Data Center ToR Switch: Complete Guide

The ToR Switch: The Foundation of Modern Spine-Leaf Networks

A ToR (Top-of-Rack) switch is installed at the top of each server rack in a data center. It aggregates the NICs of all servers in the rack (typically 20-40 servers × 2 NICs = 40-80 links) and connects them upstream to the spine switches. It is the leaf switch in a spine-leaf architecture. Its sizing determines server density per rack and network performance.

Critical features of a modern ToR switch include 48 access ports (25G SFP28 for servers and 10G SFP+ for legacy equipment), 6-8 100G QSFP28 uplinks to the spine switches, <1µs cut-through latency, a 32-40MB buffer to absorb storage and backup bursts, and native EVPN-VXLAN support.

Leading ToR Switches in 2026

Cisco Nexus 93180YC-FX

• 48× 10/25G SFP28 + 6× 40/100G QSFP28.
• ~1µs cut-through latency.
• 40MB shared buffer.
• NX-OS or ACI.
• OPTINOC price: ~€12,000-15,000 excl. VAT.
• The world's most widely deployed ToR switch. A Cisco standard since 2017.

Juniper QFX5120-48Y

• 48× 10/25G SFP28 + 8× 100G QSFP28.
• 550ns latency (best on the market).
• 32MB buffer.
• Junos OS + Apstra compatibility.
• OPTINOC price: ~€10,000-13,000 excl. VAT.
• 20% less expensive than the Nexus, with lower latency and a strong alternative overall.

Aruba CX 8325-48Y8C

• 48× 25G SFP28 + 8× 100G QSFP28.
• AOS-CX with a native REST API.
• Active-active VSX MLAG.
• OPTINOC price: ~€14,000 excl. VAT.

Sizing by Rack

• Standard 42U rack: 20-40 2U servers or 40-80 1U servers.
• NICs per server: 2× 25G with dual-homing for redundancy.
• Required ToR access ports: 40-80 25G ports.
• Uplinks: 6 to 8× 100G to the spine switches (3:1 to 2:1 oversubscription).
• Dual ToR: 2 ToR switches per rack for redundancy. Each server connects to 2 different ToR switches through MLAG/vPC.

Oversubscription: A Practical Calculation

A Nexus 93180YC-FX provides 48× 25G = 1200 Gbps of downlink capacity to servers. With 6× 100G uplinks, it provides 600 Gbps to the spine switches. Oversubscription = 1200/600 = 2:1. This means that, in the worst-case scenario where every server transmits at full line rate simultaneously, only half of the traffic can leave the rack at any given time.

• 2:1: Suitable for conventional workloads such as web applications, APIs, and standard databases.
• 1.5:1: Recommended for bursty storage and backup workloads.
• 1:1 (non-blocking): Required for AI/ML training and HPC.

OPTINOC Recommendations

• Conventional enterprise DC: Nexus 93180YC-FX, an established standard with widely available expertise.
• Highly automated DC: QFX5120-48Y + Apstra. 20% lower price and better latency.
• AI/ML workloads: Target a 1:1 non-blocking design. Consider a 64-port 100G ToR switch for GPU clusters.
• Highly resilient DC: Dual ToR is mandatory. Use a Nexus pair with vPC or QFX switches in a Virtual Chassis.

OPTINOC supplies Nexus, QFX, and Aruba CX switches with manufacturer warranties and matching DAC/fiber cables. Free sizing is available based on your server count.