Will AI Backend Networks Exceed $100 B in Spending by 2029? The Ethernet Surge Explained

AI Workload Growth and Network Evolution

Rapid growth of artificial‑intelligence workloads is driving a disruptive transformation in the network industry. In this new era AI requires four distinct networks to operate efficiently.

The Four AI Networks

Compute Fabric : Connects AI accelerators, GPUs, CPUs and other components inside a server. It is designed for short‑distance, high‑speed communication, typically using proprietary copper interfaces such as PCIe or NVLink.

Backend Network : Links multiple servers into an AI cluster using multi‑layer switches and optical modules. Protocols may be InfiniBand or Ethernet, both carried over optical links.

Front‑end Network : Connects the AI cluster to the cloud data‑center for storage and data exchange. Each server’s CPU has its own NIC that attaches to an optical module, and the front‑end uses fiber‑based Ethernet.

Data Center Interconnect (DCI) Network : Uses 100‑km or longer links to connect one data‑center with others in the same region.

Why New AI Backend Networks Are Needed

As the number of accelerators in large AI clusters grows exponentially, building new AI backend networks that provide efficient high‑speed connectivity has become an urgent industry need. Investment in AI backend networks has reached unprecedented levels, while the traditional front‑end network remains a critical foundation.

Market Forecasts

Analysts predict that AI backend network spending will exceed $100 billion by 2029, with Ethernet showing strong momentum. DeepSeek’s resource‑efficient model has heightened concerns about sustainable accelerator infrastructure spending, yet major vendors such as Google, Amazon, Microsoft and Meta continue to maintain high AI‑infrastructure capex.

According to Dell’Oro Group’s “AI Networks for AI Workloads” report, switch spend in AI backend networks is expected to surpass $100 billion between 2025 and 2029. By 2025 most AI backend switch ports will reach 800 Gbps, 1,600 Gbps by 2027, and 3,200 Gbps by 2030.

Ethernet’s growth is outpacing InfiniBand, and Dell’Oro now expects Ethernet to overtake InfiniBand by 2027. In 2024 the Ethernet market was led by Celestica, Huawei and Nvidia; by 2025 vendors such as Accton, Arista, Cisco, Juniper and Nokia are projected to capture larger shares, reshaping the competitive landscape.

Front‑end network spending is also expected to rebound strongly after 2024, driven by the need for >100 Gbps connections and premium pricing for high‑speed links. Dell’Oro forecasts total Ethernet data‑center switch sales to exceed $1.8 trillion from 2025 to 2029, with the same vendors benefiting.

Over the next five years, front‑end networks will deploy more than 100 million 800 Gbps and 1,600 Gbps switch ports, leveraging upcoming 51.2 Tbps and 102.4 Tbps chips. Backend network shipments are expected to be at least twice that volume.

Future Outlook

Continued expansion and diversification of AI workloads will require both front‑end and backend network infrastructures to evolve toward higher‑speed Ethernet. The year 2025 is likely to be pivotal for the Ethernet data‑center switch market.

Reference links: Dell’Oro Report Original Article (Chinese)