CEOInterviews.AI
Start App
Get an email when Matthew Murphy gives a new interview
One email per new interview. Unsubscribe in one click.
Matthew Murphy
Chief Executive Officer & Chairman, Marvell Technology

Marvell CEO Matt Murphy Keynote at COMPUTEX 2026

📅 Jun 09, 2026 Marvell Technology 56 MIN 3578 VIEWS 14 SEGMENTS · 4 SPEAKERS
At COMPUTEX 2026, Marvell Chairman and CEO Matt Murphy will deliver a high-visibility keynote address: Scaling Data Center Infrastructure to Drive AI Innovation. In his address, Murphy discusses how Marvell technologies enable every critical link in modern AI data center infrastructure allowing customers to deploy AI-optimized systems with unprecedented performance and scale. He also highlights how Marvell and its broad partner ecosystem help the industry unlock the next wave of AI innovation. #compute2026 #datacenter #ai #hyperscalers #compute2026 #custom #customcompute #optical #copper #in...

What Matthew Murphy said

Written from the verified transcript and checked against it. Every figure links to the moment it was said.

Matthew Murphy, CEO of Marvell Technology, opened his Computex 2026 keynote by emphasizing that AI infrastructure performance is fundamentally a connectivity challenge, not just compute or memory. He recounted Marvell's decade-long transformation into a pure-play data infrastructure semiconductor company, including $36 billion in acquisitions and a leap to 5nm. Murphy highlighted Marvell's partnership with Nvidia, which invested $2 billion, and their collaboration on NVLink Fusion and optics. He detailed Marvell's connectivity portfolio: coherent optics for long distances, PAM4 for inside data centers, a new 100T Ethernet switch, and SerDes at 200G with 400G demonstrated. He explained the 'copper wall', at 400G, copper cannot fully connect a rack, driving optics inside the rack via co-packaged optics (CPO). He showcased a 51.2T CPO switch with 16 3.2T optical engines. Murphy also discussed Taiwan's unique semiconductor ecosystem with ASE's Tien Woo, noting its 40 years of experience and 350,000 semiconductor employees.

Key takeaways

  1. Marvell will sample the world's first 1.6T 2nm coherent solution later this year.
  2. Marvell announced a new 100T Ethernet switch for AI data centers with the industry's lowest power.
  3. At 400G, copper can no longer fully connect a rack, pushing optics inside via co-packaged optics.
  4. Marvell has demonstrated 400G SerDes and leads in 200G SerDes.
  5. Marvell's data center business is over 75% of revenue, growing rapidly.

Numbers and commitments

FigureWhat it refers toTypeAt
$36 billion Total investment in acquisitions like Cavium, Avera, Inphi metric 1:43
75% Share of revenue from data center business metric 1:43
$2 billion Nvidia's investment in Marvell commitment 1:43
1.6T World's first 2nm coherent solution to sample later this year timeline 27:14
100T New Ethernet switch for AI data centers other 27:14
200G Leading SerDes speed metric 27:14
400G Demonstrated SerDes speed metric 27:14
2.5 meters Cable length at 200G, at rack limit metric 27:14
51.2T CPO-based switch with 16 3.2T optical engines other 27:14
350,000 Semiconductor employees in Taiwan metric 49:29

Chapters

  1. 0:00Connectivity as the core of AI infrastructure
  2. 19:34Nvidia's view on agents and connectivity
  3. 25:28Copper to optics transition
  4. 27:14Marvell's connectivity portfolio
  5. 46:58Taiwan ecosystem discussion with ASE
  6. 51:30All-optical data center future
Host 1:26 ↗
Please welcome Marvell Chairman and CEO Matt Murphy.
Matthew Murphy 1:43 ↗
Morning everybody. It's great to be here to kick off day one at Computex and to be back in Taiwan. The first time I came here was nearly 30 years ago, and many of the young technology companies I visited then have become world leaders. Today, the future of AI infrastructure is being built here. What defines the performance of AI infrastructure? It's not just the processor or process node, nor just memory. At scale, computing is fundamentally a connectivity challenge. The next major wave of innovation will come from the underlying connectivity of these systems, as connections move from copper to optical. At Marvell, we've been preparing for this for nearly a decade. When I joined as CEO 10 years ago, we had a vision to build a pure-play company focused on semiconductor solutions for data infrastructure. We divested non-core businesses and made strategic acquisitions like Cavium, Avera, Inphi, and others, investing roughly $36 billion. We skipped 7nm and jumped to 5nm, becoming a process node leader. Today, our data center business is over 75% of revenue, growing rapidly. The next phase is all about connectivity. We've seen the industry solve compute and memory bottlenecks; now connectivity is the next challenge. Our customers are reimagining network architectures, and optics is the way forward. Marvell is unique: the vast majority of our revenue comes from connectivity. We partner deeply with compute and memory companies, and we are the Switzerland of the industry. A great example is our strategic partnership with Nvidia, which invested $2 billion in Marvell. We're expanding our partnership across optics, photonics, and NVLink Fusion. I'm thrilled to welcome Jensen Huang to the stage.
Jensen Huang 19:34 ↗
Yeah, that's really great. Useful AI has arrived, which is why demand is through the roof. This new computing pattern is called agents, which is disaggregated and distributed. That requires connectivity, which is why Matt and Marvell are doing so well. We've distributed and disaggregated computing across enormous clusters, and connectivity makes it possible.
Matthew Murphy 20:35 ↗
We're seeing it. The next trillion-dollar company? We got a little work to do, but we're on our way. Let's talk about scale. We used to talk about tens of GPUs; now thousands, maybe millions. As you scale compute and connectivity, how do you think about connectivity across and within data centers, and what technologies are important?
Jensen Huang 21:10 ↗
At the foundation, the agent computing pattern requires an orchestration system for large language models to think, reason, plan, and use tools. That requires a lot of connectivity. We introduced Vera Rubin for running agents, which includes Vera CPUs for orchestration and Vera CX for storage acceleration. We also partner on NVLink Fusion, which allows customers to use the same system architecture with their custom chips and our interconnect silicon photonics. This creates a disaggregated, distributed, heterogeneous data center.
Matthew Murphy 23:19 ↗
I think about the partnership and the investment, which we really appreciate. NVLink Fusion and working together on optics—the era of agents and your new platform is ideally suited. We had this idea years ago, but now the time is right for Marvell and Nvidia to enable our customers to have flexibility and scale our platforms together.
Jensen Huang 24:34 ↗
Ultimately, if you buy nothing but Nvidia, it's okay. But if you must design your own ASIC, we're happy to have Nvidia inside that data center. Between us, you have the benefit of a general-purpose, high-efficiency system with Vera Rubin, and you can specialize as well. That's why Nvidia is in AWS and Marvell is in all the clouds.

7 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Murphy, M. (2026, June 9). Marvell CEO Matt Murphy Keynote at COMPUTEX 2026 [Interview transcript]. Marvell Technology. CEOInterviews.AI. https://ceointerviews.ai/interview/1050997/

MLA

Matthew Murphy. "Marvell CEO Matt Murphy Keynote at COMPUTEX 2026." Marvell Technology, 9 Jun. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/1050997/.

BibTeX
@misc{murphy2026_1050997,
  author       = {Matthew Murphy},
  title        = {Marvell CEO Matt Murphy Keynote at COMPUTEX 2026},
  howpublished = {Interview transcript, Marvell Technology. CEOInterviews.AI},
  year         = {2026},
  month        = {jun},
  url          = {https://ceointerviews.ai/interview/1050997/},
  note         = {Speaker-attributed transcript with timestamps}
}