Nvidia launches six-chip AI Rubin platform

Nvidia has unveiled its new Rubin platform, featuring six chips aimed at powering next-generation AI systems.

The Rubin lineup includes the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet Switch.

Nvidia said the platform improves efficiency for AI training and inference, citing reductions in GPU requirements and cost per token compared to the previous Blackwell platform.

The company plans to make Rubin-based products available through partners in the second half of 2026.

Major cloud providers including AWS, Microsoft, Google Cloud, and Oracle Cloud Infrastructure are expected to deploy Rubin systems.

Microsoft will use Rubin hardware in its upcoming Fairwater AI data centers.

CoreWeave, an AI cloud service provider, will also integrate Rubin systems into its offerings.

Nvidia said the Rubin platform supports large-scale AI workloads and incorporates new interconnect and computing technologies.

🔗 Source: Nvidia

🧠 Food for thought

Implications, context, and why it matters.

Rubin CPX prioritizes efficiency over exotic packaging, with no published cost-per-token vs Blackwell

Rubin CPX uses commodity GDDR7 (Graphics Double Data Rate 7) memory and a monolithic die (a single large chip rather than multiple chiplets) without CoWoS (Chip-on-Wafer-on-Substrate) packaging or NVLink interconnects (Nvidia’s high-speed chip-to-chip fabric) 1. Analysts peg its build cost at about 25% of standard GPUs in the Rubin line with roughly 60% of the compute 1.
Rubin CPX delivers 30 petaFLOPS of NVFP4 (Nvidia 4-bit floating-point) compute with 128 GB GDDR7 memory 2. Nvidia has not posted measured efficiency gains or cost-per-token cuts versus Blackwell, so outside teams cannot confirm the claims yet 2. Cost-per-token means the expense to process a unit of text from a large language model.
The disaggregated design runs compute heavy prefill on Rubin CPX and routes memory bandwidth bound decode to Rubin GPUs 3. Prefill is the initial pass over the input prompt, while decode is the token by token generation phase, which helps avoid unified GPU waste that can add up to a 37.5% premium on underused resources 1.

Rubin networking lifts demand for 800G and 1.6T optics

Rubin Spectrum‑6 Ethernet and Quantum‑X800 InfiniBand (a high performance datacenter networking technology) lift module demand 4. 800G (800 gigabits per second) datacenter optics shipments could grow 60% in 2025 4, with the 800G transceiver market topping $2.5 billion within five years 5.
Optical transceiver component suppliers plus manufacturers such as Acacia, Innolight, Coherent (optics makers) should ready more capacity 5. 1.6T (1.6 terabits per second) modules will enter volume output for select Nvidia uses in 2025 4, though units stay under 1 million at first.
Cabling plus interconnect distributors for AI data centers should map 2026 supply chains to the shift from a 400G (400 gigabits per second) baseline to 800G or 1.6T standards 6. Large cloud providers (hyperscalers) plan a broad move to 1.6T optics in the second half of 2025 4.