Networks & InfrastructureEthernetAI NetworkingScale Up NetworkingNvlink
Ethernet to Overtake Nvidia's NVLink in AI Scale-Up by 2029
Dell'Oro forecasts Ethernet will dominate AI scale-up networking by 2029 as back-end switch spending nears $1 trillion through 2030 and port speeds quadruple to 3.2 Tb/s.
Networks & InfrastructureWhy it matters
- Back-end data center switch spending approaches $1 trillion from 2026 to 2030, per Dell'Oro Group.
- Ethernet is forecast to become the dominant scale-up networking technology by 2029.
- InfiniBand held over 80% of scale-out share in 2023; Ethernet overtook it by 2025.
- UALink Consortium targets commercial deployments in 2027.
- 3.2 Tb/s switches ship in 2028-2029 and will hold the majority of ports by 2031.
The story
Data center operators will spend close to $1 trillion on back-end network switches between 2026 and 2030, and Dell'Oro Group expects Ethernet to displace Nvidia's proprietary NVLink as the dominant scale-up networking technology by 2029. The forecast lands as AI builders discover that high-bandwidth, low-latency networks now matter as much as the GPUs they connect.
AI networking splits into three layers. Scale-up connects accelerators inside a rack or across a few racks. Scale-out connects racks within a single data center. Scale-across links separate data centers, sometimes across vast distances.
What is happening in scale-up networking?
Scale-up networks tie up to hundreds of accelerators into one tightly coupled compute domain, a requirement that grows more critical for AI inference workloads. Nvidia's NVLink dominates today because Nvidia dominates the GPU market. That grip is loosening.
Dell'Oro VP Sameh Boujelbene said market forces are pushing operators toward open, Ethernet-based alternatives such as Ethernet for Scale-up Networking (ESUN) and Scale Up Ethernet (SUE), both overseen by the Open Compute Project Foundation. Broadcom claims more than five hyperscalers already run Ethernet scale-up in production, with more than 10 vendor partners building racks around it.
The UALink Consortium — whose members include AMD, Astera Labs, Intel, AWS, Meta and Microsoft — is developing an open interconnect designed to match NVLink performance and scale beyond it. Moor Insights & Strategy VP Matt Kimball cautioned that UALink is not yet available. The consortium targeted commercial deployments for 2027. Operators running non-Nvidia GPUs can also use PCIe switches, a slower option, according to Kimball.
Three forces drive Ethernet's rise in scale-up, Boujelbene said:
- Vendor diversity, as AI clusters grow and strain supply.
- Proliferation of non-Nvidia GPUs and custom accelerators demanding open architectures.
- Continuous improvements in Ethernet performance, congestion management and reliability.
How did scale-out flip from InfiniBand to Ethernet?
Scale-out connects up to hundreds of thousands of GPUs across racks into one large AI compute cluster. In 2023, InfiniBand held more than 80% of that market. By 2025, Ethernet took the lead, driven by the same forces reshaping scale-up, Boujelbene said.
Ethernet options now include Nvidia's Spectrum-X, RoCEv2 and the emerging Ultra Ethernet Consortium (UEC) specification. Arista Networks, Cisco, Dell and Hewlett Packard Enterprise compete in scale-out gear; Cisco and Dell support all three Ethernet standards. Broadcom, Marvell and Cisco back RoCEv2 and UEC in silicon.
Who needs scale-across?
Scale-across links hundreds of thousands to potentially millions of GPUs across campus buildings or longer distances. Hyperscalers and neoclouds running large AI training workloads build most of these networks, Kimball said. Large enterprises need them only if they operate like hyperscalers. The vendor field mirrors scale-out: Arista, Cisco and HPE/Juniper, plus hyperscalers' own custom solutions.
How should enterprises build an AI network?
Boujelbene advises starting with workload requirements, not networking technology. Take a phased approach:
- Scale-up: determine how much tightly coupled GPU compute is needed.
- Scale-out: build a high-performance Ethernet fabric with room to expand.
- Scale-across: evaluate connecting AI resources across buildings, data centers or clouds.
Where is the market headed?
Scale-out networks carry more than two-thirds of today's AI back-end networking market, with scale-up below one-third, Boujelbene said. That balance flips by 2030, when scale-up should exceed half the market.
Port speeds will double and then quadruple. Most switch ports today run at 800 Gb/s across all three layers. Switches at 1.6 Tb/s began shipping in the second half of 2026 and will comprise more than half of ports by 2027. Switches at 3.2 Tb/s start shipping in 2028 and 2029 and will hold the majority of ports by 2031, Boujelbene said.
The economics are straightforward. "The network needs to operate at 100% to maximize the utilization rate of the expensive GPU resources," she said. "Therefore, increasing network speed would have a direct impact on AI application performance."
Also reported
Source: sdxcentral.com
More from Daniel Okafor
Show full bio
Senior reporter covering regulation, policy and their impact on Carrier business, 5G and 6G, fiber and broadband at Telecom Gazette.
582 articles