The AI industry's biggest competition may no longer be about building the smartest model. Elon Musk's latest comments suggest the next phase of the race is about delivering more intelligence using less energy—a shift that could ultimately make AI cheaper to run and more practical to deploy at scale.

Replying to a post by Replit CEO Amjad Masad, Musk wrote that "Intelligence/Joule will keep improving," echoing a growing view that AI progress should be measured not only by how capable models become, but also by how efficiently they deliver those capabilities.

The post referenced new research showing a sharp improvement in "intelligence per joule" over the past 16 months, driven by advances in both AI models and hardware.

AI Is Entering an Efficiency Race

For much of the past three years, the AI conversation has centered on larger models, more parameters and record-breaking benchmarks. But as companies move from training models to running them in real-world applications, the economics of inference are becoming increasingly important.

A recent research paper by Avanika Narayan introduces "intelligence per watt" as a way to measure how much useful AI performance can be delivered for a given amount of energy. The researchers found that local AI systems improved this metric more than fivefold between 2023 and 2025, with gains coming from both better model architectures and more efficient hardware.

They also reported an 18-fold improvement in "intelligence per joule" over a 16-month period when combining advances in models and accelerators.

The takeaway is straightforward: AI isn't just becoming smarter—it is becoming dramatically more efficient.

Why Nvidia and AI Infrastructure Could Benefit

The idea isn't entirely new. In a recent exclusive interview, DigitalOcean CEO Paddy Srinivasan said AI companies are increasingly optimizing for "intelligence per dollar" by routing workloads across different models instead of relying exclusively on expensive frontier systems.

Musk's comments point to the same broader trend from a different angle. If companies can generate more useful AI with fewer watts of power, the cost of running AI applications can fall even as model capabilities improve.

That has implications across the AI ecosystem.

Companies such as Nvidia Corp (NVDA) have increasingly emphasized performance improvements that allow customers to complete more AI work with the same infrastructure, while cloud providers are looking for ways to lower inference costs as AI adoption expands.

At the same time, growing demand for electricity has made power availability a key constraint for AI data centers, making efficiency gains even more valuable.

What Investors Should Watch Next

Musk's post wasn't simply an observation about AI hardware. It reflects a broader shift in how the industry may define progress.

The next winners may not be the companies with the biggest models alone, but those that can deliver the most useful intelligence at the lowest energy and computing cost.

For investors, that means AI efficiency—not just AI capability—could become one of the most important metrics to watch over the next phase of the industry's growth.