Opinion | Network & Cabling
Compute without connectivity is just expensive silicon waiting for data
The AI sector loves to talk about compute, obsessing over floating-point operations, GPU allocations, and the sheer volume of silicon being crammed into data halls. As AI workloads scale out across thousands of accelerators, networking deserves far more attention than it currently receives because the individual chip is no longer the story; the architecture connecting them is.
The 1.6T reality check: Breaking the laws of physics.
The numbers tracking this shift are becoming extraordinary – while 800G connectivity is actively moving into commercial deployment now, the industry is already engineering and testing 1.6T optical links, but simply doubling bandwidth does not solve the underlying infrastructure problem.
At these extreme speeds, physical laws push back, and push back hard. Signal integrity, optical loss, erratic thermal behaviour, escalating power consumption, and vendor interoperability have all become brutal engineering challenges. Testing specialists are already warning operators that validation can no longer happen in isolation; testing must cover the complete, integrated system rather than individual links.
This is where the traditional distinction between the “network” and the “server” completely breaks down.
The cost of waiting ? Idle clusters bleed cash
AI accelerators do not operate as isolated units; they function as a single, massive distributed system. When thousands of processors participate in a single large language model training run, they must remain perfectly synchronised – if one part of a cluster is forced to wait for data from another, then millions of dollars of compute capacity sits idle.
Network latency and bandwidth are no longer back-office performance metrics, they are directly tied to the fundamental economics of AI. The resulting race to move data faster – and with less energy – is rewriting the physical layout of the rack. Of course copper still has a vital role to play over short distances but it’s physical limitations at extreme bandwidths are rapidly pushing workloads toward optical technologies, co-packaged optics, and liquid-cooled switches.
The thermal trade-off: Scaling optics, compounding heat
Optical transitions present an entirely new tradeoff – Optics consume massive amounts of power.
At hyperscale, operating millions of optical interfaces across a campus turns a facility’s networking layer into a major line item on the energy budget – effectively transforming the industry’s networking problem into a power problem, a cooling problem, and a fundamental rack-design problem.
This is precisely why the next generation of AI infrastructure cannot be designed in engineering silos: The network architect must understand the accelerator architecture, the electrical engineer must understand the switch power profile, the cooling engineer needs to know exactly where those high-heat switches sit, and the data centre designer needs to orchestrate all of this before the foundation is even poured.
Beyond the four walls : the 500km data hall
This connectivity challenge is expanding geographically
A recent demonstration by Arelion and Nokia, pushed 800G optical technology across a 500km route between Amsterdam and London – proving that high capacity optical networking is no longer confined to the data hall – it is actively driving inter-data-centre connectivity, which matters, because the future of AI infrastructure will be inherently distributed; Massive training runs, rapid inference, storage tiers, backups, and specialised compute nodes do not need to share a physical roof.
The network has become the vital connective tissue keeping these distributed systems alive – so is it time to update the terminology and stop talking about the network as just something that connects data centres? In the age of AI, the network is the data centre.
