Why Are 800G and 1.6T Optics So Hard to Get? The 2026 Shortage

If you are building an AI cluster in 2026, the transceivers can turn out to be harder to source than the GPUs. 800G modules are the workhorse of current GPU fabrics, 1.6T modules are entering volume production for the newest switches and NICs, and both depend on a narrow supply chain for lasers and signal-processing chips. This guide explains why the shortage is happening, how long analysts expect it to last, and how to plan optics for an AI fabric so they do not become the last item holding up a deployment.

Why is there a shortage of 800G and 1.6T transceivers?

Demand grew faster than component capacity. TrendForce expects the AI optical transceiver market to grow from $16.5 billion in 2025 to $26 billion in 2026, and identifies tight supply of EMLs (electro-absorption modulated lasers) and CW (continuous-wave) lasers, along with precision manufacturing steps such as optical alignment, as the main limits on output (TrendForce). In unit terms, TrendForce projected shipments of 800G-and-faster modules rising from 24 million in 2025 to nearly 63 million in 2026, and noted that NVIDIA had secured capacity at key EML suppliers, pushing lead times for other buyers out beyond 2027 (Communications Today). Every GPU in a scale-out fabric typically needs at least one high-speed NIC port, a switch port and an optic at each end, so optics demand scales directly with GPU shipments.

What is the indium phosphide bottleneck?

The lasers used in single-mode 800G and 1.6T modules are built on indium phosphide wafers, and both the substrates and the laser fabs are concentrated in a few companies. TrendForce counts five EML suppliers globally: Lumentum, Coherent, Mitsubishi, Sumitomo and Broadcom (Communications Today). In August 2026, TrendForce reported Lumentum's CEO saying the InP shortage had surpassed those in DRAM and NAND, that customers were ordering further ahead and paying substantial deposits, and that analysts expect the imbalance to persist beyond 2027, with Chinese export controls on related materials adding pressure (TrendForce). LightCounting estimated in August that demand for high-speed optical components was outstripping supply by roughly 30% (IEEE ComSoc Technology Blog).

Who has the supply locked up?

• NVIDIA: in March 2026 it committed $2 billion each to Lumentum and Coherent, alongside multibillion-dollar purchase commitments and capacity access rights for advanced lasers and optical products (Optics & Photonics News).
• Hyperscalers: Coherent told investors in February 2026 that most of calendar 2026 was already booked and 2027 was filling quickly, with a data center book-to-bill above 4x (Futurum).
• Vendor diversification: LightCounting reports hyperscalers widening their supplier pools from two or three module makers to five to seven, which spreads demand but also absorbs capacity at second-tier suppliers.

Why are 1.6T modules even tighter?

1.6T modules need faster lasers (200G per lane) and new DSP chips, and both are early in their production ramps. DIGITIMES reported in April 2026 that 1.6T modules were nearing mass production for AI data centers, with Broadcom and Marvell as the main DSP suppliers (DIGITIMES). Coherent says 1.6T is ramping across multiple customers and that it was about 80% of the way to doubling its internal InP wafer capacity by Q4 of calendar 2026 (Futurum). Capacity is growing, but the largest buyers are first in line for it. Unless your switches and NICs actually run at 1.6T, an 800G or 400G design is far easier to source this year.

What lead times should you expect for 800G optics?

There is no single reliable lead-time figure for enterprise buyers. Fusion Worldwide reported lead times past 50 weeks on some NVIDIA transceiver SKUs in September 2026 (Fusion Worldwide), but quotes vary widely. Lead time depends on the module type (multimode SR8 versus single-mode DR8 or 2xFR4), the coding you need, the quantity, and whether the supplier holds allocation. What the evidence supports is that supply is allocated, that large buyers are placing orders well ahead with deposits, and that prices are rising. Ask for a committed ship date on each line of the quote, not just a price, and treat optics as a long-lead item in the project plan, on the same timeline as the GPU servers. Our guide to managing GPU hardware lead times covers the scheduling side.

How should you plan optics for an AI fabric?

• Count every link at design time. For each server, list the NIC ports, the switch ports they land on, the distance, and the optic or cable at each end, then add leaf-to-spine links and spares. Our 400GbE fabric design guide shows how the port counts add up.
• Match form factors exactly. 800G switches such as the NVIDIA SN5600 use OSFP cages, while many NICs use OSFP or QSFP112 at 400G, and twin-port and flat-top versions are not interchangeable.
• Order optics with the servers and switches, not after the racks arrive. A cluster without cables is a cluster that cannot be accepted.
• Order a small stock of spares for the module types you use most, because a failed optic during burn-in will otherwise wait in the same queue as the original order.
• Use 400G where the workload allows it. Storage, management and inference front-end networks rarely need 800G, and 400G optics and cables are easier to obtain.

Can DAC or AEC cables replace optics for short runs?

For in-rack and adjacent-rack links, copper is the most effective way to cut optics demand. NVIDIA's LinkX documentation for its 800Gb/s twin-port OSFP family lists passive DAC up to 3 meters and active copper (ACC) at 4 to 5 meters (NVIDIA). Active electrical cables with retimers go further: Credo markets 800G AECs up to 7 meters (Credo). Copper uses no lasers, draws less power than a pair of optics, and is generally easier to source. The trade-offs are cable bulk, bend radius and weight in dense racks, and a hard reach limit, so it works best when switches sit in or next to the GPU racks. Beyond about 5 to 7 meters, you need optics or active optical cables. Our cabling guide for high-density GPU racks covers layout.

Should you use OEM or compatible optics?

Compatible optics from reputable manufacturers, coded for your switch or NIC, are a legitimate way to widen supply, and many AI fabrics run them. The risks are coding mismatches, firmware updates that tighten vendor checks, support disputes, and modules of unknown origin sold on the grey market during a shortage. Approve compatible part numbers as alternates in advance, test a sample in your own switches and NICs, and buy from a supplier who can show where the modules came from. Our guide to third-party versus OEM optics goes through what to check, and our 800G switch guide explains where 800G is genuinely needed.

Nexus Compute supplies new 800G optics and cables, passive DACs, AOCs and switches, sourced through authorized distribution with full manufacturer warranty. We check coding and compatibility against your switches and NICs before dispatch, quote within 48 business hours, and offer purchase orders and net terms for qualified organizations. For the wider picture on memory, storage, CPU and GPU supply, read our 2026 AI hardware shortage playbook and our guide to securing AI compute.