:::writing{variant="document" title="Where Did the $700 Billion Go? | How AI Capex Flows Into Chips, Power, Cooling, and Land"}
# Where Did the $700 Billion Go?

## How AI Capex Flows From Big Tech Into Chips, Power, Cooling, and Land

**SOVEREIGN PRICING WORLD 001**

The easiest mistake to make about artificial intelligence is to imagine that it lives in the cloud.

Models have no visible weight. Tokens appear instantly. Inference feels like software. Agents look like abstractions.

From the user’s perspective, AI seems almost immaterial.

But follow the money far enough and the picture reverses.

Every token ultimately requires a physical machine. Every machine requires networking. Every rack requires power and cooling. Every data hall requires transformers, switchgear, fiber, substations, permits, and land. Every megawatt has to come from somewhere.

AI may appear on the screen as software.

Economically, it is becoming one of the largest physical capital formation projects of the modern era.

That is the starting point of this series.

Not:

**Will AI change the world?**

That question is already too broad.

A better question is:

> **Who is financing the world AI requires, where does the money go, and who ultimately captures the cash flow?**

In 2026, Amazon, Microsoft, Alphabet, and Meta alone are collectively operating around the scale of **$700 billion-plus in annual capital expenditure plans**, depending on guidance and accounting definitions.

Not all of that spending is purely AI.

But AI infrastructure is increasingly the marginal force driving the expansion.

The number itself is extraordinary.

Yet “$700 billion of Capex” tells us almost nothing.

The real work begins when we ask:

**What does that money become?**

---

## I. Capex Is Not Just Spending

Capex means **capital expenditure**.

When a company pays employee salaries, electricity bills, or marketing expenses, those costs largely pass through the income statement as operating expenses.

But when Microsoft builds a data center or buys GPU servers, the economic structure is different.

Cash leaves the company today.

In exchange, the company receives an asset expected to produce economic value over multiple years.

The flow is roughly:

**Cash → Capex → PP&E → Depreciation → Future Productive Capacity**

PP&E means property, plant, and equipment.

This distinction matters because AI infrastructure is fundamentally an attempt to purchase future production capacity.

A hyperscaler is not simply “spending money on AI.”

It is converting liquid capital into machines capable of manufacturing future compute.

And this is where the first major shift becomes visible.

For much of the internet era, software was celebrated because it was asset-light.

Build the software once.

Distribute it globally.

Marginal cost approaches zero.

AI is partially reversing that model.

The intelligence may be digital.

The production system underneath it is increasingly industrial.

---

# II. The First Destination: Compute

The most obvious destination of AI capital is the semiconductor stack.

But “GPU spending” is too crude a description.

An AI compute system contains far more than a GPU.

It requires accelerators, CPUs, HBM memory, NICs, switches, optics, storage, servers, racks, and software orchestration.

Only then does a cluster begin to exist.

This is why Nvidia increasingly talks about the **AI Factory** rather than simply selling chips.

The economic object is no longer the isolated processor.

The economic object is an entire production system for intelligence.

That distinction matters for investors.

A company can have enormous AI exposure without necessarily capturing enormous value.

The question is not merely:

**How much compute will the world buy?**

The harder question is:

> **Where inside the compute stack does scarcity remain durable enough to support pricing power and high returns on capital?**

Nvidia currently occupies one of the strongest positions because advanced AI accelerators remain technologically difficult to reproduce at equivalent performance and ecosystem depth.

But pricing is always dynamic.

Custom ASICs are advancing.

Google has TPUs.

Amazon has Trainium.

Meta is building its own silicon.

Technological progress reduces cost per unit of compute.

So even at the strongest point in the chain, the question eventually becomes:

**How long does scarcity last?**

That is the difference between identifying a good industry and pricing an asset.

---

# III. Networking: A Million GPUs Are Not Automatically One Computer

The next layer is less visible.

Networking.

A single GPU can compute.

A million GPUs must communicate.

The larger the cluster becomes, the more difficult the problem shifts from raw computation toward data movement.

Latency matters.

Bandwidth matters.

Optical connectivity matters.

Memory architecture matters.

Switching matters.

The system must increasingly behave like a single giant computer.

This transition changes the economic structure from:

**Chip Scale**

to

**Rack Scale**

to

**Cluster Scale**

to

**Data Center Scale.**

As compute scales, networking becomes a tax on intelligence.

The more machines calculate together, the more data must move between them.

That brings companies such as Broadcom and the broader high-speed networking ecosystem into the same capital flow.

The insight is simple:

> **AI does not merely increase demand for computation. It increases demand for coordination between computation.**

And coordination requires infrastructure.

---

# IV. Cooling: All Computation Eventually Becomes Heat

Then AI reaches physics.

Electricity enters a server.

Computation occurs.

Most of that energy eventually leaves the system as heat.

So as rack density rises, another variable becomes increasingly important:

**How do you remove the heat?**

Traditional data centers were built around relatively moderate rack densities and could rely heavily on air cooling.

AI systems push far beyond those historical assumptions.

The result is a growing thermal infrastructure stack:

Chillers.

Coolant Distribution Units.

Heat exchangers.

Pumps.

Direct-to-chip liquid cooling.

Rear-door systems.

Fluid management.

This is why a company such as Vertiv becomes so important to the AI infrastructure map.

Vertiv does not build foundation models.

It does not manufacture GPUs.

Yet as compute density rises, its products become increasingly difficult to exclude from the physical production chain.

This creates a direct capital transmission mechanism:

**Hyperscaler Capex**

→ higher rack density

→ higher thermal density

→ greater liquid cooling requirements

→ orders

→ revenue

→ margins

→ free cash flow.

The money leaves Microsoft or Amazon.

It enters the order books of physical infrastructure suppliers.

This is what it means to follow capital rather than narrative.

---

# V. Power Equipment: AI Meets the Industrial World

Eventually, AI runs into something far older than machine learning:

the electrical grid.

A server cannot simply plug directly into a power plant.

Electricity must travel through multiple layers:

**Generation**

→ transmission

→ substation

→ transformer

→ medium-voltage switchgear

→ UPS

→ power distribution

→ rack

→ GPU.

This produces one of the strangest economic facts of the AI era:

> **The most advanced artificial intelligence in the world may have to wait for a transformer.**

This is why companies such as Eaton have become structurally relevant to AI infrastructure.

The technology may look less glamorous than an accelerator chip.

But economic value is not determined by glamour.

It is determined by bottlenecks.

Imagine a data center project with billions of dollars already committed.

The buildings exist.

The GPUs have been ordered.

The customers are waiting.

But one critical electrical component has a long lead time.

Suddenly, the economic value of that component is no longer proportional to its apparent technological sophistication.

Its value derives from the amount of capital it prevents from becoming productive.

This is **bottleneck pricing**.

And bottleneck pricing is one of the most important concepts in infrastructure investing.

---

# VI. The Unit of Analysis Changes From GPUs to Gigawatts

At first, the AI boom was discussed in chips.

Then racks.

Then megawatts.

Now increasingly:

**gigawatts.**

This change in units is not cosmetic.

It means AI has crossed from the technology sector into the energy system.

Once a campus requires hundreds of megawatts or multiple gigawatts, investors can no longer understand the industry by reading semiconductor earnings calls alone.

They must understand:

natural gas,

gas turbines,

nuclear power,

transmission,

substations,

interconnection queues,

power purchase agreements,

capacity markets,

grid reliability.

AI has become an electricity question.

And electricity is a physical system with long lead times, regulatory constraints, geography, and enormous capital requirements.

Software can scale globally in seconds.

Power plants cannot.

That mismatch may become one of the defining economic tensions of the AI era.

---

# VII. Why GE Vernova Is Suddenly Part of the AI Story

A company like GE Vernova once looked far removed from artificial intelligence.

It manufactures gas turbines, grid equipment, and electrification systems.

But the moment AI factories require reliable 24/7 power at industrial scale, the connection becomes obvious.

The causal chain looks like this:

A model gains adoption.

Cloud providers require more capacity.

They order more accelerators.

Accelerators require larger facilities.

Facilities require more electricity.

Local grids lack available capacity.

New generation and grid infrastructure become necessary.

Eventually:

**AI demand appears in GE Vernova’s backlog.**

This is the crucial lesson.

An industry is not a list of companies.

An industry is a **causal transmission system**.

If you cannot explain how demand at one node becomes revenue at another node, you do not yet understand the chain.

---

# VIII. Land: The Scarce Asset Is Not Land, but Powered Land

Eventually the capital reaches the ground.

Literally.

Data centers need land.

But raw land is not the scarce resource.

The scarce resource is land with usable access to:

large quantities of power,

grid interconnection,

fiber,

permitting,

water or viable cooling alternatives,

and a credible path to construction.

This creates a more useful concept:

**Powered Land.**

A thousand acres without available electricity may have little strategic value for AI.

A smaller parcel with secured power, fiber, permits, and substation infrastructure may become extraordinarily valuable.

This changes the economics of real estate.

The land is no longer merely land.

It is an option on future compute capacity.

That is why the AI infrastructure boom increasingly involves infrastructure funds, real estate capital, private credit, project finance, and structured leases.

Big Tech balance sheets alone may not be sufficient — or economically optimal — to finance every future data center.

So the financing structure is beginning to evolve.

The old model:

**Big Tech → Capex → Data Center**

is becoming:

**Big Tech + Infrastructure Funds + Private Credit + Leases + Project Finance + Customer Commitments**

→ AI infrastructure.

AI is beginning to financialize.

---

# IX. Oracle Shows What Happens When Demand Outruns the Balance Sheet

Oracle offers a useful case study.

A company can have extraordinary future contracted demand and still experience severe pressure on present free cash flow.

Why?

Because infrastructure must be built **before** the revenue is fully recognized.

This creates a timing mismatch.

The customer says:

“I want enormous compute capacity.”

The provider must first finance:

land,

construction,

servers,

networking,

power,

cooling.

Only after the infrastructure becomes operational can much of the revenue be recognized over time.

This means AI infrastructure behaves increasingly like older capital-intensive systems:

railways,

telecommunications networks,

power plants,

pipelines.

Large upfront investment.

Long-duration asset lives.

Utilization risk.

Financing risk.

And eventually, hopefully:

recurring cash flows.

That word matters:

**hopefully.**

Because contracts are not the same as economic returns.

---

# X. Why EPS Is No Longer Enough

For years, investors in mega-cap technology companies could focus heavily on revenue growth, operating margins, and earnings per share.

AI requires a deeper financial lens.

The key relationship becomes:

**Operating Cash Flow − Capital Expenditure ≈ Free Cash Flow**

A company may report spectacular revenue growth.

It may simultaneously consume enormous amounts of capital.

This does not mean the investment is bad.

It means the investor now has another question to answer.

Not:

**Is revenue growing?**

But:

> **What return will today’s incremental dollar of capital generate tomorrow?**

That moves us into the language of **ROIC — Return on Invested Capital.**

Suppose a company invests $100 billion.

If that capital eventually generates $30 billion of sustainable incremental economic profit, the investment may be exceptional.

If it generates only enough cash to maintain competitive parity, the same $100 billion can have very different economic meaning.

Therefore:

> **High Capex is not inherently bullish.**

The key variable is **incremental ROIC**.

---

# XI. The Ultimate Test: ROIC vs. WACC

Every capital cycle eventually confronts the same equation.

Does the return on invested capital exceed the cost of capital?

If:

**ROIC > WACC**

capital creates value.

If:

**ROIC < WACC**

capital destroys value.

This is where the entire AI infrastructure cycle will ultimately be judged.

The $700 billion is not the answer.

It is the wager.

The wager is that these GPUs, servers, power plants, cooling systems, transformers, networks, and data centers will create enough economic utility to justify the capital consumed.

The physical infrastructure is arriving first.

The monetization must follow.

AI therefore contains a giant temporal bet:

> **We are building tomorrow’s productive capacity before tomorrow’s revenue fully exists.**

Capital markets are deciding whether that bet will pay.

---

# XII. This Is Where Expectations Enter

Now we move from industry analysis into pricing.

Suppose Vertiv grows revenue 30%.

Is that bullish?

We still do not know.

If the market expected 15%, 30% may be extraordinary.

If the stock price already assumes five years of 35% growth, 30% may disappoint.

The same applies to Nvidia.

Eaton.

GE Vernova.

Oracle.

Any company.

The market does not trade reality in isolation.

It trades:

> **Reality relative to what was already expected.**

That is why strong earnings can produce falling share prices.

And weak numbers can produce rallies.

The asset was never priced against zero expectations.

Every price contains a story about the future.

The investor’s job is to discover what story is already embedded inside the price.

---

# XIII. $700 Billion Will Not Be Shared Equally

One of the most common mistakes in thematic investing is:

“This industry is enormous, therefore every company in the industry will win.”

No.

Capital leaks.

Competition destroys margins.

Some companies require enormous reinvestment just to maintain position.

Some suppliers experience revenue growth without meaningful free cash flow.

Some sectors become commoditized quickly.

Others preserve extraordinary pricing power.

So the correct question is not:

**Who has AI exposure?**

The correct question is:

> **Who captures value?**

Exposure means AI spending passes near you.

Value capture means:

**When AI spends another $100, how many dollars ultimately remain as your sustainable free cash flow?**

That is a much harder question.

It is also a far more useful one.

---

# XIV. Three Bottlenecks Worth Watching

At this stage of the cycle, three structural categories appear especially important.

The first is **technological scarcity**.

Advanced accelerators.

HBM.

High-speed networking.

These markets can support extraordinary economics when technological barriers remain high and supply expansion remains difficult.

The central risk is:

**How quickly does competition catch up?**

The second is **physical infrastructure scarcity**.

Liquid cooling.

Transformers.

Switchgear.

UPS systems.

Gas turbines.

These technologies may be more mature, but manufacturing capacity, certification, supply chains, and lead times can still create significant bottlenecks.

The central risk becomes:

**When does supply finally catch demand?**

The third is **scarce asset control**.

Powered land.

Grid connections.

Permitted sites.

Energized megawatts.

Here, the moat may not be technology at all.

It may be geography.

Time.

Permitting.

Transmission access.

The central question becomes:

> **Who controls the megawatts that can actually turn on?**

This may become one of the most important questions of the next phase of AI infrastructure investing.

---

# XV. Every Capital Boom Contains Its Own Bust Mechanism

A real industry does not guarantee a good investment.

Railways changed civilization.

Railway investors still lost fortunes.

Telecommunications changed the world.

Telecom bubbles still existed.

The internet was real.

The dot-com bubble was also real.

AI can transform global productivity and still produce periods of massive capital misallocation.

Several variables could eventually break the current cycle:

1. AI monetization grows more slowly than infrastructure spending.
2. GPU utilization falls.
3. Token prices decline faster than demand expands.
4. New architectures shorten the economic life of existing hardware.
5. Too much manufacturing capacity destroys current bottleneck pricing.
6. Power constraints prevent compute assets from becoming productive.
7. Higher interest rates or wider credit spreads make long-duration infrastructure economics unattractive.

This is why serious infrastructure investing requires permanent verification.

Not faith.

---

# XVI. The Dashboard I Want to Build

From here, this series will increasingly track a small group of variables.

At the hyperscaler level:

Capex.

Operating Cash Flow.

Free Cash Flow.

Cloud growth.

AI monetization.

At the compute level:

GPU revenue.

HBM supply.

Custom ASIC penetration.

Networking growth.

At the physical infrastructure level:

Vertiv orders and margins.

Eaton backlog.

Transformer lead times.

Switchgear capacity.

Liquid cooling penetration.

At the power level:

Gas turbine backlog.

Grid equipment orders.

Interconnection queues.

Natural gas demand.

Nuclear contracting.

At the asset level:

Powered land.

Megawatts under development.

Energized gigawatts.

Construction cost per MW.

Financing cost.

But all of these variables ultimately collapse into one question:

> **Is Capex turning into economic return?**

---

# XVII. Look Again at the $700 Billion

At the beginning, $700 billion looked like a headline.

Now it should look different.

It is a river.

**Advertising / Cloud / Software Cash Flow**

↓

**Hyperscaler Balance Sheets**

↓

**Capex**

↓

GPU / CPU / HBM

↓

Networking / Storage

↓

Servers / Racks

↓

Power Distribution

↓

Liquid Cooling

↓

Data Centers

↓

Transformers / Switchgear

↓

Grid

↓

Gas Turbines / Generation

↓

Powered Land

↓

**Energized MW**

↓

AI Compute

↓

Tokens

↓

AI Products

↓

Revenue

↓

Cash Flow

↓

**ROIC**

And then capital markets discount those future cash flows back into the present:

↓

**Valuation**

↓

**Stock Price**

That is the full pricing loop.

---

# XVIII. An AI Factory Is a Machine That Converts Dollars Into Tokens

There is a more useful way to define an AI factory.

An AI factory is not simply a data center.

It is an industrial system that converts:

**capital + electricity + physical equipment**

into:

**compute**

and then into:

**tokens.**

But tokens themselves are not the final economic output.

Tokens must eventually create:

advertising revenue,

cloud revenue,

subscriptions,

enterprise productivity,

automation,

software value,

or cost savings.

Only then do tokens return to the language capital understands:

**cash flow.**

The complete loop becomes:

> **Dollar → Physical Infrastructure → Compute → Token → Economic Utility → Revenue → Cash Flow → Dollar**

Capital leaves as money.

It travels through the physical world.

It becomes intelligence.

The intelligence must create economic utility.

Then the utility must return as cash.

If the chain breaks, Capex becomes stranded capital.

If the chain closes, AI infrastructure may become one of the foundational production systems of the next era.

---

# The AI Revolution Is Gaining Weight

For several years, the AI conversation focused on models.

GPT.

Gemini.

Claude.

Benchmarks.

Context windows.

Parameters.

Reasoning scores.

But the next decade may increasingly be governed by slower, heavier variables:

**MW.**

**GW.**

**Transformers.**

**Chillers.**

**Switchgear.**

**Backlog.**

**Capex.**

**FCF.**

**ROIC.**

**WACC.**

Models exist on screens.

Civilization does not.

Every unit of intelligence must eventually land on silicon.

Silicon must land inside a rack.

The rack must land inside a building.

The building must land on real ground.

The ground must connect to real electricity.

And somebody must commit real capital before any of it can exist.

That is what changed in 2026.

The AI revolution is gaining weight.

The $700 billion is simply the first moment when that weight becomes difficult to ignore.

The useful question is no longer:

**Will AI change the world?**

The useful questions are now:

Who is financing the infrastructure?

Where is the money flowing?

Where is scarcity actually located?

Who owns the bottleneck?

Who has revenue growth but weak free cash flow?

Who can convert an infrastructure boom into durable high ROIC?

And above all:

> **How much of that future is already embedded in today’s stock price?**

That is where AI commentary ends.

And pricing begins.

---

**SOVEREIGN PRICING WORLD 001**

*Definition → Mechanism → Industry → Company → Financials → Pricing → Verification*

The first beam of the framework is simple:

> **Capex is not merely a number.**
>
> **It is the physical trace of a future world under construction.**

The task from here is to follow the capital all the way down.

Until we can see:

**who receives the dollars,**

**who builds the assets,**

**who produces the cash flow,**

and finally,

**who deserves the capital.**

— **Pavel | Sovereign**
:::