Nvidia's third-generation AI platform promises 10 times the agent throughput of its predecessor, turning five racks into a single supercomputer for the agentic era.
Back
Nvidia Vera Rubin enters full production with 10x agentic AI throughput
Nvidia's third-generation AI platform promises 10 times the agent throughput of its predecessor, turning five racks into a single supercomputer for the agentic era.
Nvidia's Vera Rubin platform is ramping into full production, with more than 150 Taiwanese supply chain partners manufacturing systems that deliver 10 times the agentic AI throughput of the Grace Blackwell generation, the company said Sunday.
"Agentic AI is a new kind of workload. One prompt can launch a thousand-step journey of reasoning, retrieval, tool use and response generation," Jensen Huang, founder and chief executive officer of Nvidia, said. "Vera Rubin was built for this moment — an AI factory engine that delivers intelligence at scale."
The platform integrates five purpose-built racks — Vera Rubin NVL72 systems, Vera CPUs, Groq 3 LPX GPUs, BlueField-4 STX storage and Spectrum-6 SPX Ethernet — operating as one unified AI supercomputer. It introduces Spectrum-X Ethernet Photonics, the industry's first co-packaged-optics switches with 200-gigabit-per-second SerDes, now in production. Nvidia claims the photonics-based networking delivers 5 times better power efficiency and 5 times longer AI uptime versus traditional transceivers, a critical advantage as AI factories scale toward million-GPU clusters.
Production shipments begin this fall, with Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro among the system builders already manufacturing Vera Rubin-based systems. The ramp comes as Nvidia forecasts $1 trillion in cumulative sales from its Blackwell and Rubin chip generations through 2027, and as Huang told a Taipei audience last week the company "will be worth even more in three to five years." Nvidia shares trade at 32 times trailing earnings, roughly half their five-year median of 61.
The Vera Rubin ramp marks the third generation of Nvidia's MGX rack-scale design, an open-source blueprint that allows hundreds of partners across 350 factories in 30 countries to manufacture the systems. Each Vera Rubin system contains close to 2 million components and involves roughly 150 Taiwanese supply chain partners spanning advanced packaging, power supplies and thermal management.
The platform was purpose-built for agentic AI workflows, which differ fundamentally from traditional model training. Where a single prompt in a conventional large language model might trigger one inference pass, agentic systems can spawn hundreds or thousands of sequential reasoning steps, each requiring retrieval, tool calls and response generation. Vera Rubin's 10x throughput gain addresses that computational explosion directly.
Supply Chain Deepens in Taiwan
Nvidia's commitment to Taiwan is deepening alongside the Vera Rubin ramp. Huang said last week the company plans to spend roughly $150 billion annually in the country, up from $10 billion to $15 billion four to five years ago. A new corporate headquarters in Taipei, set to break ground in mid-2027 and employ 4,000 people, reinforces the island's role as what Huang called "the epicenter of the AI revolution."
The concentration carries risk. Taiwan Semiconductor Manufacturing Co., Nvidia's sole foundry for advanced chips, operates in a region that China claims as its own territory. The U.S. has imposed export controls limiting China's access to Nvidia's most advanced AI processors, though Huang said the $200 billion CPU market opportunity he outlined on the company's earnings call includes China. The U.S. has licensed H200 chip sales to about 10 Chinese firms, but Chinese officials have not yet approved those shipments.
Security at Rack Scale
Vera Rubin also introduces full-stack confidential computing at the pod level, encrypting data across high-speed interconnects and providing hardware-level attestation to ensure systems are tamper-proof. The Nvidia DOCA software platform enforces zero-trust policies, runtime threat detection and multi-tenant isolation at speeds of up to 800 gigabits per second without taxing host CPU resources. Cloud providers including CoreWeave, Microsoft Azure, IBM Cloud and Lambda are adopting the security framework.
For investors, the Vera Rubin ramp represents the next demand cycle for Nvidia's sprawling supply chain. The company's data-center revenue reached $75.2 billion in its most recent fiscal year, up 85% from a year earlier, and management guided for roughly $91 billion in the current quarter. Bank of America carries a $320 price target on Nvidia, implying about 45% upside from current levels, while Morgan Stanley's Joseph Moore called the current valuation "a surprisingly good entry point" in a recent client note.
This article is for informational purposes only and does not constitute investment advice.
