NVIDIA Says Continuous Agent Training Will Become a Major Computing Workload

NVIDIA described continuous post-training as one of the major future workloads for data centres. A traditional model is created through initial training on a large dataset and later adaptation. For agents, however, changes do not end at deployment because software, rules, databases and error patterns continue to evolve. The vendor uses the term “intelligence per dollar.” In addition to the cost of one million tokens, it is intended to capture the cost of creating and continuously improving a model. If an agent stops working correctly after a business application changes, the company must gather new examples, run additional training and retest the entire workflow. NVIDIA argues that initial training primarily gives a model language fluency. The ability to code, plan multiple steps, use search or recover from an error is said to arise mainly during post-training. This is a simplified marketing distinction, but it illustrates the growing importance of data from real operations. The article uses the open Nemotron 3 Ultra model as an example. It has 550 billion parameters and a mixture-of-experts architecture. According to NVIDIA, it scored 71.7 percent on SWE-bench Verified, producing a working fix for roughly seven out of ten defects taken from real public software projects. The Vera Rubin platform is expected to perform post-training of the largest models with one quarter of the GPUs required by the Blackwell generation. NVIDIA does not provide a complete methodology for every possible configuration, electricity price or acquisition cost. The figure therefore cannot automatically be translated into a fourfold saving for the entire project. Hardware savings may be especially important in reinforcement learning, where the model repeatedly attempts a task, receives an evaluation and generates new trials. NVIDIA gives an illustrative scenario with 20 billion tokens in these attempts, derived from about 1.2 million runs of the earlier Nemotron 3 Super model. “Cost per token measures operational throughput,” NVIDIA says. “Intelligence per dollar” is intended to answer how much it costs to create a model worth operating. The distinction is useful, but it is also an argument for buying an entire computing platform from one supplier. Companies will have to determine who owns the data used for post-training. During operation, an agent may encounter contracts, internal code, personal data or trade secrets. If those examples are used to adapt the model, the company needs clear rules on data separation, retention and later deletion. Continuous training creates a validation problem as well. Every change may improve one type of task and degrade another. Organisations therefore cannot track only an average score. They need a permanent set of critical cases covering safety, legal compliance and the possibility of irreversible harm. From an accounting perspective, AI is shifting from a one-off project to a permanent operating expense. Companies pay not only for the model and integration but also for new examples, evaluation, computing runs and repeated approvals. This may change the economics of projects that initially counted only the cost of ordinary responses.

NVIDIA Says Continuous Agent Training Will Become a Major Computing Workload

The full article is available to users registered at Hard Skills