NVIDIA published a detailed technical post on July 17 arguing that its next-generation Vera Rubin platform reframes the central metric for AI infrastructure from raw compute speed to “intelligence per dollar”, the efficiency of every dollar spent refining and deploying an AI model after its initial training.
Why post-training has become the critical workload
Post-training, the phase that refines a model after initial training on raw data, is no longer a one-time finishing step. It is continuous, because the environment that agentic models operate in shifts fast.
A model is no longer simply asked for an answer; it is given a goal and must keep adapting as environments shift, edge cases emerge and tools change. Unlike a generative model responding to a prompt, an agentic model must plan, use different tools, and recover from problems it encounters mid-run.
NVIDIA is extending the core value proposition of its Vera Rubin platform from inference cost to model training efficiency, betting on “intelligence per dollar” as a new metric as continuous post-training becomes the core computing demand in the era of agentic AI.
The phrase bundles training cost, inference efficiency, and the value of each token into one figure. A cheap token is of limited value if the model repeatedly fails a workflow requiring five tool calls and a reliable rollback.
A costly post-training cycle is equally hard to justify if the resulting agent cannot be served economically at scale.
The GPU efficiency claim and platform architecture
In its July 17 post, NVIDIA said Vera Rubin could train a specified 10-trillion-parameter mixture-of-experts model on 100 trillion tokens in one month using one-quarter as many GPUs as a Blackwell NVL72 deployment. The company attributes this to what it calls extreme codesign.
Extreme codesign is the foundation of the Vera Rubin platform: GPUs, CPUs, networking, security, software, power delivery, and cooling are architected together as a single system rather than optimized in isolation, treating the data center, not a single GPU server, as the unit of compute.
The NVIDIA Vera Rubin platform comprises seven chips across five rack-scale systems, announced at GTC 2026. The Vera Rubin NVL72 packs 72 Rubin GPUs and 36 Vera CPUs into a single liquid-cooled rack. The platform also integrates ConnectX-9 SuperNICs and BlueField-4 DPUs, scaling up with NVLink 6 and scaling out with Quantum-X800 InfiniBand and Spectrum-X Ethernet.
The Rubin GPU is rated at 50 petaflops in FP4 precision, up from 20 petaflops in Blackwell, while Rubin Ultra will double that figure to 100 petaflops.
The important qualification is in the setup. This is NVIDIA’s modeled comparison for a specific large-model training target, not a universal promise that every AI workload suddenly needs 75 percent fewer accelerators. Nor has NVIDIA published a rack price or full cost-of-ownership figure that would let customers independently turn the GPU-count claim into a procurement saving.
Nemotron 3 Ultra and early enterprise partners
NVIDIA’s own Nemotron open model family illustrates the post-training approach in practice. NVIDIA Nemotron 3 Ultra is an open-weight, 550-billion-parameter mixture-of-experts model that offers verifiable benchmarks and a fully disclosed post-training recipe run on NeMo RL.
It scores 71.7 percent on SWE-bench Verified, producing a working fix for roughly seven in ten real software bugs from open-source projects, each checked against the project’s own tests.
Prime Intellect, Perplexity, and Together AI are among the organizations using NVIDIA infrastructure for distributed post-training and reinforcement-learning workflows. When comparing realistic reinforcement-learning sandbox workloads against alternative x86 architectures, Prime Intellect found that the Vera CPU delivers, on average, 30 percent greater throughput per CPU.
Perplexity’s reinforcement-learning post-training stack runs asynchronously across hundreds of NVIDIA GPUs, with an RDMA-based weight transfer engine that synchronizes trillion-parameter models in under two seconds between training and inference compute nodes. The resulting post-trained models are served on NVIDIA GB200 NVL72 systems.
Deployment momentum and what comes next
The NVIDIA Vera Rubin platform is ramping into full production, with Taiwan’s top server makers and global supply chain leaders manufacturing at scale and shipping Vera Rubin-based systems to fuel AI labs, cloud providers, and hyperscalers.
Microsoft will deploy NVIDIA Vera Rubin NVL72 rack-scale systems as part of next-generation AI data centers, including future Fairwater AI superfactory sites, designed to provide the foundation for Microsoft’s next-generation cloud AI capabilities.
AWS, Google Cloud, Microsoft, Oracle Cloud Infrastructure, CoreWeave, Lambda, Nebius, and Nscale are all confirmed first-wave deployment partners. The Vera Rubin generation is itself a stepping stone: at NVIDIA GTC 2025, the company announced that Rubin will be followed by an improved Rubin Ultra architecture in 2027.
For enterprises evaluating the platform, NVIDIA’s full technical breakdown is available on the NVIDIA blog, alongside the Nemotron enterprise partner details covered previously by GCN.
