Which data center switch should you choose?

Key criteria for a data center switch

A data center switch differs fundamentally from a campus switch in five non-negotiable areas: high-speed port density (25G/100G/400G instead of 1G), latency below 2µs (compared with 5-10µs for campus switching), buffering capacity designed for traffic bursts (40-80MB), native support for overlay protocols (VXLAN and EVPN with a BGP control plane), and full hardware resilience (redundant hot-swap power supplies and N+1 fans). If your switch does not meet all five requirements, you are not buying a data center switch—you are buying a poorly positioned campus switch.

The choice then comes down to three factors: its role in the fabric (leaf ToR vs spine vs border), target throughput per port (25G server access, 100G uplink, 400G spine), and the automation ecosystem (Cisco ACI/NDFC, Juniper Apstra, or standalone NX-OS/Junos with Ansible).

Cisco Nexus 9300: the market benchmark

Positioning and key models

The Nexus 9300 series is the de facto standard for enterprise data centers in France and Europe. Cisco holds approximately 45% of the global data center switching market. The 9300 series covers leaf and border roles in spine-leaf fabrics, with fixed 1U models ranging from 48×10G to 36×400G.

Most widely deployed models in 2024-2026:

• N9K-C93180YC-FX (C93180YC-FX): 48× 10/25G SFP28 + 6× 100G QSFP28. The universal leaf switch. Latency: ~1µs. Shared buffer: 40MB. OPTINOC price: ~€12,000-15,000 excluding VAT.
• N9K-C9336C-FX2: 36× 40/100G QSFP28. The standard spine switch for 100G fabrics. It can also serve as a high-density 100G leaf switch. Price: ~€18,000-22,000 excluding VAT.
• N9K-C9332D-GX2B: 32× 400G QSFP-DD. The next-generation 400G spine switch. For fabrics targeting 400G today. Price: ~€45,000-55,000 excluding VAT.
• N9K-C9364C-GX: 64× 40/100G QSFP28. Very high density, ideal for spine or super-spine deployments. Price: ~€30,000-35,000 excluding VAT.
• N9K-C9316D-GX: 16× 400G QSFP-DD. A compact 400G switch, ideal for DCI or as a spine in a small DC. Price: ~€35,000-40,000 excluding VAT.

Standalone NX-OS vs ACI

Nexus 9300 switches support two operating modes. In standalone NX-OS mode, the switch operates independently—you configure it through the CLI or with Ansible/Terraform, like a traditional router. This is the recommended mode for EVPN-VXLAN fabrics managed by NDFC (Nexus Dashboard Fabric Controller) or in-house automation.

In ACI (Application Centric Infrastructure) mode, the switch is controlled by 3 APIC controllers that manage the fabric end to end using an intent-based model (EPGs, contracts, and VRFs). ACI is powerful for very large DCs (>50 racks), but it has a steep learning curve (allow 6-12 months to master ACI). For small and medium-sized DCs (<50 racks), standalone NX-OS + NDFC is more pragmatic.

EVPN-VXLAN capabilities on Nexus

• VXLAN with BGP-EVPN: Type-2 (MAC/IP), Type-5 (IP prefix), and multi-tenancy through VRF/VNI. Standard since NX-OS 9.2.
• vPC (Virtual Port-Channel): MLAG between 2 Nexus switches for server dual-homing. Peer-keepalive + peer-link are mandatory.
• ITD (Intelligent Traffic Director): L4 load balancing integrated into the switch—eliminating the need for an external load balancer in certain use cases.
• Streaming telemetry: continuous export of metrics to Kafka/InfluxDB/Prometheus. Essential for DC observability.
• On-box Python (Guest Shell): Python scripts can run directly on the switch. Designed for automation, monitoring, and troubleshooting.

Juniper QFX5120/5130: the Junos alternative

Positioning and key models

Juniper QFX is the leading challenger to Nexus, with one major advantage: a unified Junos OS across the entire portfolio (MX, SRX, EX, and QFX), featuring atomic commits and native rollback. For teams already managing Juniper routing platforms (MX among service providers) or campus networks (EX), QFX delivers unmatched operational consistency.

• QFX5120-48Y: 48× 25G SFP28 + 8× 100G QSFP28. Latency: ~550ns (the lowest on the market in this segment). Buffer: 32MB. The quintessential modern leaf switch. OPTINOC price: ~€10,000-13,000 excluding VAT.
• QFX5120-32C: 32× 100G QSFP28. A high-density 100G spine or leaf switch. Price: ~€15,000-18,000 excluding VAT.
• QFX5130-32CD: 32× 400G QSFP-DD. Juniper's 400G spine switch. Latency: ~800ns. Buffer: 64MB. Price: ~€40,000-48,000 excluding VAT.
• QFX5700: 4-slot chassis supporting up to 144× 400G. For super-spines or very high-density DCs. Price: ~€80,000-100,000 excluding VAT.

Junos advantages for the data center

• Atomic commits: configuration changes are applied as a complete set or not at all. There is no partial configuration, as can occur on NX-OS if a line fails.
• Configuration rollback: revert to any configuration version from the last 49 commits (rollback 25 = 25 commits ago).
• Commit confirmed: protection against lockout—if you do not confirm within N minutes, the configuration automatically rolls back. Essential for remote administration.
• RPL (Routing Policy Language): the most powerful and expressive BGP filtering language on the market. Critical for multi-tenant EVPN fabrics.
• Native NETCONF since 2003: automation with Ansible/Nornir/Salt without a wrapper—the switch communicates through NETCONF natively.

Apstra: multi-vendor fabric automation

Juniper Apstra (acquired in 2021) is an intent-based platform for automating EVPN-VXLAN fabrics. Its unique advantage is multi-vendor support. Apstra natively manages QFX, as well as Cisco Nexus, Arista EOS, SONiC, and NVIDIA Cumulus. If you operate a mixed fabric or are considering a migration from Cisco to Juniper (or vice versa), Apstra is the logical orchestration solution.

Apstra Blueprint: you describe the required fabric (number of spines, leaf switches, VRFs, and tenants), and Apstra generates and deploys the configuration for each switch. Apstra continuously compares the actual state against the intent and alerts you to any deviation. Troubleshooting is automated: “your VLAN 100 is isolated on leaf 12 → cause: missing EVPN Type-2 route on spine 2.”

Technical comparison: Nexus vs QFX

Across the criteria that matter in the data center:

• Cut-through latency: QFX5120 = 550ns, Nexus 93180YC-FX = 1µs. QFX has the advantage for HFT and latency-sensitive GPU/storage clusters.
• Buffer: Nexus 93180YC = 40MB shared, QFX5120-48Y = 32MB. Nexus has the advantage for burst-heavy workloads (storage and backup).
• 400G density: N9K-C9332D = 32×400G, QFX5130-32CD = 32×400G. Equal.
• Automation ecosystem: Cisco offers ACI + NDFC + NX-API. Juniper offers Apstra (multi-vendor) + native Junos NETCONF. Juniper has the advantage for automation purity, while Cisco leads in integrated ecosystem capabilities.
• Skills availability in the French market: Cisco leads by a wide margin (80%+ of DC engineers have Cisco experience). Juniper is gaining ground, but recruitment is more challenging.
• Price: QFX5120 is 15-20% less expensive than Nexus 9300 with an equivalent configuration. On a 32-switch project, this represents savings of €50-80K.

Sizing a spine-leaf fabric

Calculating the number of spine switches

The fundamental rule is that every leaf switch must connect to ALL spine switches. The number of spine switches determines north-south bandwidth (to external networks) and resilience (losing one spine causes a proportional performance reduction, not an outage).

• Oversubscription formula: OS = (total downlink bandwidth) / (total uplink bandwidth to spine switches).
• Example: a 48×25G leaf switch provides 1200G of downlink bandwidth. With 6× 100G uplinks to 6 spine switches, it provides 600G of uplink bandwidth. OS = 1200/600 = 2:1.
• Standard target: OS 3:1 is acceptable for conventional workloads (web, API, and databases). OS 2:1 is recommended for storage/backup. OS 1:1 (non-blocking) is mandatory for AI/ML training and HPC.

Calculating the number of leaf switches

The number of leaf switches depends on the number of servers and rack density:

• Standard: 1 leaf ToR per rack. Each rack contains 20-40 servers connected at 25G.
• High-density: 2 leaf switches per rack (dual-ToR) for redundancy. Each server has 2 NICs connected to 2 different leaf switches (MLAG/vPC).
• Collapsed: 1 leaf switch for 2-3 small racks (lower cost, but a single point of failure).

Practical example: a DC with 200 servers, 10 racks, and dual-ToR. → 20 QFX5120-48Y leaf switches + 4-6 QFX5130-32CD spine switches. Optical module budget: ~€80K-120K. Switch budget: ~€300K-400K excluding VAT.

Essential capabilities in 2026

• EVPN-VXLAN with BGP: non-negotiable. It is THE protocol for modern data center fabrics. If the switch does not support it, do not buy it.
• ECMP (Equal Cost Multi-Path): distributes traffic across all spine paths. Without ECMP, your fabric is not a fabric.
• BFD (Bidirectional Forwarding Detection): sub-second failure detection (50ms) between leaf and spine. Essential for fast convergence.
• Streaming telemetry (gNMI, gRPC): continuous export of metrics. SNMP polling is obsolete in the modern DC.
• ZTP (Zero Touch Provisioning): automated deployment of new switches. On a 32-switch project, ZTP saves 2-3 days of installation work.
• 256-bit MACsec: hop-by-hop L2 encryption on inter-switch links. Required for DCs that comply with NIS2, PCI-DSS, and HDS.

OPTINOC recommendations by DC size

Small DC (< 8 racks, < 100 servers)

• 4-8 QFX5120-48Y or Nexus 93180YC-FX leaf switches
• 2 QFX5120-32C or Nexus 9336C-FX2 spine switches
• Automation: Ansible + CLI or Apstra lite
• Switch budget: €80-150K excluding VAT

Medium-sized DC (8-50 racks, 100-1000 servers)

• 16-50 leaf switches + 4-6 100G spine switches
• NDFC (Cisco) or Apstra (Juniper) for orchestration
• Dual-ToR recommended for critical racks
• Switch budget: €300K-800K excluding VAT

Large DC (50+ racks, 1000+ servers)

• 50-200 leaf switches + 8-16 spine switches + super-spines
• ACI (Cisco) or full Apstra (Juniper) is mandatory
• 400G spine-to-spine, 100G leaf-to-spine
• Switch budget: €1M+ excluding VAT

Conclusion: Nexus or QFX?

If your team holds Cisco certifications (CCNP/CCIE DC) and you already operate within the Cisco ecosystem, stay with Nexus 9300. The ACI/NDFC ecosystem, TAC support, and availability of skilled professionals in France justify the 15-20% premium.

If you are starting from scratch, prioritize automation purity (Junos + NETCONF + Apstra), or have strict budget requirements, Juniper QFX5120/5130 offers the best balance of quality, price, and operational efficiency on the market in 2026.

OPTINOC distributes both product lines with manufacturer warranties and eligibility for TAC/JTAC support. Request a comparative quote based on your target architecture—we provide sizing services free of charge.