AI accelerator clusters move beyond single data centres due to power limits
Researchers at Supercomputing 2026 say power and cooling limits are forcing AI systems with more than 100,000 accelerators to stretch across multiple data centres and regions. New scale-across networking technologies from Google, Cisco and Marvell aim to connect these distributed clusters.

Key takeaways
- Power, cooling and space limits are pushing AI clusters with more than 100,000 accelerators beyond single data centres.
- Researchers at Supercomputing 2026 described a new scale-across networking domain that links separate facilities and regions.
- Google Virgo, along with Cisco and Marvell technologies, is being developed to connect geographically distributed AI clusters.
- Network delays, congestion and software design matter as much as raw link speed for large AI training workloads.
In this article
What changed at SC26
AI accelerator clusters are being pushed beyond single data centres by power and cooling limits, according to Tom's Hardware. Researchers from The Ohio State University presented the findings at Supercomputing 2026 on 20 November 2026, describing how systems with more than 100,000 accelerators must now stretch across multiple facilities and regions. Clusters now hold more than 100,000 AI accelerators and potentially millions of CPU endpoints, and faster AI GPUs do not necessarily deliver proportional gains because communication capabilities are not keeping pace. Power availability, cooling needs and physical space make it hard or impossible to keep all of that hardware in a single data centre. The industry is therefore looking at a third type of network, called scale-across, that links facilities in different buildings, campuses or regions.
How scale-across networking works

Traditional systems use scale-up networks inside a server, such as Nvidia NVLink, AMD Infinity Fabric, Intel UPI or UALink, and scale-out networks inside a data centre, such as Ethernet, InfiniBand, HPE Slingshot or Ultra Ethernet. Scale-across extends high-performance links between separate locations. Google has introduced its Virgo network architecture to treat a whole campus of computing facilities as one system, while Cisco and Marvell develop routers, switching chips and optical interconnects for the same purpose. Long-distance links are difficult because even optical networks cannot remove the time it takes light to travel between regions.
Who is affected
Operators of large AI infrastructure are directly affected by the shift to multi-site designs. Network vendors must supply hardware that can handle long-distance links without losing performance. Companies running AI training workloads also face new demands because accelerators constantly exchange model parameters, gradients and intermediate results. Delays that affect individual accelerators can dramatically reduce the performance of the entire cluster.
What this means for the industry
Hardware is becoming more sophisticated, with network cards that can carry out Remote Direct Memory Access, which lets them read and write memory directly without involving the main processor. Technologies such as Nvidia SHARP and HPE Slingshot hardware collectives move communication work into the network itself. SmartNICs and data processing units are specialised cards that handle infrastructure tasks in dedicated hardware, reducing the load on host processors. Software matters too, with frameworks such as libfabric, MPI and UCX shaping how data moves, while peak speeds like 800 Gb/s or 1.6 Tb/s do not tell the full story because real performance depends on topology, congestion management and software cooperation.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



