NVIDIA Corporation unveiled a series of advancements in artificial intelligence at Computex 2025 in Taipei today, featuring an AI marketplace that connects global developers with GPU computing and humanoid robotics.

The company also introduced NVIDIA NVLink, a novel silicon solution enabling partners to construct semi-customized AI infrastructures using NVLink - the company's existing high-speed interconnect technology.

DGX Cloud Lepton: Bridging Developers and Worldwide Computing

Lepton is an AI platform offering a compute marketplace that links global developers with tens of thousands of GPUs from NVIDIA's cloud partner network, supporting today's agent and physical AI applications.

NVIDIA's cloud partners include prominent firms such as CoreWeave Inc., Crusoe Energy Systems LLC, Firmus Technologies Inc., Foxconn GMI Cloud Solutions Ltd., Lambda Labs Inc., Nscale GmbH, SoftBank Corp., and Yotta Data Services Private Ltd.

With Lepton, developers can access GPU computing power on-demand or long-term based on their operational needs in specific regions.

"NVIDIA DGX Cloud Lepton connects our worldwide network of GPU cloud providers with AI developers," said Jensen Huang, founder and CEO of NVIDIA. "Together with our NCPs, we are building an AI factory on a planetary scale."

In the backend, Lepton streamlines the procurement and management of computing for developers while allowing partners to offer GPU capacity through a unified experience. NVIDIA states this enables developers and enterprises to focus on experimenting and building AI capabilities while flexibly receiving the GPU workloads they need to test, train, and deploy AI models when required.

Delivering Humanoid Robotics Compute with GR00T

NVIDIA announced today the first update to its humanoid robotics foundation model, Isaac GR00T N1.5.

Physical AI combines AI with real-world physical interactions, enabling machines to perceive, understand, and act upon their environment. Foundation models like GR00T serve as the AI brains of robots, empowering them to perceive the world and reason, thus enabling them to perform tasks like humans, such as recognizing objects, navigating spaces, and picking up items.

GR00T N1.5 has been upgraded over its predecessor to better adapt to new environments and workspace configurations. NVIDIA claims it significantly enhances the model's success rate in common material handling and manufacturing tasks such as sorting and placing items. It can be deployed on the company's Jetson Thor robotic computer, scheduled for release later this year.

Early adopters of the GR00T N model include companies like AeiRobot, Foxlink Lightweel, and NEURa Robotics.

Training robots requires vast amounts of data, most of which must come from human actors who can only provide limited datasets. These datasets, known as "trajectories," show how robots should move to grasp, handle, and move objects to perform tasks in different environments.

Human actors can only generate data for limited periods each day. Thus, NVIDIA created GR00T-Dreams, a method for generating videos of robots performing new tasks in new environments. GR00T-Dreams can use a single image as input to create videos for robots to learn how to execute new tasks.

In AI terminology, this is a method for generating synthetic data for robot AI training, providing abundant additional digestible data. GR00T-Dreams complements Isaac GR00T-Mimic blueprint, a tool for generating additional synthetic motion data from a few human demonstrations. Mimic uses Omniverse and Cosmos to augment existing data, whereas GR00T-Dreams generates entirely new data.

NVLink Fusion Paves the Way for Larger AI Factories

NVIDIA announced today NVLink Fusion, calling it industry-validated custom CPUs and accelerators at scale.

"A tectonic shift is underway: for the first time in decades, data centers must fundamentally be rearchitected - AI is being integrated into every computing platform," said Huang.

Fusion aims to provide cloud providers with a way to scale to millions of GPUs within NVIDIA's rack-level systems and end-to-end networking platform using any application-specific integrated circuits, delivering throughput up to 800 gigabits per second.

The company stated that with NVLink Fusion, CPUs from Fujitsu Ltd. and Qualcomm Technologies, Inc. can also integrate with NVIDIA GPUs to build high-performance NVIDIA AI factories.

NVIDIA's AI chip manufacturer partners, including MediaTek Inc., Marvell Technology, Alchip Technologies Ltd., Astera Labs Inc., Synopsys Inc., and Cadence Design Systems Inc., are creating custom AI computing solutions deployable on NVIDIA NVLink.

"HPC and AI workload demands are unique and rapidly evolving; hyperscalers building