Jensen Huang, the CEO of Nvidia, made a bold declaration that his company’s AI chips are advancing at a pace exceeding the historical rates set by Moore’s Law—the benchmark that defined computing progress for decades.
“Our systems are progressing way faster than Moore’s Law,” said Huang during an interview, a day after delivering a keynote address to a 10,000-person crowd at CES in Las Vegas.
The Rise and Slowdown of Moore’s Law
Coined in 1965 by Intel co-founder Gordon Moore, Moore’s Law predicted that the number of transistors on a computer chip would double approximately every year, resulting in doubled performance and drastically reduced costs. For decades, this principle drove revolutionary advancements in computing.
However, as physical and technical limitations have emerged, Moore’s Law has significantly slowed in recent years. Yet, Huang asserts that Nvidia’s AI chips are outpacing this slowdown with remarkable speed. According to the company, its latest data center superchip is more than 30 times faster for running AI inference workloads compared to its previous generation.
Breaking Barriers Across the Full Stack
Huang attributes Nvidia’s ability to surpass Moore’s Law to its end-to-end innovation across hardware and software.
“We can build the architecture, the chip, the system, the libraries, and the algorithms all at the same time,” Huang explained. “If you do that, you can innovate across the entire stack and move faster than Moore’s Law.”
AI Chips: Powering the Future of Computing
This announcement comes as concerns grow over whether AI’s rapid progress is stalling. Nvidia’s chips are the cornerstone of leading AI labs like Google, OpenAI, and Anthropic, enabling the training and execution of their state-of-the-art AI models. Advancements in Nvidia’s hardware could translate directly to breakthroughs in AI capabilities.
Huang’s remarks echoed statements he made in the past, where he suggested that the AI industry is moving towards “hyper Moore’s Law,” a term highlighting the accelerated evolution of AI technologies.
Three Scaling Laws of AI
Rejecting the notion that AI progress is slowing, Huang introduced a new framework: three active scaling laws for AI development. These include:
- Pre-training: The initial phase where AI models learn patterns from vast datasets.
- Post-training: Fine-tuning the model’s responses using methods such as human feedback.
- Test-time compute: The inference phase, where AI models “think” and process inputs in real time.
Huang emphasized that advancements in inference will drive down computing costs—a trend reminiscent of Moore’s Law’s impact on general computing.
“The same thing will happen with inference,” said Huang. “We will drive up performance, which in turn will lower the cost of inference.”
AI Inference and Cost Challenges
Nvidia’s AI chips, including the H100, have dominated the market for training AI models. However, the industry’s current focus has shifted towards inference—the process of running and applying trained AI models. This transition has raised questions about the affordability of Nvidia’s high-end chips.
AI inference models, particularly those using test-time compute, remain expensive to operate. For instance, OpenAI’s cutting-edge “o3” model, which uses scaled-up test-time compute, incurs costs of nearly $20 per task to achieve human-level intelligence in certain benchmarks. By comparison, a ChatGPT Plus subscription costs just $20 for an entire month of usage.
Image Credits: Nvidia
The GB200 NVL72: A Game-Changing Superchip
During his CES keynote, Huang unveiled Nvidia’s latest superchip, the GB200 NVL72, to the audience. Holding it aloft like a shield, he emphasized its groundbreaking capabilities. The GB200 NVL72 delivers performance 30 to 40 times greater than Nvidia’s best-selling H100 chips when running AI inference workloads.
This leap in performance, according to Huang, will significantly reduce the cost of running advanced AI reasoning models like OpenAI’s o3 over time, making these models more accessible to businesses and users worldwide.
The Long-Term Vision: Lower Costs, Higher Performance
Huang highlighted that improving chip performance is key to lowering the cost of AI operations in the long run.
“The direct and immediate solution for test-time compute, both in terms of performance and affordability, is to increase our computing capability,” said Huang. He also noted that advancements in inference could pave the way for creating better data for the pre-training and post-training phases of AI development.
Over the past year, the cost of running AI models has significantly declined, thanks to breakthroughs from hardware innovators like Nvidia. Huang expects this trend to continue, even as initial versions of AI reasoning models remain expensive.
Faster Than Moore’s Law: A Ten-Year Leap
Looking back, Huang claimed that Nvidia’s AI chips today are 1,000 times more powerful than those the company developed a decade ago. This rate of improvement far outstrips the trajectory established by Moore’s Law and underscores Nvidia’s central role in the future of AI.
With no signs of slowing, Nvidia’s advancements are setting the stage for the next era of computing—one where AI innovation outpaces every previous benchmark. Huang’s vision signals not just an evolution of technology, but a reshaping of how humanity leverages computing power for generations to come.
