Scaling AI Factories with Co-Packaged Optics for Better Power Efficiency
Vendor-stated absolute figures the GTC release lacked: pluggables 'often 30W per interface' vs CPO 'as low as 9W' (about 3.3x, not the 3.5x claimed); electrical loss up to 22 dB (200 Gb/s channels) vs ~4 dB, an 18 dB gap = ~63x, which is where the 63x/64x signal-integrity figure comes from. Availability restated: Quantum-X early 2026, Spectrum-X 2H 2026.
Read depth: full text of the post (about 1,100 words). The figures themselves (Figures 1–5) are images; only their captions were readable. Every number below is vendor-stated, as of 2025-08-18, about products that were not yet shipping.
Why this source was added
The NVIDIA press release already on this rung says 3.5x better power efficiency for co-packaged optics, and its close read notes "No absolute power figure (watts per port, pJ/bit) anywhere in the release". This blog, by the same company, gives the watts.
The numbers, as published
| Item | Pluggable optics | Co-packaged optics (CPO) |
|---|---|---|
| Power | "often 30W per interface" | "as low as 9W" |
| Electrical signal loss, switch chip to optics | "up to 22 dB for 200 gigabit-per-second channels" | "only ~4 dB" |
| Stated multipliers | — | 3.5x power efficiency, 10x resiliency, 1.3x faster time-to-operation |
Figure captions read: "Spectrum-X Photonics delivers 64x better signal integrity" and "3.5x power-saving with Spectrum-X Photonics".
Products and stated availability (as of 2025-08-18):
| Switch | Capacity | Ports | Availability stated |
|---|---|---|---|
| Quantum-X Photonics (InfiniBand, Q3450) | 115 Tb/s | 144 × 800 Gb/s | early 2026 |
| Spectrum-X Photonics SN6810 (Ethernet, one Spectrum-6 chip) | 102.4 Tb/s | 128 × 800 Gb/s | second half of 2026 |
| Spectrum-X Photonics SN6800 (Ethernet, four chips) | 409.6 Tb/s | 512 × 800 Gb/s | second half of 2026 |
All are liquid-cooled. Quantum-X also carries 14.4 teraflops of in-network computing (SHARP, fourth generation).
What the watts do and do not settle
- The 30 W and 9 W do not quite give 3.5x. 30 ÷ 9 is about 3.3x (our arithmetic). The blog never shows how its 3.5x is derived, and "often" and "as low as" make the two figures a typical case against a best case. So the 3.5x is not a ratio of the two numbers it prints. Broadcom's "70% reduction" on the rung is likewise 3.3x, not the "more than 3.5x" it also claims. Both vendors round up.
- "Per interface" is not defined. The post does not say whether an interface is an 800 Gb/s port or something else. If it is an 800 Gb/s port, the figures would work out to about 37 pJ/bit for pluggables and about 11 pJ/bit for CPO [unverified: depends on that assumption, which the source does not confirm].
- The signal-integrity multiplier is a unit conversion, not a marketing composite. A loss of 22 dB against 4 dB is a gap of 18 dB, and 18 dB is a power ratio of 10^1.8, about 63x. That matches the "63x" in the press release and the "64x" here (our arithmetic). The existing press release close read called the multiplier a composite "with no defined unit"; it has one. It is still a best case, and the 22 dB is quoted as an upper bound ("up to").
- The dates slipped. The March 2025 release said Quantum-X InfiniBand "later in 2025" and Spectrum-X Ethernet "in 2026". Five months later this post says early 2026 and second half of 2026. Nothing on the rung says whether either has shipped since.
Through-line: token price → task price
A site has a fixed power feed. At the figures stated, one 512-port SN6800 would draw about 10.8 kW less for its optics than the same ports on pluggables (512 × (30 − 9) W, our arithmetic, and only if every port ran at the typical and best-case figures respectively). That is power left over for GPUs, which is the only way this saving reaches a token price. How much it matters depends on how many switch ports a site runs per GPU, which no source here gives.
Limitations
- Vendor source, vendor product. NVIDIA is describing its own switches. No independent test.
- Unshipped at the time. Every figure describes products not yet available on the post's date.
- Typical against best case. "Often 30W" against "as low as 9W".
- No definition of "interface", so no confirmed per-bit figure.
- Same company as an existing source. This adds numbers, not a second independent voice. The rung's independence still rests on Broadcom being a different vendor.
- The post promises a second part on the photonics engines; that part was not read.