Networking & Interconnect — Research Frontier

What's genuinely new and where the rung is heading. First compiled 2026-09-24 from four sources, two of them in scope; extended the same day with NVIDIA's CPO blog, which supplies the watts.

The binding constraint, named

Power per bit moved, inside a fixed site feed. The Broadcom close read names it directly: bandwidth figures answer nothing on their own, and the power cost of moving that bandwidth is the constraint that binds in a data centre with a fixed feed, which is what ties this rung to energy and datacenters (Broadcom).

The number this rung owns

About 3.5x lower optical link power for co-packaged optics than for pluggable modules, stated by both Broadcom (2025-10-08, sampling) and NVIDIA (2025-03-18, roadmap) (Broadcom; NVIDIA). Both are vendor-stated and neither names its baseline. The watts now exist, from NVIDIA alone: "often 30W per interface" for pluggables against "as low as 9W" co-packaged, which is about 3.3x, not 3.5x (NVIDIA CPO blog). Broadcom's 70% is also 3.3x. So the number the rung can defend is about 3.3x, vendor-stated, typical against best case; the 3.5x headline is the vendors rounding up.

Active Frontiers

1. Co-packaged optics at the switch

Status: Steady Key sources: Broadcom, NVIDIA, NVIDIA CPO blog

Both large switch vendors have moved optics into the switch package and claim the same order of power saving. NVIDIA has now put watts on it (30 W against 9 W), and its "63x/64x signal integrity" turns out to be the 22 dB to 4 dB loss gap expressed as a ratio. NVIDIA's own stated dates slipped between March and August 2025. The open step is from sampling and roadmap to measured deployment.

Open problems:

• Watts per port or picojoules per bit from anyone other than the vendor, with "interface" defined.
• Whether NVIDIA's switches shipped on the restated dates (early 2026 InfiniBand, second half of 2026 Ethernet).
• Whether the saving is large enough to matter at the level of the whole system, since it is scoped to the optical links only.

2. Replacement or expansion of optical content

Status: Early stage Key sources: Broadcom, NVIDIA

Broadcom's part is evidence that co-packaged optics displaces pluggable modules at the switch face. The thesis on this rung needs to know whether total optical content per system rises or falls. See theses.

Open problems:

• Total optical content per rack before and after, on any shipping system.

3. Optics inside the rack

Status: Early stage Key sources: NVIDIA Vera Rubin

The silicon photonics page argues optical links are what make rack-scale bandwidth physically possible. It rests on one vendor blog.

Open problems:

• Cost premium over copper, and where it crosses over.
• Yield of co-packaged photonics at volume.

What this rung does not yet hold

• Anything on fabric economics. Scale-up versus scale-out, NVLink versus InfiniBand versus Ethernet, bandwidth per dollar, topology and links between sites: no sources.
• Any independent measurement. Every in-scope figure is vendor-stated, including the new watts.
• In-scope evidence under the thesis. Its original sources are about photonic computing, which is out of scope.

Standing

Read it. Nothing here has been measured in-house. The thesis has a live position behind it; that is exposure, not standing.