NVIDIA Reveals the Next AI Race: Building Smarter Models Instead of Bigger Ones
NVIDIA believes the next breakthrough in artificial intelligence won’t come from simply building larger models—it will come from teaching AI systems to keep learning long after they’re deployed.
In a new deep dive into the future of agentic AI, the company explained why continuous post-training is becoming the most important workload in modern AI development, introducing a new performance metric called “intelligence per dollar.”
The shift could redefine how companies train AI assistants, coding agents, autonomous systems, and enterprise AI over the coming years.
Why Agentic AI Changes Everything
Traditional generative AI models simply respond to prompts.
Agentic AI takes things much further.
Instead of producing a single answer, these AI systems are given a goal and must:
- Plan multiple steps ahead
- Use external tools
- Adapt to changing environments
- Recover from mistakes
- Learn from new situations
Because real-world environments constantly evolve, NVIDIA says these models require continuous post-training, not a one-time update after pretraining.
As new software, policies, and edge cases appear, AI agents must keep improving just like professional athletes reviewing every game to perform better in the next one.
Post-Training Is Becoming AI’s Biggest Compute Challenge
NVIDIA describes post-training as the stage where AI actually becomes intelligent.
During this phase, models learn practical skills such as:
- Writing better code
- Solving complex reasoning tasks
- Using search tools
- Planning workflows
- Correcting their own mistakes
This learning happens through reinforcement learning (RL), where AI repeatedly attempts tasks, receives feedback, and updates itself through millions of training cycles.
Unlike traditional AI training, these cycles never truly end.
Introducing “Intelligence Per Dollar”
The company argues that the AI industry should begin measuring success differently.
Instead of focusing only on cost per token—the cost of generating AI responses—NVIDIA proposes intelligence per dollar as the metric that matters most.
The concept measures how efficiently companies can build smarter AI models while keeping training costs under control.
Lower inference costs improve profitability, while better post-training increases the value of every AI response the model generates.
Nemotron 3 Ultra Shows What Continuous Learning Can Do
To demonstrate the concept, NVIDIA highlighted Nemotron 3 Ultra, its open-weight 550-billion-parameter mixture-of-experts (MoE) model.
According to NVIDIA, the model achieved an impressive:
- 71.7% score on SWE-bench Verified
That benchmark measures whether an AI can successfully fix real software bugs from open-source projects using verified tests.
The result suggests the model can correctly solve roughly 7 out of every 10 real programming issues, placing it among the strongest open AI coding models available.
Blackwell and Vera Rubin Are Built for Constant AI Training
NVIDIA says its latest AI hardware is specifically designed for the new post-training era.
Blackwell
The Blackwell platform lowers training costs, making continuous reinforcement learning financially practical for AI companies.
Vera Rubin
The upcoming Vera Rubin platform pushes efficiency even further.
According to NVIDIA, it can train frontier-scale AI models using only one-quarter of the GPUs required by Blackwell, dramatically reducing infrastructure costs while supporting larger reinforcement learning workloads.
Major AI Companies Are Already Adopting the Strategy
Several leading AI companies are already embracing NVIDIA’s continuous post-training approach.
Prime Intellect
Prime Intellect is using NVIDIA Blackwell to continuously improve open AI models and plans to expand reinforcement learning with Vera Rubin. The company reports that Vera CPUs deliver around 30% higher throughput per CPU than comparable x86 systems for reinforcement learning workloads.
Perplexity
Perplexity has built an asynchronous reinforcement learning pipeline running across hundreds of NVIDIA GPUs. Its system synchronizes trillion-parameter models in under two seconds, allowing rapid movement between training and inference.
Together AI
Together AI now offers post-training as a service, supporting supervised fine-tuning, reinforcement learning, and direct preference optimization while preparing to integrate NVIDIA’s Vera Rubin platform.
Why It Matters
The AI race is no longer just about creating larger language models.
Instead, companies are competing to build AI systems that continuously improve after deployment, adapt to changing environments, and deliver more intelligence without dramatically increasing costs.
If NVIDIA’s vision becomes the industry standard, future AI assistants, coding tools, autonomous robots, and enterprise agents may never stop learning.
That could mark the beginning of a new era where continuous intelligence—not model size—becomes the defining advantage in artificial intelligence.