Intel, Marvell and Nvidia outline AI data center visions

TAIPEI (TVBS News) — The race to build AI isn't just about chips anymore. At Computex 2026, which opened Tuesday (June 2), executives from Marvell, Intel and Nvidia revealed that the real battle has shifted to something far less glamorous: the wires, cables and light beams connecting millions of processors into a single thinking machine.

Marvell Technology Chairman and CEO Matt Murphy kicked off the day's keynotes with a morning address focused on high-speed connectivity. Murphy argued that as AI models grow larger and data centers expand, the real bottleneck is no longer processors or memory — it is the connections linking thousands, and eventually millions, of chips. "It's a data center without distance, where compute, memory, networking, and photonics operate as one unified system," he said.

Murphy explained Marvell has invested approximately US$36 billion (approximately NT$1.13 trillion) over the past decade to transform itself into a data infrastructure company. That includes about US$22.5 billion (approximately NT$706 billion) in acquisitions and US$18 billion (approximately NT$565 billion) in internal research and development. The segment now accounts for more than 75% of Marvell's revenue, up from less than 10% a decade ago, he said.

Intel CEO Lip-Bu Tan (陳立武) took the stage in the afternoon for his first Computex appearance since becoming chief executive. Tan said demand for central processing units is surging as agentic AI systems require new architectures for reinforcement learning and orchestration. Supply is constrained, he noted — over the past four weeks, numerous CEOs have called him directly requesting more CPUs.

"Taiwan is very important in terms of — I mentioned about the PC client, the server, the whole rack system, the OEM, ODM," Tan told reporters after his keynote. "Taiwan is very strong and that's why we have multiple partnerships with different Taiwan (companies) for long historical reasons. But we also going to triple down in some of these exciting partnerships that we have, and we want to expand that."

Tan also addressed Intel's relationship with Taiwan Semiconductor Manufacturing Co. (台積電), calling TSMC founder Morris Chang (張忠謀) and Chairman C.C. Wei (魏哲家) longtime friends. Intel remains a TSMC customer with a trusted partnership it intends to maintain, he added. Nvidia (輝達) CEO Jensen Huang (黃仁勳), who delivered his keynote Monday, held a global media Q&A session Tuesday morning, emphasizing supply chain diversification and resilience.

"Everybody's supply chain should be as diversified and as redundant, has redundant capabilities as possible so that we can have resilience," Huang told reporters. "And it's one of the reasons why Taiwan is such an incredible strategic partner for the United States."

Huang's comments came as Nvidia suppliers ramp up production. Quanta Computer (廣達) subsidiary QCT (雲達科技) said it plans to add three more factories in the United States by year-end, primarily in California, to meet demand for Nvidia's next-generation Vera Rubin architecture.

Mike Yang (楊麒令), QCT general manager and Quanta executive vice president, told CNA that power supply through 2027 appears sufficient, with most capacity through 2028 already secured. Yang acknowledged some component shortages persist, including CPUs, but said supply chains are adjusting based on customer priorities.

The chips may grab the headlines, but the executives at Computex made clear that the real race is now about what connects them. As Murphy put it: a data center without distance. Taiwan's supply chain is betting it can build exactly that. ◼