but Gurule recorded more total strikes and controlled the ground after completing two of five takedown attempts. Gurule said he planned to take Hasan down and wear him out. Hasan had been listed as a -1100 favorite.
Sources
- Vera Rubin resets math behind Kimi K3 benchmark throughput
- Google Ironwood resets expectations in MLPerf Training v6.1 results
- NVIDIA Vera Rubin NVL72 Achieves up to 2x Blackwell Ultra Performance in MLPerf Training v6.1
- NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance
- NVIDIA Vera Rubin NVL72 Leads Agentic Inference Performance on MLPerf Endpoints v1.0
- US stocks: Wall Street ends sharply lower on AI jitters ahead of Nvidia results, Anthropic IPO
- Wall Street slumps on AI jitters ahead of Nvidia results, Anthropic IPO
- AI ‘changes everything’ in business intelligence
- Older teachers lead AI use in Efekta's UK and US survey | ETIH EdTech News
- Anthropic IPO set to define AI trade
- John Gray, president of Blackstone, pointed to caution on artificial intelligence (AI) as the reason..
- Proprioception Before Posture: AI Agents Acting at Machine Speed Inherit Data Awareness
- AMD finalizes $8.2 billion World Labs acquisition, adds 'godmother of AI' as EVP and chief scientist
- Agora reports 19% revenue growth in Q3 2026 with rising AI adoption and positive net income
- UFC 334: Luis Gurule celebrates upset win over Bilal Hasan with AI callout: 'ChatGPT don't know s***'
Comments
Please log in to post a comment.