The California Gold Rush created fortunes not only for prospectors, but also for the people selling picks and shovels.
At CHM Live, computer architecture pioneers Bill Dally of NVIDIA and Norm Jouppi of Google joined CHM Fellow David Patterson to discuss the modern version of that story: the explosive demand for AI hardware.
The event was made possible by the generous support of Mark and Mary Stevens.
CHM Fellow David Patterson moderates a conversation with NVIDIA's Bill Dally and Google's Norm Jouppi on the forces behind today's AI "gold rush."
Asked what makes today's AI boom different from previous waves of computer architecture innovation, Dally pointed to three factors: "intense economic demand," rapidly evolving applications, and the relative simplicity of AI workloads compared with earlier supercomputing applications.
He contrasted today's environment with the supercomputing boom of the 1990s, when many startups were built on expectations of future markets that never fully materialized. By comparison, Dally said, AI is already creating value across industries, leading to what he described as "insatiable" demand for more compute.
Jouppi highlighted the scale of investment. Technology companies are now spending tens of billions of dollars on infrastructure, creating what he characterized as an unprecedented race to build capacity.
When Patterson asked who the prospectors and pickaxe sellers are in this modern gold rush, Dally suggested that entrepreneurs finding new applications for AI are the prospectors. Companies like Google and NVIDIA, he said, are providing the infrastructure.
Patterson raised a theory that modern AI accelerators have undergone a kind of "convergent evolution," arriving at similar designs despite being developed independently.
Both speakers acknowledged that AI workloads naturally drive designers toward certain architectural choices. Modern systems need massive matrix computation, large amounts of memory bandwidth, and fast interconnects.
But Dally argued that important differences remain. He pointed to innovations in numerical formats, sparsity, and networking as examples of architectural decisions that significantly affect performance.
Jouppi agreed that there are common requirements but noted that Google's TPU program benefited from being designed as a supercomputer from the beginning. He also highlighted Google's long-standing investment in optical and large-scale networking technologies.
One theme surfaced repeatedly throughout the conversation: building successful AI systems requires far more than designing a chip.
"The product is the whole system," Dally said.
That system includes processors, memory, networking, software, power delivery, cooling, and the operational expertise needed to make everything work reliably at scale.
Both speakers suggested that this systems knowledge is one advantage large technology companies hold over newer competitors. Drawing on lessons from decades of large-scale computing, they argued that deploying AI infrastructure successfully requires experience that extends well beyond semiconductor design.
Audience members were asked what they believe is the biggest bottleneck facing future AI deployment. The overwhelming answer was energy availability.
The panelists largely agreed.
Jouppi discussed the challenges of supplying power to increasingly large data centers and described efforts to use carbon-free energy sources where possible. Dally noted that building AI infrastructure increasingly requires balancing land, power, and facility capacity, while demand for generation equipment has surged.
The discussion underscored how AI's future may depend as much on energy infrastructure as on advances in chip technology.
Asked what advice they would offer young engineers, both guests emphasized fundamentals.
Jouppi encouraged students to learn the history of computer architecture, noting that many ideas have already been tried multiple times and that understanding past successes and failures remains valuable.
Dally offered three recommendations: master an application domain, develop broad technical knowledge, and learn to work with AI as a partner.
The most important skill, he suggested, is increasingly deciding what should be built rather than manually implementing every detail.
As the evening concluded, both speakers conveyed a sense that the current AI boom is more than another technology cycle. Unlike earlier computing frenzies, Dally argued, this one is being driven by real demand and real applications.
For the companies building the hardware behind AI, the gold rush is already underway.
