Nvidia Says Vera Rubin Trains Agentic AI Models With Four Times Fewer GPUs

Nvidia unveiled details of its Vera Rubin platform on July 17, which the company is marketing as a way to drastically cut the cost of post-training agentic models. According to materials published on the company's blog, the new hardware lets the largest models be trained using four times fewer GPUs than the previous Blackwell generation required.

What Was Announced

Vera Rubin is the successor to the Blackwell architecture, pairing a new Vera CPU with Rubin GPUs into a single integrated system. Nvidia says the key to the savings isn't the raw compute power of a single chip, but the combination of hardware and software the company calls 'extreme co-design.' In practice, this is meant to mean that companies training large agentic models need far fewer physical chips to achieve the same result.

For post-training, the stage where finished models are fine-tuned for specific tasks and behaviors, Nvidia points to a separate set of tools: NeMo RL for distributed reinforcement learning, NeMo Gym as training environments, and NVIDIA Dynamo for inference orchestration. The company stresses that post-training, rather than initial model pretraining, is becoming the main cost in the era of agentic models, which need to learn to repeatedly correct their own mistakes.

The Numbers Behind the Announcement

Nvidia presented results from the Nemotron 3 Ultra model, with 550 billion parameters, which scored 71.7 percent on the SWE-bench Verified test, which measures the ability to fix real-world code bugs. That means the model correctly resolves roughly seven out of ten real software bug reports from the test set.

The company also released figures about the infrastructure itself. The new Vera CPU is said to deliver 30 percent higher throughput than x86 architecture on reinforcement learning workloads, a training method in which the model learns from rewards for correct actions. This matters because this method dominates the fine-tuning of agentic models that are meant to carry out multi-step tasks independently.

Early Partners

Prime Intellect, a company specializing in distributed model training, currently runs post-training on Blackwell and plans to move its reinforcement learning scaling to Vera Rubin. Perplexity has already launched an asynchronous reinforcement learning stack and serves post-trained models based on Qwen3 235B, adding that it can synchronize trillion-parameter models across compute nodes in under two seconds. Together AI, which offers post-training as a service to external customers, announced it will deploy Vera Rubin in its offering.

These three companies represent different segments of the market, from distributed infrastructure to AI-powered search to cloud services for developers, which Nvidia presents as evidence that the new platform has applications across the entire agentic model production chain, not just in a narrow research niche.

Why It Matters

The cost of training and fine-tuning large language models has been one of the industry's main talking points for months, since infrastructure bills determine which companies can afford to develop their own agentic models and which must rely on off-the-shelf services. The claim of four times lower GPU requirements at the same training scale, if confirmed in practice by customers, would mean a measurable cost reduction for companies building their own agentic systems, whether large corporations or smaller startups relying on the cloud.

For Polish companies and institutions that are just beginning to invest in AI agents, more affordable post-training infrastructure could lower the barrier to building their own fine-tuned models instead of relying solely on ready-made APIs from foreign providers. Domestic tech companies using public clouds typically feel infrastructure price changes with a delay, once cloud providers update their own pricing after deploying new hardware.

Some caution is warranted, however, toward figures provided by the hardware maker itself. The claimed fourfold reduction in GPU count comes from Nvidia's own promotional materials rather than independent lab tests, and the actual savings customers see will depend on specific workloads, configurations, and how quickly cloud partners roll out the new hardware in their data centers.

What's Next

Systems based on Vera Rubin are expected to reach major cloud providers, including AWS, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure, in the second half of 2026. Only then will it be possible to verify whether the claimed savings translate into real service prices for companies using ready-made infrastructure rather than building their own data centers.