Scientific computing got its own hardware. NVIDIA announced a rack-scale Vera CPU platform for high-performance computing, pairing what it calls its first fully custom data center CPU with its HPC networking, and added Rubin HPC, a GPU in the Vera Rubin family aimed at simulation, engineering and analytics work. It launched Nemotron for Science for building research agents, worked with healthcare and life sciences groups on protein design, drug discovery and precision patient care, and expanded its CUDA-Q quantum platform with CUDA-Q Logical, alongside a record simulation of a fault-tolerant quantum processor on the JUPITER supercomputer at the Jülich Supercomputing Centre.
Several of the quarter's deals were with large customers and partners. Meta plans a wider rollout of Vera Rubin, Vera CPUs, Spectrum-X networking and BlueField-4 for its AI agent Muse and related models, work that NVIDIA paired in October with plans to put $3 billion into Meta's new share sale. NVIDIA also unveiled DSX POD, a blueprint for whole Vera Rubin AI factories that partners assemble inside modular containers, integrated and validated before they ship, reported European adoption of its DSX technology to cut power needs and nearly eliminate cooling water use, and agreed to work with Australia's data center sector on a planned buildout of up to 2 gigawatts by 2027.
The company has agreed to acquire Hugging Face, which it said would scale that platform and make intelligence more open, capable and accessible; the release describes an agreement, not a completed purchase. NVIDIA said it won every MLPerf Training and Endpoints benchmark, and that Vera Rubin NVL72 led on inference in its first MLPerf Inference preview. It also released an Open Agent Safety Platform and an OpenShell Reference Architecture, in which BlueField-4 processors run NVIDIA Sentry to watch agent behavior and enforce security rules in hardware.
To keep up with demand, NVIDIA set out plans with Amkor Technology, IBIDEN, King Yuan Electronics, Samsung Electro-Mechanics, Siliconware Precision Industries and Unimicron Technology Corporation to expand manufacturing capacity. It deepened its MediaTek partnership, taking its rack-scale platform to custom XPUs through NVLink Fusion, planned a new optical-module format called USFP with Cisco, teamed up with Palantir on sovereign AI for critical supply chains and partnered with Firebird to scale total potential AI infrastructure capacity.
On the edge side, RTX Spark laptops went on sale from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, and DLSS 5 arrived in NBA 2K27 and Hogwarts Legacy, with Where Winds Meet to follow. A 64GB DGX Spark desktop model joined the local AI range, with faster performance for models from NVIDIA Nemotron, Black Forest Labs, Liquid AI and Qwen and more support for Hermes, OpenClaw and Perplexity.
In vehicles and robotics, NVIDIA added Alpamayo tools and the Halos SAFE framework, widened its work with Mercedes-Benz and Volvo Cars, and said Bolt with Lucid, Volkswagen Commercial Vehicles and Einride are among European groups showing Hyperion uses. It also updated its Isaac robotics platform and Cosmos world models, while Agile Robots unveiled a reference humanoid on Jetson Thor and Isaac GR00T, and T-Mobile completed its first AI-RAN call on a commercial network as part of work with Nokia.
Shareholders received $28.1 billion through buybacks and dividends in the quarter. In September, the board added $150 billion to the share repurchase authorization, with no expiry date, leaving about $227.3 billion available when the quarter closed, against roughly $99.0 billion three months earlier. The next quarterly dividend of $0.25 per share will be paid on December 23, 2026, to holders of record on December 2. Operating activities generated $48.3 billion of cash in the quarter, while spending on property, equipment and intangible assets rose to $4.4 billion from $1.6 billion a year earlier.
The fourth-quarter outlook covers an extra week. Besides revenue of $130.0 billion, plus or minus 3%, NVIDIA expects gross margins of 71.0% to 72.0% on both GAAP and non-GAAP bases, below the third quarter's level, with operating expenses of about $11.0 billion under GAAP and $10.7 billion adjusted. As in August, the forecast includes no Data Center compute revenue from China, so any such sales could come on top of the forecast, and the company expects full-year tax rates of 16.0% to 18.0%.
Kress discussed the quarter with analysts and investors on a call at 2 p.m. Pacific time on Tuesday, and her written commentary sits on NVIDIA's investor relations site, where a recording of the webcast stays available until the fourth-quarter call. Over the first nine months of fiscal 2027, NVIDIA's revenue reached $293.7 billion, almost double the $147.8 billion of the same period a year earlier, and net income rose to $183.3 billion from $77.1 billion.