Nvidia’s announcement positions Vera Rubin as a rack-scale, full-stack “AI factory,” not merely a faster GPU generation: it combines custom GPUs, CPUs, scale-up and scale-out networking, DPUs, and power-management software to lower the energy and cost of serving agentic-AI workloads. The headline performance figures are Nvidia’s claims from a vendor-backed benchmark, not independent, peer-reviewed validation. 8

• Vera Rubin NVL72 platform: A liquid-cooled rack integrates 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs, BlueField-4 DPUs, and sixth-generation NVLink switching. It is designed to operate as one 72-GPU scale-up domain, while connecting outward through Quantum-X800 InfiniBand or Spectrum-X Ethernet. 36

• Agentic-AI efficiency claim: Nvidia says Vera Rubin NVL72 delivers up to 30× more agentic-workload throughput per megawatt than GB300 NVL72, and up to 35× lower token cost in the cited comparison. 89 Nvidia has separately claimed up to a 10× reduction in inference-token cost versus Blackwell for Rubin generally. 11

• Benchmark context: The measurement used SemiAnalysis’s AgentX workload, which replays production-style agentic coding sessions, with DeepSeek V4 Pro as the model. The comparison emphasizes long-context, interactive inference, where tool calls, code execution, orchestration, and repeated model turns make CPU work, memory bandwidth, and interconnect efficiency material—not just GPU FLOPS. 89

• Hopper comparison and token-use figures: The material available here substantiates the GB300 comparison but does not provide a primary-source, directly comparable Hopper figure or an exact published “cost per one million tokens” dollar value. Nvidia’s broader point is that agentic workflows consume substantially more tokens than one-shot chat because they involve iterative reasoning and tool use; the cited claims should therefore be read as workload- and configuration-dependent, not universal serving prices. Insufficient evidence is available here to state a verified exact Hopper multiplier or a dollar-per-million-token figure.

• DSX MaxLPS: Nvidia presents DSX as an end-to-end AI-factory design spanning chips, systems, software, facilities, and partner technology. Its DSX MaxLPS software is intended to maximize token throughput per unit of power and thereby minimize token cost, rather than simply run hardware at maximum instantaneous power. 1215 The practical implication is better power allocation across the AI-factory stack—useful when data-center power delivery and cooling, rather than accelerator availability alone, constrain deployment.

• Vera CPU: Vera is Nvidia’s custom Armv9.2-compatible CPU for agent-era infrastructure. It has 88 Olympus cores and 176 Spatial Multithreading threads, aimed at high-concurrency CPU tasks surrounding inference: data processing, orchestration, retrieval, tool execution, sandboxed code, and simulation. 210

• Memory and interconnect: Vera uses SOCAMM2 LPDDR5X memory, delivering up to 1.2 TB/s aggregate bandwidth, or up to 14 GB/s per core. Its second-generation Scalable Coherency Fabric connects the 88 cores, shared L3 cache, and memory subsystem with up to 3.4 TB/s bisection bandwidth; Nvidia says the chip has a 164 MB unified L3 cache. 78 Compared with Grace, Nvidia says Vera provides 2.4× memory bandwidth, 3× memory capacity, and twice the NVLink-C2C bandwidth for coherent CPU–GPU operation. 1

• “Scale-In” networking’s role: Nvidia’s terminology distinguishes three layers:

  • Scale-in / scale-up: NVLink 6 makes the 72 GPUs behave like a tightly coupled rack-scale computer. Nvidia specifies 3.6 TB/s bidirectional GPU-to-GPU bandwidth per GPU and 260 TB/s at rack level. 3
  • Scale-out: Quantum-X800 InfiniBand or Spectrum-X Ethernet connects racks into larger clusters. 3
  • Data and infrastructure offload: ConnectX-9 and BlueField-4 handle networking, storage, security, and data-path functions. 36

• SpaceXAI partnership: Nvidia says SpaceXAI will deploy Vera CPUs for agentic workloads and expand Grok infrastructure using Vera Rubin as it targets gigawatt-scale compute capacity. 14 The companies also plan to extend the architecture to the planned Starmind orbital-computing effort, with Nvidia hardware forming the basis of a space-optimized compute payload; this is a planned deployment, not evidence of an operational orbital AI data center today. 1412

• Strategic meaning: If Nvidia’s workload claims translate to customer deployments, the main gain is not simply more model performance but more useful agent work per constrained megawatt—potentially reducing electricity, cooling, networking, and capital cost per generated token. Nvidia is also extending its competitive boundary from GPUs into custom Arm CPUs, coherent CPU–GPU systems, rack-scale switching, DPUs, and AI-factory operating software. That tighter integration may strengthen its position against GPU rivals and incumbent CPU vendors, but it also increases customer dependence on Nvidia’s architecture and software stack.

Nvidia’s announcement positions Vera Rubin as a rack-scale, full-stack “AI factory,” not merely a faster GPU generation: it combines custom GPUs, CPUs, scale-up and scale-out networking, DPUs, and power-management software to lower the energy and cost of serving agentic-AI workloads. The headline performance figures are Nvidia’s claims from a vendor-backed benchmark, not independent, peer-reviewed validation. 8

• Vera Rubin NVL72 platform: A liquid-cooled rack integrates 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs, BlueField-4 DPUs, and sixth-generation NVLink switching. It is designed to operate as one 72-GPU scale-up domain, while connecting outward through Quantum-X800 InfiniBand or Spectrum-X Ethernet. 36

• Agentic-AI efficiency claim: Nvidia says Vera Rubin NVL72 delivers up to 30× more agentic-workload throughput per megawatt than GB300 NVL72, and up to 35× lower token cost in the cited comparison. 89 Nvidia has separately claimed up to a 10× reduction in inference-token cost versus Blackwell for Rubin generally. 11

• Benchmark context: The measurement used SemiAnalysis’s AgentX workload, which replays production-style agentic coding sessions, with DeepSeek V4 Pro as the model. The comparison emphasizes long-context, interactive inference, where tool calls, code execution, orchestration, and repeated model turns make CPU work, memory bandwidth, and interconnect efficiency material—not just GPU FLOPS. 89

• Hopper comparison and token-use figures: The material available here substantiates the GB300 comparison but does not provide a primary-source, directly comparable Hopper figure or an exact published “cost per one million tokens” dollar value. Nvidia’s broader point is that agentic workflows consume substantially more tokens than one-shot chat because they involve iterative reasoning and tool use; the cited claims should therefore be read as workload- and configuration-dependent, not universal serving prices. Insufficient evidence is available here to state a verified exact Hopper multiplier or a dollar-per-million-token figure.

• DSX MaxLPS: Nvidia presents DSX as an end-to-end AI-factory design spanning chips, systems, software, facilities, and partner technology. Its DSX MaxLPS software is intended to maximize token throughput per unit of power and thereby minimize token cost, rather than simply run hardware at maximum instantaneous power. 1215 The practical implication is better power allocation across the AI-factory stack—useful when data-center power delivery and cooling, rather than accelerator availability alone, constrain deployment.

• Vera CPU: Vera is Nvidia’s custom Armv9.2-compatible CPU for agent-era infrastructure. It has 88 Olympus cores and 176 Spatial Multithreading threads, aimed at high-concurrency CPU tasks surrounding inference: data processing, orchestration, retrieval, tool execution, sandboxed code, and simulation. 210

• Memory and interconnect: Vera uses SOCAMM2 LPDDR5X memory, delivering up to 1.2 TB/s aggregate bandwidth, or up to 14 GB/s per core. Its second-generation Scalable Coherency Fabric connects the 88 cores, shared L3 cache, and memory subsystem with up to 3.4 TB/s bisection bandwidth; Nvidia says the chip has a 164 MB unified L3 cache. 78 Compared with Grace, Nvidia says Vera provides 2.4× memory bandwidth, 3× memory capacity, and twice the NVLink-C2C bandwidth for coherent CPU–GPU operation. 1

• “Scale-In” networking’s role: Nvidia’s terminology distinguishes three layers:

  • Scale-in / scale-up: NVLink 6 makes the 72 GPUs behave like a tightly coupled rack-scale computer. Nvidia specifies 3.6 TB/s bidirectional GPU-to-GPU bandwidth per GPU and 260 TB/s at rack level. 3
  • Scale-out: Quantum-X800 InfiniBand or Spectrum-X Ethernet connects racks into larger clusters. 3
  • Data and infrastructure offload: ConnectX-9 and BlueField-4 handle networking, storage, security, and data-path functions. 36

• SpaceXAI partnership: Nvidia says SpaceXAI will deploy Vera CPUs for agentic workloads and expand Grok infrastructure using Vera Rubin as it targets gigawatt-scale compute capacity. 14 The companies also plan to extend the architecture to the planned Starmind orbital-computing effort, with Nvidia hardware forming the basis of a space-optimized compute payload; this is a planned deployment, not evidence of an operational orbital AI data center today. 1412

• Strategic meaning: If Nvidia’s workload claims translate to customer deployments, the main gain is not simply more model performance but more useful agent work per constrained megawatt—potentially reducing electricity, cooling, networking, and capital cost per generated token. Nvidia is also extending its competitive boundary from GPUs into custom Arm CPUs, coherent CPU–GPU systems, rack-scale switching, DPUs, and AI-factory operating software. That tighter integration may strengthen its position against GPU rivals and incumbent CPU vendors, but it also increases customer dependence on Nvidia’s architecture and software stack.