The AI data centre has a space problem – it also has a power problem
As power and physical space become harder to secure, Arista thinks the next generation of AI infrastructure will have to spread across sites — putting the network at the centre of the architecture.
As AI clusters grow from thousands of accelerators towards hundreds of thousands and beyond, just adding more GPUs to the same facility is becoming increasingly difficult – Arista’s latest thinking is the answer may be to stop treating the data centre as a single physical location; In a recent paper, they argued that restrictions on available power and physical space are pushing AI infrastructure towards what it calls “scale-across” – extending AI fabrics across multiple locations rather than building everything into one giant cluster.
This sounds like a networking problem, is it becoming a compute problem? Arista thinks so.
The data centre is getting bigger than the building
Traditional data centre interconnect is designed largely around moving applications and data between facilities – AI changes the requirements.
Training workloads generate enormous, highly synchronised and bursty communication between accelerators, Arista argues that simply connecting two sites with a conventional DCI link is not enough when parts of an AI fabric may be distributed across different locations – the network therefore has to do more than provide bandwidth; It has to manage congestion, maintain predictable latency, steer workloads and protect AI jobs when something goes wrong.
Their proposed architecture uses its 7800 platform as the spine for these distributed fabrics, with what the company calls REACH capabilities covering routing intelligence, congestion protection and high-radix scaling. Arista says its approach is designed to extend AI fabrics towards up to one million XPUs across locations.
Networking is moving into the rack
The scale-across argument follows another important shift in Arista’s strategy.
In June, the company introduced its 7060XE7 family of 1.6Tb/s networking platforms, designed for rack-scale AI infrastructure. They position the network as an integrated component of the AI system rather than a separate infrastructure layer, with air- and liquid-cooled configurations aimed at increasingly dense deployments. The platforms also include congestion-management and telemetry features designed around the bursty communication patterns of AI workloads, which matters because the networking problem is no longer simply “How much bandwidth do we have?” It is increasingly becoming “How efficiently can we connect compute?”
The new AI infrastructure map
For operators, this creates an interesting change in how AI capacity could be built.
Instead of waiting for one enormous site with sufficient power, space and cooling, future AI infrastructure could be assembled across multiple facilities – provided the network can make those locations behave more like one coordinated system.
Arista’s latest announcements point towards that model: higher-speed networking inside the AI rack, and intelligent networking between AI sites. They still sell networking hardware and software, but the strategic shift is bigger: AI is starting to push the boundaries of what counts as a data centre and the next AI factory may not be one building at all – It may be a network.
