Networking, Fabrics & Optics

In a synchronous AI cluster the network sets the pace of every accelerator: the job runs at the speed of its slowest collective and stalls on its longest tail.

The scale-up domain, where accelerators talk at memory speed over a fabric an order of magnitude faster than the back-end, sets your tensor- and expert-parallel ceilings, your MoE economics and your blast radius. Size it wrong and you pay in idle accelerators or larger restarts.

The switch ASIC, NIC and DPU set the ceiling for every fabric above them. Confirm their SerDes generation, buffer architecture and offload engines fit the topology before the purchase is frozen.

A scale-out protocol is a 3–5 year bet on who supplies your switches, how much link rate survives as goodput, and whether you can leave the vendor whose firmware you depend on.

Topology and oversubscription turn a GPU count into a bill of materials, blocking factor and MFU ceiling. The 2026 mistake is one ratio for the whole fabric, or one read off the workload's label instead of its measured locality, exposure and failure headroom.

A non-blocking Clos delivers its rated bandwidth only if congestion control and load balancing both work, and a mistuned fabric can strand a large share of it. In-network reduction speeds up the collectives that support it; it does not gate the fabric.

AI clusters need an out-of-band network that reaches a wedged node while the data plane is down, and a timing plane that keeps logs from every device on one timeline, inside a measured error bound, through the GNSS outage you plan for.

Scale-across is an optional tier that links sites for portfolio, failover and overflow work. Split one job across sites only if its algorithm, placement and failure plan survive the latency, bandwidth and shared-failure boundary; otherwise you pay for fiber the job cannot use.

Every link bets on the cheapest, lowest-power medium that still meets its loss and error budget. Faster lanes eat copper margin, so your host, module, operating temperature and repair path decide where that bet gives way to optics.

OS2 single-mode is the only fiber with 1.6T Ethernet optics near standardization; on structured trunks that pass each new link’s fiber-count, loss and polarity test, an upgrade is a re-patch, not a re-pull. Moving the laser onto the switch package trades watts for repair time.