Optical modules running at 800Gb/s or faster will make up more than 60% of global optical-module shipments in 2026, up from 19.5% in 2024, according to a February forecast from TrendForce. The Taiwanese market researcher puts much of that jump down to AI data centres, and to Google in particular.
Optical modules, or transceivers, are the small units that plug into switches and servers and turn electrical signals into light for fibre cables. Every link between racks in a large AI cluster needs one at each end. As clusters grow to tens of thousands of accelerators, the number of modules, and their speed, has become a planning constraint in its own right.
Why AI clusters run on light
A port running at 800Gb/s moves 100 gigabytes of data a second. Data-centre links have mostly run at 100G or 400G, which was enough for web serving and databases. AI training is different: the chips in a cluster exchange results with one another at every training step, so the network is effectively part of the computation. If the links are too slow, expensive accelerators sit idle waiting for data.
Copper cables can carry those speeds only over short distances before the signal degrades, which is why links between racks and across halls use fibre. A module contains lasers that turn electrical signals into light, photodetectors that turn light back into electrical signals and, in most designs, a digital signal processor that encodes and cleans up the data. Linear-drive designs leave that processor out to save power and cost, relying on the switch chip’s own electronics instead.
The size of current projects multiplies the demand. TSN reported in March that Microsoft would take over a 900MW data-centre project originally planned for Oracle and OpenAI, and every accelerator in a site of that kind needs network ports, often several. For readers newer to the technology, our guide to machine learning and deep learning explains how the models trained on these clusters work.
Google’s TPUs alone need millions of modules
TrendForce’s forecast centres on Google’s Ironwood TPU. Google connects these chips in a 3D torus, a network in which each chip links directly to neighbours in three dimensions, and uses its Apollo optical circuit switches to reconfigure links between groups of chips. TrendForce expects about 4 million Google TPUs to ship in 2026, driving demand for more than 6 million 800G-and-faster modules.
The optical circuit switches matter for power. TrendForce says one draws about 100 watts, against about 3,000 watts for a comparable electrical packet switch, because it steers light with mirrors rather than converting signals. It also says the design lets Google move from 800G to 1.6T by swapping modules rather than rebuilding the network. Two Chinese suppliers, Innolight and Eoptolink, hold about 80% of Google’s 800G-plus orders, by TrendForce’s count, while Lumentum supplies the optical-switch hardware.
The standards are partly in place
Ethernet at 800Gb/s already has a finished standard. IEEE 802.3df-2024, published on 15 March 2024, added an 800Gb/s Ethernet MAC and 400G and 800G physical layers. The next step, 1.6Tb/s Ethernet under IEEE P802.3dj, was still listed as a draft standard on 28 October 2026, with its project authorised in December 2022.
Longer links are being standardised too. In April 2025 the Optical Internetworking Forum published its 800LR implementation agreement, a single-wavelength coherent interface for 800G over up to 10km without amplifiers. That is the range that connects buildings on a data-centre campus.
Standards matter because modules from different vendors have to work in the same switches. Where form factors or management interfaces differ, buyers risk being tied to a single supplier, and each module adds power and heat that a large site has to budget for across tens of thousands of ports.
Pluggables now have a rival inside the switch
The alternative to pluggable modules is co-packaged optics (CPO), which puts the optical engines in the same package as the switch chip. In March 2025 NVIDIA announced Spectrum-X Photonics and Quantum-X Photonics switches that use CPO. It claimed 3.5 times better power efficiency and four times fewer lasers than pluggable designs, with Spectrum-X configurations of up to 512 ports at 800Gb/s. NVIDIA said it would keep supporting pluggables from suppliers including Coherent, Eoptolink, Fabrinet and Innolight.
By July 2026, TrendForce reported that NVIDIA was shipping Spectrum-X CPO switches to select partners and that Broadcom’s 51.2T Bailly CPO switch was in volume manufacturing, with up to 70% lower power than pluggables and validation with Meta. Celestica has since introduced a 102.4Tbps switch with co-packaged optics. Pluggables are not standing still either: NVIDIA and Cisco have announced plans for USFP, a pluggable form factor at 400Gb/s per lane, with no ship date yet.
Optics now sets the pace of AI build-outs
TrendForce’s numbers turn a networking detail into a supply question. If one customer needs more than 6 million high-speed modules in a year, and two suppliers fill most of those orders, the cluster timetable depends on their factories as much as on chip output. The same pressure is showing up further down the chain: TrendForce has also forecast an optical-fibre shortfall of 180 million fibre-kilometres in 2026. For buyers, the practical result is that optics has moved from a line item to something negotiated early, alongside GPUs and power. It also gives smaller photonics firms an opening. In the UK, Salience Labs has raised $125 million for optical switching, and Openreach has added national 800G services with Ciena, showing 800G spreading beyond hyperscale data centres into telecoms networks.
Co-packaged optics moves the bottleneck into packaging
CPO promises lower power, but it changes who controls supply. A pluggable module can come from several vendors and be swapped on site. A co-packaged engine is built into the switch, which ties optics to advanced chip packaging and silicon photonics capacity. TrendForce names exactly those, optical engines, silicon photonics and advanced packaging, as the bottlenecks, and expects volume to ramp only in 2027 and 2028. That timing suggests pluggable 800G and 1.6T modules will carry most AI traffic for at least the next two years, while CPO grows in the largest switch deployments.
The next fixed marker is the IEEE’s approval of 802.3dj, which would give 1.6Tb/s Ethernet the finished standard that 800G gained in 2024.
Sources
- TrendForce: 800G+ optical modules to exceed 60% of shipments in 2026 (10 Feb 2026)
- TrendForce: co-packaged optics shipments and bottlenecks (27 Jul 2026)
- IEEE Standards Association: IEEE 802.3df-2024
- IEEE Standards Association: IEEE P802.3dj
- NVIDIA: Spectrum-X and Quantum-X silicon photonics networking switches (18 Mar 2025)
- OIF: 800LR coherent implementation agreement (22 Apr 2025)
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- Microsoft’s 900MW Power Grab: The AI Infrastructure War Just Escalated
- Machine Learning vs Deep Learning: The Complete Guide (With Pizza)
- NVIDIA and Cisco Plan USFP, a New Pluggable Optics Form Factor for 400Gb/s Per Lane
- Celestica Introduces DS6005, a 102.4Tbps AI Switch With Co-Packaged Optics
- Optical Fibre Prices Jump Severalfold as AI Demand Outruns Supply Until 2028
- UK Photonics Start-up Salience Labs Raises $125 Million to Scale Optical Switches for AI Data Centres
- Openreach Adds National 800G Services With Ciena’s WaveLogic 6, and Names 112 More Exchanges for Copper Stop-Sell
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X



