The standards published in September 2026 decide how AI data centers will be cabled through 2030 — and the installation details are where projects succeed or fail

Chapter 1: Why the September 2026 Standards Cycle Changes Your Spec

Data center cabling is governed not by one document but by a stack of complementary frameworks — and the AI era has made the differences between them matter. TIA-942-C sets the minimum floor a facility must meet. BICSI 002-2024 is the best-practice ceiling, covering design, commissioning and operations. ISO/IEC 24764 extends the same logic internationally, and EN 50600 adds the European regulatory context. When the conference season drops new proposals and clarifications, specifiers have to decide — quickly — which ones are adoptable today and which are still on the roadmap.

Two events this month carried most of the news:

The practical takeaway is that the two halves of an AI data center cable plant are being pushed in opposite directions. Copper is being asked to reach further (165 m) for large horizontal campus deployments, while fiber is being asked to carry more — more lanes, higher rates, and lower latency. Your specification has to reconcile both.

The fiber-density figure is the one that reframes everything. A single 72-GPU rack using 800G-DR4 optics can consume 864 individual single-mode fibers — roughly 10× the optical fiber of a conventional server rack in the same footprint. Multiply that by 100 racks and you have more than 86,000 fibers before a single spine link is added. The installation discipline that a plant like this needs — polarity control, bend radius, pathway capacity, fire rating, and testing — is what the rest of this guide addresses. If you want the broader design context first, see our article on what is changing in structured cabling for AI data centers.

Chapter 2: BICSI Fall 2026 — Copper Standards Updates Installers Must Apply

BICSI Fall 2026 concentrated on the copper side, where three proposals have direct installation consequences. None of them are ratified standards yet — but a good specifier tracks them now so the design does not need to be reopened later.

2.1 GigaREACH XL — Cat 6A to 165 m

The headline proposal is an extension of Cat 6A to 165 m over 4-pair Class FA channels. This closes a long-standing gap: today, if a horizontal run exceeds 100 m, you must jump to Cat 8.1 (limited to 30 m) or move to fiber — neither of which is a clean answer for a large campus building where the actual requirement is simply "a bit more than 100 m." If ratified, GigaREACH XL lets a single balanced-twisted-pair channel serve distances that previously forced a media conversion.

For AI-adjacent campus deployments — where the GPU hall is a long walk from a management or out-of-band IDF — this directly reduces fiber or extender hardware previously specified to bridge the gap.

2.2 Cat 8.2 and the TERA Connector

There is a continuing push to standardize the TERA connector for data center environments that need 40G at 30 m over copper. The traditional RJ45 interface struggles with the higher-frequency performance Cat 8.2 demands, and TERA's four-pair construction handles the bandwidth more cleanly. For AI racks where a short, robust copper link at 40G is genuinely useful (top-of-rack console, management, or short intra-row runs), TERA is a specifier's option to watch — but it requires a coherent end-to-end ecosystem, so you cannot deploy it piecemeal.

2.3 MPTL Testing — TIA-568.2-D Rev C

The third update is the least visible and, for installers, the most immediate. TIA-568.2-D Revision C is expected to clarify MPTL (Modular Plug Terminated Link) testing procedures for 10 Gbit infrastructure in emerging single-pair architectures. If your AI deployment uses MPTL anywhere — a direct-terminated link from a panel to a device — this revision determines how you set up the tester, which adapters you use, and what constitutes a passing record. Getting the test method wrong means either false failures or, worse, false passes that surface only under production traffic.

Chapter 3: ECOC 2026 — Fiber Innovations You Can Specify Today

ECOC 2026 in Málaga — more than 8,000 industry experts and 340 exhibitors — produced three developments with near-term specification impact.

3.1 Hollow-Core Fiber Reaches Production Readiness

Multiple vendors showed hollow-core fiber (HCF) products now capable of production deployment, with attenuation below 0.5 dB/km and latency below 1.0 ns/m. Because light propagates through an air core rather than a glass core, HCF is roughly 47% faster in terms of propagation velocity, translating to about 30–32% lower propagation latency than conventional single-mode fiber. Microsoft has already deployed HCF in production, reporting a 47% speed increase and 32% latency reduction. For AI all-reduce collective-communication patterns, where total training time is partly set by how long it takes a signal to cross the fabric, that latency reduction is a direct performance feature — not a curiosity.

3.2 1.6T Optical Modules Advance

IEEE study groups continue to push 1.6TbE across both single-mode and multimode fiber, evaluating 16 × 100G and 8 × 200G lane options. As we detailed in our analysis of what 800G and 1.6T trends mean for data center cabling in 2026, the 8-lane parallel-optic path keeps the fiber count identical between 800G and 1.6T — which is what makes a 16-fiber MPO-16/Base-16 plant future-proof without recabling.

3.3 C+L-Band Dual-Window Transmission

Dual-band C+L EDFAs let operators transmit simultaneously in the C and L bands, effectively doubling fiber capacity without installing new fiber. This is largely a long-haul and DCI development, but it matters to the data center edge: it changes the economics of when a new duct or a new cable run is actually necessary, and it favors single-mode OS2 as the backbone medium of record.

Hollow-core fiber and 1.6T optics are converging on the same single-mode connectivity interfaces — the installation complexity is in the details, not the connector

Chapter 4: Hollow-Core Fiber — Installation Rules for the New Latency Layer

If you are going to specify HCF anywhere, specify it in the right place and install it by a stricter rulebook than OS2. HCF is not a "faster cable" in the sense of more bandwidth; it is a lower-latency medium. Its value is concentrated in paths where propagation delay matters: GPU scale-up links, all-reduce paths, time-sensitive control, and latency-critical DCI.

4.1 Where HCF Belongs

Use Case	Latency Sensitivity	Recommended Medium
AI scale-up fabric (all-reduce, NVLink-class)	Very high	HCF where certified; otherwise OS2
Scale-out / spine-leaf leaf uplinks	High	OS2 (G.657.A1/A2 bend-insensitive)
Latency-critical DCI	High	HCF or OS2 depending on distance & budget
Storage, management, out-of-band	Low	OM4/OM5 multimode or Cat 6A copper
General intra-facility backbone	Low–medium	OS2 single-mode

4.2 The Installation Rules for HCF

Chapter 5: Getting 1.6T-Ready — MPO-16 / Base-16 Installation Rules

The 1.6T discussion is really a connector discussion. If the dominant AI optic stays on eight parallel lanes, then the fiber connector for 800G-DR8/SR8 and 1.6T-DR8/SR8 is the same 16-fiber interface — carried today by MPO-16 and by the smaller-form-factor MMC16 (designated by OCP). Specify a Base-16 plant now and the 1.6T upgrade becomes a transceiver swap, not a recabling project. Our dedicated guide, MPO Fiber Solutions: Choosing 8, 12, or 24 Fibers, covers the fiber-count decision in depth; here are the installation rules that follow from it.

High-density termination enclosures with MPO-16/MMC interfaces let a Base-16 plant carry both 800G today and 1.6T tomorrow through the same fiber infrastructure

Chapter 6: OFNP & Fire-Rating Compliance — The Code Item Inspectors Catch

As AI racks push fiber density up, more fiber is routed through overhead return-air and raised-floor supply plenums — which makes the fire rating of the cable a first-order compliance item, not a footnote. Under NEC Article 770, optical fiber carries its own rating hierarchy:

The decisive rule is simple: if air can freely circulate around the cable pathway, the space is a plenum and requires OFNP. A higher-rated cable may always substitute for a lower-rated one, but never the reverse. So OFNP may be used in a riser or general space; OFNR may not be used in a plenum.

6.1 What Inspectors Actually Check

• Jacket markings — the rating must be printed legibly on the jacket at intervals not exceeding 24 inches (610 mm), with a UL listing mark. Faded or missing markings are grounds for rejection.
• Rating vs. location — inspectors verify plenum-rated cable is present in any ceiling cavity where HVAC air returns, and they check the transition points where cable leaves a conduit into open plenum.
• Pathway compliance — trays and supports must suit the rated cable and respect bend radius and fill limits.
• Documentation — cut sheets, UL listings and country-of-origin records on regulated projects.

6.2 OFNP Is Not LSZH — and Not CPR

The good news is that high-density AI cables are increasingly available in plenum-rated construction. Recent OFNP certifications for ribbon fiber cables in the 48–576 fiber range specifically target AI infrastructure deployments — high fiber counts, plenum-rated, ready for overhead pathways. For a deeper dive on jacket materials and ratings, see our practical guide to fiber optic cable types: OS2, OM3, OM4, OFNR and OFNP.

Plenum-rated cable is a code decision, not a performance upgrade — and the jacket marking is what the inspector verifies on site

Chapter 7: The Updated AI Data Center Cabling Checklist

Pulling the standards updates into one field checklist makes the gap between a compliant plant and a network that actually passes cleanly at 800G/1.6T. Use this as a project-commissioning reference.

Two of these items deserve special emphasis for AI builds. First, reading the Fluke test reports before accepting a delivery is what catches defective links while they are still the vendor's problem. Second, tight loss budgets at 400G/800G mean that a link which passed at 10G will not necessarily pass now — test against the specific loss budget of the optic you actually intend to deploy.

Certification testing is the acceptance gate — every AI data center link must be verified against its own transceiver loss budget, not a legacy 10G figure

Chapter 8: Putting It Together — A 1.6T-Ready Specification

Here is how the September 2026 standards updates resolve into a single, coherent AI data center cabling specification. It is deliberately conservative: it adopts what is ready and stages what is coming.

Chapter 9: Key Takeaways

The standards set in September 2026 point in one direction: single-mode, high-density, plenum-rated fiber as the backbone of the AI-era data center

Chapter 10: Frequently Asked Questions