3 min read
NVIDIA unveiled an end-to-end AI ecosystem with AI GPU capabilities, heterogeneous silicon integration, and embodied intelligence in humanoid robotics in their Computex 2025 today.
DGX Cloud Lepton is the magnum opus in the form of an AI GPU computing marketplace, offering certified third-party systems and Blackwell GPU clusters with metered, on-demand access.
DGX Cloud Lepton: Democratizing GPU Compute
NVIDIA’s DGX Cloud Lepton marketplace offers developers worldwide instant, metered access to Blackwell GPU instances hosted by NVIDIA and its partners. Third‑party systems certified under the DGX Ecosystem program appear alongside NVIDIA’s own clusters, ensuring compatibility and performance guarantees.
By standardizing APIs and billing through NVIDIA’s NGC registry, DGX Cloud Lepton removes traditional procurement hurdles and accelerates time‑to‑market for AI projects.
NVLink Fusion: Heterogeneous AI on a Unified Fabric
NVLink Fusion extends NVIDIA’s NVLink interconnect to include partner silicon, allowing integration of NVIDIA GPUs, Grace CPUs, and third‑party AI accelerators into a single coherent memory space. Partners such as MediaTek, Marvell, Alchip, Astera Labs, Synopsys, and Cadence have committed to building semi‑custom chips compatible with NVLink Fusion.
While Fujitsu and Qualcomm plan factory‑scale deployments coupling their CPUs with NVIDIA GPUs. This open approach meets customer demand for flexible AI infrastructure and breaks from NVIDIA’s historical GPU‑only stance by embracing mixed‑vendor ecosystems.
Physical AI: From DGX Spark to Isaac GR00T
NVIDIA introduced Isaac GR00T (Generalist Robot 00 Technology) as an open research platform featuring foundation models and data pipelines that accelerate humanoid robotics R&D. The GR00T N1 model, a 2 billion‑parameter multimodal transformer, ingests language, vision, and state inputs to output control policies.
These policies generalize manipulation skills across different robot embodiments without retraining. GR00T’s synthetic data pipelines generate massive motion datasets via Omniverse and Cosmos, enabling 780,000 synthetic trajectories (equivalent to nine months of human demonstrations) in just 11 hours, and boosting performance by 40 percent when combined with real‑world data.
Isaac GR00T runs end‑to‑end on Jetson AGX Thor, NVIDIA’s integrated robotics compute platform, while DGX Cloud Lepton and Omniverse provide cloud‑based training and simulation.
The GR00T ecosystem includes GR00T‑Mimic for teleoperation data, GR00T‑Gen for environment randomization, and GR00T‑Dexterity for pixel‑to‑action workflows. NVIDIA has open‑sourced the GR00T N1 weights and blueprints, empowering partners like Boston Dynamics, Apptronik, and Unitree to accelerate their robotics programs.
Strategic Investments and Local Engagement
NVIDIA committed significant investments in Taiwan, including a new AI research campus in Beitou. The chipset developer behemoth plans to have a joint development of a quantum‑AI supercomputer featuring 10,000 Blackwell chips, reinforcing Taiwan’s role as a pivotal R&D and manufacturing hub amid evolving export controls.
According to experts, NVLink Fusion and DGX Cloud Lepton will grow Taiwan’s AI ecosystem, so helping TSMC and Foxconn’s Big Innovation Company among other partners.
With all these advancements via Computex, NVIDIA is steering toward a unified AI platform that spans cloud, on‑premises, and embodied intelligence.
