NVIDIA’s quarterly sales more than double to $96.2 billion

NVIDIA reported quarterly revenue of $96.2 billion, up 106% from a year earlier. The results, announced August 26, cover the quarter ended July 26. Revenue was up 18% from the preceding quarter.

Net income reached $59.7 billion, up 126% from a year earlier.

Data Center revenue reached $89.0 billion, up 117% year over year. NVIDIA also said its Vera Rubin computing platform was entering full production, with racks running at partners including CoreWeave, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure.

For the next quarter, NVIDIA forecasts $108 billion in revenue, plus or minus 2%. That outlook assumes no Data Center compute revenue from China.

Top Stories#

IBM released three sizes of model for coding, command-line tasks and search. IBM released Granite 4.2, a family of language models trained for reasoning, software tool use and multistep workflows. The models have 3 billion, 8 billion and 30 billion parameters.

IBM says the two larger versions receive additional training for software engineering, command-line tasks and search.

The main releases use Apache 2.0, allowing commercial use and modification. A separate variant trained on restricted-use data is non-commercial.

Anthropic let three outside research groups study aggregate Claude usage. In its account of the pilot, the groups designed studies covering roughly 250,000 Claude.ai or Claude Code conversations from April and May.

Anthropic performed privacy-preserving data collection, and the researchers analyzed the outputs independently. Aggregate project data was released.

Researchers accessed outputs through Anthropic's analysis system rather than unrestricted raw conversations. Contractual review covered privacy, confidential information, misuse-related information and research accuracy. One study found consequential tasks in more than half of conversations and human direction in nearly three-quarters.

Google DeepMind is piloting evaluations that keep test questions confidential. The pilot places confidential benchmarks and a proprietary model inside a cryptographically protected environment. It tests a Gemini Flash Lite model with the Singapore AI Safety Institute, OpenMined, AVERI and MLCommons.

Google's stated goal is to prevent evaluation questions from becoming available for later model optimization. A model that has already seen the test questions can earn a score that no longer means what evaluators intended.

Top Links#

• Geographic predictions from a written question (research.google): Google's experimental system handles data discovery, cleaning, model training, evaluation and reports, with company-reported workflow times falling from weeks to minutes on selected geographic tasks.
• More control over generated video (blog.google): Google's new extension controls use up to 10 seconds of preceding footage and add 10-second segments, reaching 40 seconds total.
• A large model, some pelicans (simonwillison.net): An exploratory Qwen3.8-Flash-Next test runs 72.5 GB and 78.9 GB compressed versions of the model on an NVIDIA DGX Spark computer and shares the resulting pelican illustrations.