· via dev.to (home feed)
SC26: power and cooling limits push 100,000-accelerator AI clusters toward scale-across networking
Ohio State researchers told SC26 that power, cooling and space constraints are pushing 100,000-accelerator AI clusters beyond single data centers, with Google, Cisco and Marvell building scale-across networks.

On November 20 at Supercomputing 2026, researchers from The Ohio State University laid out an argument that the next constraint on AI infrastructure is not the accelerator itself but the building around it. As summarized in coverage by Tom's Hardware and Chip Dispatch, modern AI clusters routinely hold more than 100,000 accelerators plus potentially millions of CPU endpoints, and accelerator performance is improving faster than the networks connecting them. The upshot is that network efficiency increasingly decides how much of a cluster's compute is actually usable.
A third networking domain
AI networking has traditionally been split into two tiers. Scale-up fabrics such as NVLink, AMD Infinity Fabric, Intel UPI and the emerging UALink tie accelerators together inside a single system. Scale-out technologies such as Ethernet, InfiniBand, HPE Slingshot and Ultra Ethernet connect racks within one data center. The SC26 presentation adds a third domain, scale-across: links between buildings, campuses and even regions that must behave, from the application's point of view, as one system pursuing one performance target.
The driver is physical
According to the researchers, this is not architectural fashion but necessity. Concentrating every accelerator at a single site may be difficult or impossible given limits on power, cooling or floor space, so combining sites is often the only route to a larger system. Tom's Hardware reports that Google introduced its Virgo network architecture earlier in 2026 for exactly this scenario: AI workloads that outgrow one facility's space and power budget, with an entire campus treated as a single system and separate networks for accelerators, servers and data centers. The same coverage reports that Cisco and Marvell are developing scale-across technology, from high-capacity routers to switching silicon and optical interconnects for distributed clusters, though so far without commitments on capacity, shipment schedules or pricing.
Distance still bites
The researchers were candid about the limits. Optical links cannot eliminate propagation delay, and congestion, packet loss and synchronization overhead all eat into effective performance. Two GPU clusters joined by fiber do not automatically behave as one tightly integrated machine. Training workloads feel this first, because accelerators constantly exchange model parameters, gradients and intermediate results through collective operations, and delays affecting even a single accelerator can drag down the whole cluster. Mitigation is increasingly built into the network itself: NICs that handle RDMA without continuous CPU involvement, collective offload in HPE Slingshot and NVIDIA SHARP, and the behavior of software layers such as libfabric, MPI and UCX.
Bandwidth alone is not the metric
The industry is already shipping 800 Gb/s and 1.6 Tb/s Ethernet, but the presentation argued that peak transfer rates alone say little about a system's real capability, since application performance depends on topology, congestion management, protocols and how well software cooperates with the fabric. Demand is climbing regardless. IDC survey data published by Cisco, drawn from 293 organizations, found that 89% expect within-data-center and 91% expect between-data-center bandwidth to grow by at least 11% over the next year, while 36% of enterprises anticipate between-data-center growth above 51%.
Why it matters
If the binding constraint on AI capacity is now power and cooling rather than silicon, capacity planning has to expand beyond the single facility. Buying faster ports does not guarantee faster training: cross-site latency, congestion and collective communication performance have to be modeled before a multi-site design is signed. For operators, campus- and region-scale systems have become legitimate design targets, and vendors from Google to Cisco and Marvell are already positioning products for a world where the network, not the data center, defines the size of the machine.
- #networking
- #ai-infrastructure
- #data-centers
- #supercomputing
- #cloud