Rubin is not a single graphics card. NVIDIA introduced it on January 5, 2026, as a codesigned data-center platform combining processors, accelerators, networking, storage infrastructure and rack-scale systems. The initial CES announcement listed six chips; a March announcement described a seven-chip Vera Rubin platform after adding the Groq 3 inference processor.
NVIDIA says Rubin is intended for AI training, inference, agentic workloads and scientific computing. Its performance and cost figures are NVIDIA claims, and the announcements do not provide independent benchmark validation, final system pricing or a complete current availability list.
What NVIDIA Rubin is
Rubin is an integrated AI-computing architecture for enterprise and research data centers. NVIDIA presents the parts as a coordinated AI supercomputer rather than as interchangeable consumer components. The named systems include GPU racks, CPU racks, inference-accelerator racks, storage racks and Ethernet networking racks.
NVIDIA has separately described DGX Vera Rubin NVL72 as a training and inference system and DGX SuperPOD as a deployment blueprint. Those products target organizational infrastructure; nothing in the announcements establishes a consumer retail GPU or an Amazon-searchable Rubin product.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →#1 Best Overall
Why the chip count changed from six to seven
| Announcement | Chip count | Components or change |
|---|---|---|
| January 5, 2026, CES | Six | Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet Switch. |
| March 2026, Vera Rubin platform update | Seven | Added the Groq 3 LPU, an inference accelerator acquired into the broader platform description. |
The two descriptions are not contradictory snapshots of the same announcement. Six chips was the January launch configuration; seven chips is the later March configuration that incorporates Groq 3.
What the seven-chip platform includes
| Component | Purpose in NVIDIA’s platform design |
|---|---|
| Vera CPU | Host and general-purpose computing for the AI system. |
| Rubin GPU | Accelerated computation for model training and inference. |
| NVLink 6 Switch | High-speed interconnection for scaling GPU resources within a system. |
| ConnectX-9 SuperNIC | Data-center networking between compute resources. |
| BlueField-4 DPU | Data-processing infrastructure, including storage and network services. |
| Spectrum-6 Ethernet Switch | Ethernet fabric for connecting racks and data-center systems. |
| Groq 3 LPU | Inference acceleration, particularly for serving and agentic workloads. |
In the March organization, NVIDIA grouped the hardware into Vera Rubin NVL72 GPU racks, Vera CPU racks, Groq 3 LPX inference accelerator racks, BlueField-4 STX storage racks and Spectrum-6 SPX Ethernet racks. NVIDIA says the combination is designed for pretraining, post-training, test-time scaling and agentic inference.
Rank #2
- Ideal for Gifting
- Ideal for a bookworm
- Compact for travelling
What performance does NVIDIA claim?
The figures below come from NVIDIA announcements. They are not independently audited results, and each comparison depends on the stated baseline and workload.
| Metric | NVIDIA’s published claim | Comparison or qualification |
|---|---|---|
| Inference token cost | Up to 10× lower | Compared with NVIDIA Blackwell; the announcement does not establish a universal cost for every model or deployment. |
| Mixture-of-experts training | 4× fewer GPUs | Compared with NVIDIA Blackwell for the stated MoE training comparison. |
| Agent throughput | 10× higher at scale | Compared with the previous-generation NVIDIA Grace Blackwell platform. |
| AI for science | More than 7 exaflops | NVIDIA’s figure for its scientific-computing systems. |
| Native FP64 | 5 petaflops | NVIDIA’s stated double-precision performance for the scientific-computing configuration. |
| GPU density | Up to 144 GPUs per rack | For custom high-density scientific-computing systems, not a general specification for every Rubin rack. |
NVIDIA CEO Jensen Huang said on January 5 that “Rubin arrives at exactly the right moment, as AI computing demand for both training and inference is going through the roof.” That is a company statement about demand, not an independently measured market finding. In a May 31 statement about agentic AI, Huang described a prompt as potentially launching “a thousand-step journey of reasoning, retrieval, tool use and response generation.”
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWhich workloads and systems is Rubin aimed at?
AI training and inference
DGX Vera Rubin NVL72 is presented as a system for both training and inference. The broader platform separates GPU computation, CPU hosting, networking, storage and inference acceleration so operators can build rack-scale deployments around their workload mix.
Agentic and test-time workloads
NVIDIA’s March platform description specifically names post-training, test-time scaling and agentic inference in addition to conventional pretraining. Groq 3 is included in that seven-chip description as the inference-focused element.
Rank #4
Scientific computing
In a June 22 announcement, NVIDIA positioned Vera Rubin for climate modeling, computational fluid dynamics, quantum chemistry and energy exploration. It linked those uses to native double-precision performance, CUDA-X libraries and the wider NVIDIA AI platform. NVIDIA named the Leibniz Supercomputing Centre, NERSC and Los Alamos National Laboratory in planned scientific-computing deployments; the announcement describes intended deployments rather than completed independent evaluations.
When will Rubin be available?
NVIDIA said on May 31, 2026, that Vera Rubin was ramping into full production. It reported manufacturing across more than 350 factories in 30 countries and 150 partners in Taiwan. Those are company-reported supply-chain figures.
Recommended Free Tools
Best Value
- It can be a gift option
- Comes with secure packaging
- Helpful in various ways
The January announcement said Rubin-based products were expected from partners in the second half of 2026. It named AWS, Google Cloud, Microsoft and Oracle Cloud Infrastructure, along with NVIDIA Cloud Partners CoreWeave, Lambda, Nebius and Nscale, among providers expected to deploy Rubin instances in 2026. That announcement is an expectation, not proof that a particular instance, configuration or region is available to every customer. Check each provider’s current catalog, region, hardware profile and contract terms.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How an organization should evaluate a Rubin purchase
Because final pricing and a comprehensive availability matrix were not provided, a buyer should compare a confirmed configuration rather than rely on the platform name alone.
- Choose the deployment route: compare an owned DGX or rack-scale system with a cloud instance or hosted service.
- Define the workload: separate pretraining, post-training, inference, agentic serving and scientific FP64 computing; they can require different racks and economics.
- Confirm the scale: document GPU count, CPU and accelerator mix, memory, NVLink topology, storage, networking and rack density.
- Match the metric to the decision: ask whether the vendor is quoting tokens per dollar, training time, throughput, latency, FP64 performance or another measure, and preserve the comparison baseline.
- Check facility requirements: validate power, cooling, floor space, network fabric, security controls and failure-recovery design.
- Obtain commercial proof: request a written price, delivery date, supported region, service-level terms and the exact software stack before committing capital.
What the announcements do not establish
- They do not provide final Rubin system pricing.
- They do not independently validate the headline performance or cost claims.
- They do not confirm that every named cloud provider has a generally available Rubin instance in every region.
- They do not turn Rubin into a consumer product or establish a retail replacement path for a desktop GPU.
For finance and procurement teams, the practical conclusion is to treat Rubin as a data-center platform whose economics depend on the complete system, facility and software deployment—not as a single chip with a standard retail price.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minute
