The Power–Compute–Connectivity Loop: A Framework for Investing Across AI Infrastructure

Sample post. This is placeholder content created to preview the SectorVantage design. It is written as a generic, illustrative framework. It contains no real data, statistics, quotes, or sources, and nothing here is investment advice.

The thesis in brief

AI is often discussed as a software story. In practice, every AI workload runs on physical infrastructure: chips that have to be designed and manufactured, electricity that has to be generated and delivered, buildings that have to be sited and cooled, and networks that have to carry data in and results out.

Those requirements link four sectors that are usually analyzed separately: AI, semiconductors, power, and telecom. We think of them as a loop. Demand in one link creates requirements in the next, and the link with the least slack sets the pace for the whole system. That link is the binding constraint, and it is where pricing power, project delays, and new business opportunities tend to concentrate.

The aim of this framework is not to predict which link is constrained at any moment. It is to give investors and builders a repeatable way to ask the question, and to notice when the answer changes.

How the loop works

1. AI → Semiconductors. More model training and more inference require more accelerators, memory, and networking silicon. Chip supply, and the cost of each unit of compute, in turn shape how fast AI can be deployed.
2. Semiconductors → Power. Dense compute concentrates electrical load in a small footprint. That load needs generation, transmission, interconnection to the grid, and cooling.
3. Power → Telecom. Large facilities tend to be sited where power is available and high-capacity fiber can reach. Increasingly, power and connectivity are evaluated together during site selection.
4. Telecom → AI. Training clusters need high-capacity links within and between facilities, and inference has to reach users through metro, long-haul, last-mile, and wireless networks. Networks are AI's distribution layer.

The data center sits in the middle of all four, which is why it appears in almost every SectorVantage analysis.

A framework for locating the bottleneck

For each link, we ask three questions: What limits supply? How long does it take to add capacity? What signals would show the constraint tightening or easing?

Link	What typically limits it	Time to add capacity	Signals to watch
Compute (semiconductors)	Manufacturing and packaging capacity, design cycles	Long: new capacity is planned years ahead	Capacity expansion plans, lead times, pricing commentary
Power	Generation, transmission, grid interconnection, permitting	Often the longest: physical and regulatory steps	Interconnection processes, utility planning, new generation contracts
Connectivity (telecom)	Fiber routes, rights-of-way, construction crews	Medium: construction-bound	Network build announcements, interconnect demand, construction capacity
Demand (AI)	Adoption, business models, cost of compute	Fast to change, in either direction	Usage and pricing trends, enterprise deployment, capex guidance

The key insight is the mismatch in timelines. Demand can change quickly; physical capacity cannot. When demand outruns the slowest link, that link becomes the constraint, and the slowest links tend to be the ones involving permits, construction, and the grid.

Three illustrative scenarios

These are not forecasts. They are a way to think through how value could shift around the loop.

• Scenario A: Demand keeps outrunning supply. Constraints show up in the slowest-to-build links. Owners of ready capacity (powered land, interconnection rights, available fiber) may gain leverage, and projects that can bring their own power or connectivity may move faster.
• Scenario B: Supply catches up. As capacity comes online across the loop, scarcity premiums compress. Efficiency, utilization, and cost discipline matter more than access, and businesses built on a temporary shortage come under pressure.
• Scenario C: Demand pauses. If adoption or monetization slows, the longest-lived assets carry the most risk of under-utilization, while flexible, contracted, or multi-use assets hold up better.

What would change our mind

A framework is only useful if it can be wrong. Signals that would weaken the "single loop" view include:

• AI workloads becoming dramatically more efficient, reducing the link between AI demand and power or chip demand.
• Compute moving substantially to locations or devices that do not depend on large centralized facilities.
• Power and connectivity becoming abundant enough that they stop influencing where compute is built.

We will revisit this framework as conditions change and note when we think the binding constraint has moved.

How to use this framework

Use the table above as a checklist when you read news in any of the four sectors. For each development, ask which link it affects, whether it tightens or loosens that link, and who benefits from the change. The read-throughs below translate the framework into starting points for three types of readers.