The Hidden Cost of AI Networking: Why Cables and Switches Can Make or Break Your GPU Investment

They rarely break the budget on their own, but they can make or break how much work the GPUs deliver. For a single GPU server, the switch, cables and optics are a modest share of the bill. As a cluster grows, they multiply. A non-blocking 400G fabric for 1,024 GPUs needs about 48 switches and roughly 2,560 optical transceivers, about 2.5 per GPU, costing around US$2.5 million in optics alone at third-party list prices and drawing about 35 kW. Estimates of networking's share of total AI cluster cost range from about 10% to 30%. The bigger risk is quality: in published training runs, cables, links and network adapters cause a measurable share of failures and slowdowns. Choosing the right cable for each distance, and planning the fabric before ordering, is where the savings are.

What do AI cluster cables and optics cost?

Source: NVIDIA LinkX 100G-PAM4 product overview (27 Nov 2025) and DAC/ACC overview for reach and power; Credo (10 Oct 2024) for AEC; third-party list prices from NADDOD and FS.com fetched 6 Nov 2026. NVIDIA-branded parts and volume pricing differ. NVIDIA offers no AOC at 100G-PAM4 speeds.

The pattern is simple: copper is cheap and almost free in power, but only reaches a few metres. Optics reach across a hall or a campus, but each module costs hundreds to over a thousand US dollars and draws 9–17 W. NVIDIA's guidance, as quoted by The Next Platform, is "use copper cabling when you can and use optical cabling when you must."

How many cables and optics does a cluster need?

More than most buyers expect, because a non-blocking fabric connects every GPU twice: once from its network adapter to a leaf switch, and once from that leaf to a spine. For a two-tier fat tree with one 400G port per GPU and switches of radix k:

• Leaf switches = GPUs ÷ (k ÷ 2)
• Spine switches = GPUs ÷ k
• Cables = 2 × GPUs

With 64-port 400G Quantum-2 switches and 1,024 GPUs, that gives 32 leaf and 16 spine switches and about 2,048 cables. NVIDIA's own H100 SuperPOD reference design for 1,016 GPUs matches: 32 leaf, 16 spine, 1,020 compute cables and 1,024 spine-leaf cables. Because most of these runs are longer than copper reaches, they are optical. Using the counts above, the fabric needs about 1,024 400G optics on the GPU side and about 1,536 twin-port 800G optics in the switches, about 2.5 modules per GPU.

At the NADDOD list prices in the table, that is about US$1.33 million for the NIC-side optics and US$1.20 million for the switch optics: about US$2.53 million, or S$3.2 million at USD/SGD 1.2774, before fibre, patch panels and spares. At NVIDIA's maximum ratings the optics draw about 35 kW. In Singapore colocation, at CBRE's average of US$403 per kW per month, that power allocation alone would cost about US$14,000 a month. Add 48 switches at a third-party list price of about US$35,000 each, roughly US$1.68 million, and the fabric comes to about US$4.2 million. That is about 7.6% of the cost if the 1,024 GPUs were 128 HGX B200 servers at a reseller list price of US$401,609 each. These figures are illustrative; real quotes vary widely.

What share of an AI cluster's cost is networking?

There is no single agreed figure. In May 2024, The Next Platform's Timothy Prickett Morgan paraphrased NVIDIA's argument that InfiniBand is about 20% of cluster cost, and wrote that 10% or lower is normal for Ethernet clusters. In June 2025, writing in The Register, he argued it may be as high as 30% of the real cost. LightCounting estimates optics sales to the top five cloud providers at 2.7% of their capital spending in 2025, forecast at 3.1% in 2026 and 4.1% by 2031, worth about US$16.5 billion and US$26 billion for AI clusters in those two years. The Next Platform also notes that transceivers and fibre make up the bulk of network cost, which matches the worked example above. Our post on InfiniBand vs Ethernet compares the switches themselves.

A cheap cable that drops a link once a week can cost more in lost GPU time than the entire cable budget.

How do cables and links affect reliability?

Directly. In a 54-day snapshot of Llama 3 405B training, Meta logged 419 unexpected interruptions: 35 were attributed to network switches and cables and 7 to network adapters, together about 10%. ByteDance's MegaScale team traced "link flapping" to poor link quality between network adapters, active optical cables and switches, and fixed it with better cables, signal tuning and longer communication timeouts. Alibaba reported that 0.057% of its adapter-to-switch links failed each month; if all 2,048 links in the 1,024-GPU fabric above failed at that rate, it would see about one link failure a month. Alibaba estimated one failure on a 3,000-GPU job costing US$20,000 an hour at about US$30,000.

The practical lessons: buy optics and cables qualified for your switches and adapters, keep spares, label and document every run, and monitor link error counters rather than waiting for a job to fail. Our post on messy network cabinets covers the office-scale version of the same discipline.

Why does NVL72 use copper inside the rack?

To save power and cost where distances are short. For GB200 NVL72, Jensen Huang said optical transceivers and retimers for the NVLink spine would have needed about 20 kW, and that the rack uses about 5,000 copper cables, about two miles in total; SemiAnalysis counts 5,184. GB300 NVL72 keeps the same copper design. Copper does not remove scale-out optics, though: each GPU still has its own 800 Gb/s network port, which leaves the rack over optics. See how NVL72 links 72 GPUs.

Will new optics cut the cost?

Possibly the power bill first. Linear pluggable optics remove the signal processor from the module; Semtech, which sells LPO components, cites 4–8 W against 12–16 W for a conventional 800G module. Co-packaged optics put the optical engines beside the switch chip. NVIDIA claims 3.5 times the power efficiency, and cites 9 W per co-packaged port against 30 W for a 1.6T pluggable. NVIDIA's Quantum-X Photonics switch was still an engineering sample in June 2026, and Spectrum-X co-packaged switches had begun shipping to select partners. All of these are vendor claims, and pricing for the new parts is not yet public.

How can buyers avoid hidden networking costs?

• Lay out the racks first. Keeping GPU servers and leaf switches in adjacent racks lets more links use copper.
• Count every network. Compute, storage, front-end and management fabrics each need ports, cables and optics.
• Match speeds deliberately. A 200G fabric costs less but halves per-GPU bandwidth compared with 400G.
• Price optics with the switches, including spares, and check compatibility with your adapters.
• Size for the cluster you will grow into. Adding a spine layer later is harder than planning ports now.

For smaller setups, see 400Gbps Networking for AI: Do You Really Need Such an Expensive Switch? and 400Gbps Networking on a Desk. For a budget-level view, read Could Your S$1 Million AI Cluster Be Wasting Money Because of Poor Networking?

Buying in Singapore

For short copper links, SourceIT lists the HP ZGX 0.4 m QSFP112 400G DAC at S$365.00, the D-Link 100G QSFP28 1 m DAC at S$255.60 and the D-Link 100G-to-4×25G breakout DAC at S$309.60. For switching: the 4-port 400G MikroTik CRS804-4DDQ at S$2,255.00 for linking a pair of systems, the NVIDIA SN3700 32 × 200GbE switch at S$28,523.25, and the D-Link DQS-5000-32Q28 32 × 100G switch at S$20,186.40 (all listed 6 Nov 2026). We do not currently list 400G or 800G optical transceivers, InfiniBand switches or ConnectX adapters; ask us to source them for your design. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

What is the difference between a DAC and an optical transceiver? A DAC (direct attach copper) is a fixed copper cable with connectors on each end. It is cheap and draws almost no power but reaches only about 2–3 m at 400G/800G. Optical transceivers plus fibre reach 50 m to 2 km but cost far more and draw 9–17 W each.

How much does an 800G transceiver cost? Third-party list prices seen in November 2026 were about US$780–1,000 for 800G SR8 multimode, US$880–900 for DR8 and US$1,440–1,690 for 2xFR4. NVIDIA-branded and volume prices differ.

How many optics does an AI cluster need? In a non-blocking two-tier 400G fabric, roughly 2.5 modules per GPU: one at each GPU's network adapter plus twin-port modules on the leaf and spine switches.

Can I use copper for everything? Only within a few metres. Copper suits server-to-leaf links in the same or an adjacent rack; longer leaf-to-spine runs need active cables or optics.

Do cheap cables really cause problems? They can. Meta attributed about 10% of unexpected Llama 3 training interruptions to switches, cables and network adapters, and ByteDance traced link flapping to poor cable and adapter link quality.