Inside Nvidia’s Vera Rubin: Seven-Chip Supercomputers for the Age of Agentic AI

Nvidia’s latest data center platform, Vera Rubin, is not a single chip but a tightly integrated seven-silicon stack, already in full production and backed by OpenAI, Anthropic, Meta, Mistral AI and every major cloud provider. For AI infrastructure engineers and cloud architects, it represents Nvidia’s bid to redefine how compute, memory, networking and storage are organized for what the company calls the age of “agentic AI.”

At GTC, CEO Jensen Huang framed Vera Rubin as “a generational leap” and the start of “the greatest infrastructure buildout in history.” The company claims up to 10x higher inference throughput per watt and one-tenth the cost per token versus its still-new Blackwell systems, and it is pairing the hardware with a broad software and ecosystem story that stretches from open models to digital-twin blueprints for “AI factories.”

The seven-chip Vera Rubin stack and why it matters

Vera Rubin is built from seven distinct chips that Nvidia positions as a single, rack-scale supercomputing platform:

• Vera CPU
• Rubin GPU
• NVLink 6 Switch
• ConnectX-9 SuperNIC
• BlueField-4 DPU
• Spectrum-6 Ethernet switch
• Groq 3 LPU (inference accelerator)

These components are not presented as standalone products but as building blocks of five “interlocking” rack systems that together behave as one logical AI supercomputer. AWS, Google Cloud, Microsoft Azure and Oracle Cloud will offer the platform, and more than 80 manufacturing partners are building systems around it.

Nvidia says Vera Rubin delivers up to 10x better inference throughput per watt and one-tenth the cost per token compared with Blackwell-based systems. Those numbers are not yet backed by independent benchmarks, and details of test conditions are not public in this announcement, but the directional claim is clear: Nvidia is attempting to reset the efficiency bar for large-scale inference, particularly for long-context, high-traffic workloads.

The endorsement list is also notable from an infrastructure-planning perspective. Anthropic’s Dario Amodei says the platform gives the company “the compute, networking and system design” it needs to keep scaling while pursuing safety and reliability; OpenAI’s Sam Altman says Vera Rubin will help it run “more powerful models and agents at massive scale” for “hundreds of millions of people.” For engineers, that signals that the requirements of leading model labs—latency, scale, and safety controls—have been front-loaded into the design.

Breaking down the rack-scale design

Nvidia organizes Vera Rubin into five main rack types, each optimized for a segment of the end-to-end agentic AI workload:

NVL72 rack (training and heavy reasoning)
This flagship rack integrates 72 Rubin GPUs and 36 Vera CPUs, connected by NVLink 6. Nvidia claims it can train large mixture-of-experts (MoE) models with one-quarter the GPUs required on Blackwell. If validated in production, that would shift total cost of ownership (TCO) curves for frontier-scale training and complex inference, especially for organizations exploring MoE architectures to trade off performance and cost.

Vera CPU rack (agent execution at scale)
Packed with 256 liquid-cooled Vera CPUs, this rack is pitched as the first processor specifically designed for agentic AI and reinforcement learning. Each chip has 88 custom Olympus cores and LPDDR5X memory delivering 1.2 TB/s of bandwidth, at roughly half the power of conventional server CPUs, according to Nvidia. A single rack can support more than 22,500 concurrent CPU environments—essentially, sandboxes where agents execute code, run tools, and validate results. For architects, this reshapes the assumed balance between CPU and GPU: CPU becomes an intentional, high-density substrate for tool use and environment simulation rather than just “glue” around accelerators.

Groq 3 LPX rack (low-latency inference)
This rack houses 256 Groq 3 inference processors, each with 128 GB of on-chip SRAM, targeting trillion-parameter models with million-token context windows. The emphasis is ultra-low latency on very large contexts—a key requirement for conversational and agent workloads that maintain long histories and reason over extended sessions.

BlueField-4 STX rack (context memory)
Nvidia positions this as “context memory” for agentic systems—a storage tier optimized for high-speed access to large key-value caches. As context windows stretch into the millions of tokens and agents accumulate extensive intermediate state, this kind of storage becomes a first-order design parameter, not merely a backing store.

Spectrum-6 SPX Ethernet rack (fabric)
The networking layer uses co-packaged optics that Nvidia says offer 5x better optical power efficiency than traditional transceivers. For data center designers trying to stay within power envelopes while scaling bandwidth, that claim—again, still Nvidia’s—signals an attempt to keep network power from becoming the limiting factor as cluster sizes and traffic grow.

From chatbots to agentic AI: Nvidia’s workload thesis

The organizing idea behind Vera Rubin is Nvidia’s view that the industry is moving from single-turn chatbots to “agentic AI”: systems that can reason autonomously for hours or days, call tools, write and run code, and continually improve. Huang’s definition encompasses workloads like drug discovery pipelines, large-scale software debugging and multi-step orchestration tasks.

For infrastructure teams, this is less about branding and more about resource patterns. A prompt–response chatbot might occupy GPUs for milliseconds per query and do minimal CPU work. An agent orchestrating a full workflow can run continuously, saturating CPU for tool execution, GPUs for reasoning, and storage for long-lived context. The implication is that designing only for peak GPU FLOPS is no longer sufficient; the platform must balance persistent compute, large working sets, and dense networking.

Vera Rubin’s seven-chip design reflects this thesis: dedicated CPU density for agent execution, GPU density for training and reasoning, a specialized inference layer, storage explicitly tuned for context and KV caches, and a power-efficient network fabric to connect it all.

Agent Toolkit, NemoClaw and Dynamo: rethinking the software stack

Nvidia is coupling its hardware story with a re-architected software stack for agentic workloads.

Agent Toolkit and OpenShell
The Nvidia Agent Toolkit includes OpenShell, an open-source runtime that enforces security and privacy guardrails for autonomous agents. Enterprises including Adobe, Atlassian, Box, Cadence, Cisco, CrowdStrike, Dassault Systèmes, IQVIA, Red Hat, Salesforce, SAP, ServiceNow, Siemens and Synopsys are integrating these capabilities into their platforms.

From an operations point of view, this suggests Nvidia is trying to make secure, always-on agents a first-class deployment target, not a bespoke, per-company integration. OpenShell’s emphasis on guardrails hints at a future where multi-tenant agent environments and tool access policies need the same rigor as traditional service isolation.

NemoClaw and local-to-cloud symmetry
NemoClaw is an open-source stack that packages Nvidia’s Nemotron models with OpenShell into a single-command install for secure, always-on assistants, from RTX laptops to DGX Station. For architects, that offers a consistent runtime from development to high-end deskside to data center, reducing friction when scaling prototypes into production clusters.

Dynamo 1.0 as “OS for inference”
Nvidia describes Dynamo 1.0 as the first “operating system” for AI inference at factory scale: software that orchestrates GPU and memory resources across clusters. Already adopted by AWS, Azure, Google Cloud, Oracle, Cursor, Perplexity, PayPal and Pinterest, it reportedly boosted Blackwell inference performance up to 7x in Nvidia’s benchmarks.

While details are limited in this announcement, the intent is clear: inference is being treated as a schedulable, optimizable workload at cluster scale, not a collection of independent deployments. For teams running large fleets of models, a system like Dynamo could become central to utilization, SLO management and cost control—though third-party validation of its performance claims will be important.

Nemotron coalition: Nvidia’s open-model power play

On the model side, Nvidia introduced the Nemotron Coalition, a global collaboration of AI labs that will co-develop open frontier models on Nvidia’s DGX Cloud. Initial members include Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam and Thinking Machines Lab (led by former OpenAI executive Mira Murati). The first jointly developed model, from Mistral AI and Nvidia, will underpin the Nemotron 4 family.

Huang described open models as “the lifeblood of innovation and the engine of global participation in the AI revolution.” Strategically, this serves two purposes for infrastructure buyers:

• It grows a developer and research ecosystem around open models that run best on Nvidia hardware.
• It positions Nvidia as a platform provider rather than a direct competitor to AI labs, even as its own models gain capabilities.

Nvidia also expanded its own model portfolio: Nemotron 3 Ultra (frontier-level intelligence with 5x throughput efficiency on Blackwell, according to Nvidia), Nemotron 3 Omni (multimodal across audio, vision and language) and Nemotron 3 VoiceChat (real-time, simultaneous conversations). It previewed GR00T N2, a robotic foundation model that Nvidia says more than doubles success on new tasks in new environments relative to leading alternatives, ranking first on MolmoSpaces and RoboArena benchmarks.

For infrastructure teams, these models are less about lock-in to a single stack and more about defining target workloads: multimodal assistants, high-throughput inference and physical AI/robotics, all of which have demanding and distinct system profiles.

Real-world deployments: healthcare, autonomy and space

Nvidia coupled the Vera Rubin and software announcements with a wide set of domain-specific use cases that help clarify expected workload mixes.

Healthcare and pharma
Roche is deploying more than 3,500 Blackwell GPUs across hybrid and on-prem environments in the U.S. and Europe, focused on biological foundation models, drug discovery and digital twins of manufacturing sites, including its new GLP-1 facility in North Carolina. Nearly 90% of eligible Genentech small-molecule programs now incorporate AI, and the company reports one oncology molecule designed 25% faster and a backup candidate delivered in seven months instead of over two years. While these figures come from Roche and Nvidia, they illustrate the kind of long-running, simulation-heavy workloads Nvidia expects its platforms to serve.

Nvidia also introduced what it calls the first domain-specific physical AI platform for healthcare robotics, centered on Open-H—a 700+ hour surgical video dataset. CMR Surgical, Johnson & Johnson MedTech and Medtronic are among adopters, again pointing to high-fidelity, low-latency inference and control workloads.

Autonomous vehicles
BYD, Geely, Isuzu and Nissan are building Level 4–ready vehicles on Nvidia Drive Hyperion. Nvidia and Uber expanded their partnership with a plan to launch autonomous vehicles in 28 cities across four continents by 2028, beginning in Los Angeles and San Francisco in the first half of 2027.

Nvidia launched Alpamayo 1.5, a reasoning model for autonomous driving already downloaded by more than 100,000 automotive developers, and Halos OS, a safety architecture based on ASIL D-certified foundations. Together, these define a stack where perception, reasoning and safety logic all demand reliable, deterministic compute—an environment where the Vera Rubin-style integration of hardware and software is directly relevant.

Orbital computing
The Vera Rubin Space Module targets orbital inference, with Nvidia claiming up to 25x more AI compute than its H100 GPU in space applications. Aetherflux, Axiom Space, Kepler Communications, Planet Labs and Starcloud are building on the module. For infrastructure teams looking at edge and remote inference, this is a signal that Nvidia’s architecture vision extends beyond terrestrial data centers to constrained, latency-sensitive environments.

DGX Station and the deskside AI supercomputer

Beyond the data center, Nvidia quietly expanded its hardware footprint into the enterprise with the DGX Station, a deskside system built on the GB300 Grace Blackwell Ultra Desktop Superchip. The system delivers 748 GB of coherent memory and up to 20 petaflops of AI compute, and Nvidia says it can run open models of up to one trillion parameters from a desk.

Snowflake, Microsoft Research, Cornell, EPRI and Sungkyunkwan University are early users. The system supports air-gapped configurations for regulated industries, and applications developed on DGX Station are designed to move to Nvidia’s data center systems without rearchitecting.

Nvidia also updated DGX Spark, a more compact system that can be clustered up to four units into a “desktop data center” with linear performance scaling. Both DGX Station and DGX Spark ship preconfigured with NemoClaw and the Nvidia AI software stack, supporting models such as Nemotron 3, Google Gemma 3, Qwen3, DeepSeek V3.2 and Mistral Large 3.

For technical decision-makers, these systems offer a bridge between local experimentation and large-scale deployment, using a consistent software stack and model portfolio—important when trying to avoid divergent environments between labs and production.

AI factories and DSX: Nvidia’s end-to-end infrastructure blueprint

Perhaps the clearest expression of Nvidia’s long-term vision is the Vera Rubin DSX AI Factory reference design: a blueprint for entire facilities built to “manufacture” intelligence tokens.

The DSX AI Factory defines how compute, networking, storage, power and cooling should be integrated to maximize “tokens per watt,” Nvidia’s chosen metric for end-to-end efficiency. Alongside it, Nvidia offers an Omniverse DSX Blueprint for digital twins of AI factories, allowing operators to simulate facility performance before construction.

The accompanying software includes:

• DSX Max-Q for dynamic power provisioning, which Nvidia says allows 30% more AI infrastructure in a fixed-power data center.
• DSX Flex to connect AI factories to power-grid services and tap an estimated 100 GW of stranded grid capacity.

Energy players including Emerald AI, GE Vernova, Hitachi and Siemens Energy are using the architecture. Nscale and Caterpillar are building what Nvidia describes as one of the world’s largest AI factories in West Virginia based on the Vera Rubin reference design. Partners such as Cadence, Dassault Systèmes, Eaton, Jacobs, Schneider Electric, Siemens, PTC, Switch, Trane Technologies and Vertiv are contributing simulation-ready assets and platform integrations, while CoreWeave is using DSX Air to run operational rehearsals of AI factories in the cloud before hardware delivery.

Huang’s formulation—“intelligence tokens are the new currency, and AI factories are the infrastructure that generates them”—captures how Nvidia wants operators to think: less about servers and clusters, more about production capacity and yield.

What’s proven, what isn’t, and how competitors fit

Nvidia’s story is unusually coherent: custom silicon, systems architecture, networking, storage, inference software, open models, agent frameworks, safety runtimes, digital twins and domain applications are all framed as parts of a single strategy for agentic AI. No other vendor currently matches that breadth.

However, the most aggressive performance claims for Vera Rubin—10x inference throughput per watt, one-tenth cost per token, 4x GPU savings for large MoE training versus Blackwell—are Nvidia’s own and remain unverified by independent benchmarks. The core bet that long-running, tool-using agents will dominate compute workloads is also not yet proven at scale, even if many organizations are moving in that direction.

Nvidia’s expansion up the stack into models, software and reference architectures raises familiar questions for infrastructure buyers about concentration risk and vendor dependence. At the same time, competition is real: AMD is closing the gap on data center GPU performance, while Google’s TPUs and Amazon’s Trainium are powering some of the largest AI training runs inside their respective clouds, and a wave of startups is targeting pieces of the AI infrastructure stack.

Even so, none of those competitors arrived this week with seven new chips in full production, public endorsements from the CEOs of Anthropic and OpenAI, and a complete “AI factory” reference architecture. For AI infrastructure engineers and cloud architects, Vera Rubin is not just another GPU generation; it is Nvidia’s attempt to define the blueprint for how the next decade of AI infrastructure will be built. Whether that blueprint becomes the industry default will depend on how well the claims stand up under real workloads—and how comfortable enterprises are with Nvidia’s increasingly central role in every layer of the stack.