Powering AI Data Centres Without Compromise in the 1.6T Era

Artificial Intelligence (AI) has had nothing short of a seismic impact on data centres, especially when it comes to buildouts. In the US alone, spending on the construction of infrastructure has tripled over the last three years¹, with operators looking to solutions that can withstand the sheer data throughput that comes with training AI models.

At the same time, optical transceivers are being spotlighted for their role in offering low insertion loss and strong signal integrity for these facilities. But with AI scaling bandwidth demands faster than traditional workloads ever have, it’s time that these transceivers must also take the next step to keep pace.

From Cloud Traffic to AI Fabric

Historically, transceiver innovation has closely mirrored the evolution of data centre architectures. Take traditional cloud environments, where traffic patterns moved in a north-south direction and consisted mostly of users accessing cloud applications and databases to respond to queries. 100G and 400G transceivers were perfectly suited for these requirements, before the exponential growth of AI and cloud environments resulted in a jump to 800G.

800G handled the new east-west, server-to-server, communication – a more challenging feat given the vast amount of modern data centre traffic, and the intense focus on high-volume, low latency communication. Its capability in supporting efficient, high-performance network infrastructures was strong, until the sheer amount of data from Large Language Models (LLMs) pushed the transceivers to capacity.

As a result, the 800G transceivers which were once considered cutting-edge are quickly becoming an unwanted bottleneck in AI data centres. Without sufficient bandwidth and ultra-low latency provided by transceivers, General Processing Units (GPUs) simply cannot operate at full efficiency.

Pair this with current Dynamic Random Access Memory (DRAM) pressures causing hyperscalers to look at efficiencies in their network², and solutions that can deliver this bandwidth are of paramount importance for data centre operators. This, in turn, has led to the development of 1.6T transceivers.

What are 1.6T Transceivers?

Designed for AI and Machine Learning (ML) environments, 1.6T transceivers are fundamental in supporting the next generation of AI. What may initially appear to be just the next strategic jump from 800G, actually reflects a monumental performance leap for the sector.

For example – using 224G Pulse Amplitude Modulation (PAM4) lanes for superior density – 1.6T transceivers double the speed of 800G to maximise bandwidth for AI and cloud via the switch to offer far more throughput. This is done without increasing the data centre’s physical footprint, meaning fewer switch layers, shorter data paths, reduced hop count, and lower overall latency for faster synchronisation between GPUs for AI training clusters.

1.6T is also a greener choice for high-capacity networks given its ability to reduce power consumption. Despite electricity consumption expected to reach ~990 terawatt-hours³ (TWh) in global data centres by 2030, 1.6T delivers four times the bandwidth without four times the power, shifting efficiency benchmarks across network infrastructure.

However, one problem still persists: the demand for high-speed optics which is outpacing supply. Data centre operators are looking to leading manufacturers such as NVIDIA, Juniper Networks, and Cisco for answers, but their inventories are thin and their lead times are stretched, creating a space for a new wave of ‘alternative’ suppliers such as AddOn Networks to thrive.

Adding 1.6T to the AddOn Networks Offering

As the world’s largest independent manufacturer of compatible fibre-optic transceivers, we recently added 1.6T transceivers⁴ to our extensive portfolio to answer global data centre demands and bring seamless interoperability with all major manufacturers.

AddOn’s new 1.6T OSFP 2xDR4 transceivers and Direct Attach Cables (DACs) bring high-speed, reliable data transmission for short-reach networking environments, guaranteeing low insertion loss and strong signal integrity while complying with OSFP and IEEE standards. Our solutions directly interlink AI clusters and high-performance computing (HPC) environments reliably and efficiently.

Within these architectures, operators are deploying switching platforms from multiple providers in order to optimise performance and cost. Now, with the addition of multi-vendor compatible 1.6T transceivers, deployments without vendor lock-in and the seamless integration into existing fabrics can be achieved. This empowers infrastructure teams to introduce next-generation speeds into existing environments with confidence, maintaining operational continuity while preserving full control over their network strategy.

Future Proofing AI Data Centres

In essence, 1.6T transceivers are foundational infrastructure for the AI era. With the addition of AddOn Networks solutions, operators can build AI data centres that are faster, flatter, and free from vendor constraints.

Expect to see widespread 1.6T deployments through 2026 and beyond, before the inevitable leap to 3.2T - widely expected to be broadly available in 2027 – which will mark the next phase of network evolution. AddOn Networks will continue to redefine performance ceilings in AI infrastructure, no matter the environment.