AI Infrastructure Capex Passthrough

AI Infrastructure Capital Allocator

The AI buildout is transitioning from pure GPU silicon procurement into physical power, substation transformers, direct-to-chip cooling, and high-density networking. Model institutional capex distribution and evaluate supply chain bottlenecks across five critical sub-industries.

Scenarios:

Capex & Physical Assumptions

Hyperscaler Annual AI Capex $280 Billion

Combined AI infrastructure spend across Microsoft, Google, Meta, AWS & Oracle.

Cluster Rack Power Density 45 kW / rack

Air-cooled limit ~35-40 kW. Above 50 kW mandates 100% direct-to-chip liquid cooling.

Grid Interconnect Lead Time 36 Months

Utility backlog drives pricing power for transformers, gas turbines & SMR nuclear deals.

Custom ASIC vs Merchant GPU Mix 25% ASIC / 75% GPU

In-house silicon (Trainium, TPU, Maia) shifts spend from chip margins to optical networking & packaging.

Optical Fabric Architecture 800G Optics

Optics-to-compute ratio scales with cluster size and distributed parameter count.

Supply Chain Pricing Power Tight Capacity (High Pricing)

Dictates EBITDA margin expansion for bottleneck equipment makers.

Modeled Infrastructure Passthrough

Dynamic Equilibrium v4.2

Physical Infra TAM

$118.2B

Non-GPU power, cooling & shell

Power + Thermal Share

Share of total hyperscaler capex

Network / Optical Spend

$36.4B

Scale-up & scale-out fabrics

Bottleneck Severity

78 / 100

High Pricing Power

Sub-Industry Revenue Capture ($ Billions & % Share)

Total Hyperscaler Capex Deployment $280.0B

Sub-Industry Exposure & Beneficiaries

Key ticker universe & operational moats

Infrastructure Tier	Modeled TAM	Share	Moat / Bottleneck Driver	Sample Public Bellwethers	Cycle Phase

Model calibrated to hyperscaler Q3/Q4 earnings disclosures and grid interconnection queue reports.

The Electrical Substation Moat

High-voltage step-down transformers and backup natural gas microturbines face 3-to-4 year order backlogs. Even if GPU compute is delivered in 12 weeks, gigawatt clusters cannot energize without bespoke utility interconnection gear.

Thermal Physics at 100+ kW

Next-generation clusters (e.g. Blackwell GB200 NVL72) dissipate up to 132 kW per rack—far beyond air convection limits. Coolant Distribution Units (CDUs), direct-to-chip water plates, and manifold quick-disconnects become mandatory capex.

Optical Fabric Multipliers

As clusters scale from 16k to 100k+ chips, the ratio of optical transceivers and digital signal processors (DSPs) to compute chips rises from 2:1 to 5:1. Optical fabric spend outpaces pure compute wafer growth.

Why does the fund manager state the AI infrastructure trade is "still not over"?

Phase 1 of the generative AI boom was pure accelerator hoarding (merchant GPUs). Phase 2 is the physical energization and thermal bottleneck phase. Hyperscalers have committed over $1 trillion in cumulative 5-year capital expenditures, yet less than 25% of required data center power substations, advanced liquid cooling loops, and high-density electrical grid gear have been energized.

How does the Custom ASIC transition affect infrastructure suppliers?

When hyperscalers shift 10% of workloads to internal silicon (e.g., Google TPU v5p, AWS Trainium2, Meta MTIA), pure merchant GPU margins fall, but demand for high-bandwidth memory (HBM3e/HBM4), advanced foundry packaging (CoWoS), high-radix optical switches, and custom power delivery systems actually accelerates.

How are these calculations derived?

This allocator utilizes a structural passthrough matrix calibrated against hyperscaler capital expenditure 10-K filings, utility grid interconnection queue studies, and equipment supplier bill-of-materials (BOM) cost breakdowns. Changing power density, grid lead times, and optics architectures dynamically recomputes the component capture rate across each supply chain tier.