One prompt can launch a thousand-step journey of reasoning, retrieval, tool use and response generation.
At a moment when artificial intelligence is evolving from passive text generation toward autonomous, multi-step reasoning, NVIDIA has moved its Vera Rubin platform into full global production — a third-generation architecture designed not merely to answer questions, but to power machines that think in sequences, delegate tasks, and act at scale. Announced at GTC Taipei by CEO Jensen Huang, the platform represents a coordinated bet by hundreds of manufacturers across thirty countries that the next era of computing will be measured not in model size, but in the depth and speed of agentic thought. The infrastructure being assembled — million-GPU factories secured by hardware-level trust and connected by light itself — suggests that the industry is preparing for a kind of intelligence that has not yet fully arrived, but whose arrival is being treated as inevitable.
- The shift from language models to autonomous agents demands a fundamentally different kind of hardware — one that can sustain thousands of reasoning steps per prompt without collapsing under the weight of its own complexity.
- NVIDIA's answer is Vera Rubin, a five-rack supercomputer-in-a-box delivering ten times the agent throughput of its predecessor, already being assembled by 350 factories across 30 countries in one of the most distributed manufacturing mobilizations in AI history.
- A breakthrough in networking — co-packaged optical switches embedded directly into silicon — promises five times better power efficiency and makes the dream of million-GPU AI factories physically and economically viable for the first time.
- Security has become a structural concern: with cloud providers like Microsoft Azure, Oracle, and CoreWeave running multiple customers' sensitive workloads on shared infrastructure, Vera Rubin bakes encryption, zero-trust isolation, and hardware attestation into every rack at 800 Gb/s.
- The DSX software platform aims to standardize the entire operational stack so that deploying an AI factory becomes a repeatable, reliable process rather than a bespoke engineering challenge.
- Production shipments begin fall 2026 — and whether the market will consume infrastructure at this scale remains the open question hanging over an otherwise confident announcement.
NVIDIA announced this week that its Vera Rubin platform is entering full production, marking a significant turn in the company's vision for artificial intelligence — away from models that respond to prompts and toward autonomous agents capable of chaining thousands of reasoning steps together. The announcement came at GTC Taipei, where Jensen Huang framed Vera Rubin as the engine of what NVIDIA calls agentic AI factories.
Vera Rubin is the third generation of NVIDIA's rack-scale MGX systems, and it arrives as a unified supercomputer spanning five purpose-built racks. It integrates Vera Rubin NVL72 compute nodes, Vera CPUs, Groq 3 LPX processors, BlueField-4 storage controllers, and Spectrum-6 Ethernet into a single coherent machine — delivering ten times the agent throughput of the previous Grace Blackwell generation.
The manufacturing mobilization behind the platform is striking in its breadth. One hundred fifty partners in Taiwan and 350 factories across 30 countries are already building Vera Rubin systems. Dell, HPE, Lenovo, and Supermicro are in full production alongside dozens of others, reflecting NVIDIA's strategy of releasing Vera Rubin as an open reference design rather than a proprietary product.
At the heart of the platform's networking layer is Spectrum-X Ethernet Photonics — the first production deployment of co-packaged optics at scale, embedding optical components directly into switch silicon. The result is five times better power efficiency, five times longer uptime, and a 30 percent reduction in deployment time, while supporting the data throughput needed to connect AI factories with a million GPUs or more.
Security is treated as a first-class architectural concern. Confidential Computing encrypts data across high-speed interconnects, while BlueField-4 data processing units enforce multi-tenant isolation and zero-trust policies at 800 Gb/s — without taxing the main CPU. Cloud providers including CoreWeave, Oracle, Microsoft Azure, and IBM Cloud are among the first to adopt these protections, reflecting the reality that shared AI infrastructure increasingly carries sensitive, regulated, and proprietary workloads.
NVIDIA is also releasing DSX, a software platform offering reference designs, simulation tools, and operational guidance to standardize the full deployment stack. Major manufacturers are already adopting it to accelerate their Vera Rubin rollouts.
Shipments are scheduled to begin in fall 2026. The scale of the infrastructure being assembled — and the confidence of the partners building it — suggests the industry has already decided that agentic AI is coming. Whether demand will match the ambition of the supply being prepared remains the one question the announcement leaves open.
NVIDIA announced this week that its Vera Rubin platform is moving into full production, a milestone that signals the company's bet on a new kind of artificial intelligence workload — one where a single prompt can trigger thousands of reasoning steps, tool calls, and responses in sequence. The announcement came at NVIDIA's GTC conference in Taipei, where CEO Jensen Huang positioned Vera Rubin as the engine for what the company calls agentic AI factories — massive computational facilities designed to run autonomous agents at scale.
Vera Rubin is the third generation of NVIDIA's MGX rack-scale systems, and it represents a significant leap in capability. The platform bundles together five purpose-built racks that operate as a single supercomputer, integrating NVIDIA's Vera Rubin NVL72 systems, Vera CPUs, Groq 3 LPX processors, BlueField-4 storage controllers, and Spectrum-6 Ethernet infrastructure into one unified machine. The result is a system that delivers ten times the agent throughput of NVIDIA's previous-generation Grace Blackwell platform — a substantial performance jump designed specifically for the demands of agentic workloads rather than traditional large language model inference.
The manufacturing ramp is already underway. One hundred fifty partners in Taiwan alone, along with 350 factories across 30 countries, are now building Vera Rubin systems at scale. The list of manufacturers reads like a who's who of the server and infrastructure world: Dell Technologies, HPE, Lenovo, and Supermicro are all in full production, alongside dozens of other vendors including ASUS, Foxconn, GIGABYTE, Pegatron, Quanta Cloud Technology, and Wistron. This distributed manufacturing footprint reflects NVIDIA's strategy of making Vera Rubin an open-source design that can be built and customized by a broad ecosystem rather than a proprietary black box.
A critical innovation in Vera Rubin is its networking layer. The platform introduces Spectrum-X Ethernet Photonics, the first production deployment of co-packaged optics switches at scale — a technology that embeds optical components directly into the switch silicon rather than relying on separate transceiver modules. This approach delivers five times better power efficiency than traditional networking, five times longer uptime for AI workloads, and cuts deployment time by 30 percent. The switches support 200 gigabits per second of data transfer per lane and are designed to enable AI factories with a million GPUs or more, a scale that would have been impractical with older networking technology.
Security is woven throughout the platform's design. Vera Rubin incorporates NVIDIA's full-stack Confidential Computing architecture, encrypting data across high-speed interconnects and providing hardware-level attestation to ensure systems cannot be tampered with. The BlueField-4 data processing units embedded in each rack enforce multi-tenant isolation, zero-trust security policies, and runtime threat detection at speeds up to 800 gigabits per second without burdening the main CPU. This matters because AI factories increasingly process proprietary data, regulated content, and mission-critical models in shared or cloud environments where multiple customers' workloads run on the same infrastructure. Cloud providers including CoreWeave, Lambda, Oracle Cloud Infrastructure, Microsoft Azure, and IBM Cloud are among the first to adopt these security features.
NVIDIA is also releasing DSX, a software platform that provides reference designs, simulation tools, infrastructure software, and operational guidance for building and running Vera Rubin AI factories. The goal is to standardize the entire stack — from silicon to lifecycle management — so that customers can deploy factories faster and operate them more reliably at scale. Dell, HPE, Lenovo, Supermicro, and several Asian manufacturers are already adopting DSX to accelerate their Vera Rubin deployments.
Production shipments are scheduled to begin in the fall of 2026. The announcement positions Vera Rubin as the infrastructure foundation for the next phase of AI development, one where autonomous agents handle complex, multi-step tasks rather than simply generating text in response to a prompt. Whether the market will demand this capability at the scale NVIDIA is building for remains to be seen, but the company's manufacturing partners and cloud providers are clearly betting that it will.
Notable Quotes
Agentic AI is a new kind of workload. One prompt can launch a thousand-step journey of reasoning, retrieval, tool use and response generation. Vera Rubin was built for this moment — an AI factory engine that delivers intelligence at scale.— Jensen Huang, NVIDIA founder and CEO
The Hearth Conversation Another angle on the story
What exactly is an agentic AI workload, and why does it need different hardware than what came before?
An agentic workload is fundamentally different from traditional inference. Instead of a user sending a prompt and getting a response, an agent receives a task and then autonomously decides what to do — it might retrieve information, call a tool, reason about the result, and then take another action. One prompt can trigger hundreds or thousands of steps. That's computationally very different from just running a single forward pass through a model.
So Vera Rubin is optimized for that kind of work. But what does ten times the throughput actually mean in practical terms?
It means you can run ten times as many agents in parallel on the same hardware, or you can run the same number of agents ten times faster. For a cloud provider or AI lab, that translates directly to either serving more customers or delivering results faster. At scale, that's the difference between a factory that's profitable and one that isn't.
The networking piece — co-packaged optics — sounds like a big deal. Why does that matter for AI factories?
Traditional networking uses separate optical transceivers plugged into switch ports. That adds latency, power consumption, and complexity. Co-packaged optics puts the optics directly on the switch chip. For a million-GPU factory, you're moving staggering amounts of data between GPUs. Even small improvements in power efficiency and latency compound across the entire system. Five times better power efficiency means your electricity bill drops dramatically.
And the security architecture — is that a response to real threats, or is it more about compliance and customer comfort?
Both. If you're running proprietary models or processing regulated data in a shared cloud environment, you need to know that another tenant's workload cannot access your data or models. Confidential Computing with hardware attestation provides that guarantee at a level that software alone cannot. It's becoming table stakes for enterprise AI infrastructure.
What happens if the market doesn't actually need million-GPU factories?
Then NVIDIA has built a platform that's far more capable than demand requires — which is not uncommon in infrastructure. But the company is betting that agentic AI will drive that demand. If it's right, Vera Rubin will be the standard. If it's wrong, the platform will still be useful, just not at the scale NVIDIA is designing for.