NVIDIA Vera Rubin 30x Throughput, OpenAI Codex Adoption Surges

NVIDIA Vera Rubin achieved up to 30x per-megawatt throughput and 35x lower token cost on real agentic coding tasks vs GB300 NVL72.
OpenAI Codex adoption surged 108x in legal, 41x in sales and recruiting, showing AI expanding beyond tech.
Marin launched a 535B-parameter open-source training run on 18.75T tokens using 11 GB200 NVL72 systems for ~3 months.
Open-source AI token share jumped from 28% to 62% in two months, per Yann LeCun.
Andrew Gordon Wilson joined Perplexity AI as research lead for continual learning, synthetic data, and long-horizon RL.

NVIDIA unveiled silicon-verified performance for Vera Rubin: on SemiAnalysis AgentX real agent coding tasks with DeepSeek V4 Pro, per-megawatt throughput improves up to 30x and token cost drops as much as 35x vs GB300 NVL72. The company stressed that agentic sessions differ from chat or summarization workloads, with context growing to hundreds of thousands of tokens across steps. — via 1
SpaceX will deploy NVIDIA Vera for accelerated orchestration, code execution, and data processing to support its next-generation agentic AI. NVIDIA frames this as bringing a unified architecture from gigawatt-scale AI factories to orbit, keeping GPUs fed and agents fast. — via 1
NVIDIA announced NVIDIA Groq 3 LPX is in full production, designed to pair with Vera Rubin NVL72 for faster, smarter agents. It relies on close co-design of seven chips and five dedicated racks, forming what NVIDIA calls the broadest AI factory platform. — via 1

Greg Brockman said OpenAI has built the capacity for long-term research bets, citing a member's first-hand account that the gpt-live series of full-duplex models had full management support. This signals OpenAI's willingness to commit to ambitious, multi-year AI research directions. — via 1
Brockman highlighted Codex adoption growth as evidence of a new way to complete knowledge work: 108x in legal, 41x in sales, 41x in recruiting, 26x in marketing, and 24x in healthcare. These figures show AI power users expanding well beyond the tech sector. — via 1

Andrew Ng framed the Marin project as a valuable demonstration of openness in model training, publishing code, data, recipes, and even experimental results. Marin kicked off its largest training run this week: a 535B-A23B model planned for pretraining and mid-training on 18.75T tokens, using 11 GB200 NVL72 systems for about three months, preceded by scaling-law ladder experiments with smaller models to debug and predict the full run. — via 1
Yann LeCun noted that open-source token share rose from 28% to 62% in two months, predicting closed frontier models will ultimately account for only 15-25% of tokens but 60-90% of economic value. He argued the open-source share gain is favorable for infrastructure demand, as open-source AI is likely to drive more compute usage. — via 1

Andrew Gordon Wilson joined Perplexity AI as research lead, reporting to Denis. He will lead new research directions including continual learning, synthetic data, long-horizon RL environments, and architecture, and Perplexity is hiring for the team. — via 1
Runway introduced Ruby, a feature that converts content from any model into delivery specifications such as 16-bit EXR, 10/12-bit ProRes, and HEVC, with support for Seedance 2.5, Gen-4.5, MiniMax H3, among others. This seems aimed at simplifying post-production pipelines for generative video. — via 1
Hamel Husain outlined two approaches to AI evaluation: top-down, where Claude excels at designing evals, and bottom-up, where users must aggregate feedback themselves. The distinction is useful for teams deciding how to build eval frameworks. — via 1
Yann LeCun urged precision in AI terminology: VLM is a multimodal extension capturing static semantics, while a world model is a dynamics model—and they should not be conflated. This distinction matters as the field debates path to embodied intelligence. — via 1

Discover