# AI Factory Architecture

One-line thesis

AI is moving from a chip-led buildout into full-stack AI factories, shifting value capture toward system bottlenecks such as networking, interconnect, optics, memory, packaging, power, cooling, rack deployment, and enterprise data infrastructure.

Summary

GTC 2026 reinforced that the next phase of AI infrastructure is about making clusters scale economically, not just supplying more accelerators. The stack is broadening toward system-level architecture, where bandwidth, memory movement, optical interconnect, thermal efficiency, power density, and install speed all become investable chokepoints. This widens the opportunity set beyond NVDA alone while still strengthening NVDA’s role as the lead stack owner.

Core thesis points

• Networking remains a chokepoint, especially as the battle shifts toward scale-up and scale-out architecture.
• Cluster economics are the next bottleneck: power, heat, bandwidth, memory movement, and deployment efficiency now matter as much as raw chip availability.
• Co-packaged optics is a real clue that electrical interconnect limits are becoming binding.
• AI infrastructure value capture is broadening beyond compute into memory, packaging, networking, optics, cooling, and power infrastructure.
• Enterprise AI remains dependent on data accessibility and structured data infrastructure.
• Physical AI is becoming more credible as a future demand layer, but it is not yet core to the thesis.

Supporting evidence from GTC 2026

• NVLink versus Ethernet/InfiniBand points to interconnect architecture as a strategic control point.
• Spectrum X supports the view that network/fabric ownership is becoming more valuable.
• Co-packaged optics / photonics is the strongest clue that optical bandwidth and system efficiency may be the next scarcity layer.
• Rubin system design reinforces rack-level integration as part of the moat.
• Liquid cooling / hot-water cooling / cable-free racks suggest deployment speed and thermal efficiency are becoming real differentiators.
• LPDDR5 CPU and system-level efficiency messaging support the view that memory architecture and perf-per-watt are increasingly central.

Beneficiary layers

Near-term winners

• Compute: NVDA, AMD, AVGO
• HBM / memory: MU, SK Hynix, Samsung
• Advanced packaging / foundry: TSM, AMKR
• Networking / interconnect: AVGO, MRVL, ANET
• Optics / photonics: COHR, LITE, AAOI, CIEN
• Power / thermal / cooling: VRT, ETN, TT, JCI
• See also linked child theme: theme-photonics-interconnect

Second-order winners

• Data infrastructure / agentic layer: SNOW, MDB, AMZN, MSFT, GOOGL, Databricks (private)

Optionality bucket

• Physical AI / robotics / edge: NVDA, SYM, ROK, ABB, TER
• See also: theme-physical-ai-supply-chain for the downstream material/component stack

Next bottlenecks to watch

• power density
• heat / cooling efficiency
• bandwidth / interconnect
• optics / photonics
• memory movement / packaging
• rack deployment speed
• structured enterprise data access

Hype / discard

High-signal / investable

• NVLink vs Ethernet/InfiniBand
• Spectrum X
• co-packaged optics
• Rubin system design
• liquid cooling
• photonics

Medium-signal

• Grok volume production

Low-signal / likely exclude

• OpenClaw section
• graphics / neuro rendering as core investment thesis
• robotics as immediate earnings driver

Route

• Primary destination: theme
• Secondary destination: worldview
• Tertiary destination: NVDA stock note
• Supporting destination: beneficiary board

Status

Done as first persisted seed object from the GTC 2026 brain-dump flow.