可以。今天这篇不再写成“研究心得”，而是直接按未来可付费产品的形态来做，只是先免费开放。

核心主题就用你最近刚刚真正接通的一条线：

**AI CapEx → Capital Demand → Rates → WACC → ROIC → Asset Pricing**

这已经足够形成一篇有明确研究价值的 Substack 首发型交付。

:::writing{variant="document" title="AI Is Becoming a Macro Variable" id="43157"}
# AI Is Becoming a Macro Variable
## The next phase of the infrastructure cycle will be decided by whether returns can outrun the rising cost of capital

**Pavel | Sovereign**  
*Independent AI Infrastructure Research*

For most of the AI boom, investors have focused on a relatively simple chain:

**AI adoption → compute demand → GPU demand → semiconductor earnings.**

That framework worked.

It is also becoming incomplete.

The AI infrastructure buildout is now large enough that it is beginning to interact with the macro system that finances it.

The next phase of the cycle may therefore look very different from the first.

AI spending can strengthen corporate fundamentals while simultaneously raising capital demand, financing costs, and discount rates.

That creates a paradox:

> **The stronger the AI infrastructure boom becomes, the more it may begin to raise the cost of capital used to finance — and value — the boom itself.**

This changes the central research question.

The question is no longer simply:

**How much AI infrastructure will be built?**

It is increasingly:

# Can incremental AI returns remain above the rising cost of capital?

That spread — **Incremental ROIC versus WACC** — may become one of the most important variables in the next phase of AI infrastructure investing.

---

# 1. AI Is Moving From a Technology Cycle Into a Capital Cycle

AI is often discussed as a software revolution.

Physically, however, it looks increasingly like an industrial buildout.

Every additional layer of AI demand eventually travels downward into the real economy.

AI adoption creates token demand.

Token demand creates compute demand.

Compute demand requires:

- accelerators
- networking
- servers
- cooling
- power systems
- data centers
- electricity
- transmission
- grid capacity

And all of those require capital.

The chain increasingly looks like this:

**AI Adoption**  
↓  
**Token Demand**  
↓  
**Compute Demand**  
↓  
**AI CapEx**  
↓  
**Physical Infrastructure**  
↓  
**Capital Demand**

This matters because capital is not free.

Once the physical requirements of AI become large enough, the industry stops being only a beneficiary of the macro environment.

It begins to participate in creating that environment.

---

# 2. There Are Now Two AI Infrastructure Stacks

The AI infrastructure story is usually described through the physical stack.

## The Physical Stack

**GPU / Accelerators**  
↓  
**Networking**  
↓  
**Cooling & Power Management**  
↓  
**Data Centers**  
↓  
**Electricity**  
↓  
**Grid**

This is where companies such as NVIDIA, Arista Networks, Broadcom, Vertiv, Equinix, Constellation Energy and Vistra enter the picture.

But there is another stack that deserves equal attention.

## The Capital Stack

**Treasury Yields**  
↓  
**Credit Spreads**  
↓  
**Corporate Financing Costs**  
↓  
**WACC**  
↓  
**Funding Capacity**

The physical stack determines whether infrastructure can be built.

The capital stack determines whether it is economically rational to build it.

The two meet at one point:

# Project Economics

A data center can have available land.

It can have GPUs.

It can secure cooling equipment.

It can even find power.

But if the expected return on the project no longer exceeds the cost of financing it, the capital cycle eventually changes.

That is why the next constraint in AI infrastructure may not be purely physical.

It may be financial.

---

# 3. The AI Equity Paradox

The interaction between these two stacks creates an unusual equity setup.

AI spending can improve fundamentals:

**AI Demand ↑**  
↓  
**Revenue ↑**  
↓  
**Margins / EPS / FCF ↑**

But the same investment cycle may also increase demand for capital:

**AI CapEx ↑**  
↓  
**Capital Demand ↑**  
↓  
**Long-Term Rates / WACC ↑**  
↓  
**Valuation Multiples ↓**

The result is a tension between:

# Fundamental Strength

and

# Valuation Pressure

This is why the next phase of AI investing cannot be reduced to:

> “AI demand is strong, therefore AI stocks should rise.”

A more useful framework is:

# Can EPS growth outrun multiple compression?

A company can continue producing excellent operating results while the stock underperforms if the market reduces the multiple it is willing to pay for those earnings.

The reverse can also occur.

This is the point where industry research and asset pricing finally meet.

---

# 4. The Bottleneck Is Migrating

The first phase of the AI cycle was dominated by compute scarcity.

The most obvious constraint was:

**Can companies obtain enough advanced accelerators?**

That constraint has not disappeared.

But as compute supply expands, the bottleneck is gradually moving downstream.

Increasingly important constraints include:

- networking
- rack power density
- liquid cooling
- electrical equipment
- transformers
- interconnection
- generation capacity
- transmission
- financing

The important research question is therefore not:

> Where is AI demand growing?

It is:

# Where is the next binding constraint moving?

Because profit pools often migrate toward scarcity.

When one bottleneck eases, another part of the stack can gain bargaining power, pricing power or capital intensity.

That is why AI infrastructure should be studied as a moving system rather than a static list of “AI beneficiaries.”

---

# 5. NVIDIA: The Question Is Moving Beyond GPU Demand

NVIDIA remains the clearest upstream beneficiary of AI compute demand.

But over time, the most important question may move beyond:

> How many GPUs can hyperscalers buy?

The more important question becomes:

> What economic return are customers generating from the GPUs already deployed?

The full chain is:

**AI CapEx**  
↓  
**GPU Deployment**  
↓  
**Utilization**  
↓  
**AI Revenue**  
↓  
**Customer Cash Flow**  
↓  
**Incremental ROIC**  
↓  
**Next CapEx Budget**

As long as customers can earn attractive returns on additional AI infrastructure, spending can continue.

If incremental returns begin converging toward the cost of capital, spending discipline becomes more likely.

For NVIDIA, the ultimate long-term demand variable is therefore not merely unit demand.

It is the **economics of customer AI investment**.

---

# 6. Vertiv: More Infrastructure Per Unit of Compute

Vertiv represents a different part of the cycle.

Its structural opportunity is not simply that more data centers are being built.

It is that each new generation of AI compute can require more physical infrastructure per unit of deployed capacity.

Higher rack density means:

**More Compute per Rack**  
↓  
**More Heat per Rack**  
↓  
**More Power Density**  
↓  
**Greater Cooling Complexity**  
↓  
**More Infrastructure Content**

This creates a potential second-order beneficiary.

Even if the number of data centers were unchanged, higher compute density could increase the amount of power and thermal infrastructure required inside each facility.

For Vertiv, the highest-value variables are therefore not simply “AI demand” or even headline data-center construction.

They include:

- infrastructure content per MW
- liquid-cooling penetration
- organic growth
- margin expansion
- working-capital conversion
- hyperscaler CapEx intentions

The key question is whether the company can turn the physical complexity of AI into durable FCF and ROIC expansion.

---

# 7. Networking: Scaling Compute Requires Scaling Communication

Larger AI clusters do not only require more accelerators.

They require those accelerators to communicate at enormous speed.

As cluster size increases, the value of networking rises.

The chain is straightforward:

**Larger AI Clusters**  
↓  
**More East-West Traffic**  
↓  
**Higher Networking Intensity**  
↓  
**Demand for High-Speed Ethernet / Custom Silicon**

This is why companies such as Arista Networks and Broadcom sit inside the same infrastructure cycle as NVIDIA despite occupying very different parts of the technology stack.

The key variable is not simply total AI spending.

It is:

> How much networking intensity is required for each additional unit of compute?

That determines how much economic value migrates from compute into connectivity.

---

# 8. Power May Become the Longer-Duration Constraint

Compute equipment can eventually be manufactured at larger scale.

Electricity systems operate on very different timelines.

Generation, transmission, transformers, interconnection and permitting can take years.

That creates a structural asymmetry:

> **Digital demand can grow much faster than physical energy infrastructure.**

This is why power is becoming increasingly central to the AI investment thesis.

The chain is:

**Token Demand**  
↓  
**Compute Demand**  
↓  
**Electricity Demand**  
↓  
**Power Scarcity**  
↓  
**Higher Value of Existing Generation / New Capacity**

For companies such as Constellation Energy and Vistra, AI therefore creates a potentially favorable fundamental channel.

But there is a complication.

Power assets are themselves capital intensive.

Higher long-term rates increase:

- discount rates
- financing costs
- required returns

So the equation becomes:

**AI-driven cash-flow uplift**

versus

**WACC uplift**

Once again, the central variable is not simply demand.

It is the spread between economic return and capital cost.

---

# 9. Equinix Sits Directly at the Intersection

Data-center operators sit almost exactly where the physical stack meets the capital stack.

They benefit from strong infrastructure demand.

But their business models also require large amounts of capital.

That makes a simple “AI demand is strong” framework inadequate.

A more useful question is:

# Incremental Development Yield − Cost of Capital

If the return generated by new capacity remains comfortably above financing costs, growth can continue to create value.

If the spread narrows, strong demand alone may not be enough to generate attractive equity returns.

This distinction will matter more as infrastructure becomes increasingly expensive to finance.

---

# 10. The Underwatched Variable: Incremental AI ROIC

The market spends enormous time tracking:

- GPU shipments
- hyperscaler CapEx
- model launches
- AI adoption
- data-center construction

These are important.

But they all occur before the most important economic test.

After infrastructure is deployed, it must produce something economically useful.

The chain eventually becomes:

**AI CapEx**  
↓  
**Utilization**  
↓  
**AI Monetization**  
↓  
**Operating Cash Flow**  
↓  
**Incremental ROIC**

The strongest version of the AI infrastructure thesis requires more than continuously rising spending.

It requires evidence that:

> **Each additional dollar of AI CapEx produces enough future economic value to justify another dollar of AI CapEx.**

This is where the infrastructure narrative eventually meets accounting reality.

---

# 11. The Next Constraint May Be the Cost of Capital

The AI industry has spent several years discussing physical shortages:

GPU constraint.

HBM constraint.

Networking constraint.

Cooling constraint.

Power constraint.

Grid constraint.

The next important constraint may be:

# Cost of Capital Constraint

This would represent an important transition.

When infrastructure is small relative to the wider economy, financing is mostly an external variable.

When infrastructure spending becomes enormous, the industry begins competing with:

- government borrowing
- housing
- industrial investment
- energy investment
- other corporate borrowers

for the same global capital pool.

In that regime, AI investment does not merely respond to interest rates.

It may contribute to sustaining them.

That leads to a deeper structural possibility:

> **AI may eventually lower the cost of intelligence while raising the near-term price of capital required to build the infrastructure that produces it.**

---

# 12. A Capital Gravity Well

One way to describe the current cycle is as an:

# AI Capital Gravity Well

The AI buildout is increasingly absorbing:

- equity capital
- debt capital
- electrical equipment
- energy
- engineers
- land
- cooling systems
- networking equipment
- construction capacity

The larger this capital gravity well becomes, the more it can alter the price of resources outside the AI ecosystem.

That creates potential crowding effects.

AI-related activity can remain extremely strong even while other areas of the economy experience:

- higher financing costs
- weaker investment
- lower asset valuations

This may help explain why a powerful AI cycle can coexist with a highly uneven broader economy.

---

# 13. What Could Break the Cycle?

The most important bearish signal is not simply:

“AI stocks fall.”

Nor is it:

“AI spending is too large.”

The deeper failure condition is economic.

The framework weakens if:

- compute utilization disappoints
- AI monetization lags infrastructure spending
- token prices fall faster than usage rises
- hyperscaler FCF deteriorates materially
- incremental project returns decline
- financing costs continue increasing
- hyperscalers begin reducing forward CapEx commitments

The critical crossover would be:

# Incremental ROIC ≤ WACC

At that point, the marginal AI infrastructure project begins destroying rather than creating economic value.

Capital discipline would then return rapidly.

---

# 14. What I Am Watching

For the next several quarters, the highest-value signals are not unlimited.

They can be compressed into a relatively small monitoring system.

## Demand

- hyperscaler CapEx guidance
- enterprise AI adoption
- sovereign AI investment
- inference growth

## Physical Constraints

- networking demand
- rack density
- liquid-cooling penetration
- data-center power availability
- grid interconnection

## Economics

- AI revenue monetization
- utilization
- margin progression
- FCF conversion
- incremental ROIC

## Capital

- 10Y and 30Y Treasury yields
- investment-grade credit spreads
- hyperscaler debt issuance
- infrastructure financing costs

## Pricing

- Forward EPS revisions
- valuation multiples
- expectations embedded in current prices

Everything else is secondary unless it changes one of these variables.

---

# 15. The Investment Framework

The AI infrastructure cycle should not be treated as one trade.

Different nodes have different economic exposures.

### First-order CapEx beneficiaries

They get paid when infrastructure is built.

Examples:

**NVDA / VRT / ANET / AVGO**

### Scarce physical asset owners

They may benefit when existing constrained assets become more valuable.

Examples:

**CEG / VST**

### Capital-intensive intermediaries

They benefit from infrastructure demand but remain highly sensitive to financing conditions.

Example:

**EQIX**

The same AI cycle can therefore improve the economics of one node while compressing returns somewhere else.

That is why the central research task is not to find “AI stocks.”

It is to understand:

# Where the constraint is moving, where the profit pool follows, and whether the market has already priced it.

---

# Conclusion

The first phase of the AI infrastructure cycle was relatively easy to describe.

There was not enough compute.

Demand exploded.

The companies supplying compute captured extraordinary economic value.

The next phase will be more difficult.

The physical infrastructure stack is expanding.

The capital stack is becoming more important.

Power and grid constraints are rising.

The cost of capital is no longer an external detail.

And the ultimate economic question is moving downstream from spending to returns.

The research framework therefore changes from:

**AI Demand → More CapEx → Higher Stocks**

to:

# Demand → Constraint → Capital → Cash Flow → Return → Price

The most important question in the next phase is no longer:

> How much will companies spend on AI?

It is:

> **Can the return on the next dollar of AI infrastructure remain above the cost of financing that dollar?**

As long as the answer remains yes, the infrastructure cycle can continue to compound.

When that answer changes, the cycle changes with it.

---

## SOVEREIGN SIGNAL

**Structural regime:** AI is evolving from a technology theme into an industrial capital cycle.

**Constraint migration:** Compute → Power / Grid / Capital.

**Core variable:** Incremental AI ROIC.

**Macro risk:** Higher-for-longer cost of capital.

**Equity question:** Can EPS growth outrun multiple compression?

**Failure condition:** Incremental ROIC falls toward or below WACC.

---

**Pavel | Sovereign**  
*Independent AI Infrastructure Research*

*Researching the intersection of AI infrastructure, energy, capital and asset pricing.*
:::

这篇今天就可以作为你的 **Free Substack Issue 001**。

我建议免费阶段先不要急着“藏结论”。真正应该建立的是一种稳定预期：读者每次打开你的文章，都知道能得到 **变量压缩 + 公司传导 + 预期差 + 失效条件**，而不是另一篇 AI 评论。

以后每篇都可以保持这个产品骨架，只换研究问题。等开始付费后，免费层保留 **Structural Thesis + 核心变量**；付费层再放 **Company Read-Through、Expectation Gap、Bull/Base/Bear、具体监控指标和定价判断**。这样从免费到付费，不需要重新创造一种产品。