How Much Fiber Do I Need for an AI Data Center GPU Cluster? A Calculator and a Method You Can Defend

How much fiber does your cluster need? Run the numbers

Enter the numbers you know and the page returns fiber-km, cable-km, cable assemblies and a buy-by date. Start with GPU count and fibers per link from your bill of materials, then set tiers and oversubscription from your fabric design. Open the advanced assumptions to change run lengths and buffers. Every default is an illustrative placeholder.

Where does the fiber go in a GPU cluster?

Almost all of it sits in the compute fabric, and the number of links, not the number of racks, drives it. Each GPU has its own fabric link, and a non-blocking leaf-spine design adds roughly one more link per GPU between leaf and spine. At the defaults, 1,024 GPUs sit in 128 nodes of eight and are served by 32 leaf switches with 32 GPU-facing ports each. Each node needs 71 fibers with spares, a 72-fiber cable, and each leaf needs 282, a 288-fiber cable toward the spine. The spine side adds 1,024 links, one per GPU.

Figure 1 shows the structure. NVIDIA’s DGX SuperPOD reference architecture for DGX H100 describes a rail-optimized, full fat-tree compute fabric with eight 400 Gb/s connections per DGX H100 node, and groups nodes into scalable units of 32. Its network fabrics page lists 1,024 spine-leaf cables and 1,020 compute and UFM cables for the four-unit, 1,016-GPU configuration, so spine-leaf links roughly equal GPU count, as the 1,024 links in this model do. Storage and in-band management each use a pair of ports per node on separate fabrics. At the defaults, the leaf-to-spine layer holds 58% of the order. For conventional racks, compare fiber cabling per data center rack.

Figure 2 puts the count in context. Adding each node’s compute links, its storage and management ports and its share of the leaf-to-spine layer, a GPU node lights 136 fibers at these defaults, or 17 per GPU. A conventional server with two duplex ports lights 4. The 34 times ratio sits close to the 36 times that an STL executive was quoted as saying for AI data centers, as Tom’s Hardware reported in May 2026. The match depends on assuming 8 fibers per link, so read it as a sanity check, not proof. Links between sites are sized like a metro ring.

Fiber type and optics are the other half of the specification. NVIDIA’s documents do not state the cable type or the fibers per link, so read the optic datasheet. A 400G DR4 module uses 8 fibers, four transmit and four receive, on an MPO-12 connector, according to a transceiver vendor’s explainer, which is a secondary source. The Ethernet standard behind 400G, IEEE 802.3bs, was approved in December 2017, and single-mode fiber is defined by ITU-T G.652. Tom’s Hardware also reported that manufacturers have shifted output from G.652D toward G.657A fiber, so confirm which category your supplier can deliver.

How much extra should you order beyond the fibers you light?

Order for the lit fibers plus five adjustments, and test where your numbers sit against the standard cable sizes. In this example the adjustments take 819.2 lit fiber-km to a 1184.4 fiber-km order, and raising spares from 10% to 13% can add 68% because it crosses a cable-size step.

Re-pulls are the part of contingency teams rarely price, because every failed link test sends a technician back with a new cable, and a cluster has thousands of links. The Technician Checklist Generator turns your acceptance tests into a checklist crews actually follow, which is the cheapest way to lower that number.

Figure 3 builds the order one step at a time. The fibers actually lit come to 819.2 fiber-km; the order is 1,184.4. Slack adds 142.8 fiber-km, the most of any step, because the runs are long and every one gets the 15%. Standard sizes add only 46.1 because 71 and 282 fibers land just under the 72 and 288 steps. That is luck, not safety.

Figure 4 moves one input at a time. Fibers per link dominates: 20% fewer (6 instead of 8) cuts the order by 14.5%, while 20% more (10 instead of 8) adds 67.7%, because 32 uplinks of 10 fibers with spares no longer fit one 288-fiber cable. Spares are the trap. At 8% or 12% they change nothing, yet at 13% the leaf cable doubles to two 288-fiber cables and the order rises 68%, from 1,184 to 1,987 fiber-km.

The fabric design moves the answer as much as the optics: a third tier adds about 58% and a 2:1 oversubscribed fabric saves about 29%. Fiber itself is a small part of the cost. At the roughly $6.30 per fiber-km that Tom’s Hardware reported, the 1,184 fiber-km in this order is under $8,000 of bare fiber, so the risk is delivery, not the glass price.

When do you need to buy the fiber?

Work backwards from cluster start: buy-by date equals start date minus supplier lead time minus a shipping and QA buffer, and treat the lead time as a number to confirm, not assume. The reports below describe the market in general and some rest on a single source, so use them to set expectations and then get written quotes.

In May 2026, Tom’s Hardware reported that fiber demand from data centers grew 76% in 2025, citing CRU data, that prices rose about 70%, and that lead times reached up to 20 weeks for large-volume buyers and up to a year for smaller ones. It reported that two large Chinese suppliers, Hengtong and FiberHome, were booked into early 2027. CRU Group’s OFC 2026 review names preform as the binding constraint and says ribbon cables for AI data centers, from 1,000 to over 13,000 fibers, carry lead times over a year. Those are market reports, not quotes to you.

Figure 5 applies those lead times to an example plan. If you order on 21 October 2026 with a 20-week lead time and a four-week buffer, the earliest start is 7 Apr 2027, only 8 days ahead of a 15 April 2027 cluster start. At 52 weeks the earliest start is 17 Nov 2027, about 31 weeks late. With so little margin, any slip in a quote moves the cluster date, which is why the buy-by date matters more than the exact fiber-km.

Figure 6 shows one way to cut exposure: one call-off per 256-GPU scalable unit. NVIDIA’s table shows spine-leaf cables growing by 256 with each unit, from 256 to 1,024, which supports building and cabling in units. In this example, four call-offs cut the peak quantity on site from 1,184 to 296.1 fiber-km. The catch is commitment: phasing only protects you if the supplier agrees in writing to hold allocation for every call-off.

How do you explain the number to a board, a customer or a network team?

Write the consequence, not the parameter. Every figure on this page becomes a sentence someone else has to accept: a board approving a purchase order, a customer told a go-live date has no slack, a network team told why a spare allowance was capped. Engineers tend to write these as parameters. Readers need to know what changes for them and what you have secured. When a fabric change goes wrong, the Outage Message Generator drafts the customer update in the same plain style. The examples below use this page’s illustrative numbers.

What should you check before placing the order?

Confirm eight things, and write down who owns each.

• GPU count and phasing match the approved design, by scalable unit, not the sales forecast.
• Fibers per link come from the optic datasheet and the bill of materials, not from a reference design that does not state them.
• Fabric tiers and oversubscription are fixed, because a third tier or a changed ratio changes every leaf cable.
• Every cable sits clear of a standard-size step, or the step is a recorded decision.
• Node-to-leaf and leaf-to-spine runs are measured along the tray route, not drawn as straight lines.
• Lead time is in writing for each cable type and states whether the date is shipment or delivery.
• Call-offs are planned by scalable unit, and the supplier has committed allocation for each one.
• Link acceptance tests are agreed before the first delivery. The Technician Checklist Generator can turn them into a checklist crews use.

Frequently asked questions