Jensen Huang, CEO, Nvidia
Nvidia
Nvidia chief government Jensen Huang used his GTC Taipei keynote on June 1 to declare that the age of autonomous brokers has arrived, and he backed the declare with new {hardware} throughout the info heart, the desktop and the bodily world. Huang introduced that the Vera Rubin platform, Nvidia’s subsequent information heart system, has reached full production, and he framed practically each product the corporate revealed in Taipei round software program brokers that observe, cause, plan and act with little human enter.
The keynote on the Taipei Music Heart mirrored a change in how Nvidia describes itself. Huang stated the corporate now sells AI infrastructure moderately than chips alone, and he argued that compute has develop into a direct income for the companies that purchase it. He pointed to coding platforms the place developer commits have practically tripled within the first months of 2026 as proof that brokers are already doing helpful work. For expertise leaders, the bulletins map the place information heart spending, enterprise software program and private computing are heading.
Vera Rubin Reaches Full Manufacturing
Vera Rubin is a five-rack system that Nvidia treats as one giant laptop for agentic workloads. The platform combines Vera Rubin NVL72 methods, the brand new Vera CPU, Groq 3 LPX inference trays, Spectrum-6 Ethernet racks and Vera BlueField-4 STX storage. Every NVL72 rack hyperlinks 36 Vera CPUs and 72 Rubin GPUs via the sixth-generation NVLink swap, with ConnectX-9 community playing cards and BlueField-4 information processing items dealing with site visitors and safety.
Nvidia says the rack delivers as much as 10 occasions greater inference efficiency per watt and 10 occasions decrease value per token than its prior technology, and that pairing it with Groq 3 LPX raises throughput per watt as a lot as 35 occasions for trillion-parameter fashions. The cable-free, hose-free and fanless tray design cuts meeting from two hours to 5 minutes per compute tray, and a completely liquid-cooled design runs at 45 levels Celsius so it matches current information facilities.
Huang stated the Vera Rubin provide chain is twice the dimensions of the prior Grace Blackwell effort, spanning 150 companions in Taiwan and greater than 350 factories throughout 30 international locations. Manufacturing shipments start this fall.
A CPU And Software program Stack Constructed For Brokers
The Vera CPU is Nvidia’s first standalone information heart processor, with 88 cores and an on-chip material the corporate designed for brokers moderately than human operators. Huang argued that billions of brokers will run repeatedly and demand far decrease latency than folks do, making a processor market that didn’t exist earlier than.
Nvidia additionally moved its Spectrum-X Ethernet Photonics networking into manufacturing, describing it as the primary 200 gigabit per second Ethernet swap with co-packaged optics and naming CoreWeave, Lambda and Oracle Cloud Infrastructure amongst early adopters. A separate framework known as DSX helps operators design and run AI factories, and Nvidia stated one configuration matches 40% extra GPUs inside the similar energy price range.
On the software program facet, Nvidia launched an Agent Toolkit that bundles fashions, an agent harness and an enterprise runtime, alongside a safe runtime known as OpenShell that isolates every agent and enforces coverage. The corporate launched Nemotron 3 Extremely, a 550-billion-parameter combination of consultants mannequin it says runs inference 5 occasions sooner and prices about 30% lower than main open alternate options. Verified Nvidia agent expertise are actually accessible contained in the Claude Code plug-in market and the Hermes Abilities Hub.
Nvidia Enters The PC Market
Huang introduced RTX Spark, a chip constructed with MediaTek that brings 1 petaflop of AI efficiency to Home windows laptops and compact desktops. It pairs a Blackwell RTX graphics processor that has 6,144 CUDA cores with a 20-core Grace CPU, and Nvidia positioned it as the muse for private computer systems that run brokers domestically moderately than calling a cloud server.
The corporate unveiled a Home windows lineup that features a laptop computer, an always-on desktop agent field and a deskside DGX Station for Windows able to operating frontier fashions as much as 1 trillion parameters on the desk. Companions together with Asus, Dell, Gigabyte, HP, MSI and Supermicro start transport DGX Station methods this month. Adobe is rebuilding Photoshop and Premiere for RTX Spark, with variations Nvidia says run twice as quick and work with brokers.
Bodily AI Strikes To The Foreground
Nvidia prolonged its agent message into robotics and autos. It launched Cosmos 3, an open world basis mannequin constructed on a mixture-of-transformers design that learns from teleoperation, simulation and re-projected video so robots can cause about their environment. The corporate stated its Drive Hyperion automobile platform now reaches providers representing about 97% of the world’s mobility market, and it launched Alpamayo 2 Tremendous, an open reasoning mannequin for self-driving analysis paired with a reinforcement studying coach and a situation generator. For robotics labs, Nvidia launched an open humanoid robot reference design constructed on its Jetson Thor module. A set of media instruments rounded out the day, together with an artificial video detector Nvidia says flags AI-generated footage with about 92% accuracy in 22 milliseconds.
The Limitations
Many of the efficiency numbers come from Nvidia and haven’t been independently examined, and Vera Rubin won’t ship in quantity till the autumn, so patrons can’t but validate the cost-per-token claims in their very own workloads. The brand new Home windows machines and RTX Spark methods additionally arrive later this yr, which leaves their software program ecosystem and agent tooling unproven exterior managed demos. Enterprise agent runtimes increase governance and safety questions that merchandise akin to OpenShell handle in precept however haven’t confronted at manufacturing scale. Competitors is intensifying as nicely, with AMD pushing its Intuition accelerators and cloud suppliers increasing customized silicon akin to AWS Trainium, Google Ironwood and Microsoft Maia.
For expertise determination makers, the keynote sharpened a selection that can outline AI budgets. Huang argued that efficiency per watt and the runtime that surrounds the mannequin now matter as a lot because the chip itself, which implies structure choices revamped the subsequent yr will form each functionality and value lengthy after the {hardware} lands.
