The next phase of artificial intelligence (AI) advancement hinges entirely on rewriting the rules of hardware architecture and material science, a panel of prominent industry experts said at the COMPUTEX 2026 Forum at Taipei Nangang Exhibition Center Hall 2 on June 4.
The explosive growth of generative AI is pushing global data center infrastructure to its absolute physical limits, requiring a radical power and cooling revolution to prevent an impending energy crisis, panelists said.
Jeff Morroni, vice president and head of research and development at Texas Instruments’ Kilby Labs, said the physical power required to generate data has changed dramatically over the past 25 years, even though the foundational purpose of data centers remains unchanged.
To bypass massive power distribution losses, the industry is migrating from lateral power delivery to vertical power delivery, Morroni said.
Addressing the energy challenge, Adam White, division president of power and sensor systems at Infineon, said shaping AI’s future requires efficient power across all levels, from infrastructure to data centers and physical AI systems.
“The more we use AI, the more compute power is required — and therefore, the more energy we need. Scale is absolutely essential,” White said, highlighting the industry’s transition from legacy 20-kilowatt (kW) racks toward high-density 500kW systems.
Davood Yazdani, executive vice president and chief product officer of Endura Technologies, warned that software efficiency has officially run out of room.
The next phase of the AI revolution hinges entirely on hardware — specifically, how power is delivered to silicon, Yazdani said.
“The core challenge in scaling AI isn’t just generating more power. It’s how efficiently we consume the power budget we already have,” Yazdani said. “AI made power sexy again.”
Addressing flaws in legacy power architecture, Yazdani said the solution lies in a transition to vertical power delivery and integrated voltage regulators.
“The physics demand it, and the economics make sense,” he said. “The only question is who gets there first.”
Power is the most critical factor in processor performance, said Jinghai Zhou, senior vice president of cloud computing at Monolithic Power Systems, stressing that the path toward high-power delivery will not be easy.
“In December 2022, everything changed — AI became real,” Zhou said.
Citing recent data, Zhou said a single processor package will exceed 10kW by 2029, up from the current 2.5kW per processor. Data center electricity usage is doubling every three to four years, with worst-case future power scenarios tracking eight times higher than best-case projections, he added.
As AI workloads explode globally, heat has shifted from a mere hardware byproduct into the ultimate bottleneck for computing, said Myra Zai, director for research and development and technical service at Dow Co.
“We are at a defining moment in the evolution of computing,” Zai said. “If we do not manage it well, performance, reliability, and operating costs will all start melting down.”
Industry analysts predict that by 2030, the electricity consumption of AI infrastructure will double compared to 2025 levels, representing an annual growth rate of 18 percent driven by AI adoption, edge computing, and cross-industry digitalization, Zai said.
Air cooling has reached its limit, making liquid cooling an absolute necessity, she added.
The experts delivered a unified message: without an immediate, systemic overhaul of underlying power and cooling infrastructure, the physical limits of electricity and heat will inevitably stall the AI revolution.
