Scale-across networking is emerging as the third way to wire up AI infrastructure, because clusters of more than 100,000 accelerators are outgrowing the power, cooling and floor space of a single data center. Google, Cisco and Marvell are all building for it, and researchers at The Ohio State University laid out the challenges at the Supercomputing 2026 conference (SC26).
The short version: faster chips don’t help much if the network can’t keep up.
Why one building is no longer enough
Modern AI clusters can hold more than 100,000 accelerators and potentially millions of CPU endpoints. That puts heavy demands on bandwidth, latency and congestion control.
There is also a mismatch inside the machines. An accelerator’s compute performance keeps rising faster than its ability to communicate. A faster GPU therefore doesn’t always deliver a matching gain in application speed, and the quality of the network decides how much of the compute is actually used.

Scale-up and scale-out, the two existing domains
High-performance computing has leaned on two kinds of interconnect so far.
- Scale-up networks join processors and accelerators inside tightly integrated systems. Examples include AMD’s Infinity Fabric, Intel’s UPI, Nvidia’s NVLink and the newer UALink.
- Scale-out networks join servers or rack-scale systems into larger clusters, using Ethernet, InfiniBand, HPE Slingshot or Ultra Ethernet.
The two have different needs for bandwidth, latency and reach, so they serve different jobs.

What scale-across adds
Scale-across extends high-performance connectivity between buildings, campuses or even regions. Scale-out usually stays within one data center. Scale-across doesn’t.
The appeal is practical. Instead of packing an ever larger number of accelerators into one site, which may be hard or impossible given power supply, cooling and space, an operator can combine resources from several locations to reach a target level of performance.
The shift has started. Google introduced its Virgo network architecture earlier this year for AI workloads that exceed the space and power capacity of any one data center. The company plans to treat an entire campus of facilities as a single system, with specialized networks linking accelerators, servers and data centers. Cisco and Marvell are developing high-capacity routers, switching chips and optical interconnects for geographically distributed clusters.

The physics problem
Connecting two big GPU clusters with fast links doesn’t turn them into one tightly integrated machine. Scale-up networks give very high bandwidth and low latency over short distances. Links between data centers have to deal with long-distance limits instead.
Even optical networks can’t remove propagation delay. Congestion, packet loss and synchronization overhead cut effective performance further.
AI training makes this worse. Accelerators constantly exchange model parameters, gradients and intermediate results, and a delay affecting a single accelerator can drag down the whole cluster. Efficiency therefore depends on topology, congestion management, communication protocols and software, as well as raw bandwidth.
Smarter adapters, switches and software
To cope, network hardware is getting more capable. Modern network interface controllers can run Remote Direct Memory Access (RDMA) operations without constant CPU involvement. HPE Slingshot hardware collectives and Nvidia SHARP push some communication operations into the network itself. SmartNICs and data processing units take selected network and infrastructure tasks off the host processors.
Software matters just as much. Communication frameworks such as libfabric, MPI and UCX decide how applications exchange data and use the adapters. GPU-aware communication and GPUDirect RDMA skip needless copies through host memory, and optimized collective operations reduce communication overhead across large numbers of accelerators, according to the researchers.
Peak speeds don’t tell the full story
The industry keeps introducing faster links, including Ethernet versions with nominal speeds of 800 Gb/s and 1.6 Tb/s. But a headline transfer rate says little about real application performance. That depends on how well hardware and software work together to move data, avoid congestion and keep thousands of processors in step.
For buyers and operators, the lesson is to measure whole workloads rather than link speeds alone.
What this means for AI infrastructure planning
As single accelerators get faster and clusters outgrow their buildings, communication becomes as important as processor performance. The next generation of AI systems will depend on network adapters, switches and, perhaps most of all, software that can tie distributed resources together without excessive overhead.
The vendors named above are already shipping or developing pieces of that stack. Whether scale-across clusters can match the behavior of a single-site system is the open question, and the physical limits of distance will decide how closely they get.
Source: Tom's Hardware