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 — accelerators talking 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.

The switch ASIC, NIC and DPU set the ceiling on every fabric above them, so their SerDes generation, buffer architecture and offload engines must close with 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.

Oversubscription and topology turn a GPU count into a bill of materials, a blocking factor, and an MFU ceiling; the 2026 error is setting one ratio for the whole fabric, or taking it from the workload's label instead of its measured locality, exposure and failure headroom.

A non-blocking Clos only delivers its bandwidth if congestion control and load balancing both work; lose either and an untuned fabric leaves a large fraction of its rated throughput on the floor, while in-network reduction accelerates the collectives that support it rather than gating the fabric.

An AI cluster needs an out-of-band network to reach a wedged node when the data plane is dead, and a timing plane — PTP where the event budget requires its hardware path — to place cross-device logs on one timeline within a measured error bound through the declared outage.

Scale-across is an optional portfolio-connectivity tier; decide whether sites need DCI and, only for eligible work, whether a job may cross the latency, bandwidth, algorithm, placement, and correlated-failure boundary.

Every link bets on the cheapest, lowest-power medium that closes the required channel and error budget; faster lanes consume copper margin, so the selected host, module, temperature and repair path decide where that bet gives way to optics.

OS2 may preserve useful attenuation and dispersion headroom, but every future application still needs a declared and certified fiber-count, connector/polarity, loss/reflectance, reach, module and host migration map; the other live fork is whether the laser stays on the faceplate or moves onto the switch package.