AI systems · 26 August 2026.

AI systems

House expert call · Wednesday, August 26, 2026

Optical intensity still steps up. Mix, not just attach rate, is what books the revenue.

The call’s central claim is architectural, not cyclical: hyperscalers are abandoning NVIDIA’s canonical rail-optimized SuperPOD (optics all the way to the GPU) in favor of adjacent-rack, copper-heavy topologies. That design choice is the largest near-term driver of active copper (AEC/ACC) and a reason CPO on Spectrum-X ramps slower than NVIDIA’s supply-chain prep implies.

Working model

• GB200: 400G, ~1:1. GB300: 800G, 1:1 — the step the street already owns.
• VR200: 2×800G per GPU via ConnectX-9 (2:1 intensity). Switch side one 1.6T port; CPO optional. Front-panel OSFP remains the customer default.
• Rubin Ultra likely stays 2-die, not the 4-die plan of record. 2027 is when optical scale-up (CPO/NPO) hits production.

Positioning

• AEC is an adjacent-rack custom-cluster cycle (Meta Mini-Pack already visible), not SuperPOD optics-to-the-GPU. Mix walks off DAC as rates rise.
• Google TPU v9 ICI moves to 300G/lane PAM4 — keeps PAM4 and a 2.4T-class optic alive. Marvell near-term merchant DSP plus custom ICI.
• UALink is losing the room. Ethernet scale-up (TH6) is what OpenAI and AMD are actually building. Amazon is the residual.
• CRDO constructive on AEC. AVGO TH6 2027 better than 2026; CPO/DSP design-win rich, shipment poor. Pluggable optics cycle intact — more 800G than CPO.

Do not model VR200 optical intensity as 1:1. A CPO miss is not an optical-cycle miss. Not a recommendation.

House synthesis of a dated expert call. Not a recommendation to buy or sell any security.

House AI systems deep dive, August 26, 2026. House notes.