Summary

Artificial intelligence is increasingly being analysed as a software boom when its economic footprint is beginning to resemble an industrial supercycle. The scarce resources are no longer just models and programmers. They are GPUs, advanced memory, electricity, transformers, cooling, land, grid connections and capital.

That distinction matters for investors. The greatest economic rents from AI may not ultimately sit with the companies charging the highest price per token. They could migrate toward whoever controls the bottlenecks required to produce intelligence at scale.

The result is a new macroeconomic variable: compute scarcity.

Intelligence Has Become Physical

The arrival of GPT-6 Astra on September 3 offered another reminder of how quickly the AI frontier is moving. OpenAI describes Astra as its most capable model yet, spanning coding, computer use, scientific work and complex professional tasks. It is also the company’s first broadly deployed model to reach its “Critical” cybersecurity capability threshold.

But the economically important part of increasingly capable AI may be less visible.

Intelligence now has an industrial supply chain.

Every additional unit of machine intelligence ultimately requires some combination of semiconductor fabrication, advanced memory, networking equipment, electricity, cooling, land, construction, financing and software capable of operating the infrastructure efficiently.

That makes the AI revolution fundamentally different from the software boom that preceded it.

The cloud transformed computing into an on-demand service.

AI is transforming electricity and silicon into cognition.

And that may be one of the most consequential commodity conversions since hydrocarbons became the foundation of the industrial economy.

The Market May Be Looking at the Wrong AI Bubble

There is a reasonable case that parts of AI are overvalued.

That is not the same as saying the infrastructure boom is fictional.

The Bank for International Settlements estimates that the five largest hyperscalers are on course to spend more than $1 trillion on AI-related capital expenditure across 2025 and 2026. The BIS also warns that the spending is beginning to outrun earnings and free cash flow, increasing the importance of debt financing.

That sounds like classic bubble material.

But bubbles and genuine technological revolutions are not mutually exclusive.

Railways transformed economies and still produced railway manias. Electrification created enormous productivity gains while destroying capital invested at the wrong price. The internet exceeded almost every long-term technological expectation and still generated the dot-com collapse.

The better question is therefore not:

Is AI a bubble?

It is:

Which layer of the AI capital structure is being overbuilt, and which bottlenecks are still being underbuilt?

That is where the alpha may be.

The Numbers Are Already Industrial

Nvidia’s latest results illustrate the scale.

For its quarter ended July 26, 2026, Nvidia reported $96.2 billion of revenue, up 106% from a year earlier. Data-center revenue reached $89 billion, up 117%.

This is no longer a niche semiconductor cycle.

Microsoft spent roughly $41 billion on capital expenditure in a single quarter during its fiscal fourth quarter, with about two-thirds directed toward shorter-lived assets including CPUs and GPUs. It also recorded $5.6 billion of finance leases, primarily for large data-center sites.

Meta expects 2026 capital expenditure of $130 billion to $145 billion, after repeatedly increasing its infrastructure spending expectations during the year.

The important observation is not simply that technology companies are spending heavily.

It is what they are spending on.

The largest companies in the digital economy are becoming some of the world’s largest builders of physical infrastructure.

The New AI Stack

The investable AI economy can increasingly be understood as seven interconnected layers.

1. Models

OpenAI, Anthropic, Google, Meta, xAI and competing laboratories produce intelligence.

This is the layer that attracts the most attention.

It may not ultimately capture the most durable margins.

2. Compute

GPUs, accelerators and alternative chip architectures transform electrical energy into model training and inference.

Nvidia currently occupies an extraordinary position here, but competition will intensify.

3. Memory and networking

AI workloads are increasingly limited not merely by computation but by the ability to move enormous amounts of data between processors.

That makes HBM, advanced packaging, networking and interconnects strategically important.

4. Data-center infrastructure

Servers require buildings, racks, cooling systems, backup systems, fibre and increasingly enormous campuses.

5. Electricity

At sufficient scale, AI becomes an energy industry.

6. Capital

Gigawatt-scale infrastructure costs billions of dollars and increasingly requires debt, structured finance, project finance and private credit.

7. Applications

Finally, intelligence reaches businesses and consumers through agents, enterprise software, robotics, research systems and consumer applications.

The mistake is assuming all seven layers will earn similar returns.

They will not.

The First Proprietary Thesis: Follow the Bottleneck, Not the Brand

Commodity economics provides a useful framework.

When demand increases faster than supply, economic rent tends to migrate toward the constrained input.

During an oil shortage, value moves toward oil producers.

During a shipping shortage, freight rates capture the scarcity.

During semiconductor shortages, fabrication capacity and equipment become unusually valuable.

AI should be analysed similarly.

If models proliferate while high-quality inference capacity remains scarce, value moves toward compute.

If GPUs become abundant but electrical connections become scarce, value migrates toward power.

If electricity is available but grid connections take four years, value migrates toward permitted sites.

If capacity becomes plentiful but capital becomes expensive, financing becomes the bottleneck.

The question investors should continuously ask is:

What prevents the next dollar of AI demand from being served?

Whatever the answer is may be the most valuable part of the stack.

The Electricity Constraint Is Becoming Real

The US electricity system spent more than a decade experiencing almost no demand growth.

That era has ended.

The US Energy Information Administration says electricity demand grew around 1.7% annually between 2020 and 2025, compared with just 0.1% annually between 2005 and 2019. It identifies data centers as a major driver.

The Department of Energy now describes America’s power system as moving from decades of stagnant demand toward unusually rapid load growth driven partly by hyperscale AI data centers.

Lawrence Berkeley National Laboratory scenarios highlighted by the DOE suggest data centers could represent approximately 11.8% of total US electricity consumption by 2030, with a range of roughly 9.5% to 15.3%.

The EIA goes further in its long-term scenarios. Servers alone accounted for an estimated 7% of commercial-sector electricity consumption in 2025 and could represent 22% to 33% by 2050.

This creates an inversion worth watching.

For decades, software companies were largely indifferent to geography because digital products could be reproduced almost anywhere.

AI infrastructure is deeply geographic.

Cheap electricity matters.

Grid availability matters.

Cooling matters.

Permitting matters.

Fibre matters.

Political stability matters.

Water availability can matter.

Natural gas pipelines can matter.

AI is making geography economically important to technology again.

Compute Could Become the New Oil

The analogy is imperfect but useful.

Oil powered machines.

Compute powers increasingly autonomous intelligence.

Oil created geopolitical chokepoints.

Advanced compute already has chokepoints across semiconductor manufacturing, lithography, advanced packaging, memory, power and networking.

Oil-rich states accumulated strategic power.

Regions with abundant low-cost electricity and AI infrastructure may acquire a different form of strategic advantage.

Oil futures became a macroeconomic indicator.

Compute pricing could eventually become one too.

A future central bank may care about GPU prices, token consumption, data-center electricity demand and semiconductor lead times for the same reason central banks currently monitor oil.

They can reveal demand pressure before traditional statistics do.

AI Is Already Becoming a Macro Variable

This is not theoretical.

The BIS says AI investment helped support global growth despite tariffs and geopolitical shocks. It also says AI enthusiasm helped sustain favourable financial conditions through higher equity valuations.

More strikingly, recent BIS-hosted analysis estimates that AI-connected companies account for roughly 40% of S&P 500 market capitalization, more than 30% of the MSCI Emerging Markets Asia Index, around half of US investment-grade corporate bond issuance and 87% of new venture-capital funding.

AI is therefore becoming several things simultaneously:

a productivity story,

an equity-market story,

an investment-cycle story,

an electricity story,

a credit story,

and increasingly a monetary-policy story.

That concentration creates enormous upside if AI monetization accelerates.

It also creates a new transmission mechanism for financial shocks.

The Great AI Contradiction

Here is perhaps the most important contradiction.

AI is supposed to reduce costs.

Yet building AI is currently increasing demand for some of the world’s most capital-intensive resources.

Models become more efficient.

Chips become faster.

Inference gets cheaper.

Normally, those developments should reduce infrastructure demand.

They may instead increase it.

This resembles the Jevons paradox, in which improvements in efficiency lower the effective cost of using a resource enough that total consumption rises.

If an AI task becomes ten times cheaper, businesses may not spend one-tenth as much.

They may run one hundred times more AI tasks.

Agents could make this effect much stronger.

A human may make several dozen meaningful software or research requests in a day.

An autonomous agent could generate thousands.

A network of agents could generate millions.

The unit cost of intelligence can therefore fall while aggregate demand for intelligence rises dramatically.

This is why cheaper AI is not necessarily bearish for compute.

It may be the opposite.

Open Source Could Compress Models and Expand Compute

The same logic applies to open-source AI.

Markets often interpret cheaper or open models as threats to expensive frontier-model providers.

That may be correct at the model layer.

It does not automatically follow that it is negative for the infrastructure layer.

If open models push the price of intelligence lower, more developers can deploy them.

If companies can customize models for specific workflows, inference volume can increase.

If routers dynamically choose between proprietary and open models, applications can consume more intelligence while paying less per unit.

This creates an important possibility:

AI model margins fall while compute demand rises.

That would move economic value downward through the stack.

Model providers might experience software-style margin compression while infrastructure owners experience commodity-style scarcity rents.

This is one of the most important distinctions the market may currently be missing.

The Second Proprietary Thesis: AI’s Marginal Dollar May Move Down the Stack

The first stage of the AI boom rewarded whoever demonstrated intelligence.

The next stage may reward whoever can deliver intelligence cheaply.

That changes the economic hierarchy.

Initially:

Model capability > everything else

Eventually:

cost × latency × reliability × availability > raw capability

Once several models become sufficiently intelligent for a task, customers can substitute between them.

That makes inference increasingly price sensitive.

But applications still require physical computation somewhere.

The likely result is a redistribution of margins.

Potential margin losers

Generic SaaS

thin AI wrappers

undifferentiated model APIs

applications without proprietary workflow or data advantages

Potential beneficiaries

high-utilization compute

advanced memory

AI networking

efficient inference infrastructure

power generation

grid equipment

data-center land

cooling

nuclear infrastructure

natural gas infrastructure

high-quality AI financing

The AI trade may therefore broaden even while individual AI stocks collapse.

Why SaaS Faces a Different Problem

Traditional software economics relied on unusually attractive characteristics.

Build once.

Distribute almost infinitely.

Serve the next customer at extremely low marginal cost.

Charge recurring subscription fees.

Generate 70%, 80% or even higher gross margins.

AI changes part of that equation because every intelligent action has a meaningful inference cost.

It also changes customer expectations.

A software company growing steadily according to an old five-year plan may still destroy relative value if AI-native competitors are increasing output or reducing costs exponentially faster.

The benchmark itself is moving.

The danger for incumbent software companies is therefore not merely that customers will “vibe code” replacements.

The deeper risk is that software value migrates from owning interfaces toward owning outcomes.

Customers may eventually stop paying primarily for seats.

They may pay for work completed.

That would represent a profound repricing of enterprise software.

But There Is a Real Bear Case

None of this means the AI investment cycle cannot break.

It can.

In fact, the scale of the infrastructure boom increases the consequences if it does.

The BIS warns that AI capital expenditure is increasingly moving beyond internal cash flows toward debt financing. If expected returns disappoint, financing could retreat quickly, converting today’s investment boom into a prolonged capex downturn.

Its July research also argues that AI complicates monetary policy because it stimulates demand through investment while potentially increasing supply through productivity.

This produces several credible failure modes.

1. Model progress slows

If additional computation stops producing economically valuable improvements, demand assumptions must fall.

2. AI revenue fails to catch capex

This is arguably the most important metric.

Revenue eventually has to justify investment.

3. Compute becomes genuinely abundant

Persistent declines in utilization, rental prices and lead times would indicate the scarcity thesis is weakening.

4. Financing breaks

A significant rise in corporate spreads or private-credit stress could stop projects even if long-term demand remains intact.

5. Electricity becomes politically constrained

AI infrastructure increasingly competes with households and other industries for electricity. Political resistance can become a genuine supply constraint.

Texas offers a striking example. The EIA lowered its 2027 electricity-demand forecast after the state announced a pause on new data-center development while reviewing projects.

6. Efficiency outruns the rebound effect

The bullish infrastructure thesis assumes falling inference costs create enough new consumption to offset efficiency gains.

That cannot simply be assumed.

It must be measured.

What Investors Should Measure Instead of AI Headlines

The next stage of AI investing requires different indicators.

Model benchmark scores still matter.

But they are increasingly insufficient.

The more useful telemetry may be:

GPU utilization

Are processors genuinely scarce?

GPU rental pricing

What is the marginal price of compute?

HBM and DRAM pricing

Is memory becoming the limiting resource?

Token consumption

Is actual AI usage accelerating?

Hyperscaler operating cash flow

Can current businesses finance the buildout?

AI capex growth

Is investment still accelerating?

AI credit spreads

What does the bond market think about the economics?

Data-center power pricing

Is electricity capturing scarcity rents?

Grid connection queues

How quickly can new capacity actually come online?

Revenue per unit of compute

Is monetization improving faster than infrastructure costs?

These variables together tell investors considerably more than asking whether the latest model is impressive.