Beyond Bandwidth Podcast

Beyond Bandwidth Season 2, Episode 1: AI Data Centers Introduction

 

Introduction to the AI Data Center podcast series

[00:04] Roy Chamberlain: Hello everyone, and welcome to Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[00:13] Roy: I am your host from Leviton Network Solutions, Roy Chamberlain.

[00:17] Roy: We have a brand new season for you where we uncover the mystery of the AI data center.

[00:23] Roy: These data centers are the lifeblood of artificial intelligence applications and in this series, we will focus on the design and infrastructure from the physical layer perspective of these unique works of technology.

[00:36] Roy: For our first episode, we are joined by today's subject matter expert, Leviton's own Mike Connaughton.

 

Meet the Expert: Mike Connaughton of Leviton Network Solutions

[00:45] Roy Chamberlain: Well, here we are again, Mike. Welcome to Beyond Bandwidth.

[00:48] Mike Connaughton: Thanks, Roy, I appreciate it.

[00:50] Roy: It's been a little while since last season. This season you have a very significant role because of your background, relationships, and job in the industry. Can you tell us a little bit about your role at Leviton?

[01:06] Mike: Sure. Technically, I'm a Senior Product Manager, but that's not really what I do day to day. From a strategic standpoint, I focus on major industry trends, particularly data centers, and help prepare Leviton with the right products and services to meet future demands.

[01:44] Mike: I also manage technical alliances with companies such as Cisco and Arista. We exchange insights about developments in cabling, optics, switching, and emerging technologies.

[02:24] Roy: That's a good explanation. Working with you, I see those influences and the information gathered through your industry relationships and associations.

 

What You'll Learn about AI Data Centers this Season

[02:43] Roy Chamberlain: This season focuses on AI data centers. What do you hope listeners will learn?

[03:14] Mike Connaughton: AI is the topic everyone is discussing right now. Growth in the AI space has been incredible over the last few years. The pace of change is so fast that standards organizations simply can't keep up.

[03:59] Mike: Our goal is to speak with experts from across the industry and gather their perspectives on how they're helping organizations implement AI in data centers.

[04:21] Mike: Eventually standards and best practices will emerge, but we're not there yet. Right now there are many different approaches to solving the same problems.

[04:39] Mike: Hopefully listeners will hear a variety of ideas and find approaches that are relevant to their own environments.

 

AI Data Center Experts and Industry Perspectives

[04:49] Roy Chamberlain: What types of guests can listeners expect this season?

[05:07] Mike Connaughton: We'll have experts from manufacturing, distribution, consulting, and market intelligence organizations.

[05:34] Mike: Consultants often have a unique bird's-eye view of all the moving parts in AI data center projects.

[05:56] Mike: We'll also discuss supply chain challenges. Getting products where they need to be, when they need to be there, remains one of the industry's biggest challenges.

[06:16] Mike: We'll speak with people who analyze market trends and gather intelligence to help understand how these developments affect everyday decision-making.

 

How AI Data Centers Differ from Traditional Data Centers

[06:41] Roy Chamberlain: AI data centers are groundbreaking and evolving quickly. How does this compare to previous technology transitions?

[07:26] Mike Connaughton: The biggest difference is the number of disciplines involved. Everyone is asking questions and collaborating because so much is changing simultaneously.

[07:41] Mike: Five years ago, most people understood their role within a traditional data center environment. Today, there are far more unknowns.

[08:06] Mike: One of the biggest concerns in AI data centers is power consumption driven by advanced GPUs.

[08:19] Mike: That leads directly to changes in cooling technology, which in turn impacts cabling infrastructure.

[08:33] Mike: For example, liquid immersion cooling introduces entirely new considerations. Now we're asking whether networking products can operate while submerged in fluid, what temperatures they'll encounter, and what lifespan expectations will be.

[08:55] Mike: Everything inside the data center is changing. The challenge is understanding how those changes affect one another.

 

AI Data Center construction trends and infrastructure challenges

[09:30] Roy Chamberlain: Are organizations building AI infrastructure in existing facilities, or are they building new data centers?

[09:42] Mike Connaughton: So far, most AI deployments are being built in new spaces.

[09:46] Mike: However, AI data centers mean different things to different organizations. For some, it might be a single rack. For others, it could be a massive facility like the AI-focused projects making headlines.

[10:16] Mike: The largest organizations are leading AI infrastructure development today, while many enterprise data center operators are watching closely and preparing for AI technologies to become more accessible.

[10:38] Mike: There are many unanswered questions, and hopefully this season will help address some of them.

 

How Leviton Tracks AI Data Center Technology Trends

[10:55] Roy Chamberlain: How do Leviton's product managers stay ahead of rapidly evolving technologies like AI data centers?

[11:29] Mike Connaughton: We start by identifying major industry trends and assigning ownership for tracking those trends across the organization.

[12:02] Mike: Data centers are one of those focus areas, and AI data centers are a major subset of that category.

[12:14] Mike: The process involves researching trends, speaking with experts, working with customers, and gathering insights from the marketplace.

[12:29] Mike: We then share that information across product teams so they can adapt products and strategies accordingly.

[13:12] Mike: We rely on a combination of customer conversations, industry events such as BICSI, OFC, and Data Center World, along with feedback from field teams, sales professionals, strategic account managers, and specification engineers.

 

Season Preview: AI Data Center Panel Discussions and Expert Interviews

[13:51] Roy Chamberlain: That's a great explanation. It gives us valuable insight into how the organization stays informed. What can listeners expect in the upcoming episodes?

[14:05] Mike Connaughton: At a recent BICSI event in Florida, we hosted a panel discussion that became part of this podcast series.

[14:19] Mike: The presentations and audience Q&A provided excellent insights into the current state of AI data centers and serve as a great kickoff for the season.

[14:33] Roy: That's great. I'm ready to get started. I look forward to working with you this season, and we'll see you in the first episode.

 

About Beyond Bandwidth by Leviton Network Solutions

[14:44] Roy Chamberlain: Beyond Bandwidth is a podcast series produced by Leviton Network Solutions.

[14:48] Roy: For more information about today's topics, enterprise and data center network infrastructure content, and additional episodes, visit leviton.com/beyondbandwidth or listen wherever you get your podcasts.

[15:08] Roy: If you enjoyed this episode, we'd appreciate an honest review in your podcast application and your word-of-mouth recommendations.

[15:15] Roy: Leaving a review helps others discover the show.

[15:18] Roy: I'm Roy Chamberlain, and this was Beyond Bandwidth.

[15:21] Roy: Until next time.

Season 2, Episode 2: AI Data Center Reference Architecture

 

Introduction

[00:04] Roy Chamberlain: Hello, everyone, and welcome to Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[00:13] Roy: I'm your host from Leviton Network Solutions, Roy Chamberlain.

[00:17] Roy: This conversation shines the spotlight on AI data centers.

[00:21] Roy: Back in February, we recorded a panel discussion just outside the BICSI Winter Conference in Orlando.

[00:29] Roy: The panel consists of some of the top internationally recognized data center technology experts in our industry.

[00:36] Roy: Leviton's Mike Connaughton was the host for this discussion.

[00:40] Roy: In Part 1 of this four-part special edition to our series, we hear from Jonathan Jew with J&M Consulting. The topic: AI data center reference architecture.

 

Meet the Expert: Jonathan Jew

[00:54] Roy Chamberlain: Our first panelist is Jonathan Jew with J&M Consultants. He's based out of San Francisco.

[01:00] Roy: Jonathan has over 30 years of experience with more than 100 data center projects totaling well over 2 million square feet.

[01:07] Roy: He's an internationally recognized expert in data center telecommunications infrastructure and telecommunications administration.

[01:15] Roy: He holds leadership positions on several U.S. national and international standards committees and has received awards from organizations including BICSI, TIA, and ANSI.

[01:25] Roy: For today's discussion, his most notable role is serving as editor of the TIA-942 Standard on data center infrastructure.

[01:34] Roy: Today, Jonathan will discuss cabling and architecture changes that support AI clusters.

[01:40] Roy: Without further ado, Jonathan Jew.

 

BICSI Data Center Standards and Industry Experience

[01:42] Jonathan Jew: Great. Thanks a lot, Mike. I appreciate this opportunity to speak to you all.

[01:46] Jonathan: Since we're at a BICSI conference, I'd be remiss if I didn't mention that I'm also the Chair of the BICSI Data Center Design Subcommittee, which published BICSI 002, the industry's data center design best practices standard.

[01:58] Jonathan: We published that in 2024 at about the same time that TIA-942-C came out.

[02:04] Jonathan: I've worked on a lot of data centers over the past 30-plus years.

[02:15] Jonathan: I've been working on data centers since the mainframe days back in the 1980s, and the industry has changed quite a bit.

[02:24] Jonathan: Most of my current work involves designing data center cabling and physical infrastructure, such as pathways, cabinets, and racks for AI data centers.

 

AI Data Center Reference Architectures

[02:35] Jonathan Jew: The majority of the AI market includes three major GPU manufacturers: NVIDIA, AMD, and Intel.

[02:47] Jonathan: NVIDIA has the vast majority of the market. AMD has a much smaller share, and Intel is also a player.

[02:56] Jonathan: All three manufacturers provide what they call reference architectures, which are recommended network designs for servers using their GPUs.

[03:10] Jonathan: These architectures help ensure reliable performance, consistent operation, and scalability as GPU networks grow.

[03:21] Jonathan: While the vendors have slightly different approaches, all three share the same four primary network types.

 

GPU Network Design for AI Clusters

[03:32] Jonathan Jew: The primary network is the GPU network, which operates as a non-blocking network.

[03:44] Jonathan: In NVIDIA's architecture, each server typically contains eight GPUs and eight 400G ports connected to a non-blocking network, usually a fat-tree architecture.

[04:05] Jonathan: AMD similarly provides eight GPUs with eight 400G connections.

[04:12] Jonathan: Intel's Gaudi 3 architecture uses six GPU network connections running at 800 gigabits per second.

[04:29] Jonathan: NVIDIA has already announced switches and network interface cards that support 800G.

[04:40] Jonathan: We expect NVIDIA and AMD to transition from 400G to 800G connections as the market evolves.

 

AI Storage Network Architecture

[04:54] Jonathan Jew: The second network is the storage network, which connects servers to storage systems through dedicated storage servers.

[05:03] Jonathan: In NVIDIA and AMD architectures, that's typically two 400G connections.

[05:08] Jonathan: In Intel's architecture, it's two 100G connections.

[05:18] Jonathan: NVIDIA uses a four-to-three network with some level of aggregation.

[05:28] Jonathan: AMD allows aggregation but doesn't explicitly specify the level.

[05:41] Jonathan: Intel requires the storage network to remain non-blocking.

 

In-Band and Out-of-Band Management Networks

[05:48] Jonathan Jew: The third network is the in-band management network.

[05:51] Jonathan: This network manages server software and consists of two 100G connections across all three GPU vendor architectures.

[06:01] Jonathan: NVIDIA and AMD generally utilize aggregation, while Intel prefers a non-blocking design.

[06:12] Jonathan: The fourth network is the out-of-band management network used to manage physical server hardware.

[06:24] Jonathan: These typically run through 1G ports and are usually copper connections.

[06:30] Jonathan: In addition, most environments deploy a fifth network that connects switches together using serial console connections running at relatively low speeds.

 

AI Server Connectivity Requirements

[06:42] Jonathan Jew: AI servers generally include:

• GPU network connections running at 400G or 800G (six to eight links)
• Storage network connections (two links)
• In-band management connections (two 100G links)
• Out-of-band management connections (typically one 1G copper link)

 

Fiber Optics and MPO Connectivity for AI Data Centers

[07:08] Jonathan Jew: On GPU and storage networks running at 400G, we exclusively use array connectors.

[07:14] Jonathan: Whether using multimode or single-mode fiber, we're deploying angled connections.

[07:23] Jonathan: Historically, that wasn't always the case, but today's GPU network transceivers require angled connections regardless of fiber type.

 

Non-Blocking Network Topology Explained

[07:36] Jonathan Jew: In a non-blocking network, switches typically have the same number of ports facing servers as they do facing upper-layer network tiers.

[07:48] Jonathan: That's one of the reasons AI clusters require such a large number of connections.

[08:05] Jonathan: A non-blocking architecture means the bandwidth available going up the network equals the bandwidth going down.

[08:11] Jonathan: For example, a 48-port leaf switch may dedicate 24 ports to servers and 24 ports to spine switches.

[08:23] Jonathan: Similarly, a spine switch might dedicate 24 ports downward to leaf switches and 24 ports upward to the super spine layer.

[08:29] Jonathan: That's a tremendous amount of fiber.

[08:32] Jonathan: We install thousands upon thousands of fiber connections in these environments.

[08:35] Jonathan: High-density fiber infrastructure is essential for AI deployments.

 

The Future of 800G and 1.6T Ethernet

[08:39] Jonathan Jew: Today's 800G connections typically use two angled MPO connectors.

[08:50] Jonathan: That requires 16 fibers across two MPO connections.

[08:52] Jonathan: When IEEE 802.3dj becomes available, scheduled for March 2026, 800G will operate over just eight fibers, though only in single-mode implementations.

[09:05] Jonathan: The benefit is significant because it cuts fiber requirements in half.

[09:13] Jonathan: It will also enable support for 1.6 terabit Ethernet over 16 fibers.

[09:24] Jonathan: That's going to be a major advancement.

 

Structured Cabling vs. Direct Attach Cables in AI Data Centers

[09:26] Jonathan Jew: When comparing structured cabling with direct attach cables (DACs) or active optical cables (AOCs), we generally use DACs and AOCs between leaf switches and servers because those distances are relatively short.

[09:41] Jonathan: However, from leaf-to-spine and spine-to-super spine layers, we recommend structured cabling.

[09:48] Jonathan: Structured cabling makes future upgrades from 400G to 800G easier because you don't need to remove and replace large quantities of active optical cables.

[09:57] Jonathan: It also helps with manageability.

[10:03] Jonathan: Once you reach the distribution layer, the amount of fiber becomes overwhelming.

[10:06] Jonathan: Without structured cabling, managing fiber capacity within optical fiber raceways becomes extremely difficult.

[10:10] Jonathan: For upgrades, troubleshooting, and capacity management, we strongly recommend structured cabling beyond the leaf-to-server connections.

 

About Beyond Bandwidth

[10:27] Roy Chamberlain: Beyond Bandwidth is a podcast series produced by Leviton Network Solutions.

[10:32] Roy: For more information on the topics discussed today, as well as additional enterprise and data center network infrastructure content, visit leviton.com/beyondbandwidth or listen wherever you get your podcasts.

[10:52] Roy: If you enjoyed this episode, we'd appreciate an honest review in your podcast application.

[10:59] Roy: Along with word of mouth, reviews help others discover the show.

[11:02] Roy: I'm Roy Chamberlain, and this was Beyond Bandwidth.

[11:05] Roy: Until next time.

Season 2, Episode 3: AI Data Center Fiber Technologies, Optics, and Emerging Infrastructure Trends

 

Introduction

[0:04] Roy Chamberlain: Hello everyone and welcome to Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[0:13] Roy: I am your host from Leviton Network Solutions, Roy Chamberlain.

[0:18] Roy: This conversation shines the light on AI data centers.

[0:22] Roy: In Part 2 of our four-part special edition to this series, Leviton's Mike Connaughton hosts the panel discussion as Brooke Ford of Leviton Network Solutions breaks down prevailing fiber optic and transceiver technologies. 

 

Meet the Expert: Brooke Ford, Director of Product Management

[0:38] Mike Connaughton: Bring up Brooke Ford, Director of Product Management at Leviton Network Solutions based in the Fuquay-Varina facility outside of Raleigh, NC.  Specifically, he runs product management for fiber cable and connectivity, where he's responsible for portfolio strategy and bringing valued fiber solutions to data center and enterprise customers.

[0:57] Mike: He has over 25 years of experience in technical roles in telecom, data centers, cable MSOs, U.S. federal government markets.

[1:06] Mike: And most importantly, he is my boss. So speak well of me, all of you. I would appreciate it.

[1:11] Mike: Brooke's going to talk about emerging technologies ranging from new fiber types to transceiver deployments.

[1:18] Brooke Ford: Yeah. Thanks, Mike. 

 

AI Data Center Network Scale and Fiber Density Challenges

[1:20] Brooke Ford: If we take an example, Jonathan talked about reference architectures. Looking at an NVIDIA SuperPOD, they break these down into scalable units consisting of eight racks with 32 nodes per scalable unit.

[1:38] Brooke: They can put four scalable units together for a SuperPOD.

[1:42] Brooke: You end up with about 32 racks, which doesn't seem like a lot in traditional data center terms.

[1:47] Brooke: But connecting compute, storage, in-band, and out-of-band networks generates tens of thousands of fibers and thousands of MTP connectors in a very small space.

[2:08] Brooke: Data center operators may only have one or two rack units available to make all of these connections.

[2:15] Brooke: These networks require significant scale and extremely high density to support AI workloads. 

 

Small Form Factor Fiber Connectors: Increasing Port Density

[2:22] Brooke Ford: One emerging solution is very small form factor connectors, particularly multi-fiber versions.

[2:36] Brooke: Two industry leaders are currently driving this technology.

[2:38] Brooke: Senko offers the SN-MT connector, while US Conec offers the MMC connector.

[2:44] Brooke: Unfortunately, those connector types are not compatible with each other.

[2:50] Brooke: These connectors have a lower profile and slimmer design than traditional MTP connectors.

[3:00] Brooke: Their biggest advantage is density. They can triple the density available from traditional MTP connectivity.

[3:07] Brooke: Today's ultra-high-density patch panels support 72 MTP connectors in 1RU.

[3:15] Brooke: Small form factor connectors increase that to 216 connectors in the same rack space.

[3:22] Brooke: Using 16-fiber connectors, that equals 3,456 fibers in a single rack unit.

Challenges of Ultra-High-Density Connectivity

[3:38] Brooke Ford: There are several deployment considerations.

[3:49] Brooke: Port access becomes more difficult when 216 connectors occupy a single rack unit.

[3:59] Brooke: Fiber management becomes critical to avoid cable congestion.

[4:09] Brooke: Labeling is also challenging because of limited front-panel space.

[4:25] Brooke: Availability remains a concern as manufacturers scale production to meet demand.

 

Multi-Core Fiber Technology for Future Data Centers

[4:40] Brooke Ford: Beyond small form factor connectors, another long-term solution for fiber density is multi-core fiber.

[4:53] Brooke: Traditional fiber contains a single core, whether it's 9-micron single-mode or 50-micron multimode fiber.

[5:01] Brooke: Multi-core fiber contains multiple cores within a single strand of glass.

[5:07] Brooke: Those fibers can contain two, four, or even eight cores.

[5:12] Brooke: For example, a 12-fiber array cord could effectively provide the equivalent of 48 fibers without increasing cable size.

Benefits and Limitations of Multi-Core Fiber

[5:31] Brooke Ford: The technology is currently seeing deployment primarily in submarine networks.

[5:45] Brooke:  While the fiber can be spliced and terminated, specialized fan-in/fan-out devices are required.

[6:02] Brooke: These devices separate the individual cores into usable fiber connections.

[6:13] Brooke: Manufacturing processes still require maturation.

[6:17] Brooke: Loss performance remains a challenge.

[6:27] Brooke: Costs also remain high.

[6:28] Brooke: Additional development is needed before the technology becomes mainstream within enterprise and hyperscale data centers. 

 

Reducing AI Data Center Power Consumption with Optical Technologies

Co-Packaged Optics (CPO)

[6:39] Brooke Ford: Power consumption continues to be one of the biggest challenges for AI data centers.

[6:46] Brooke: Optics and transceivers are among the largest contributors to power draw within networking equipment.

[7:04] Brooke: One emerging technology is Co-Packaged Optics, or CPO.

[7:10] Brooke: In a traditional Ethernet switch, data travels from the ASIC across a PCB trace to the transceiver.

[7:27] Brooke: The transceiver cleans, amplifies, and converts the electrical signal into an optical signal.

[7:37] Brooke: Co-Packaged Optics moves transceiver functionality directly into the switch ASIC.

Benefits of Co-Packaged Optics

[7:51] Brooke Ford: Benefits include improved signal performance.

[8:02] Brooke: Latency can be reduced because components are removed from the signal path.

[8:02] Brooke: Power consumption may be reduced by 30% to 50%.

Challenges Facing Co-Packaged Optics

[8:16] Brooke Ford: The technology is complex and often vendor-specific.

[8:25] Brooke: Thermal management, manufacturing, testing, and reliability remain significant challenges.

[8:35] Brooke: Replacing failed optical channels is especially difficult because the optics are integrated into the switch architecture.

 

Linear Drive Pluggable Optics (LPO)

[8:54] Brooke Ford: A more near-term solution is Linear Drive Pluggable Optics.

[9:08] Brooke: LPO pursues many of the same power-saving goals as Co-Packaged Optics while maintaining removable transceiver modules.

[9:15] Brooke: It reduces power consumption by eliminating the digital signal processor from the transceiver.

[9:38] Brooke: This approach may reduce power consumption by 20% to 25%.

Industry Adoption and Challenges for LPO

[9:45] Brooke Ford: LPO technologies are supported through Multi-Source Agreements (MSAs) designed to promote interoperability.

[10:05] Brooke: However, removing the DSP can negatively impact performance and reach.

[10:14] Brooke: The technology still faces manufacturing and thermal challenges.

[10:27] Brooke: Market forecasts suggest LPO transceivers may account for only about 10% of total transceiver shipments.

 

Hollow Core Fiber: Solving Latency Challenges in AI Networks

[10:46] Brooke Ford: Reducing latency remains a top priority for AI and machine learning networks.

[10:57] Brooke: Hollow core fiber is one of the most promising technologies under development.

[11:05] Brooke: Unlike conventional fiber with a solid glass core, hollow core fiber contains open space within the core.

[11:31] Brooke: Because light travels fastest through a vacuum-like environment, transmission speeds are significantly improved.

[11:40] Brooke: Traditional fiber requires approximately five microseconds per kilometer. Hollow core fiber reduces that to roughly 3.3 microseconds per kilometer.

[11:51] Brooke: That represents approximately a 46% improvement in transmission speed.

Challenges of Hollow Core Fiber Deployment

[11:55] Brooke Ford: Hollow core fiber can be spliced and terminated but requires highly specialized techniques.

[12:07] Brooke: The technology remains expensive and loss sensitive.

[12:15] Brooke: Manufacturing costs are orders of magnitude higher than standard fiber.

[12:20] Brooke: Production volumes remain low.

[12:26] Brooke: Most hollow core solutions currently support relatively low fiber counts.

[12:41] Brooke: Further technology maturation is needed before mainstream deployment becomes practical. 

 

Microsoft's Investment in Hollow Core Fiber

[12:49] Brooke Ford: Despite these challenges, hollow core fiber offers additional advantages.

[13:03] Brooke: Microsoft recognized this potential when it acquired Lumenisity, a leader in hollow core fiber technology.

[13:14] Brooke: One key advantage is the ability to expand the geographic footprint of data centers while maintaining latency requirements.

[13:20] Brooke: By locating facilities farther apart, organizations can reduce land and power costs.

[13:35] Brooke: These operational savings may offset the higher cost of the fiber itself.

[13:41] Brooke: Microsoft continues working with the Lumenisity team to further mature the technology. 

 

Conclusion and Additional Resources

[13:54] Roy Chamberlain: Beyond Bandwidth is a podcast series produced by Leviton Network Solutions.

[13:58] Roy: For more information on the topics discussed today, as well as additional enterprise and data center networking content, or to listen to all available episodes of Beyond Bandwidth, visit leviton.com/beyondbandwidth or your preferred podcast platform.

[14:18] Roy: If you enjoyed this episode, we'd appreciate an honest review in your podcast application. Along with word of mouth, reviews help others discover the show.

[14:28] Roy: I'm Roy Chamberlain and this was Beyond Bandwidth.

[14:31] Roy: Until next time.

Season 2, Episode 4: AI Data Center Cooling Technologies and Challenges

 

Introduction

[00:04] Roy Chamberlain: Hello everyone and welcome to Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[00:13] Roy: I am your host from Leviton Network Solutions, Roy Chamberlain.

[00:17] Roy: This conversation shines the light on AI data centers.

[00:21] Roy: In Part 3 of our four-part special edition series, Leviton's Mike Connaughton hosts the panel discussion as Nick Taylor of Networks Centre in the UK takes his turn by speaking on cooling technologies and challenges in the data center space.

 

Meet the Expert: Nick Taylor, Networks Centre UK

[00:41] Mike Connaughton: Following Jonathan Jew will be Nick Taylor, who is the Technical Director at Networks Centre in the UK.

[00:46] Mike: Nick provides pre-sales and technical assistance to ensure the right products and services are selected by talking to customers and understanding project requirements.

[00:55] Mike: With over 25 years in the industry and RCDD and TS certifications, along with regular standards committee attendance, he ensures customers receive the best advice.

[01:06] Mike: Pertinent to today's discussion, Nick spent a number of years working for a Professional Services Group on cooling optimization, using CFD to resolve cooling issues in existing data centers.

[01:16] Mike: Nick is going to be talking about developments in cooling technologies as the industry heads toward AI networks.

 

AI Data Center Infrastructure Challenges in Europe

[01:23] Nick Taylor: Thank you very much and thank you for inviting me here today to talk about this.

[01:27] Nick: I'd start by saying that the information I'm going to present is based on what we have in the UK, although it likely applies to other countries as well.

[01:34] Nick: Across Europe, there are significant problems with power availability.

[01:37] Nick: Many of you may have heard of FLAP-D: Frankfurt, London, Amsterdam, Paris, and Dublin.

[01:44] Nick: There's also Madrid.

[01:45] Nick: These financial centers of Europe are all saying they do not have enough power.

[01:51] Nick: Although today's discussion focuses on cooling, I want to briefly highlight some of the major challenges we face in the UK.

[01:58] Nick: The first is power generation and power infrastructure.

[02:00] Nick: It's not just producing power. It's delivering that power from generation sites to where it's needed.

[02:11] Nick: The UK's National Grid is aging and already under strain.

[02:13] Nick: Demand from air-source heat pumps, electric vehicles, and data centers continues to increase rapidly.

[02:21] Nick: The growth in power requirements is exponential.

[02:26] Nick: The UK is also working to meet Paris Agreement commitments.

[02:30] Nick: We cannot simply continue relying on fossil fuels.

[02:35] Nick: Renewables only work some of the time, so we're looking for technologies capable of carrying the base load.

[02:42] Nick: There are discussions around small modular reactors from companies such as Rolls-Royce.

[02:47] Nick: However, legislation and deployment timelines mean we're likely at least 10 to 20 years away from seeing widespread implementation.

 

Challenges of Supporting AI in Existing Data Centers

[02:56] Nick Taylor: Putting those broader issues aside, I want to focus on supporting AI and high-performance computing in existing data centers.

Data Center Weight and Physical Constraints

[03:13] Nick Taylor: One issue is weight.

[03:18] Nick: A rack full of immersion-cooled equipment weighs significantly more than a traditional air-cooled rack.

[03:32] Nick: Existing on-premises and colocation data centers must carefully choose locations that can physically support these systems.

AI Connectivity and Cabling Challenges

[03:45] Nick Taylor: Connectivity has already been covered by Jonathan.

[03:48] Nick: The sheer volume of structured cabling and point-to-point connections is substantial.

[03:52] Nick: Additional considerations include cable sheath materials, fiber connectivity, and challenges associated with systems that do not use Active Optical Cables (AOCs).

 

Data Center Cooling Technologies Explained

[04:07] Nick Taylor: Today I'll focus primarily on cooling capacity and infrastructure.

Air Cooling Fundamentals

[04:14] Nick Taylor: Most listeners will know this already, but the majority of data centers worldwide use air cooling because it's simple and generally viewed as safer than liquid cooling.

[04:35] Nick: A common misconception is that cooling issues can be solved simply by lowering temperature.

[04:43] Nick: Of temperature and airflow, airflow is actually the more important factor.

[04:50] Nick: Compared to air, water offers significantly higher heat capacity and more effective heat transfer, making it better suited for AI heat loads.

[05:00] Nick: Although many people dislike the idea of liquid throughout data centers, it's worth remembering that water-cooled systems were common decades ago.

[05:09] Nick: IBM mainframes were water cooled, and data centers were built with drip trays and drainage systems beneath raised floors.

[05:20] Nick: The concept isn't new. The challenge is that today's facilities were not originally designed for it.

 

Air Containment Strategies for Data Centers

Cold Aisle Containment

[05:37] Nick Taylor: Air containment separates supply air from return air.

[05:43] Nick: Containment can be implemented across aisles, rows, or individual racks.

[06:25] Nick: Before containment became common, many data centers had rows of racks with little effective separation between hot and cold air.

[07:08] Nick: In practice, hot and cold air often mixed together.

[07:19] Nick: That mixing reduces Delta T and significantly lowers cooling efficiency.

[07:34] Nick: The next evolution was cold aisle containment using doors and roofs to separate airflow streams.

[07:46] Nick: However, raised floors still created airflow inconsistencies.

[08:12] Nick: Measurements frequently showed unpredictable pressure differences across floor tiles.

[08:32] Nick: Before implementing cold aisle containment, airflow studies should be performed to ensure the design works as expected.

Hot Aisle Containment

[08:45] Nick Taylor: Hot aisle containment is now used in the vast majority of new UK data centers.

[08:57] Nick: Most are built directly on slab floors with overhead pathways for power and data.

[09:04] Nick: Cold air is typically supplied from one side of the room and exhausted through chimneys into a return plenum.

[09:32] Nick: This is the dominant cooling approach seen across modern facilities.

 

Rear Door Heat Exchangers for High-Density AI Workloads

[09:51] Nick Taylor: Rear door heat exchangers help support racks requiring 30, 40, or even 50 kilowatts, far beyond traditional designs.

[10:07] Nick: These systems extend cooling directly to the rack.

[10:16] Nick: Some versions are passive, while others use fans to improve cooling capacity.

[10:34] Nick: Certain implementations support more than 150 kilowatts per rack.

 

Direct-to-Chip Liquid Cooling

[10:45] Nick Taylor: Direct-to-chip cooling typically uses a Coolant Distribution Unit (CDU) located within the rack.

[10:59] Nick: Coolant is routed directly to processors and high-heat components.

[11:04] Nick: Air cooling is still required because many server components continue generating heat.

 

Immersion Cooling for AI Data Centers

[11:12] Nick Taylor: With immersion cooling, entire systems are submerged in dielectric liquid.

[11:19] Nick: It's still remarkable to see servers operating while immersed.

[11:24] Nick: The liquid is non-conductive, allowing systems to function safely.

[11:28] Nick: However, immersion introduces challenges with cable jackets, connector materials, and optical fiber systems.

[11:50] Nick: These concerns help explain the increased use of Active Optical Cables in immersion environments.

 

Can Existing Data Centers Support AI?

[12:00] Nick Taylor: The key question is whether current data centers can support AI workloads.

[12:09] Nick: Most colocation providers acknowledge they can only support limited AI deployments because power availability remains constrained.

[12:22] Nick: Many facilities are considering dedicating a small area to a handful of high-density AI racks.

[12:30] Nick: Physical weight, cable routing, and infrastructure limitations all affect placement decisions.

[12:52] Nick: Most providers are looking at limited AI clusters rather than large-scale deployments.

[13:02] Nick: Platforms such as NVIDIA GB200 systems are unlikely to be fully supported in many existing facilities.

[13:09] Nick: Existing data centers can support AI, but medium and high-density AI loads generally exceed what traditional air-cooling infrastructure was designed to handle.

[13:20] Nick: Since more than 99% of UK facilities use air cooling today, substantial changes will be required moving forward.

 

Beyond Bandwidth Podcast Resources

[13:34] Roy Chamberlain: Beyond Bandwidth is a podcast series produced by Leviton Network Solutions.

[13:38] Roy: For more information on today's topics and other enterprise or data center network infrastructure content, visit leviton.com/beyondbandwidth or listen wherever you get your podcasts.

[13:58] Roy: If you enjoyed this episode, we'd appreciate an honest review in your podcast application.

[14:05] Roy: Along with word of mouth, reviews help others discover the show.

[14:08] Roy: I'm Roy Chamberlain, and this was Beyond Bandwidth.

[14:11] Roy: Until next time.

Season 2, Episode 5: AI Data Centers Q&A: Audience Questions and Expert Panel Discussion

 

Introduction

[00:04] Roy Chamberlain: Welcome back to Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[00:12] Roy: I am your host from Leviton Network Solutions, Roy Chamberlain.

[00:16] Roy: This conversation spotlights AI data centers. In part four of our special four-part series, Leviton's Mike Connaughton hosts a discussion featuring audience Q&A with industry experts.

 

AI Data Center Power Availability and Regional Infrastructure Challenges

[00:34] Mike Connaughton: I'll start with a couple of questions. Nick, while discussing UK power limitations, you mentioned challenges similar to what's happening in Northern Virginia. In the U.S., Department of Energy studies suggest total power generation isn't the issue, but regional availability is. Are similar challenges occurring across Europe?

[01:33] Nick Taylor: The short answer is yes. Norway is a good example. There can be dramatic differences in power costs between regions, largely due to limitations in transmission infrastructure. Power generation isn't the only challenge. The ability to transport power efficiently is equally important.

[02:10] Nick: In the UK, for example, wind farms in Scotland sometimes generate more electricity than transmission systems can move south into England. As a result, power is occasionally wasted. Significant infrastructure investment is still needed across Europe.

 

Small Form Factor Fiber Connectors and AI Structured Cabling

[02:41] Mike Connaughton: Jonathan and Brooke, you both discussed emerging small form factor fiber connectors. How do you see them fitting into structured cabling and AI deployments?

Structured Cabling Benefits

[03:09] Jonathan Jew: High-density patching areas supporting spine and super-spine switches are ideal candidates for small form factor connectors. They reduce space requirements and increase switch density.

[03:39] Jonathan: Long term, it would be even better if transceiver manufacturers adopted these connectors. As speeds increase to 800G and 1.6T, having the same connector on both ends simplifies installation and reduces deployment errors.

Adoption on Transceivers

[04:38] Brooke Ford: Initially these connectors will primarily be used in patch panels to improve density. Over time we may see adoption directly on transceivers.

[05:17] Brooke: Co-packaged optics architectures could further accelerate adoption because optical connectivity moves closer to the ASIC and the switch becomes more like a patch panel.

 

Standards Development for Small Form Factor Connectors

[06:05] Mike Connaughton: How is TIA approaching standardization of these connectors?

[06:24] Jonathan Jew: Standards development was delayed due to patent disputes, but those issues have largely been resolved. TIA generally adopts IEC standards, so we're waiting for IEC work to progress.

[07:34] Mike: Duplex connector standards are restarting and should move relatively quickly. Array connector standards will likely take longer.

[07:53] Jonathan: Manufacturers are already releasing products, so adoption may happen faster than standardization.

 

Structured Cabling vs. Direct Connectivity in AI Data Centers

[08:10] Mike Connaughton: NVIDIA, AMD, and Intel all have different reference architectures. Do you expect structured cabling to influence future designs?

[08:33] Jonathan Jew: Early AI deployments are often delivered as turnkey solutions using direct connections. However, as deployments scale, structured cabling becomes essential.

[09:04] Jonathan: Most second-generation deployments we're working on now use structured cabling, particularly for leaf-spine architectures.

[10:18] Jonathan: Larger deployments cannot realistically scale with direct cabling alone because of volume, management complexity, and future upgrade requirements.

[11:00] Jonathan: Eventually AI operators will face the same lessons learned decades ago in networking: structured cabling improves manageability, scalability, troubleshooting, and upgrades.

 

Designing AI Infrastructure for Future Growth

[11:28] Mike Connaughton: What information helps you create effective AI network designs?

[11:48] Jonathan Jew: Understanding business objectives and documenting assumptions is critical. Customers should define not only their immediate needs but also future growth plans.

[12:33] Jonathan: If you're planning even two to three years ahead, you need to account for 800G networking and beyond.

[13:07] Jonathan: We design pathways, patch fields, and infrastructure with enough capacity to support future upgrades even if customers don't initially deploy all required fiber.

[13:49] Jonathan: In one project, we needed to accommodate approximately 70,000 fibers entering a main distributor. Planning for scale is everything.

 

Cooling AI Data Centers: Liquid Cooling, Rear Door Heat Exchangers, and Immersion

[15:10] Mike Connaughton: Rear door heat exchangers are becoming more common, but many see them as a temporary solution. Where do you see cooling technology heading?

[15:54] Nick Taylor: Long term, immersion cooling appears inevitable. Looking ten years into the future, cooling AI clusters with air alone won't be feasible.

[16:41] Nick: The exact implementation may vary depending on architecture. Rear door heat exchangers remain useful today, especially in existing facilities.

[17:19] Nick: They're effective for many current workloads but probably aren't the ultimate long-term solution.

 

Co-Packaged Optics vs. Linear Drive Optics

[17:44] Mike Connaughton: Brooke, is linear drive optics the stepping stone toward co-packaged optics?

[18:03] Brooke Ford: Yes. Linear drive optics are likely the intermediate step. Demonstrations have already shown successful 800G performance, and the industry is moving toward 1.6T implementations.

[18:45] Jonathan Jew: Doesn't co-packaged optics effectively eliminate transceivers?

[18:57] Brooke: Yes, it does. DACs disappear, and traditional AOCs would largely disappear as well.

 

Hollow Core Fiber: Future Opportunities and Challenges

[19:55] Nick Taylor: Brooke, what is the future of hollow core fiber? Could connectorization improve enough to avoid transitioning back to traditional glass fiber?

[20:54] Brooke Ford: That's one of the industry's biggest questions. If practical connector solutions emerge, hollow core fiber becomes much more attractive, but significant technical challenges remain.

[21:16] Brooke: Realistically, widespread deployment is likely still years away.

 

Direct-to-Chip Cooling and Rack-Level CDUs

[21:23] Brooke Ford: Are customers concerned about rack space being consumed by cooling distribution units (CDUs)?

[21:53] Nick Taylor: Not significantly. Higher-density AI racks naturally reduce the total number of racks required, making room for supporting infrastructure.

[22:21] Nick: Some organizations remain uncomfortable with water-based cooling systems inside racks, but direct liquid cooling is becoming a permanent part of the industry.

 

The Future of AI Data Centers vs. Traditional Data Centers

[22:58] Mike Connaughton: Five years from now, what percentage of the market do you think will be AI data centers?

[23:45] Jonathan Jew: Most AI model training will likely remain cloud-based, while inference workloads will become widespread across enterprises, edge locations, and even personal devices.

[24:57] Jonathan: More efficient open-source AI models could make private AI infrastructure affordable for far more organizations.

[26:03] Jonathan: Greater accessibility and openness are good for the industry because they reduce dependence on a few major providers.

 

Addressing AI Latency Concerns with Structured Cabling

[27:32] Mike Connaughton: Structured cabling is sometimes criticized for adding latency. What are your thoughts?

[27:56] Jonathan Jew: The latency impact of additional connectors is extremely small and generally insignificant.

[28:42] Jonathan: Large-scale cloud providers already use structured cabling extensively because direct-connect approaches don't scale to tens of thousands of GPUs.

 

AI Network Architecture: Top-of-Rack vs. Middle-of-Row Switching

[29:18] Mike Connaughton: As AI servers continue increasing GPU density, does top-of-rack networking still make sense?

[29:53] Jonathan Jew: GPU networking introduces complexities because rail-optimized architectures often require extensive cross-rack connectivity.

[31:28] Jonathan: That's why many NVIDIA designs place switching equipment in dedicated middle-of-row cabinets rather than traditional top-of-rack positions.

 

Raised Floors vs. Slab Construction for Liquid Cooling

[33:05] Mike Connaughton: How will liquid cooling affect the use of raised floors?

[33:49] Nick Taylor: Existing facilities often route cooling infrastructure below raised floors, but increasingly dense deployments create weight concerns.

[35:23] Nick: Many raised floors simply weren't designed for the loads associated with immersion cooling tanks and modern high-density equipment.

[36:20] Nick: Planning cable pathways and plumbing access has become much more complicated than in traditional data centers.

 

Cooling Fluids and Environmental Considerations

[37:32] Mike Connaughton: Are there developments in cooling fluids that could improve cooling efficiency or environmental sustainability?

[38:05] Nick Taylor: That's outside my primary area of expertise.

[39:03] Nick: Several companies are developing dielectric immersion fluids. These solutions offer long operational lifespans and may improve cooling efficiency, but the industry is still evaluating options.

 

Cabling Considerations for Immersion Cooling

[39:51] Jonathan Jew: Immersion cooling creates new challenges for cabling.

[40:13] Jonathan: Optical connectors and transceivers can experience reliability issues if immersion fluids enter connector interfaces.

[40:39] Jonathan: Every new immersion fluid requires compatibility testing with cable jackets, connectors, and other network infrastructure components.

[41:04] Jonathan: It's an exciting area, but it introduces many new considerations for network designers.

 

Key Takeaways: The Future of AI Infrastructure

• Regional power transmission is becoming a critical AI data center challenge.
• Small form factor connectors will play an increasingly important role in high-density AI environments.
• Structured cabling is becoming the preferred long-term architecture for large AI deployments.
• 800G and 1.6T networking are rapidly approaching mainstream adoption.
• Liquid cooling and immersion cooling are expected to become increasingly important as AI densities grow.
• Open-source AI models may broaden access to private AI infrastructure.
• Cabling standards and AI reference architectures continue to evolve alongside the industry.

 

Closing

[41:45] Mike Connaughton: Thank you to everyone who participated, especially Jonathan, Nick, and Brooke.

[42:03] Roy Chamberlain: Beyond Bandwidth is produced by Leviton Network Solutions. For more information and additional episodes, visit leviton.com/beyondbandwidth.

[42:37] Roy: I'm Roy Chamberlain, and this was Beyond Bandwidth. Until next time.

Season 2, Episode 6: AI Data Center Cabling Design and Infrastructure 

 

Introduction

[0:03] Roy Chamberlain: Here we are again at Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[0:12] Roy: I am your host from Leviton Network Solutions, Roy Chamberlain.

[0:17] Roy: After completing the special edition panel discussion in our last episode, we sit down with Peter Helfrich, my colleague and fellow specification engineer here at Leviton Network Solutions.

[0:29] Roy: We jump right back in, steering our focus to go deeper into the structured cabling design and implementation for AI data centers.

 

Meet Peter Helfrich: Data Center Specification Engineering Expert

[0:40] Roy Chamberlain: So Peter, tell us a little bit about your role in the industry at Leviton Network Solutions.

[0:46] Peter Helfrich: So I'm fortunate to be part of the specification engineering group as you are, Roy.

[0:52] Peter: We are technical and business experience folks who support consultants, engineers and users in helping them understand our product set, technologies and to help our salespeople deliver those messages.

[1:09] Peter: My particular role has been to focus on international opportunities and business and clients that are doing work overseas to help interpret the specifications, standards and product choices that are different outside the United States.

[1:31] Peter: It's been an exciting 12 years here with Leviton.

[1:35] Peter: I find it challenging, but enjoyable, helping people understand technologies and develop solutions.

 

Personal Background and Industry Experience

[1:48] Roy Chamberlain: Yeah, being on the spec team is definitely a roller coaster.

[1:59] Roy: So now a little off topic, but are you a New York Yankees fan?

[2:03] Peter Helfrich: Oh, no, no, no. Let's go, Mets.

[2:06] Roy: Let's go Mets, huh?

[2:08] Roy: Have you been a lifelong Mets fan?

[2:09] Peter: Yeah. When I was 12 years old was the magical season in New York...

 

What Makes an AI Data Center Different?

[2:38] Roy Chamberlain: So back to our topic of AI data centers.

[2:45] Roy: What is your knowledge and experience when it comes to AI data centers?

[2:52] Peter Helfrich: I've been involved in designing, supplying, and building infrastructure for data centers for decades.

[3:19] Peter: More recently, with AI data centers, I've worked with global accounts that are developing and building AI infrastructure.

[4:17] Roy: In your experience, what differentiates an AI data center from other data centers?

[4:27] Peter: I think what's different is the amount of cabling that's involved.

[4:34] Peter: AI clusters require extensive mesh networking. It's not just one fabric.

[5:03] Peter: To build these neural networks or AI clusters, there's a lot of mesh cabling going on.

[5:43] Peter: With AI, there can be four or six different networks overlaid on top of each other.

[6:04] Peter: It gets more complicated in both the number of cables and their different purposes.

 

Can Existing Data Centers Be Converted for AI?

[7:14] Roy Chamberlain: Can you implement NVIDIA or other AI equipment into an existing data center structured cabling plant?

[7:23] Peter Helfrich: Probably not.

[7:30] Peter: Some existing MPO trunk cabling could potentially be repurposed.

[7:39] Peter: But the density and network design are different.

[8:01] Peter: When organizations repurpose a data hall for AI, they typically rip out and replace most of the cabling infrastructure.

[8:14] Peter: The other piece of the puzzle is that AI power densities are significantly higher.

[8:22] Peter: AI racks may consume 30 to 40 kW per rack versus 5 to 7 kW in traditional environments.

 

Major Players in the AI Data Center Market

[8:55] Roy Chamberlain: So who are the big players today in the AI market?

[9:03] Peter Helfrich: There are different categories of players.

[9:19] Peter: NVIDIA pioneered the use of GPUs for AI processing.

[9:56] Peter: NVIDIA still owns a large portion of the GPU market.

[10:08] Peter: AMD and Intel are developing competing GPU solutions.

[10:21] Peter: Server manufacturers include NVIDIA, HPE, Dell, and Supermicro.

[10:35] Peter: Networking vendors include NVIDIA, Arista, and Cisco.

[10:39] Peter: Major AI service providers include AWS, Azure, Meta, Tesla, Microsoft, and Oracle.

 

Challenges in AI Data Center Cabling Design

[11:11] Roy Chamberlain: What are some of the biggest challenges you encounter designing the physical cabling infrastructure?

[11:29] Peter Helfrich: Getting all the information from stakeholders is key.

[11:38] Peter: Networking and software teams often think differently than physical infrastructure designers.

[12:10] Peter: We often need to ask additional questions because documentation doesn't always contain everything required for a cabling design.

[12:39] Peter: We also need to determine whether to use trunks or discrete cables.

[12:50] Peter Helfrich: Not all 400G connections are the same.

[13:12] Peter: Another challenge is the aggressive construction schedule for these high-value projects.

 

AI Data Center Installation and Specialty Contractors

[13:58] Roy Chamberlain: Once the design is complete, who installs these systems?

[14:15] Peter Helfrich: Large data centers and AI facilities often rely on specialty contractors.

[14:47] Peter: AI cabling environments require expertise with MPO connectivity, testing, and deployment.

[15:06] Peter: Contractors must also understand staging strategies for large-scale data center builds.

 

AI Data Center Design Documentation and Project Planning

[15:31] Roy Chamberlain: What design documents do contractors receive?

[15:43] Peter Helfrich: They typically receive a package beginning with the data hall floor plan.

[15:58] Peter: They also receive cabling schematics and fabric-specific documentation.

[16:18] Peter: There may be separate documentation for compute, storage, in-band management, and out-of-band management networks.

[16:32] Peter: The project may also include a manufacturer bill of materials.

 

Lessons Learned from Early AI Data Center Deployments

[17:10] Roy Chamberlain: Have there been lessons learned from the industry's early AI implementations?

[17:38] Peter Helfrich: Early infrastructure relied heavily on direct-connect cabling.

[18:05] Peter: Many deployments initially bypassed patch panels entirely.

[18:24] Peter: There was concern that adding patching points would increase latency.

[18:52] Peter: In reality, patch panels introduce insertion loss rather than latency.

[19:29] Peter: Today, structured cabling is increasingly used to improve cable management and scalability.

 

How Product Managers Support AI Data Center Innovation

[20:07] Roy Chamberlain: How do Leviton's product managers support AI data center projects?

[20:29] Peter Helfrich: Product managers are eager to understand what customers are doing.

[20:45] Peter: The specification team often encounters new requirements first.

[20:58] Peter: We collaborate closely with product management to address gaps and develop new products.

[21:11] Peter: This collaboration helps solve customer challenges and support emerging technologies.

 

AI Data Center Resources, Webinars, and Training

[21:50] Roy Chamberlain: What resources does Leviton provide for those interested in AI data center design?

[22:15] Peter Helfrich: Our website includes extensive educational content beyond product information.

[22:22] Peter: We offer webinars focused on data centers, AI, and networking technologies.

[22:42] Peter: Recorded webinars and on-demand learning resources are also available.

[22:57] Peter: Many webinars include companion white papers for download.

[23:01] Peter: Customers can also work with local representatives and technical experts for training and consultations.

[23:26] Peter: We often help customers solve broader industry challenges, not just product-specific issues.

 

Conclusion: The Future of AI Data Center Infrastructure

[23:55] Roy Chamberlain: Well, Peter, you're a wealth of knowledge. It's been great going through these questions with you.

[24:02] Roy: Thank you for joining us.

[24:03] Peter: Take care.

[24:07] Roy: Beyond Bandwidth is a podcast series produced by Leviton Network Solutions.

[24:12] Roy: Visit leviton.com/beyondbandwidth for more information and additional episodes.

[24:42] Roy: I'm Roy Chamberlain and this was Beyond Bandwidth. Until next time.

Season 2, Episode 7: AI Data Center Growth and Infrastructure Challenges

 

Introduction

[0:04] Roy Chamberlain: Welcome to Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[0:11] Roy: I am your host from Leviton Network Solutions, Roy Chamberlain.

[0:15] Roy: In Episode 7 of our second season, Leviton's Mike Connaughton meets with David Potts, Director of National Data Center Business at Graybar.

[0:24] Roy: Together, they discuss the growing demand for AI data centers and the logistics required to build them.

[0:31] Roy: From electrical grid capacity constraints to the need for specialty products and large-scale resources, constructing these technological giants requires exceptional planning and organization.

[0:48] Roy: Listen as Mike and David share insights from their experience.

 

Meet the Experts: Mike Connaughton and David Potts

[0:55] Mike Connaughton: Hey everyone. Welcome back to our latest episode of Beyond Bandwidth.

[1:00] Mike: This is Mike Connaughton. I'm a Senior Product Manager at Leviton Network Solutions.

[1:05] Mike: Joining me today is David Potts, Director of National Data Center Business with Graybar.

[1:12] Mike: How are you doing today, David?

[1:13] David Potts: I'm doing well, Mike. Thank you very much for having me on today.

[1:18] Mike: Before we get started, tell us about yourself and what brought you to this role.

[1:25] David: My name is David Potts, and I am the Director of National Data Center Business for Graybar Electric...

 

AI Data Centers: Demand, Capacity, and Power Constraints

[2:26] Mike Connaughton: This season we're talking about AI. What are your customers asking about AI data centers?

[2:53] David Potts: Owners, operators, and hyperscalers are all transitioning portions of their data center capacity to AI workloads.

[3:10] David: We expect demand to outpace capacity in the near future.

[3:22] David: Customers are focused on finding power, land, water resources, and community support for new developments.

[3:42] Mike: The amount of power required by large AI data centers is incredible.

[4:00] David: A major Arizona data center operator told us that without a utility permit before 2027, available grid capacity may be unavailable for future projects.

[4:49] Mike: Northern Virginia has experienced similar challenges, creating opportunities for emerging data center markets.

[5:30] David: Today's infrastructure cannot support tomorrow's technology demands without significant upgrades.

 

The Strategic Role of Distribution in Data Center Projects

[5:49] Mike Connaughton: Can you explain the value your strategic account management team provides customers?

[6:16] David Potts: Our Strategic Account Managers build relationships with both data center owners and integrators.

[6:44] David: We help customers evaluate technologies that reduce total cost of ownership and improve project outcomes.

[7:05] David: We also help contractors grow by connecting them with owner-operators and opportunities in the data center market.

[7:40] Mike: Manufacturers often focus on specific products, but customers need solutions that integrate across the entire project.

[8:21] David: We expect our Strategic Account Managers to be the nucleus of every deal, providing strategic guidance rather than transactional support.

[8:58] David: Our role is consultative, helping customers make better decisions aligned with their long-term goals.

 

AI Data Center Supply Chains and Inventory Management

[9:18] Mike Connaughton: Are customers relying on distributors more as AI deployments accelerate?

[9:36] David Potts: Absolutely. Inventory availability has become critical again.

[9:48] David: Having products in stock can determine how quickly a project gets off the ground.

[10:16] Mike: Leviton recently expanded its fiber cable manufacturing facility to increase supply and improve responsiveness.

[10:44] David: Graybar is building dedicated distribution centers focused on supporting data center construction.

[11:00] David: These facilities will provide rapid delivery of key products and materials required for AI data center projects.

[11:32] Mike: Speed and project velocity are becoming major differentiators across the industry.

[12:08] David: Our services help accelerate project timelines through staging, logistics, and just-in-time delivery support.

 

Data Center World 2025: Industry Trends and Observations

[12:34] Mike Connaughton: We recently attended Data Center World. What stood out to you?

[12:51] David Potts: The size of the event and the volume of exhibitors and attendees were impressive.

[13:17] David: The educational opportunities and industry engagement made a strong impression.

[13:29] David: Industry consolidation through mergers and acquisitions is also reshaping the market.

[14:31] Mike: Attendance and exhibitor growth were significantly higher than previous years.

[14:52] Mike: Many companies outside the traditional data center industry are exploring opportunities in AI infrastructure.

[15:35] David: Manufacturers are increasingly looking for distribution partners to help them enter and grow within the data center market.

 

Building AI Infrastructure While Defining Best Practices

[16:24] Mike Connaughton: The industry is still determining best practices for AI cluster deployment.

[16:49] Mike: Standards organizations are actively discussing what should and can be standardized.

[17:16] David Potts: Technology is evolving so quickly that new innovations emerge almost daily.

[17:43] David: Everyone is trying to establish their role in the ecosystem, and we're effectively building the plane while flying it.

 

The Future Impact of Artificial Intelligence

[18:10] Mike Connaughton: What opportunities excite you most about AI?

[18:18] David Potts: I think about AI's potential impact across financial services, healthcare, and many other industries.

[18:45] David: AI-powered predictive analysis could significantly improve financial decision-making.

[19:06] David: Healthcare may benefit through faster diagnoses and more effective treatment recommendations.

[19:19] David Potts: We're fortunate to have a front-row seat as these changes unfold.

[19:41] Mike: Historically, major technological advances create applications that people couldn't initially imagine.

[20:17] Mike: Much like the telephone transformed communication, AI may evolve in ways we cannot yet predict.

[20:35] David: Technology continually shifts how people communicate and how businesses create value.

 

Closing Thoughts

[21:26] Mike Connaughton: Thanks for joining us and sharing your insights.

[21:36] David Potts: Thank you. It was great to participate and connect in person at Data Center World.

[22:15] Roy Chamberlain: Beyond Bandwidth is a podcast series produced by Leviton Network Solutions.

[22:19] Roy: To learn more about enterprise and data center networking topics, visit leviton.com/beyondbandwidth.

[22:39] Roy: If you enjoyed this episode, please leave an honest review.

[22:49] Roy: I'm Roy Chamberlain, and this was Beyond Bandwidth. Until next time.

Season 2, Episode 8: AI Data Centers and Structured Cabling: Designing High-Performance AI Infrastructure

 

Introduction

[00:03] Roy Chamberlain: Welcome to Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[00:11] Roy: I am your host from Leviton Network Solutions, Roy Chamberlain.

[00:15] Roy: Leviton's Mike Connaughton meets with Robert Henke today.

[00:19] Roy: Robert is a partner at the manufacturer's rep firm, WestCal Technologies.

[00:24] Roy Chamberlain: To be or not to be structured cabling? That is today's question.

[00:30] Roy: In Episode 8, Season 2 on AI Data Centers, we discuss the reasons why structured cabling is the design choice for most of these spaces.

[00:40] Roy: Latency, insertion loss, distance, pathways, airflow, and future migration all affect overall network performance and become even more critical in AI data center architectures.

[00:56] Roy: These experts are in the trenches working on these challenges every day.

[01:00] Roy: Learn from their experience so you can make the best decisions for your next data center design project.

 

Robert Henke's Background in Network Cabling and Data Center Infrastructure

[01:08] Mike Connaughton: Hey everyone, this is Mike Connaughton, Senior Product Manager here at Leviton Network Solutions. Welcome to this episode of Beyond Bandwidth.

[01:16] Mike: I'm joined today by Robert Henke of WestCal Sales. Thanks for joining us, Robert.

[01:22] Robert Henke: Thanks for the invitation, Mike. Glad to be here.

[01:25] Mike: To start, could you introduce yourself and tell listeners about your background?

[01:35] Robert: I graduated college and moved to Northern California. My first job was at Anixter Brothers in 1988, selling network cabling and connectivity solutions during the early days of computer networking...

 

The Evolution of Networking and Parallels to Today's AI Infrastructure

[03:22] Mike Connaughton: Two things you mentioned stood out to me. First, Wang computers...

[03:39] Robert Henke: We can talk about Token Ring, Twinax, IBM systems. It was the Wild West back then.

[03:53] Mike: It's much simpler now with Ethernet standardization. But that early networking environment reminds me a lot of AI today.

[04:16] Mike: Companies have reference designs from vendors such as NVIDIA, but implementation approaches are still evolving.

[04:24] Robert: Exactly. It's exciting because we're once again educating the marketplace on infrastructure options.

 

Who Is Building AI Data Centers and HPC Networks?

[04:53] Mike Connaughton: What kinds of customers are asking about AI infrastructure?

[05:00] Robert Henke: The customer base is incredibly diverse. We hear from technology companies, manufacturers, automotive firms, switch manufacturers, and organizations deploying high-performance computing.

[05:45] Robert: We also work with integrators such as those building complete AI platforms around NVIDIA, Intel, or AMD technologies.

[06:25] Robert: Then there are contractors responsible for installing the structured cabling systems inside these AI facilities.

 

Common AI Data Center Concerns: Latency, Loss, and Connectivity

[06:54] Mike Connaughton: What concerns are customers bringing to you?

[07:09] Robert Henke: Many customers hear recommendations for direct-attach connectivity and long home-run cabling.

[07:29] Robert: Data center operators are accustomed to structured cabling and want flexibility for future migration.

[07:35] Robert: We spend a lot of time discussing latency versus insertion loss because these concepts are often confused.

[08:02] Robert: Latency is fundamentally a function of distance.

[08:20] Robert: Loss is influenced by distance and the number of connections.

[08:29] Robert: Understanding these distinctions gives customers more options when designing AI infrastructure.

 

Structured Cabling vs Direct-Attach Cabling in AI Environments

[08:57] Mike Connaughton: This reminds me of early storage networking discussions where patch panels were considered the enemy.

[09:41] Robert Henke: Exactly. Many concerns stem from previous assumptions about loss and connectivity.

[10:03] Robert: Most AI deployments operate over relatively short distances where latency and loss aren't significant issues.

[10:15] Robert: Once those concerns are addressed, attention shifts to application-specific connectivity requirements.

 

NVIDIA AI Architectures and Networking Requirements

[10:22] Robert Henke: If we look at the NVIDIA H100 platform, there are multiple fabric networks involved including storage, compute, and management networks.

[10:48] Robert: Different optical connector types and architectures create new design considerations.

[11:00] Robert: Moving from H100 to Grace Blackwell introduces different connectivity requirements.

[11:18] Robert: The networking requirements continue evolving, and future platforms such as Vera Rubin will likely introduce further changes.

 

Adapting to Rapid AI Infrastructure Changes

[11:36] Mike Connaughton: Customers need to understand not just GPUs versus CPUs but also differences between AI platform architectures.

[12:06] Robert Henke: The pace of change is remarkable. Hardware generations are moving quickly and infrastructure requirements change with each generation.

[12:43] Robert: Power requirements, connectivity, and even rack architectures continue to evolve.

 

Future-Proofing AI Data Centers with Structured Cabling

[13:20] Mike Connaughton: How are organizations responding to the pace of change?

[13:51] Robert Henke: I'm seeing two strategies.

[14:01] Robert: One approach is to build quickly, knowing everything may be replaced later.

[14:45] Robert: The alternative is to build around structured cabling designed for migration and long-term ROI.

[15:14] Robert: A structured single-mode fiber infrastructure allows organizations to reuse core cabling while adapting to future connectivity changes.

[15:43] Robert: I see examples of both strategies in the market today.

 

AI Data Center Challenges: Power, Cooling, and Infrastructure Design

[16:02] Mike Connaughton: Time-to-revenue is a huge concern for operators.

[16:20] Robert Henke: Infrastructure is actually one of the easier parts of the challenge compared with cooling and power.

[16:25] Robert: The shift from air-cooled to liquid-cooled AI systems dramatically changes facility requirements.

[17:00] Robert: Structured cabling enables operators to focus on power and cooling rather than rebuilding their entire network.

 

The Future of AI Data Center Architecture

[17:15] Mike Connaughton: How do you see AI infrastructure evolving over the next several years?

[17:22] Robert Henke: Future AI systems may no longer fit traditional rack-based designs.

[18:01] Robert: We may see highly customized enclosures built specifically around cooling density and power delivery requirements.

[18:39] Robert: The scale of investment in AI computing is changing how facilities are designed.

[19:21] Robert: Processing performance is increasing at unprecedented rates, driving major infrastructure changes.

 

Return on Infrastructure Investment (ROII) in AI Networks

[20:48] Mike Connaughton: What other key conversations are you having with customers?

[21:02] Robert Henke: A major topic is ROI.

[21:12] Robert: Leviton uses the term ROII, Return on Infrastructure Investment.

[21:36] Robert: The key question is whether customers want to discard their infrastructure during upgrades or design for future migration.

[22:15] Robert: Most customers want to protect both their time and capital investments.

[22:46] Robert: Structured cabling provides a strong foundation for future AI platform generations.

 

Sustainability, Labor Efficiency, and Long-Term AI Infrastructure Planning

[23:14] Mike Connaughton: Sustainability means different things to different organizations, but labor is a major consideration.

[23:35] Mike: If you invest in infrastructure today, you shouldn't have to rebuild it again in one or two years.

[23:53] Robert Henke: That's exactly right. AI innovation isn't slowing down.

[24:08] Robert: Return on Infrastructure Investment is more important today than ever before.

 

Closing Remarks

[24:17] Mike Connaughton: I appreciate your time as always, Robert.

[24:22] Robert Henke: Keep those products and smart ideas coming.

[24:30] Mike: Thanks, Robert, and thanks to everyone for joining us today.

 

Learn More About AI Data Centers and Network Infrastructure

[24:36] Roy Chamberlain: Beyond Bandwidth is a podcast series produced by Leviton Network Solutions.

[24:40] Roy: For more information about AI data centers, structured cabling, enterprise networking, and data center infrastructure, visit leviton.com/beyondbandwidth.

[25:00] Roy: If you enjoyed this episode, we'd appreciate an honest review.

[25:10] Roy: I'm Roy Chamberlain and this was Beyond Bandwidth. Until next time.

Season 2, Episode 9: AI Data Centers, Optical Transceivers, and the Future of Network Connectivity

 

Introduction

[00:03] Roy Chamberlain: Here we are back at Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[00:11] Roy: I am your host from Leviton Network Solutions, Roy Chamberlain.

[00:16] Roy: Today, Leviton's Mike Connaughton sits down with Vladimir Kozlov.

[00:20] Roy: Vladimir is the Founder and CEO of LightCounting, an optical communications market research company focused on optoelectronic interfaces and industry forecasting.

[00:34] Roy: Listen in as Vladimir and Mike discuss AI data center growth, optical networking demand, power challenges, and the future of transceiver technology.

 

Meet the Guest: Vladimir Kozlov of LightCounting

[00:52] Mike Connaughton: Welcome to Beyond Bandwidth. I'm joined today by Vlad Kozlov of LightCounting Market Research.

[01:05] Mike: Thanks for joining us, Vlad.

[01:08] Vladimir Kozlov: Thank you for having me, Mike.

 

LightCounting's Mission and Market Research Approach

[01:11] Mike Connaughton: Can you introduce yourself and explain what LightCounting does?

[01:22] Vladimir Kozlov: I'm a laser scientist by training. After the telecom bubble, I founded LightCounting to bring more transparency to the optical communications market.

[01:50] Vladimir: We collect shipment and market data from leading transceiver and component manufacturers and use that information to forecast trends and analyze emerging technologies.

[02:28] Vladimir: Today, LightCounting has analysts across China, Japan, Europe, and the United States.

 

Optical Market Forecasting and Industry Transparency: Lessons Learned from Decades of Forecasting

[03:00] Mike Connaughton: One thing that impressed me about your presentations was your willingness to discuss forecasting mistakes and explain why they happened.

[03:59] Vladimir Kozlov: We work hard to develop a scientific approach to forecasting. Every six months we compare historical forecasts against actual market data and learn from the differences.

[04:48] Vladimir: One recurring lesson is that pricing tends to decline faster than expected.

 

AI Data Centers Driving Demand for Optical Connectivity: How AI Changed the Optics Market

[05:36] Mike Connaughton: This season we're discussing artificial intelligence and AI data centers. From an optics perspective, this is one of the biggest industry shifts we've ever seen.

[06:24] Vladimir Kozlov: We've tracked data center optics since 2007, beginning with Google's deployment of 10G transceivers.

[06:49] Vladimir: Google was also among the first companies deploying optics inside AI clusters.

[07:19] Vladimir: The launch and adoption of ChatGPT accelerated demand dramatically and helped drive unprecedented growth in optical transceiver shipments.

[08:09] Vladimir: Last year the transceiver market grew nearly 100%, which exceeded our expectations.

 

Tariffs, Supply Chains, and Market Uncertainty: How Geopolitical Factors Affect Optical Networking

[08:37] Vladimir Kozlov: One challenge this year is uncertainty surrounding trade policies and tariffs.

[09:57] Vladimir: Transceiver suppliers have spent years preparing for tariffs by diversifying manufacturing into countries such as Thailand, Malaysia, and Vietnam.

[10:40] Vladimir: The biggest concern isn't necessarily tariff costs. It's uncertainty and how that affects investment decisions by companies like Google, Meta, and Amazon.

[11:46] Vladimir: Even minor slowdowns at the top of the supply chain can have amplified effects throughout the optical ecosystem.

 

Power Consumption Challenges in AI Networks 

Why AI Clusters Require New Optical Technologies

[13:25] Mike Connaughton: When people discuss AI infrastructure, power consumption is always part of the conversation.

[13:54] Vladimir Kozlov: Power is a major challenge. The biggest contributors to transceiver power consumption are digital signal processors and modulators.

[14:26] Vladimir: Current 1.6T transceivers provide little improvement in power efficiency compared to 800G solutions.

Scale-Up Networks and NVLink Demands

[15:06] Vladimir Kozlov: AI clusters typically contain front-end, scale-out, and scale-up networks.

[15:35] Vladimir: NVIDIA's NVLink architecture requires roughly nine times more bandwidth than traditional Ethernet or InfiniBand networks.

[16:05] Vladimir: Data center operators want to maximize GPU density, making power-efficient connectivity increasingly important.

 

Co-Packaged Optics (CPO): The Next Evolution - Is Co-Packaged Optics Ready for Deployment?

[17:15] Mike Connaughton: Co-packaged optics remains a major topic among switch vendors and hyperscale operators.

[17:52] Vladimir Kozlov: We've seen limited deployments of co-packaged optics for many years, particularly in government-funded supercomputing projects.

[19:00] Vladimir: Large-scale deployments are beginning now, especially through integrated AI systems from companies like NVIDIA.

[20:02] Vladimir: Within the next 12 to 18 months, we expect broader adoption of CPO technology in scale-up AI networking applications.

 

The Evolution of 400G, 800G, and 1.6T Optical Transceivers - Customization Accelerates Innovation

[20:24] Mike Connaughton: What changes do you attribute specifically to AI-driven networking growth?

[20:52] Vladimir Kozlov: The industry is accelerating adoption of higher-speed technologies and creating highly customized transceiver designs.

[21:29] Vladimir: There are already numerous variations of 800G optics deployed by hyperscale customers.

[22:25] Vladimir: Many optical suppliers reported record profitability last year due to AI-driven demand.

[22:58] Vladimir: By the end of the decade, annual shipments could approach 100 million high-speed optical transceivers.

 

AI vs. the Dot-Com Bubble: Similarities and Differences - Comparing Today's AI Boom to the Telecom Bubble

[23:36] Mike Connaughton: How does today's AI boom compare with the internet bubble of the late 1990s?

[24:04] Vladimir Kozlov: We discussed this extensively at OFC because the industry recently marked the 25th anniversary of the telecom bubble peak.

[24:48] Vladimir: One key difference today is market transparency. We have far better visibility into deployments, spending, and demand drivers.

[26:18] Vladimir: We also benefit from decades of historical data and lessons learned from previous downturns.

[27:29] Vladimir: Major cloud providers remain financially strong, with substantial cash reserves available for continued investment.

 

Future AI Applications Driving Network Growth

Where Will Future Demand Come From?

[28:15] Vladimir Kozlov: One of the most important variables for forecasting future growth is identifying new AI applications that generate revenue.

[28:33] Vladimir: Google has successfully monetized AI through advertising and search optimization.

[28:50] Vladimir: We're watching for growth opportunities in manufacturing automation, productivity enhancement, and entirely new use cases.

AI in Everyday Business Operations

[29:12] Mike Connaughton: During a recent trip, a hotel manager told me AI software now automatically handles room pricing decisions.

[29:47] Vladimir Kozlov: That's encouraging because it demonstrates practical business value beyond the technology itself.

 

The Human Factor in AI Adoption - Why Technology Adoption Takes Time

[29:53] Vladimir Kozlov: We've seen similar excitement before with autonomous driving and robot taxis.

[30:21] Vladimir: The technology works, but societal acceptance often moves more slowly than technical development.

[31:35] Vladimir: Human behavior and trust frequently become the most significant barriers to widespread adoption.

[32:04] Vladimir: The internet ultimately transformed the world, but it took longer than most people originally predicted.

 

Final Thoughts: The Next 25 Years of Optical Networking

[32:24] Mike Connaughton: Any final thoughts before we wrap up?

[32:34] Vladimir Kozlov: We only forecast five years into the future because beyond that, technological change becomes difficult to predict.

[33:05] Vladimir: Looking back over the past 25 years, the industry has advanced at an incredible pace.

[33:28] Vladimir: Despite near-term challenges, I'm very optimistic about the next decade and beyond.

[33:53] Mike: AI remains in its early stages, making this an exciting time for the industry.

[34:14] Mike: Thank you, Vlad.

[34:15] Vladimir: Thank you.

 

About Beyond Bandwidth

[34:24] Roy Chamberlain: Beyond Bandwidth is a podcast series produced by Leviton Network Solutions.

[34:29] Roy: For more information about enterprise and data center network infrastructure topics and additional podcast episodes, visit Leviton's Beyond Bandwidth resources.

[34:58] Roy: I'm Roy Chamberlain, and this was Beyond Bandwidth. Until next time.

 

Season 2, Episode 10: AI, Data Centers, Infrastructure Planning & Construction Collaboration

 

Introduction

[00:03] Roy Chamberlain: Welcome to Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[00:10] Roy: I am your host from Leviton Network Solutions, Roy Chamberlain.

[00:14] Roy: In today's episode, Leviton's Mike Connaughton sits down with Iain Farquharson.

[00:19] Roy: Iain is with Red Bear Tech and has a dynamic history with artificial intelligence and data center technology.

[00:26] Roy: If you have a passion for AI and data centers, then you will definitely appreciate this conversation.

 

Meet the Guest: Iain Farquharson of Red Bear Technologies

[00:34] Mike Connaughton: Hi, everyone. Mike Connaughton here again, Senior Product Manager at Leviton Network Solutions, and welcome to this week's episode of Beyond Bandwidth.

[00:43] Mike: I'm joined today by Iain Farquharson of Red Bear Technologies. Thanks for joining us, Iain.

[00:50] Iain Farquharson: Hi, Mike. Thanks for having me.

[00:52] Mike: I'm hoping it's a little cooler for you in the UK than it is for us on the East Coast of the US this week.

[00:57] Iain: It's a challenge here, but the air conditioning works.

[01:10] Mike: If you don't mind taking a minute or two to tell the audience who you are, what you do, and what your background is.

 

Early Career: Data Cabling, Banking Networks, and AI Education

[01:17] Iain Farquharson: My name's Iain. I feel quite lucky in terms of ICT. My path has always been focused on information and communications technology and, later, artificial intelligence and data centers.

[01:32] Iain: When I was six or seven years old, my father was an electrician working on infrastructure upgrades for a major retail bank in the United Kingdom. That was my first exposure to data cabling.

[01:51] Iain: Data cabling has been part of my life from a very early age.

[02:11] Iain: I initially wanted to study psychology but ended up studying cognitive science.

[02:30] Iain: That led to three years studying the philosophy of AI, consciousness, and how intelligence might be replicated in digital systems.

[02:50] Iain: After university, I returned to the professional world of data cabling and eventually moved into data center projects.

[03:28] Iain: I entered the data center industry about five years ago and have worked with colocation providers, general contractors, fit-out specialists, and major hyperscale operators.

 

The Evolution of Artificial Intelligence

[03:57] Mike Connaughton: AI education in the 1990s provides an interesting perspective. Many people think AI is a recent innovation, but the concept has been around for decades.

[04:24] Iain Farquharson: Absolutely. I was fortunate to visit Project Dawn at the University of Cambridge, where I saw the historical progression from Ada Lovelace and Charles Babbage through Alan Turing and modern supercomputing.

[04:51] Iain: The perception of AI has changed dramatically. In the past, AI focused on challenges such as playing chess or passing conversational tests. Today, AI can generate content that's difficult to distinguish from reality.

[05:25] Mike: AI-generated video is a great example of exponential growth when comparing results from a year ago, six months ago, and even a few weeks ago.

[06:10] Iain: The pace of advancement is now exceeding benchmarks like Moore's Law in certain areas.

 

How AI is Transforming Data Centers

[07:11] Mike Connaughton: How has AI impacted what you do in your daily work?

[07:21] Iain Farquharson: AI isn't as new as people think, and neither are data centers. However, the industry has reached a point where certainty is disappearing.

[08:03] Iain: We once had stable, predictable data center designs. Today, technology is evolving so rapidly that predicting what a facility will look like in two years is increasingly difficult.

[08:43] Iain: That uncertainty creates challenges, opportunities, and excitement for everyone involved.

 

Cross-Discipline Collaboration in Modern Data Centers

[09:02] Mike Connaughton: Data centers still perform the same basic functions, but now every discipline impacts every other discipline.

[09:38] Mike: I need more conversations than ever with power and cooling engineers because changes in those systems directly affect network infrastructure decisions.

[09:59] Iain Farquharson: Data centers and construction are maturing simultaneously. We now have better operational data and more insights into how facilities can be designed and built.

[10:32] Iain: The ability to transform that operational data into practical construction outcomes will become increasingly important.

 

Digital Twins, BIM, and Human Problem Solving

[11:30] Mike Connaughton: How do you deal with rapidly changing design requirements and the lack of established standards?

[11:43] Iain Farquharson: The industry has powerful tools available, including BIM, digital twins, and advanced design platforms.

[11:52] Iain: I'm a big supporter of the MacLeamy Curve. It's far easier to solve problems during design than after construction.

[12:05] Iain: Despite advances in technology, human collaboration remains the most important factor.

[12:23] Iain: Some ambiguities simply cannot be resolved without people sitting down together and working through challenges.

 

AI as a Tool Rather Than a Replacement

[13:09] Mike Connaughton: There's a lot of concern that AI will replace jobs.

[13:24] Mike: At the end of the day, AI is a tool. It doesn't replace human ingenuity.

[14:13] Iain Farquharson: I've witnessed remarkable AI advances over the past few months alone.

[14:52] Iain: I see AI as an exoskeleton for the mind. It helps people perform tasks more effectively rather than replacing them.

[15:13] Iain: Human intuition and communication remain essential because AI cannot predict the future.

 

Physical Infrastructure Challenges in AI Data Centers

[15:31] Mike Connaughton: What physical infrastructure changes are you seeing compared to five or more years ago?

[15:50] Iain Farquharson: Some organizations are delaying design commitments because they're unsure what future requirements will be.

[16:29] Iain: We're seeing far greater density in the white space and unprecedented levels of connectivity.

[16:46] Iain: High-performance computing environments are creating new demands for both copper and fiber infrastructure.

[17:25] Iain: Designers must now consider everything from cable tray loading and structural support to liquid cooling and power requirements.

[18:01] Iain: The key is to begin with the intended server workloads and design backward from there.

 

Advice for Planning the Next AI Data Center

[18:46] Mike Connaughton: What advice would you give organizations planning their next data center?

[18:59] Iain Farquharson: Choose your partners carefully and collaborate with them early.

[19:31] Iain: Bring all stakeholders into the planning process as soon as possible.

[19:39] Iain: Leverage the expertise of manufacturers and technology specialists. They understand their products better than anyone else.

[20:02] Iain: When the entire ecosystem collaborates, there are very few challenges the industry cannot solve.

 

Reducing Construction Waste Through AI and Collaboration

[20:13] Mike Connaughton: Are there any final topics we haven't covered?

[20:13] Iain Farquharson: One issue that surprises people is the amount of waste that still exists in construction.

[20:30] Iain: Data centers are construction projects involving concrete, steel, mechanical systems, electrical systems, generators, and supporting infrastructure.

[20:41] Iain: In the UK alone, construction rework has been estimated at approximately £25 billion annually.

[21:12] Iain: That represents a major opportunity for improvement through AI adoption combined with human collaboration.

[21:21] Mike: Standards organizations are trying to keep pace with rapid developments in high-performance computing and AI infrastructure.

[21:53] Mike: Reducing waste and improving efficiency should be industry-wide priorities.

 

Conclusion

[22:21] Mike Connaughton: Thank you, Iain. I appreciate your time, and thanks to everyone for listening.

[22:26] Mike: Until next time, signing off.

[22:32] Roy Chamberlain: Beyond Bandwidth is a podcast series produced by Leviton Network Solutions.

[22:37] Roy: For more information on today's topics and additional enterprise and data center infrastructure content, visit leviton.com/beyondbandwidth.

[22:57] Roy: If you enjoyed this episode, we'd appreciate an honest review in your podcast application.

[23:06] Roy: I'm Roy Chamberlain, and this was Beyond Bandwidth. Until next time.

Season 2, Episode 10: AI Data Center Season Recap and Key Takeaways

 

Introduction: Reflecting on a Season of AI Data Center Innovation

[00:04] Roy Chamberlain: Welcome back to Beyond Bandwidth, Leviton's podcast series where we dive into network technology's hottest topics.

[00:12] Roy: I am your host from Leviton Network Solutions, Roy Chamberlain.

[00:41] Roy: Tune in to Beyond Bandwidth and learn about AI data centers with me, Mike Connaughton, and our amazing crew of guests.

[00:54] Roy: Today Mike and I will share our thoughts and highlights from this season in our closing episode of Season 2.

 

Season Highlights: What Stood Out Most About AI Data Centers?

[01:06] Roy: Welcome back to Beyond Bandwidth, Mike.

[01:09] Mike Connaughton: Glad to be here.

[01:10] Roy: What was the most interesting information you heard throughout this season?

 

Why the AI Data Center Expert Panel Was a Standout

[01:32] Mike: Episode 5's Q&A session was especially valuable because it brought together experts from different disciplines who were learning from each other in real time.

[02:23] Roy: I completely agree. The chemistry and discussion were very dynamic.

 

Emerging Technologies in AI Networking

[02:30] Roy: Brooke Ford's discussion about hollow core fiber technology was one of the most interesting topics for me.

[02:46] Mike: Any technology must solve a problem cost-effectively. Microsoft's acquisition of Lumenisity shows serious investment in hollow core fiber's potential.

[03:46] Roy: Jonathan Jew discussed competition among Intel, AMD, and NVIDIA. Have things changed since then?

[04:23] Mike: NVIDIA still holds a commanding market position, but AMD and Intel continue to gain traction as competition drives innovation.

 

AI Data Center Cooling Challenges and Innovations - Solving Power Consumption and Heat Management

[06:02] Roy: Are there alternatives to traditional and immersion cooling methods?

[06:52] Mike: The challenge involves both reducing chip power consumption and improving cooling efficiency. Emerging technologies like Linear Pluggable Optics (LPO) and Co-Packaged Optics (CPO) can help reduce system power consumption.

 

How Standards Organizations Are Keeping Pace

[08:44] Roy: How are organizations like IEEE, TIA, and BICSI adapting to rapid AI infrastructure advancements?

[09:25] Mike: Effective standards often need to be reactive. Real-world deployments help define best practices before standards can formalize them.

Structured Cabling vs. Point-to-Point Architectures

[09:59] Mike: The best advocates for structured cabling are often people who have experienced the challenges of point-to-point implementations.

 

Challenges Slowing AI Data Center Expansion - Supply Chain, Labor, and Power Constraints

[12:19] Roy: Are there bottlenecks affecting AI data center development?

[12:56] Mike: Chip availability remains a major constraint, along with power availability, labor shortages, and supply chain limitations.

 

AI Data Center Site Selection and Global Growth: Where New AI Data Centers Are Being Built

[14:58] Roy: Are new regions emerging as AI data center hotspots?

[15:27] Mike: In the U.S., Texas, Georgia, and Ohio continue to gain momentum. In Europe, growth is occurring in the Nordics, Spain, and Italy.

 

Leviton Resources for AI Data Center Design and Support

Engineering, Testing, and Design Assistance

[16:58] Roy: What resources does Leviton Network Solutions offer customers seeking guidance?

[17:33] Mike: Customers can access educational content, engineering consultations, testing resources, design services, and product support.

Leveraging Lab Testing for AI Infrastructure

[19:13] Roy: Testing capabilities are becoming increasingly important as performance requirements evolve.

[20:12] Mike: Our lab environment allows us to validate solutions before recommending them to customers.

Cisco Engineering Alliance and End-to-End Solutions

[20:18] Mike: Joining the Cisco Engineering Alliance strengthens our ability to provide integrated infrastructure solutions for AI data centers.

 

Market Intelligence and Industry Trends: Insights from LightCounting and Vladimir Kozlov

[21:35] Roy: Vladimir Kozlov brought a unique market perspective to this season.

[21:57] Mike: LightCounting operates at the intersection of optics and active networking technologies, helping the industry understand broader market forces.

 

AI Adoption and the Rapid Pace of Innovation: Lessons from Iain Farquharson and Red Bear Tech

[23:43] Roy: What were your biggest takeaways from your conversation with Iain Farquharson?

[24:08] Mike: His experience spans AI applications, data center operations, and network infrastructure, making him one of the most well-rounded experts featured this season.

[25:29] Mike: Organizations should work closely with subject matter experts and technology partners when solving infrastructure challenges.

 

Final Thoughts on Season 2

[26:10] Roy: That's it for Season 2. We've built strong momentum and had outstanding guests.

[26:22] Mike: Thanks everyone for listening.

[26:29] Roy: Beyond Bandwidth is a podcast series produced by Leviton Network Solutions. Visit leviton.com/beyondbandwidth for more information, additional episodes, and data center infrastructure content.

[27:03] Roy: I'm Roy Chamberlain, and this was Beyond Bandwidth. Until next time.