The AI model evolution into vast "Agent" systems is shifting data center bottlenecks from pure computing power to "connectivity," igniting a fundamental infrastructure revolution transitioning from copper cables to optical fibers.

On the second day of the Computex event in Taipei, Matt Murphy, Chairman and CEO of AI custom chips, optical communications, and data center interconnect leader Marvell Technology (NASDAQ: MRVL), delivered a keynote speech. NVIDIA CEO Jensen Huang made a surprise appearance as a special guest. The joint stage presence of these two leaders at the pinnacle of AI compute and networking brought their deep strategic partnership into the spotlight, becoming one of the event's most significant moments.

Jensen Huang set the tone with a single statement: "Ladies and gentlemen, the next trillion-dollar company"—referring to Marvell. The declaration was met with thunderous applause. This highlights the deep alliance formed months earlier when NVIDIA announced a strategic $2 billion investment in Marvell, underscoring their joint efforts in AI data center infrastructure.

Following its latest quarterly earnings, the market is intensely focused on Marvell's potential gains from the AI supercomputing cycle. Murphy presented a compelling case: a decade ago, Marvell's data center revenue was under 10% of its total, but last quarter, it exceeded 75% and is accelerating at an annual rate of approximately 40%. Based on the latest guidance, Wall Street widely expects its revenue to reach a staggering $16.4 billion next year.

Behind this performance surge, Huang and Murphy outlined the core investment thesis for AI infrastructure: after breakthroughs in compute and memory bottlenecks, "connectivity" will define a system's ultimate performance. Their shared consensus is clear: the next decisive battleground in AI infrastructure is not compute or memory, but connectivity, with Marvell positioned at the heart of this revolution.

Notably, Marvell's stock surged over 16% in after-hours trading.

The Connectivity Imperative

Why has connectivity become so critical now? Murphy explained the logical progression of infrastructure bottlenecks: compute (led by NVIDIA, becoming the world's first $5 trillion company) → memory (where three new trillion-dollar companies recently emerged) → connectivity (happening now).

"The world's top hyper-scale cloud providers are re-architecting their entire network. They recognize that scaling AI infrastructure is now the primary connectivity challenge," Murphy stated, adding this feedback comes directly from their largest customers.

Jensen Huang provided the straightforward business rationale: "Useful AI is here. It's profitable now. Token generation is profitable. When it's profitable to generate tokens, everyone wants to generate more tokens. That's why Marvell's demand is so strong, and that's why our demand is so strong."

He noted that AI is moving towards an "Agent" model, a new computing paradigm requiring tasks to be fragmented and distributed across massive compute clusters. "When you break a computing problem into many parts and distribute it across a data center, the one thing you cannot do without is connectivity," Huang said, praising Marvell as the next trillion-dollar company.

Murphy added that a single processor can no longer handle AI workloads; future systems will require millions of processors working together. "Scaling compute is fundamentally a connectivity challenge. The industry has solved the compute bottleneck, is solving the memory bottleneck, and the next bottleneck limiting infrastructure is connectivity."

Copper vs. Optics: A Pragmatic Timeline

A key market insight from the dialogue was their assessment of the copper-to-optics transition timeline.

Huang offered a clear strategic framework: "Use optics wherever you must, use copper wherever you can." He explained that copper has physical limits in bandwidth and distance; it remains a simple, low-cost, practical choice until those limits are reached. Beyond that critical point, optics take over for scaling between racks, within data centers, and across data centers.

His core conclusion: "For the next 5 to 10 years, we're going to be using a lot of copper, and we're going to be using a massive amount of optics. These data centers are now part of the infrastructure."

This "copper-and-optics, each within its boundary" outlook means Marvell, as one of the few companies offering complete solutions in both domains, stands to benefit continuously.

Murphy explained the physics: copper transmission distance is inversely proportional to bandwidth; doubling bandwidth halves the distance. The fastest production systems today achieve 200Gbps per lane, with a copper reach of about 2.5 meters—already near the limit for a standard ~2-meter rack considering internal routing. "When we go to 400Gbps, copper will not reach across the rack. The Copper Wall is moving, and it's happening now." Each shift of this wall increases connection counts by an order of magnitude, directly fueling optical demand.

To address this, Marvell is heavily investing in Co-Packaged Optics (CPO) technology, integrating optics directly into the package next to the compute chip to solve density and power challenges. At the event, Marvell launched a new 100T Ethernet switch designed for AI data centers with industry-leading low power and showcased a 51.2T switch based on CPO that eliminates copper traces at the board level. "This is not some future concept; it's happening now," Murphy stated, suggesting that once optical interconnect breaks distance barriers, future data centers will have no rigid physical boundaries between compute and memory, enabling large-scale dynamic composition based on AI model needs.

Building a Heterogeneous Ecosystem

To meet complex networking demands, NVIDIA's strategic $2 billion investment in Marvell is expanding collaboration across optics, silicon photonics, and NVLink Fusion.

NVLink Fusion aims to address cloud service providers' customization needs. Huang explained that CSPs designing their own custom ASICs still want access to NVIDIA's system architecture. "You don't have to buy everything from us, just some things. By fusing the NVIDIA technology platform with Marvell's technology and solutions, we can essentially build a disaggregated, distributed, and heterogeneous data center."

Within this ecosystem, Marvell occupies a unique and neutral position. Murphy emphasized: "We partner deeply with compute companies, and we partner deeply with memory companies. In many ways, we're like the Switzerland of the industry—we partner with everybody."

Marvell's Strategic Transformation

In his keynote, Murphy detailed Marvell's decade-long strategic pivot from a company where over 60% of revenue came from consumer electronics (like chips in a Wi-Fi-connected Barbie Dreamhouse) to a pure-play data infrastructure semiconductor leader. The vision was to build a company focused on semiconductors that transmit, store, process, and secure the world's data at massive scale.

This involved systematic execution: divesting non-core businesses (like Wi-Fi and automotive Ethernet) and strategic acquisitions totaling about $22.5 billion (including Cavium, Avera, Inphi, Innovium, Celestia AI, XCON), combined with ~$18 billion in organic R&D. The company also made a critical decision to leapfrog process nodes, moving directly from 14/16nm to 5nm, establishing itself at the technology forefront.

The results speak for themselves: from $2.3 billion in revenue in 2016, Marvell is expected to reach ~$11.4 billion this fiscal year, with Wall Street projecting ~$16.4 billion next year. Data center revenue, once under 10%, now exceeds 75% and is growing rapidly.

The Full Spectrum of Connectivity

Murphy outlined the comprehensive connectivity portfolio required for AI data centers, spanning from millimeters within a package to thousands of kilometers between data centers.

For long-haul, inter-data center links (hundreds to thousands of km), Marvell leads in coherent DSP technology, evolving from 100G to 400G, 800G, and soon sampling a 1.6T 2nm solution.

Inside the data center (spanning hundreds of meters), the company provides end-to-end solutions including PAM4 DSP chip-sets for optical modules and Ethernet switch silicon, announcing a new 100T switch at the event.

Within the rack, the focus shifts to electrical SerDes technology over copper for ultra-high-speed, any-to-any processor interconnects, with Marvell offering leading 200Gbps tech and demonstrating future 400Gbps capabilities.

Inside the package (millimeters), the challenge involves chip-to-chip interconnects using advanced SerDes and 2.5D/3D packaging technologies.

The Future: An All-Optical, Distance-Agnostic Infrastructure

The ultimate vision is an all-optical infrastructure where distance is no longer a constraint. This will enable radical architectural shifts: scale-up clusters within racks could expand from dozens or hundreds of processors to thousands, all optically connected. Servers themselves could become disaggregated, with compute (XPUs), memory, and general-purpose CPUs housed in separate, pooled resources dynamically connected via light.

This eliminates the fixed ratios and inefficiencies of today's integrated systems, allowing infrastructure to be composed on-the-fly to match specific workload needs. "Architects will for the first time be able to design AI systems around the needs of the model, not the limitations of the interconnect," Murphy concluded, framing this as the next era of computing infrastructure, with Marvell building the connective foundation to make it possible.