The Intelligence Industrial Revolution — Part I

The $1 Trillion AI Question: Who Actually Gets Paid?

Artificial intelligence is rapidly becoming the largest technology infrastructure buildout of the modern era.

And that changes the question investors should be asking.

For the last several years, the conversation revolved around models:

Which AI is smartest?

Which benchmark does it win?

Who has the largest context window?

Who can generate the best code?

Who will beat OpenAI, Anthropic, Google or Meta?

Those questions still matter.

But trillions of dollars do not ultimately flow toward benchmark scores.

They flow through factories, electrical grids, memory chips, cooling systems, networks, data centers, cloud providers, software platforms and customers.

The next phase of the AI boom is therefore less about who builds the smartest intelligence and increasingly about who captures the economics of intelligence.

That distinction may determine the real winners of the AI revolution.

J.P. Morgan Asset Management currently estimates that hyperscaler capital spending surpassed $400 billion in 2025 and is on track to approach $800 billion in 2026, roughly ten times 2019 levels. Another J.P. Morgan analysis puts 2026 hyperscaler spending at about $629 billion, illustrating how quickly estimates are being revised upward as the buildout accelerates.

Stanford’s 2026 AI Index gives another perspective: global corporate AI investment more than doubled during 2025, with private investment alone rising 127.5%.

We are no longer talking about another software cycle.

We are watching the construction of an intelligence industry.

And that brings us to the trillion-dollar question:

If the world spends a trillion dollars building AI, which companies capture the economic value—and which merely absorb the cost?

---

Follow the Money, Not the Model

A user opens an AI application and types:

“Analyze this company for me.”

The experience feels almost weightless.

Seconds later, an answer appears.

But that answer sits at the end of an extraordinary industrial chain.

Follow the dollar backward:

Customer → Application → Agent Platform → Foundation Model → Cloud → Accelerator → Memory → Networking → Cooling → Data Center → Land → Electricity

Every step extracts part of the economic value.

Every step has different:

• margins
• capital requirements
• competitors
• bottlenecks
• depreciation schedules
• pricing power

And every layer presents investors with a different bet.

That means “investing in AI” is increasingly about as precise as saying:

“I’m investing in transportation.”

Are you buying the automobile manufacturer?

The battery producer?

The oil company?

The highway operator?

The logistics company?

The dealership?

The software inside the vehicle?

The same distinction is emerging in AI.

---

Layer 1: Electricity

Before artificial intelligence is software, it is electricity.

Every training run consumes it.

Every inference request consumes it.

Every autonomous agent consumes it.

Every GPU cluster eventually converts electricity into computation—and computation into intelligence.

The International Energy Agency projects global data-center electricity consumption could roughly double to around 945 TWh by 2030, with AI the most important driver of the increase. AI-optimized data-center demand itself is projected to more than quadruple.

That creates potential AI beneficiaries most consumers would never associate with artificial intelligence:

utilities.

Power producers.

Natural-gas infrastructure.

Nuclear developers.

Transformers.

Switchgear.

Substations.

Transmission equipment.

Battery systems.

Microgrids.

Suddenly, an obscure electrical-equipment manufacturer can have meaningful exposure to the AI boom without writing a single line of AI code.

This is the first lesson of the new AI economy:

The most important AI companies may not have “AI” anywhere in their names.

---

Layer 2: Land and Data Centers

Electricity needs somewhere to go.

That means land.

But not simply cheap land.

AI infrastructure wants land near:

• enormous power connections
• fiber
• water or advanced cooling
• skilled construction
• favorable regulation
• transportation
• geopolitical stability

This turns certain locations into strategic digital-industrial territory.

The modern hyperscale data center isn’t merely a warehouse full of computers.

It increasingly resembles a factory.

NVIDIA frequently uses the phrase “AI factory” for a reason.

Inputs enter:

electricity + data.

Outputs emerge:

tokens + predictions + decisions + digital labor.

That framing is useful because it changes how we think about AI economics.

If AI infrastructure behaves like industrial infrastructure, investors eventually have to examine:

• utilization
• depreciation
• capital intensity
• energy costs
• useful equipment life
• replacement cycles
• return on invested capital

Those are factory questions.

Not software questions.

---

Layer 3: Cooling

The higher the density of computing, the harder the thermal problem becomes.

Every watt entering the system eventually becomes heat.

Hundreds of thousands of expensive accelerators operating together generate an enormous thermal load.

Traditional air cooling becomes increasingly difficult at the most demanding densities.

That pushes infrastructure toward:

• direct-to-chip liquid cooling
• cooling distribution units
• heat exchangers
• advanced pumps
• immersion systems
• industrial HVAC
• heat-reuse systems

The irony is wonderful:

Some of the world’s most sophisticated artificial intelligence may ultimately depend on extremely sophisticated plumbing.

Cooling isn’t glamorous.

That’s precisely why it is interesting.

Investors often crowd into the glamorous layer while overlooking the indispensable one.

---

Layer 4: Networking

One AI accelerator is powerful.

One hundred thousand accelerators communicating efficiently become something entirely different.

At huge scale, networking can become the difference between an expensive collection of chips and a coherent AI supercomputer.

That means AI demand increasingly benefits:

• Ethernet
• InfiniBand
• optical transceivers
• switches
• silicon photonics
• interconnect technologies

An accelerator waiting for data is wasted capital.

A cluster unable to communicate efficiently is wasted capital.

AI therefore isn’t merely a compute problem.

It is a data-movement problem.

And that theme becomes even more important at the next layer.

---

Layer 5: Memory

The GPU gets the headlines.

Memory increasingly determines how efficiently the GPU can work.

High-bandwidth memory, or HBM, has become critical because AI accelerators need extraordinary amounts of information delivered extremely quickly.

SK hynix describes HBM as a technology specifically designed to address the memory wall by feeding large amounts of data to accelerators fast enough to keep computing resources productive.

Micron similarly describes HBM as vertically stacked memory designed to deliver the immense throughput required by AI and high-performance computing.

This creates a fascinating twist in the semiconductor story.

For years:

CPU → GPU

was the transition everyone discussed.

The next conversation may increasingly be:

GPU → GPU + memory + networking as an integrated system

The accelerator matters enormously.

But the accelerator cannot perform without being fed.

That’s why companies associated with memory may become some of the most important second-order beneficiaries of AI.

---

Layer 6: Accelerators

Now we reach the obvious winner.

NVIDIA.

AI accelerated computing transformed NVIDIA from an important semiconductor company into one of the central suppliers of the new intelligence economy.

But NVIDIA’s success also teaches us something important.

It did not merely build “the best chip.”

It built an ecosystem.

Hardware.

CUDA.

Networking.

Libraries.

Developer support.

Systems.

Increasingly, AI infrastructure.

That’s the difference between selling components and controlling an architectural layer.

And those architectural layers are usually where enormous economic value accumulates.

Yet the very size of NVIDIA’s opportunity naturally attracts competitors.

AMD.

Google TPUs.

Amazon Trainium.

Microsoft custom silicon.

Meta silicon.

Specialized inference processors.

Startups.

National semiconductor initiatives.

The accelerator layer could therefore simultaneously become larger and more competitive.

---

Layer 7: Semiconductor Manufacturing Equipment

Before there is an AI chip, there must be a semiconductor fab.

Before the fab produces the semiconductor, it needs some of the most sophisticated machinery humanity has ever created.

This creates another group of toll collectors:

the companies supplying the factories that manufacture AI chips.

Lithography.

Deposition.

Etching.

Inspection.

Packaging.

Testing.

Advanced chip packaging deserves particular attention because HBM and high-performance accelerators increasingly depend on complex packaging architectures.

The AI boom therefore doesn’t just increase demand for chips.

It increases demand for the machines that make the chips—and the machines that package the chips.

These are picks and shovels behind the picks and shovels.

---

Layer 8: Cloud

Very few businesses want to build a hyperscale data center themselves.

So enormous portions of AI infrastructure are consumed through:

Amazon Web Services.

Microsoft Azure.

Google Cloud.

Oracle Cloud.

Other specialized providers.

Cloud companies therefore occupy an extraordinarily powerful position.

They can earn money from:

storage.

compute.

networking.

databases.

models.

agents.

enterprise software.

And increasingly AI-specific infrastructure.

This creates a potential flywheel:

AI increases cloud usage.

Cloud revenue finances data centers.

Data centers attract AI developers.

AI applications drive more cloud consumption.

More consumption finances more infrastructure.

That explains why the hyperscalers are willing to invest staggering amounts.

But it also creates the central financial question of the cycle.

---

When Does $800 Billion Become Profit?

J.P. Morgan notes that hyperscalers are on track to spend roughly $629 billion in 2026 after spending about $715 billion combined across 2023–2025.

Another current J.P. Morgan AI investment overview puts 2026 hyperscaler capex closer to $800 billion.

These estimates differ because methodologies, company sets and forecast timing differ.

But economically, the conclusion is identical:

The denominator is getting enormous.

Eventually markets will ask:

How much incremental revenue did all that spending create?

And even more importantly:

How much incremental profit?

Suppose the industry spends another trillion dollars.

That trillion dollars isn’t free.

Capital has a cost.

Hardware depreciates.

Chips become obsolete.

Data centers require maintenance.

Electricity must be purchased.

Employees must be paid.

Debt may need servicing.

The ultimate question isn’t:

Did AI revenue grow?

It is:

Did AI generate sufficient returns on the capital required to create that revenue?

That’s the question that separates revolutionary technology from revolutionary investment returns.

They are not the same thing.

---

Layer 9: Foundation Models

Now things become particularly interesting.

OpenAI.

Anthropic.

Google.

Meta.

xAI.

Others.

They may produce the actual intelligence layer.

And yet it isn’t guaranteed that model providers capture the majority of AI’s eventual economic value.

Why?

Because foundation models are incredibly expensive to:

• train
• operate
• improve
• distribute

And competition is fierce.

Capabilities can converge.

Open-source alternatives can pressure pricing.

Inference costs can fall.

Customers can switch models.

Cloud providers can subsidize models.

Developers can route different tasks to different models.

So one of the great paradoxes of this cycle may emerge:

The companies creating some of the world’s most economically valuable technology may operate in one of the most competitive layers of the AI stack.

AI could become indispensable while basic intelligence becomes increasingly commoditized.

Those ideas are not contradictory.

---

Layer 10: Agents

Models answer questions.

Agents perform work.

This layer may become far more economically important.

A model can tell you how to book a vacation.

An agent books it.

A model can describe a marketing campaign.

An agent:

writes it.

creates the assets.

publishes it.

monitors results.

changes targeting.

updates the CRM.

The closer AI moves toward completed economic activity, the easier monetization becomes.

Businesses don’t ultimately want tokens.

They want:

sales.

leads.

software.

support tickets resolved.

contracts analyzed.

invoices reconciled.

appointments booked.

research completed.

products manufactured.

That means one of the biggest future battles is likely to be over the agent layer between models and business outcomes.

---

Layer 11: Applications

This was supposed to be the easy part.

Build an AI app.

Charge $20 per month.

Grow exponentially.

Reality is becoming more complicated.

Many AI applications face:

• minimal switching costs
• low barriers to entry
• model dependence
• expensive inference
• aggressive competition
• features rapidly copied by incumbents

This is precisely why some AI-app valuations deserve much greater skepticism than the underlying AI technology itself.

An extraordinary technology can create mediocre businesses.

Railroads proved it.

Airlines proved it.

Telecommunications proved it.

The internet proved it.

---

The Railroad Lesson

Railroads transformed America.

They connected cities.

Opened markets.

Changed manufacturing.

Changed migration.

Changed commerce.

Railroads were unquestionably revolutionary.

But being technologically revolutionary did not guarantee every railroad investor extraordinary returns.

Too much capital chased the opportunity.

Competitors overbuilt.

Pricing deteriorated.

Companies failed.

Assets changed hands.

And society inherited the infrastructure.

That distinction is crucial:

A technology can change civilization while simultaneously destroying investor capital.

---

The Fiber-Optic Lesson

The internet boom produced a similar story.

During the late 1990s and early 2000s, companies laid enormous amounts of fiber.

Demand forecasts were astronomical.

Capital flooded into telecom infrastructure.

Then the bubble broke.

Companies collapsed.

Investors lost fortunes.

But the fiber did not disappear.

The world eventually consumed the infrastructure.

And that cheap, abundant connectivity helped enable:

Google.

YouTube.

Netflix.

Amazon Web Services.

Facebook.

Streaming.

Cloud computing.

Entire industries were built partly upon assets whose original investors sometimes suffered badly.

AI could produce a variation of this phenomenon.

---

The Most Important Distinction in AI Investing

There are three separate questions:

Will AI transform society?

Probably.

Will AI create enormous economic value?

Increasing evidence suggests yes.

Will every company spending enormous sums on AI earn attractive returns?

Absolutely not guaranteed.

Those three questions are routinely treated as if they are the same.

They aren’t.

And understanding the distinction may become one of the most important investment advantages of the next decade.

---

The Toll Booth Theory of AI

This leads to my favorite framework for thinking about the AI economy.

Imagine the artificial-intelligence highway.

Thousands of companies are racing to build the best vehicle.

Maybe OpenAI wins.

Maybe Anthropic.

Maybe Google.

Maybe Meta.

Maybe open-source models become dominant.

Maybe the ultimate winner doesn’t exist yet.

Instead of trying to predict the vehicle:

Who owns the toll road?

If OpenAI wins, it needs electricity.

Anthropic needs electricity.

Google needs electricity.

Open-source AI needs electricity.

If agents explode in popularity, they need servers.

Those servers need accelerators.

Those accelerators need memory.

Those machines need networking.

Those racks need cooling.

Those facilities need transformers.

Those transformers need electricity.

The competitive outcomes can change enormously at the top of the stack while some underlying needs remain.

That is the Toll Booth Theory.

Own the bottleneck rather than betting exclusively on the winner.

---

Where Are the AI Toll Booths?

Potential toll booths include:

Electricity
No computation without power.

Grid infrastructure
Electricity is useless if it cannot reach the data center.

Semiconductor manufacturing equipment
No advanced chips without fabs.

Advanced packaging
No modern AI accelerator system without increasingly sophisticated integration.

HBM
No high-performance AI cluster without enormous memory bandwidth.

Networking
No massive cluster without data movement.

Cooling
No dense compute without heat removal.

Cybersecurity
More autonomous AI means more systems requiring protection.

High-value proprietary data
Models become more valuable when connected to scarce information.

Not every company within these categories will win.

But the framework helps investors ask a better question.

Instead of:

“What’s the next ChatGPT?”

Ask:

“What does every ChatGPT competitor have to buy?”

That is a dramatically different way to view AI.

---

The Commodity Paradox

Another possibility deserves attention.

What if intelligence becomes incredibly cheap?

Today that sounds bearish for infrastructure.

It might be bullish.

Economists call this type of effect the Jevons paradox: increasing efficiency can lower the cost of consuming a resource and thereby increase total consumption.

Suppose AI inference becomes 90% cheaper.

Maybe companies don’t spend 90% less.

Maybe they run:

100 times more agents.

Persistent AI assistants.

Continuous analysis.

Real-time translation.

Personalized video.

Robotic control.

Millions of autonomous software processes.

Cheaper intelligence could create vastly greater demand for intelligence.

Which could mean:

cheaper token → more tokens → more infrastructure

Efficiency does not necessarily kill the AI infrastructure story.

It could accelerate it.

---

AI Is Becoming an Industrial Economy

This may be the biggest conceptual shift.

For twenty years, Silicon Valley worshipped capital-light businesses.

Build software once.

Distribute it infinitely.

Marginal cost approaches zero.

AI breaks part of that assumption.

Frontier intelligence requires:

land.

factories.

semiconductors.

memory.

servers.

fiber.

power plants.

cooling.

huge capital investment.

Software is becoming physical again.

The modern AI company increasingly sits somewhere between:

a software company,

a utility,

a semiconductor company,

and an industrial manufacturer.

That may produce entirely different economics from the SaaS era.

---

Three Eras of AI Investing

The AI investment story is already evolving.

Phase 1 — Model Mania

Who has the best AI?

Attention focused on:

GPT.

Claude.

Gemini.

Llama.

Benchmarks.

Parameters.

Context windows.

Model launches.

Intelligence itself was the scarce resource.

---

Phase 2 — Infrastructure Mania

Who can supply enough compute?

The conversation expanded toward:

NVIDIA.

HBM.

data centers.

networking.

power.

cooling.

Accelerated compute became scarce.

Capital expenditure exploded.

We are still very much inside this phase.

---

Phase 3 — Return-on-Intelligence

This is the phase that matters next.

Who can turn intelligence into profit?

Markets will increasingly ask:

How much revenue does each dollar of AI capex produce?

How much labor does AI replace or augment?

How much more can customers pay?

How sticky are AI products?

Do agents complete economically valuable work?

What happens to margins?

How quickly does hardware depreciate?

What is utilization?

What is return on invested capital?

In other words, the AI market eventually moves from:

What can the machine do?

to:

What is the machine worth?

We’re beginning to enter that transition.

---

A New Metric: Return on Intelligence

Today’s AI industry loves technical metrics.

Tokens per second.

Training FLOPS.

Parameters.

Context.

Benchmark scores.

Cost per million tokens.

Those matter.

But finance eventually demands another metric.

Call it:

Return on Intelligence

How much incremental economic value does incremental AI investment create?

At the enterprise level:

AI-generated profit
divided by
AI-related investment.

At the hyperscaler level:

incremental AI revenue
divided by
incremental AI capital expenditure.

At the employee level:

economic output with AI
versus
economic output without AI.

At the agent level:

value of completed tasks
divided by
cost of intelligence required.

That may eventually matter more than almost every leaderboard benchmark.

---

The Ultimate AI Winners May Look Boring

And this is where things get interesting.

Some of the greatest AI beneficiaries may ultimately be companies producing things like:

transformers.

electrical switchgear.

cooling equipment.

industrial pumps.

optical components.

memory.

semiconductor equipment.

fiber.

natural gas turbines.

power-management systems.

Cybersecurity.

Those companies don’t generate viral demos.

Nobody asks their transformer supplier to write a poem.

Yet without these components:

Claude goes dark.

ChatGPT stops responding.

Gemini stops reasoning.

The agent economy stops.

The robots stop moving.

This is why the next stage of AI investing could become dramatically broader than Silicon Valley.

---

The Real Bubble Question

People keep asking:

“Is AI a bubble?”

But that question may be too crude to be useful.

The better questions are:

Which layer is overbuilt?

Which layer is undersupplied?

Which layer has pricing power?

Which layer will commoditize?

Which layer captures recurring revenue?

Which assets depreciate fastest?

Which companies benefit regardless of the model winner?

Which companies are spending capital they’ll never earn back?

Those questions allow for a far more realistic possibility:

AI can simultaneously be a technological revolution and contain multiple investment bubbles.

The internet did.

Railroads did.

Telecommunications did.

AI probably will too.

---

Who Gets Paid?

Follow the trillion dollars.

Some goes to land.

Some to utilities.

Some to data centers.

Some to NVIDIA.

Some to memory makers.

Some to networking suppliers.

Some to cooling companies.

Some to semiconductor equipment manufacturers.

Some to cloud providers.

Some to model developers.

Some to AI agents.

Some to software applications.

Eventually the customer has to generate enough economic value to support the entire chain.

That’s the part of the equation that matters most.

Because underneath every trillion-dollar infrastructure cycle lies a simple truth:

Someone eventually has to make enough money to pay for it.

---

The $1 Trillion Question

Artificial intelligence increasingly looks inevitable.

Profitable artificial intelligence infrastructure is not automatically inevitable.

Those are very different statements.

The next great AI investment contest won’t simply determine who produces the smartest machine.

It will determine:

who captures the margin,

who owns the bottleneck,

who bears the capital risk,

and ultimately,

who gets paid.

“The trillion-dollar AI question is not whether intelligence becomes valuable. It is who owns the toll booths between electricity and intelligence.”

Every AI answer begins as electricity.

It travels through transformers, data centers, cooling systems, networks, silicon and memory.

Then models convert that infrastructure into intelligence.

Applications convert intelligence into utility.

And customers must finally convert that utility into money.

“Every AI answer begins as electricity, travels through silicon and memory, and ends as economic value. The investment battle is over who captures the margin in between.”

That is where the AI story is headed next.

Not merely toward smarter machines.

Toward an entire industrial economy built around intelligence.

And if the next trillion dollars really is coming, the smartest question may no longer be:

Which AI will win?

It may be:

Who gets paid no matter which AI wins?

---

Sources & further reading

The strongest primary references behind this piece are J.P. Morgan Asset Management’s current AI investment research, which estimates hyperscaler capex approaching $800 billion during 2026 and separately notes a $629 billion current 2026 run-rate estimate; Stanford HAI’s 2026 AI Index, documenting the sharp acceleration in corporate AI investment; and the IEA’s Energy and AI work, which projects data-center electricity demand rising toward roughly 945 TWh by 2030.

Next in the series: Part II — AI Has a Memory Problem: Why the Next Great AI Bottleneck May Not Be Compute.