2026 Showcase
Welcome to the NextGenInfra Data Center Networking for AI and Cloud Showcase!
As the world invests multiple hundreds of billions to trillions of dollars in data center build out for AI, what's the role of networking? Will the shift in focus from pre-training to post-training to test-time/inference-time scaling change the data center computing and networking requirements? What's the latest in scale-up, scale-out, and scale-across networking.
To help our readers sort through the rapid changes in the space, we've captured insights from the leading thinkers in the data center networking ecosystem here in our showcase. The content in the videos and our report highlight the state-of-the-art across the three domains of networking. Check it out!
Thanks to our sponsors for making the report possible!
2026 Data Center Networking
Welcome to the NextGenInfra Data Center Networking for AI and Cloud Showcase!
As the world invests multiple hundreds of billions to trillions of dollars in data center build out for AI, what's the role of networking? Will the shift in focus from pre-training to post-training to test-time/inference-time scaling change the data center computing and networking requirements? What's the latest in scale-up, scale-out, and scale-across networking.
To help our readers sort through the rapid changes in the space, we've captured insights from the leading thinkers in the data center networking ecosystem here in our showcase. The content in the videos and our report highlight the state-of-the-art across the three domains of networking. Check it out!
Thanks to our sponsors for making the report possible!
Sponsor Spotlight DriveNets
Rethinking Data Center Network Architecture
In this Director's Cut video, Dudy Cohen, VP of Product Marketing at DriveNets, explains how Ethernet is evolving through advancing speeds (800G to 3.2T), co-packaged optics (CPO) integration, and scheduling layers that address lossiness to support AI infrastructure with GPU clusters scaling from 8,000 to nearly one million units. He describes how DriveNets has adapted its scheduled fabric technology, delivering superior time-to-first-token performance and lower cost-per-million-tokens compared to proprietary technologies.
Sponsor Spotlight Marvell, AvidThink
Rethinking Data Center Network Architecture
In an extended discussion, Rishi Chugh, VP and GM, Data Center Switching at Marvell, shares how AI workloads are driving data center networking, distinguishing between bandwidth-driven scale-out networking (800Gpbs to 1.6Tbps and beyond) and scale-up fabrics that unify xPU pods requiring memory sharing and ultra-low latency. Chugh shares Marvell's commitment to deliver 100T and 200T scale-out fabrics along with 115T and 57T scale-up fabric devices supporting NVLink Fusion, UALink, and ESUN.
Sponsor Spotlight Nokia
The Optics Scaling Challenge at 1.6T
Neel Patel, GM, Optical Networking Component Solutions at Nokia, explains in our Director's Cut edition, how AI data centers face significant power and supply challenges as GPU interconnect speeds double with each generation, requiring transitions from copper to optical solutions within racks. He highlights Nokia's competitive advantages in indium phosphide (InP) technology, positioning the company as an open ecosystem component supplier for hyperscalers.
Sponsor Spotlight Upscale AI
AI Data Centers Need Purpose-Built Networks
In this director's cut video, Aravind Srikumar, SVP of Product at Upscale AI, presents the company's mission to deliver AI-specific networking solutions. Upscale AI differentiates through co-design optimization of ASIC, systems, and software to deliver deterministic performance for AI workloads while maintaining openness.
Sponsor Spotlight Aria Networks
Rethinking AI Networks from First Principles
Mansour Karam, Founder and CEO of Aria Networks, presents their Deep Networking approach centered on performance optimization of AI clusters through fine-grain telemetry. The company's full-stack architecture combines hardware innovations like microsecond-resolution telemetry embedded in ASICs with specialized AI and agentic capabilities with in-built domain expertise.
Sponsor Spotlight Arrcus
Unified Network Fabric for Distributed AI Workloads Across Data Centers
For our Director's Cut, Sanjay Kumar, Vice President of Product Management and Marketing at Arrcus, and Keyur Patel, Founder and CTO at Arrcus, explain how Arrcus provides a unified network fabric that connects GPU resources across scale-up, scale-out, and scale-across layers for distributed AI workloads. They touch on how Arrcus differentiates itself through hardware-abstracted operating system software that works across multiple silicon platforms.
Sponsor Spotlight Arrcus
How Distributed AI Workloads Are Reshaping Network Architecture
Sanjay Kumar, VP of Products and Marketing, and Keyur Patel, CTO at Arrcus, discuss how the distributed nature of AI workloads is fundamentally changing network architecture, with Ethernet becoming a first-class citizen in AI networks for both training and inferencing across increasingly heterogeneous hardware environments. They explain how Arrcus addresses the needs of telcos, colocation providers, hyperscalers, enterprises, and neoclouds by providing a unified fabric that is scalable, highly secured, fully programmatic, and orchestrated on demand across different data centers.
Sponsor Spotlight Ayar Labs
Optical Interconnects for AI
Vishal Chandrasekar, Director of Product Management at Ayar Labs, presents an optical interconnect solution that scales GPU and switch connectivity through a chip-to-system approach, featuring optical engines fabricated on TSMC's platform and multi-chip packages that integrate eight optical engines to replace copper implementations. The solution, developed with partners including FOCI, Browave, Wiwynn, Alchip, and GUC, targets 2028 deployment.
Sponsor Spotlight Broadcom
Ethernet for AI Networking at Scale: Building 100K+ XPU Clusters
Hasan Siraj, VP of Product Management for the Core Switching Group at Broadcom, explains how Ethernet has become the industry standard for AI networking at scale, enabling clusters of over 100,000 XPUs through scale-up, scale-out, and scale-across architectures. He discuses Broadcom's solution portfolio including: Tomahawk Ultra, Tomahawk 6, Thor Ultra and Jericho 4.
Sponsor Spotlight Ciena
AI Networking at Scale: GPU Cluster Interconnect Solutions for Data Centers
Helen Xenos, Senior Director of Portfolio Marketing at Ciena, presents the company's innovations for GPU cluster interconnects including the Nitro linear retimer driver for active copper cables and the Vesta 200 6.4T CPX optical engine.
Sponsor Spotlight Clockwork.io
Solving GPU Cluster Inefficiency: From 30% to Peak AI Performance
Suresh Vasudevan, CEO of Clockwork.io, explains how GPU clusters for AI training operate at only 30-50% efficiency due to poor inter-GPU communication, availability issues, and tail latency from slow straggler GPUs. He explains how nanosecond-precision clock synchronization can be leveraged to address availability and latency issues across Infiniband or RoCE networks.
Sponsor Spotlight Credo
MicroLEDs for AI Data Center Connectivity at Scale
LK Bhupathi, AVP, Product at Credo, discusses the company's expansion into microLED-based technologies. He explains that microLED solutions use dense arrays of LEDs operating at modest speeds to deliver superior reliability, resilience through spare channel failover, and low single-digit pJ per bit power consumption.
Sponsor Spotlight DriveNets
Ethernet Fabrics for AI Data Centers
Dudy Cohen, VP of Product Marketing at DriveNets, explains that while networking represents only 10% of AI infrastructure costs, it causes 80% of deployment challenges, advocating for Ethernet-based solutions with scheduling capabilities that can exceed InfiniBand performance. He emphasizes the importance of working with innovative vendors who have deep expertience.
Sponsor Spotlight ECL
Solving Power & Speed Challenges at 150-300kW Per Rack
Yuval Bachar, Co-founder and CEO of ECL, discusses how his AI data center company addresses the mismatch between rapid 9-12 month technology cycles and slower 18-24 month infrastructure timelines by building equipment and facilities on accelerated schedules while managing increasingly power-intensive racks reaching 150-300 kilowatts through a "flex grid" approach combining multiple energy sources.
Sponsor Spotlight Ethernet Alliance, Intel
Industry's First 224G Interoperability Testing
David Rodgers, Technical Business Development Manager with the Ethernet Alliance, and Sam Johnson, Link Applications Engineering Manager at Intel and High-Speed Networking Chair, discuss the second 224-gigabit Plugfest event at Keysight in Santa Clara, where over 20 companies test actual product interoperability across multiple vendors while nine compliance stations feed critical data back into IEEE specification development.
Sponsor Spotlight Lightmatter
Lightmatter's Optical Interconnects for AI Scale-Up
Steve Klinger, VP of Product at Lightmatter, presents the company's optical scale-up interconnect solutions, featuring the Passage EVK50 system with DWDM technology that delivers 16 wavelengths per fiber with 400Gbps Tx and 400Gbps Rx. The architecture offers high energy efficiency and compact integration while aligning with the recently announced OCI MSA specifications using an 8-wavelength, 4+4 band interleave model.
Sponsor Spotlight Marvell
Big Outlook for XPU-Attach
Will Chu, SVP and GM, Custom Cloud Solutions Business Unit at Marvell, discusses the company's expansion into XPU attach solutions, where Marvell customizes all components within the XPU tray beyond the XPU. The custom solutions include CXL-enabled memory for expansion and near-memory compute, security devices for AI infrastructure management, and high-performance NICs built on Marvell's SerDes and other IP platforms.
Sponsor Spotlight Nokia
Linear Pluggable Optics for Data Center Efficiency
Neel Patel, GM of Optical Networking Component Solutions at Nokia, discusses the power consumption challenges of advancing optical modules in data centers, where 1.6T modules consume nearly double the power of 800G modules, potentially adding 1KW to a 64-port switch. He explains how Linear Pluggable Optics (LPO) addresses this by leveraging switch ASIC DSPs reducing power consumption.
Sponsor Spotlight Nokia
Integration is the Real Race in AI Data Center Networking
Mike Bushong, Vice President of Data Center at Nokia, discusses how AI workloads are fundamentally changing the data center networking market, with hyperscalers driving unprecedented demand and technical challenges. He explains that Nokia's broad portfolio across switching silicon, DSPs, and differentiated optics from the Infinera acquisition positions the company to handle complex integration challenges.
Sponsor Spotlight NVIDIA
NVIDIA's AI Factory Networking Stack
Gilad Shainer, SVP of Networking at NVIDIA, shares why building AI factories requires designing four distinct infrastructure layers—scale-up using NVLink, scale-out using InfiniBand or Spectrum X Ethernet, scale-across using Spectrum XGS, and storage with NVIDIA Bluefield-4 STX architecture.
Sponsor Spotlight Synopsys
AI-Driven Multi-Die Design
Abhijeet Chakraborty, VP of Engineering at Synopsys, examines the accelerating adoption of multi-die designs for AI applications, highlighting how advanced packaging enables larger formats and better interconnect density while presenting complex engineering challenges in power delivery, thermal management, and multiphysics modeling that exceed the capabilities of traditional manual methods.
Sponsor Spotlight UALink Consortium
UALink 2.0: Open GPU Interconnect for AI Clusters
Kurtis Bowman, Chairman of the UALink Consortium, presents announces the release of UALink 2.0 specifications featuring protocol-physical layer separation, in-network compute capabilities, enhanced manageability, and the consortium's first UCIe-based chiplet specification, with the 115-member organization expecting customer solutions with 2.0 to be available in 2027.
Sponsor Spotlight UCIe Consortium
UCIe Chiplet Connectivity for Performance & Efficiency
Debendra Das Sharma, UCIe Consortium Chair at UCIe Consortium, explains how UCIe technology enables chiplet connectivity within packages and package-to-package connections using co-packaged optics for composable systems with dynamic resource allocation. UCIe delivers bandwidth densities with 1-4 orders of magnitude improvement over PCIe and Ethernet.
Sponsor Spotlight Upscale AI
AI Data Centers Need Purpose-Built Networks
Aravind Srikumar, SVP, Product at Upscale AI, explains how AI workloads require completely lossless, synchronized networks. He shares how Upscale AI delivers AI-specific networking silicon, systems, and software purpose-built for both scale-up and scale-out environments, rather than adapting general-purpose infrastructure.
Sponsor Spotlight Xscape Photonics
Optical Compute Interconnect Standardization
Vivek Raghunathan, Co-Founder and CEO of Xscape Photonics, discusses the Optical Compute Interconnect (OCI) standardization effort as an important milestone for scale-up interconnects and co-packaged optics implementations. He explains that OCI uses multi-color laser-based technology for GPU-to-memory communication and highlights Xscape's recently announced laser module.
Courtesy of our sponsors!
Complimentary Report Download
2026 Data Center Networking
Download 2026 Data Center Networking Report
By downloading this report, you agree to our Terms of Use and Privacy Policy.
Analysis by AvidThink
What the 2026 report finds
Across two dozen conversations with the vendors, standards bodies, and operators building AI data centers, five themes define networking in 2026 — the year the industry stopped bolting networks onto AI clusters and started designing them as first-class infrastructure.
Networking is ~10% of the cost but ~80% of the pain
Interconnect is a small slice of AI infrastructure spend yet the dominant source of deployment risk and idle GPUs. Getting the fabric right — not adding more compute — is what unlocks utilization, which is why so much 2026 innovation is concentrated here.
Ethernet becomes the default AI fabric
Scheduled and enhanced Ethernet — Ultra Ethernet, Broadcom's Tomahawk/Jericho, DriveNets' scheduled fabric — is displacing proprietary interconnect for scale-out, now reaching 100K+ XPU clusters. The open ecosystem is winning on cost and choice even where it trails on raw latency.
Three planes: scale-up, scale-out, scale-across
AI factories are now designed as distinct network layers, each with its own bandwidth and latency profile. Scale-up interconnect is standardizing fast — UALink 2.0, NVLink Fusion, ESUN, OCI and UCIe — and vendors increasingly differentiate on unifying all three planes into one fabric.
The rack goes optical — and power is the ceiling
As per-GPU bandwidth doubles each generation, copper runs out of reach inside the rack. Co-packaged and linear-pluggable optics, optical engines, and microLED links move light into and between packages — but a 1.6T module's power draw means energy, not bandwidth, is the binding constraint.
Efficiency and thermals decide what's usable
Training clusters routinely run at 30–50% efficiency; fine-grain telemetry, nanosecond clock synchronization, and 150–300 kW racks with liquid cooling and flex-grid power determine whether the silicon can actually be put to work. The network is now an observability and power problem as much as a bandwidth one.
Highlights from industry thought leaders
Director’s Cut
Extended, in-depth versions of select interviews.
Rethinking Data Center Network Architecture
- Ethernet is advancing from 800G to 3.2T with co-packaged optics and scheduling layers that remove lossiness for AI fabrics
- GPU clusters are scaling from 8,000 toward one million units
- DriveNets' scheduled fabric claims better time-to-first-token and lower cost-per-million-tokens than proprietary alternatives
Abstract
In this Director's Cut video, Dudy Cohen, VP of Product Marketing at DriveNets, explains how Ethernet is evolving through advancing speeds (800G to 3.2T), co-packaged optics (CPO) integration, and scheduling layers that address lossiness to support AI infrastructure with GPU clusters scaling from 8,000 to nearly one million units. He describes how DriveNets has adapted its scheduled fabric technology, delivering superior time-to-first-token performance and lower cost-per-million-tokens compared to proprietary technologies.
Rethinking Data Center Network Architecture
- Scale-out is bandwidth-driven (800G to 1.6T+); scale-up fabrics unify xPU pods needing memory sharing and ultra-low latency
- Marvell roadmap: 100T and 200T scale-out fabrics plus 115T and 57T scale-up fabric devices
- Scale-up support spans NVLink Fusion, UALink, and ESUN
Abstract
In an extended discussion, Rishi Chugh, VP and GM, Data Center Switching at Marvell, shares how AI workloads are driving data center networking, distinguishing between bandwidth-driven scale-out networking (800Gpbs to 1.6Tbps and beyond) and scale-up fabrics that unify xPU pods requiring memory sharing and ultra-low latency. Chugh shares Marvell's commitment to deliver 100T and 200T scale-out fabrics along with 115T and 57T scale-up fabric devices supporting NVLink Fusion, UALink, and ESUN.
The Optics Scaling Challenge at 1.6T
- GPU interconnect speeds double each generation, forcing copper-to-optical transitions inside the rack
- Power and component supply are the binding constraints on AI data center optics
- Nokia positions its indium phosphide (InP) technology as an open component supply for hyperscalers
Abstract
Neel Patel, GM, Optical Networking Component Solutions at Nokia, explains in our Director's Cut edition, how AI data centers face significant power and supply challenges as GPU interconnect speeds double with each generation, requiring transitions from copper to optical solutions within racks. He highlights Nokia's competitive advantages in indium phosphide (InP) technology, positioning the company as an open ecosystem component supplier for hyperscalers.
AI Data Centers Need Purpose-Built Networks
- AI workloads need deterministic, AI-specific networking rather than adapted general-purpose gear
- Upscale AI co-designs ASIC, systems, and software while maintaining openness
Abstract
In this director's cut video, Aravind Srikumar, SVP of Product at Upscale AI, presents the company's mission to deliver AI-specific networking solutions. Upscale AI differentiates through co-design optimization of ASIC, systems, and software to deliver deterministic performance for AI workloads while maintaining openness.
Video interviews
Rethinking AI Networks from First Principles
- Aria's Deep Networking approach optimizes AI cluster performance through fine-grain telemetry
- Microsecond-resolution telemetry is embedded directly in ASICs, paired with agentic AI and built-in domain expertise
Abstract
Mansour Karam, Founder and CEO of Aria Networks, presents their Deep Networking approach centered on performance optimization of AI clusters through fine-grain telemetry. The company's full-stack architecture combines hardware innovations like microsecond-resolution telemetry embedded in ASICs with specialized AI and agentic capabilities with in-built domain expertise.
Unified Network Fabric for Distributed AI Workloads Across Data Centers
- One unified fabric connects GPU resources across scale-up, scale-out, and scale-across layers
- Hardware-abstracted network OS runs across multiple silicon platforms
Abstract
For our Director's Cut, Sanjay Kumar, Vice President of Product Management and Marketing at Arrcus, and Keyur Patel, Founder and CTO at Arrcus, explain how Arrcus provides a unified network fabric that connects GPU resources across scale-up, scale-out, and scale-across layers for distributed AI workloads. They touch on how Arrcus differentiates itself through hardware-abstracted operating system software that works across multiple silicon platforms.
How Distributed AI Workloads Are Reshaping Network Architecture
- Ethernet is now a first-class citizen in AI networks, for both training and inference on heterogeneous hardware
- One unified fabric targets telcos, colos, hyperscalers, enterprises, and neoclouds
- Fabric is scalable, secured, fully programmatic, and orchestrated on demand across data centers
Abstract
Sanjay Kumar, VP of Products and Marketing, and Keyur Patel, CTO at Arrcus, discuss how the distributed nature of AI workloads is fundamentally changing network architecture, with Ethernet becoming a first-class citizen in AI networks for both training and inferencing across increasingly heterogeneous hardware environments. They explain how Arrcus addresses the needs of telcos, colocation providers, hyperscalers, enterprises, and neoclouds by providing a unified fabric that is scalable, highly secured, fully programmatic, and orchestrated on demand across different data centers.
Optical Interconnects for AI
- Optical engines fabricated on TSMC's platform; multi-chip packages integrate eight engines to replace copper
- Chip-to-system approach scales both GPU and switch connectivity
- Partner ecosystem (FOCI, Browave, Wiwynn, Alchip, GUC) targets 2028 deployment
Abstract
Vishal Chandrasekar, Director of Product Management at Ayar Labs, presents an optical interconnect solution that scales GPU and switch connectivity through a chip-to-system approach, featuring optical engines fabricated on TSMC's platform and multi-chip packages that integrate eight optical engines to replace copper implementations. The solution, developed with partners including FOCI, Browave, Wiwynn, Alchip, and GUC, targets 2028 deployment.
Ethernet for AI Networking at Scale: Building 100K+ XPU Clusters
- Ethernet has become the industry standard for AI networking, enabling clusters of 100,000+ XPUs
- Scale-up, scale-out, and scale-across are all addressed with Ethernet architectures
- Portfolio spans Tomahawk Ultra, Tomahawk 6, Thor Ultra, and Jericho 4
Abstract
Hasan Siraj, VP of Product Management for the Core Switching Group at Broadcom, explains how Ethernet has become the industry standard for AI networking at scale, enabling clusters of over 100,000 XPUs through scale-up, scale-out, and scale-across architectures. He discuses Broadcom's solution portfolio including: Tomahawk Ultra, Tomahawk 6, Thor Ultra and Jericho 4.
AI Networking at Scale: GPU Cluster Interconnect Solutions for Data Centers
- Nitro linear retimer driver extends active copper cables for GPU cluster interconnect
- Vesta 200 6.4T CPX optical engine targets cluster-scale optics
Abstract
Helen Xenos, Senior Director of Portfolio Marketing at Ciena, presents the company's innovations for GPU cluster interconnects including the Nitro linear retimer driver for active copper cables and the Vesta 200 6.4T CPX optical engine.
Solving GPU Cluster Inefficiency: From 30% to Peak AI Performance
- GPU training clusters typically run at only 30-50% efficiency
- Root causes are inter-GPU communication, availability issues, and straggler tail latency
- Nanosecond-precision clock synchronization mitigates these on InfiniBand or RoCE networks
Abstract
Suresh Vasudevan, CEO of Clockwork.io, explains how GPU clusters for AI training operate at only 30-50% efficiency due to poor inter-GPU communication, availability issues, and tail latency from slow straggler GPUs. He explains how nanosecond-precision clock synchronization can be leveraged to address availability and latency issues across Infiniband or RoCE networks.
MicroLEDs for AI Data Center Connectivity at Scale
- MicroLED links use dense arrays of modest-speed LEDs rather than fast lasers
- Spare-channel failover delivers reliability and resilience
- Power consumption is in the low single-digit picojoules per bit
Abstract
LK Bhupathi, AVP, Product at Credo, discusses the company's expansion into microLED-based technologies. He explains that microLED solutions use dense arrays of LEDs operating at modest speeds to deliver superior reliability, resilience through spare channel failover, and low single-digit pJ per bit power consumption.
Ethernet Fabrics for AI Data Centers
- Networking is ~10% of AI infrastructure cost but causes ~80% of deployment challenges
- Scheduled Ethernet fabrics can exceed InfiniBand performance
Abstract
Dudy Cohen, VP of Product Marketing at DriveNets, explains that while networking represents only 10% of AI infrastructure costs, it causes 80% of deployment challenges, advocating for Ethernet-based solutions with scheduling capabilities that can exceed InfiniBand performance. He emphasizes the importance of working with innovative vendors who have deep expertience.
Solving Power & Speed Challenges at 150-300kW Per Rack
- 9-12 month technology cycles outpace 18-24 month infrastructure builds — ECL compresses both
- Rack power densities are reaching 150-300kW
- A 'flex grid' approach combines multiple energy sources per site
Abstract
Yuval Bachar, Co-founder and CEO of ECL, discusses how his AI data center company addresses the mismatch between rapid 9-12 month technology cycles and slower 18-24 month infrastructure timelines by building equipment and facilities on accelerated schedules while managing increasingly power-intensive racks reaching 150-300 kilowatts through a "flex grid" approach combining multiple energy sources.
Industry's First 224G Interoperability Testing
- Second 224G Plugfest: 20+ companies testing real product interoperability at Keysight
- Nine compliance stations feed results back into IEEE specification development
Abstract
David Rodgers, Technical Business Development Manager with the Ethernet Alliance, and Sam Johnson, Link Applications Engineering Manager at Intel and High-Speed Networking Chair, discuss the second 224-gigabit Plugfest event at Keysight in Santa Clara, where over 20 companies test actual product interoperability across multiple vendors while nine compliance stations feed critical data back into IEEE specification development.
Lightmatter's Optical Interconnects for AI Scale-Up
- Passage EVK50 delivers DWDM with 16 wavelengths per fiber at 400Gbps Tx + 400Gbps Rx
- Architecture aligns with the new OCI MSA (8-wavelength, 4+4 band interleave)
- Targets energy-efficient, compact optical scale-up interconnects
Abstract
Steve Klinger, VP of Product at Lightmatter, presents the company's optical scale-up interconnect solutions, featuring the Passage EVK50 system with DWDM technology that delivers 16 wavelengths per fiber with 400Gbps Tx and 400Gbps Rx. The architecture offers high energy efficiency and compact integration while aligning with the recently announced OCI MSA specifications using an 8-wavelength, 4+4 band interleave model.
Big Outlook for XPU-Attach
- Marvell customizes every component in the XPU tray beyond the XPU itself
- CXL-enabled memory adds expansion and near-memory compute
- Security devices and high-performance NICs build on Marvell SerDes and IP platforms
Abstract
Will Chu, SVP and GM, Custom Cloud Solutions Business Unit at Marvell, discusses the company's expansion into XPU attach solutions, where Marvell customizes all components within the XPU tray beyond the XPU. The custom solutions include CXL-enabled memory for expansion and near-memory compute, security devices for AI infrastructure management, and high-performance NICs built on Marvell's SerDes and other IP platforms.
Linear Pluggable Optics for Data Center Efficiency
- 1.6T optical modules consume nearly double the power of 800G — up to +1kW on a 64-port switch
- Linear Pluggable Optics (LPO) cuts power by leaning on the switch ASIC's DSPs
Abstract
Neel Patel, GM of Optical Networking Component Solutions at Nokia, discusses the power consumption challenges of advancing optical modules in data centers, where 1.6T modules consume nearly double the power of 800G modules, potentially adding 1KW to a 64-port switch. He explains how Linear Pluggable Optics (LPO) addresses this by leveraging switch ASIC DSPs reducing power consumption.
Integration is the Real Race in AI Data Center Networking
- Hyperscaler AI demand is reshaping the data center networking market
- Integration — not any single component — is the real competitive race
- Nokia's portfolio spans switching silicon, DSPs, and Infinera-derived optics
Abstract
Mike Bushong, Vice President of Data Center at Nokia, discusses how AI workloads are fundamentally changing the data center networking market, with hyperscalers driving unprecedented demand and technical challenges. He explains that Nokia's broad portfolio across switching silicon, DSPs, and differentiated optics from the Infinera acquisition positions the company to handle complex integration challenges.
NVIDIA's AI Factory Networking Stack
- AI factories require four distinct network layers, each purpose-designed
- Scale-up on NVLink; scale-out on InfiniBand or Spectrum-X Ethernet
- Scale-across on Spectrum-XGS; storage on the BlueField-4 STX architecture
Abstract
Gilad Shainer, SVP of Networking at NVIDIA, shares why building AI factories requires designing four distinct infrastructure layers—scale-up using NVLink, scale-out using InfiniBand or Spectrum X Ethernet, scale-across using Spectrum XGS, and storage with NVIDIA Bluefield-4 STX architecture.
AI-Driven Multi-Die Design
- Multi-die design adoption is accelerating for AI silicon
- Advanced packaging enables larger formats and denser interconnect
- Power delivery, thermal, and multiphysics challenges now exceed manual design methods
Abstract
Abhijeet Chakraborty, VP of Engineering at Synopsys, examines the accelerating adoption of multi-die designs for AI applications, highlighting how advanced packaging enables larger formats and better interconnect density while presenting complex engineering challenges in power delivery, thermal management, and multiphysics modeling that exceed the capabilities of traditional manual methods.
UALink 2.0: Open GPU Interconnect for AI Clusters
- UALink 2.0 released: protocol-physical separation, in-network compute, enhanced manageability
- Includes the consortium's first UCIe-based chiplet specification
- 115-member consortium expects customer solutions on 2.0 in 2027
Abstract
Kurtis Bowman, Chairman of the UALink Consortium, presents announces the release of UALink 2.0 specifications featuring protocol-physical layer separation, in-network compute capabilities, enhanced manageability, and the consortium's first UCIe-based chiplet specification, with the 115-member organization expecting customer solutions with 2.0 to be available in 2027.
UCIe Chiplet Connectivity for Performance & Efficiency
- UCIe connects chiplets in-package and package-to-package, including over co-packaged optics
- Bandwidth density is 1-4 orders of magnitude better than PCIe or Ethernet
- Enables composable systems with dynamic resource allocation
Abstract
Debendra Das Sharma, UCIe Consortium Chair at UCIe Consortium, explains how UCIe technology enables chiplet connectivity within packages and package-to-package connections using co-packaged optics for composable systems with dynamic resource allocation. UCIe delivers bandwidth densities with 1-4 orders of magnitude improvement over PCIe and Ethernet.
AI Data Centers Need Purpose-Built Networks
- AI workloads demand completely lossless, synchronized networks
- Purpose-built silicon, systems, and software for scale-up and scale-out — not adapted general-purpose infrastructure
Abstract
Aravind Srikumar, SVP, Product at Upscale AI, explains how AI workloads require completely lossless, synchronized networks. He shares how Upscale AI delivers AI-specific networking silicon, systems, and software purpose-built for both scale-up and scale-out environments, rather than adapting general-purpose infrastructure.
Optical Compute Interconnect Standardization
- OCI standardization is a milestone for scale-up interconnects and co-packaged optics
- Multi-color laser technology enables GPU-to-memory optical links
- Xscape recently announced its own laser module
Abstract
Vivek Raghunathan, Co-Founder and CEO of Xscape Photonics, discusses the Optical Compute Interconnect (OCI) standardization effort as an important milestone for scale-up interconnects and co-packaged optics implementations. He explains that OCI uses multi-color laser-based technology for GPU-to-memory communication and highlights Xscape's recently announced laser module.
Reader feedback on our latest report: "Appreciate the broad coverage and unique insights." Download our 2026 Data Center Networking report now and send us your feedback!
Download our 2026 Data Center Networking Report
By downloading this report, you agree to our Terms of Use and Privacy Policy.
Continue exploring
More executive showcases across next-gen infrastructure.
2026
2026 OFC
Leading executives at OFC 2026 in Los Angeles covering optical and fiber for AI data centers
2025
2025 Data Center Networking
Infrastructure leaders discuss data center networking and impact of AI/GenAI and cloud workload. Topics: InfiniBand, UEC, NVLink, UALink
2025
2025 MPLS World Congress
Leading executives at MPLS World Congress 2025 in Paris discuss latest in IP networking, AI networking, IPv6, SRv6, converged optics