The rapid buildout of AI data centres is reshaping not only how optical connectivity is deployed, but how long copper-based interconnects can remain competitive across the connectivity stack. While copper-based Direct Attach Cables (DAC) and Active Electrical Cables (AEC) continue to serve the shortest-reach connections today, the combined pressure of rising lane speeds, increasing system density, and the emergence of optics-based architectures is progressively narrowing the space in which copper remains the natural choice.
In Part 1 of this series, we examined how optical networking is transforming data centre connectivity. This insight, constituting Part 2, examines the impact on copper-based cables, where they are currently used in the connectivity stack, the forces driving substitution toward fibre, and what the transition timeline looks like for the wider market.
Distance and speed dictate the data centre connectivity stack
The technology used to transfer data within data centres depends largely on the distance the data needs to travel and the required transfer speed. For short connections of less than three metres, such as ones between racks or on the same rack, copper-based DACs are the simplest and cheapest option. They are high-performance copper cables with no active components. As distances stretch slightly further beyond three metres, and up to seven metres, AECs add a small electronic booster to limit signal attenuation and boost transmission speeds.
Copper-based interconnects such as DAC and AEC are focused on the rack-level connections. Beyond these distances, such as between server rack clusters/nodes in a single building and links between data centres entirely, copper struggles to maintain latency requirements, and data needs to be transferred over fibre connections. Transmission over optical fibre requires the data to be converted from electrical to optical signal first. This is enabled by optical transceivers that are plugged into data centre switches. Hence, they are referred to as pluggables optics.
The challenge with AI workloads is that they require even higher speed and lower latency than what pluggable optical transceivers can deliver. That is because optical transceivers only connect at the switch level, while signals still travel on copper traces on the chip level. At scale, the signal degradation from copper traces creates a speed bottleneck where any amount of delay is crucial. In addition, the power consumption becomes a real problem when data centres are running thousands of servers together at extreme speeds.
Co-packaged optics (CPO) – including their intermediary variation, near-packaged optics (NPO) – aim to minimise the copper traces, and therefore their impact on signals, by moving the optical engine and transceiver components closer to the main chip. The result is lower power use, improved signal quality and speed that meets the performance requirement for AI applications. This development and evolution of optical networking in data centres is gaining traction, with leading players such as NVIDIA, Broadcom and Marvell investing in advancing the technology, a transition we explored in more depth in the earlier insight in this series.
Copper twinax: Near-term tailwinds, longer-term substitution pressure
The industry transition to more optical-based networking technologies will have direct implications for copper twinax cabling in data centre server rooms. DAC and AEC have remained viable in intra-rack (scale-up) and shorter reach inter-rack (scale-out) connections for two reasons – they offered a cost-effective solution for short-distance connectivity, and at the lane speeds demanded by non-AI cloud workloads, typically 25G to 100G, signal quality degradation remained within acceptable limits. However, as lane speeds rise toward 800G-1.6T and beyond in dense AI systems, insertion loss increases, thermal demands worsen and cable management becomes significantly more complex.
These dynamics are eroding copper's share of the connectivity stack through two distinct channels. First, newer AI infrastructure buildouts are increasingly being designed around optics-based architectures from the outset, meaning copper-based data cabling is simply not specified in next-generation capacity additions.
Second, the existing installed base of legacy data centres is also under pressure. As operators retrofit and upgrade facilities to meet AI performance requirements, copper interconnects within the scale-up layer are being displaced by optical alternatives. This transition is already underway at the scale-out level, where displacing copper with fibre is more manageable than at the scale-up level between GPUs. Both channels point in the same direction – a structural and accelerating shift in the connectivity stack from copper toward fibre.
Long-term demand to move more to fibre, with smaller support for copper interconnects
On timing, copper twinax is likely to continue growing through the end of the decade, supported by the first wave of AI volume deployment where DAC and AEC remain cost-competitive at current rack densities and speeds. However, growth is expected to slow in the late 2020s to early 2030s as optical substitution broadens across both new builds and legacy retrofits. Optics is better positioned to capture the next architecture-led wave, leaving copper twinax increasingly confined to shorter-reach, lower-speed, or cost-constrained segments of the market.
The pace and breadth of substitution beyond the shortest reaches will depend heavily on the economics of optics-based networking adoption. Currently, the business case is strongest in hyperscale and large AI environments where power consumption, front-panel density and signal integrity are acute constraints. By the early 2030s, we expect cost to remain prohibitive for many non-hyperscale players, such as legacy cloud and enterprise providers. Therefore, these technologies will continue to function as strategic enablers for the large hyperscale data centre operators, rather than as universally accessible solutions.
For mid-tier and enterprise and cloud data centre operators where these constraints are less acute, pluggable optical transceivers are the more likely lower-cost, intermediate step, preserving compatibility with existing infrastructure while deferring the capital commitment that full-scale optical networking requires. This means copper DAC and AEC are likely to remain relevant in those segments for longer than the headline substitution narrative implies, as the transition path runs through pluggable optics rather than jumping directly to CPO.
A further consideration is whether declining copper demand in data centre interconnects is partially offset by growth elsewhere. Rising energy consumption, driven by AI infrastructure, is accelerating investment in power distribution and grid infrastructure – markets that rely on copper in a fundamentally different form.
While this does not directly substitute for losses in copper data cable demand, the direction of the implication is clear – aggregate copper demand is likely to prove more resilient than a data-interconnect-only view would suggest. This is because volume lost in the connectivity stack is partially absorbed by rising demand in power cables, transformers, and grid infrastructure serving the same AI-driven buildout. The scale and timing of that offset warrants separate analysis, but it is a material consideration for any assessment of copper's long-term demand trajectory.
