The Week in AI
The week in brief
NVIDIA reported on 16 November that its Vera Rubin NVL72 system delivered up to 2x the per-GPU training performance of GB300 NVL72 in MLPerf Training v6.1.
This page has one day of reporting so far, Monday 16 November, and will be revised as the week goes on. The NVIDIA figures are company-reported, and MLCommons has not yet verified the inference result. OpenRouter measured the Hugging Face routing figures, and the physics record is only partly confirmed, so it is written as OpenAI says it.
Two of the three stories concern the plumbing under models. NVIDIA is showing how fast its next training and serving system runs, and Hugging Face has handed its language model routing to a third party. The third story is OpenAI's claim that a model turned a physicist's strategy into a long construction that a Lean checker accepted.
NVIDIA's Vera Rubin NVL72 posts up to 2x GB300's per-GPU training speed on MLPerf
NVIDIA says Vera Rubin NVL72 delivered up to 2x the per-GPU performance of GB300 NVL72 in MLPerf Training v6.1, and that it won every benchmark in the round.
What happened
MLPerf is the benchmark suite run by MLCommons, where vendors submit the time it takes to train or serve fixed reference models. This round is Vera Rubin's debut, and the results are early. NVIDIA reports that on DeepSeek-V3 the system scaled from 256 to 1,024 GPUs at about 97% efficiency, which cut training time from 16 minutes to about four.
NVIDIA also made the first submission to MLPerf Endpoints, a test of models served to users, which is the first such submission the records document. Its provisional Kimi K3 result shows more than 10x the throughput per watt of GB300 NVL72 at 50 tokens per second per user. The 32-accelerator Vera Rubin NVL72 system produced 41,414.2 tokens per second at 100.89 kilowatts. The largest scale in the round was 10,240 GPUs, which NVIDIA ran on GB300 systems and Azure matched with GB200.
At SC26 on the same day, NVIDIA announced Rubin HPC, a GPU for simulation and data processing with more than 50,000 CUDA cores. NVIDIA describes that as a record core count and expects Rubin HPC to be available in 2028.
Why it matters
Training time on a fixed model at a fixed scale is one of the few numbers that lets buyers compare systems directly. A 2x per-GPU gain means the same job needs about half the GPUs, or finishes in about half the time.
The serving result matters to anyone paying for inference, because throughput per watt sets the electricity bill of each token. The comparison is at a fixed 50 tokens per second per user, so it does not say what happens at other response speeds.
What we don't know
MLCommons has not verified the Endpoints number, and every figure here is NVIDIA's own. It is unknown how the results compare with submissions from other vendors, which the records do not cover. NVIDIA has not given a price or a shipping date for Vera Rubin NVL72 in the records.
Hugging Face moved language model routing to OpenRouter, reaching 80 providers
Hugging Face switched Inference Providers for language models from its own router to OpenRouter, so one Hugging Face token reaches open-weight models across more than 80 providers.
What happened
Language model requests now go through OpenRouter, which picks a provider for each call using live price, speed and reliability. If a provider fails, OpenRouter retries on another one. The router supports the OpenAI Chat Completions and Responses APIs and Anthropic's Messages API for every chat model.
OpenRouter reports 46% fewer capacity errors under the new routing, 16.9% better time to first token and 16.3% better throughput. Image, video and audio generation still run directly through Hugging Face Inference Providers. Custom provider keys no longer work for language models.
Billing is unchanged. Users pay the provider's price on their Hugging Face bill with no markup. PRO users get $2 of credits each month, and Team and Enterprise accounts get $2 per seat.
Why it matters
OpenRouter also published market figures alongside the change. Open-weight models took 71.7% of its tokens at the end of September 2026, up from 26.8% in October 2025, after a peak of 77.5% in July. OpenRouter puts the average price of open-weight models at $0.28 per million input tokens and $1.62 per million output tokens, against $1.27 and $6.36 for closed models.
These are OpenRouter's numbers about its own traffic, so they describe the developers who use OpenRouter and may differ from the market at large. The routing change means Hugging Face no longer runs its own selection logic for language models, and a developer who relied on a personal provider key has to change setup.
What we don't know
The records do not say what Hugging Face pays OpenRouter, or whether the no-markup pricing will hold. It is not stated how many providers were available through the old router, so the size of the gain in reach is unclear.
OpenAI says GPT-6 Astra led the detailed proof in a lattice physics preprint
OpenAI says GPT-6 Astra led the detailed proof in physicist Nathaniel Craig's October preprint on defining the Standard Model on a lattice.
What happened
Craig's preprint builds a gauge-invariant measure for lattice chiral fermions, a long-open piece of defining the Standard Model on a lattice. Craig says he supplied the strategy and Astra turned it into a construction. The paper reports that 16 Lean 4 modules compiled without errors.
Lean 4 is a proof assistant in which a computer checks each logical step, so a module that compiles has no gap that the checker can find. The checker confirms that the proof follows from the stated definitions, and it does not confirm that the definitions capture the physics.
The source check on this record is partial, so the account above is as OpenAI and the paper report it.
Why it matters
The division of labor here is stated plainly. A human physicist chose the approach and the model wrote the long construction. That is a different claim from a model finding a result on its own.
What we don't know
It is not yet known how physicists outside the paper judge the construction. The records do not say how much compute Astra used or how many attempts it needed.
Also worth knowing
Models and research
- Huawei Cloud released CloudRobo-WM1.2, a world action model for robots built for its Ascend platform, and says it will open-source it. A world model predicts how a robot's surroundings change after an action. According to a Pandaily summary, Huawei claims first place on RoboTwin 2.0 and LIBERO-Pro 2.0, and those rankings are not independently confirmed.
- Sarvam AI launched Token Factory, an inference service that runs on GPU clusters in India behind one OpenAI-compatible API. It serves Sarvam 105B and open-weight models including GLM 5.3, DeepSeek V4 Pro, Kimi K3 and Google's Gemma 4 31B. Sarvam says it matches or beats leading global providers on speed and price, and the page read has no figures.
- Alibaba added 10 selectable models to Qwen Office, including Qwen4-Max, which it opened in preview on 11 November, and third-party models such as Kimi-K3.1-Pro, GLM-5.3 and DeepSeek-V4-Pro. Enterprise administrators can set model permissions by department, user group or member.
- Moonshot AI is reportedly steering its Kimi models toward science as coding benchmark gains flatten, according to the South China Morning Post.
Products and safety
- OpenAI released an image classifier for vetted organizations that scores potential child sexual abuse material (CSAM), with a separate score for AI-generated material. It is free to approved organizations through the API. OpenAI says it flags 48 percent fewer images than its earlier approach, and that flagged images were twice as likely to contain abusive material. ChatGPT now also shows a warning to users seeking such material, removes the conversation and pauses the account for 24 hours. The records only partly confirm the model identifier and the role of the National Center for Missing and Exploited Children in training and evaluation.
- OpenAI says it added computer-use agents, improved clinical answers and new plugins to ChatGPT for Healthcare. Both OpenAI pages returned server errors, so how the agents work, who can use them and the price are unconfirmed.
- OpenAI is expanding ChatGPT for Academic Researchers to a goal of 200,000 researchers, double the goal from July. Applications for 50,000 free ChatGPT Plus grants open on 23 November, and each grant covers the recipient and up to two collaborators. OpenAI says a survey of more than 1,800 pilot researchers found they estimated tasks would take nearly twice as long without AI.
- OpenAI says it signed Apollo Research, METR, Transluce and Goodfire for the first phase of its embedded model assessments.
Security and policy
- US Commerce Department released a rule on remote access to AI chips, which reports describe as closing China's route through overseas data centres. The records give no text of the rule.
- Anthropic reportedly cut its coordinated disclosure window for high and critical severity bugs from 90 days to 30 days, citing AI-driven bug discovery.
- Google Threat Intelligence Group reports observing attackers using an autonomous AI framework to turn security patches into exploits. Google describes it as the first such observation.
- DeepSeek users report a daily chat quota and an input cap near 40,000 Chinese characters. DeepSeek has not confirmed either.
Funding
- Fractile, a UK chip startup, reportedly raised $720 million in a Series C.
- LangChain reportedly raised $300 million at a $3 billion valuation.
People this week
No moves were recorded on 16 November.
What did not change
Every new system with a stated access route on Monday is reached through a closed API or an app. That includes OpenAI's classifier, Sarvam's Token Factory, Hugging Face's routing and Qwen Office. Huawei says it will open-source CloudRobo-WM1.2 but has not released it.
No independent evaluator has published results for Vera Rubin NVL72. MLCommons runs the MLPerf suite, but the Endpoints result is still provisional.
What to watch next
- MLCommons verification of NVIDIA's provisional Kimi K3 result on MLPerf Endpoints.
- Whether Huawei Cloud releases CloudRobo-WM1.2 as open source, and whether the RoboTwin 2.0 and LIBERO-Pro 2.0 rankings are confirmed.
- The US pre-order opening for the motorola signature 27 on 19 November.
- Applications for the 50,000 ChatGPT Plus grants for academic researchers opening on 23 November.
- Baidu's open weights for ERNIE 5.5 Flash, promised by 25 November.
- Responses from physicists to the lattice construction that OpenAI says GPT-6 Astra led.