For the last two years, the public story of AI has mostly been a software story.
Chatbots.
Agents.
Coding assistants.
Frontier models.
Prompt engineering.
New applications.
New startups.
New workflows.
But underneath that visible software revolution, a much larger industrial story is unfolding.
AI is not only a new interface.
It is not only a new productivity tool.
It is not only a new layer of software.
AI is becoming a new demand engine for the physical world.
It needs electricity.
It needs chips.
It needs fabs.
It needs data centers.
It needs cooling.
It needs fiber, copper, optics, memory, land, zoning, turbines, capital, engineers, and geopolitical stability.
The deeper lesson from Gavin Baker’s recent conversation with Patrick O’Shaughnessy on Invest Like The Best is that the age of AI is not just digital. It is industrial.
The next phase of AI will be shaped by two constraints: watts and wafers.
Watts are energy.
Wafers are semiconductor capacity.
Everything else depends on them.
Baker’s central claim is extreme: what is happening in AI may be one of the most extraordinary moments in the history of capitalism.
The reason is the speed of value creation.
In previous software waves, iconic companies took a decade or more to build massive recurring revenue businesses. The cloud and SaaS giants compounded over years through product, distribution, enterprise sales, and operational execution.
AI labs are compressing that timeline.
The demand for frontier intelligence is appearing faster than almost any previous market demand. Companies are not merely experimenting with AI. They are consuming intelligence as fast as it can be produced.
That is the unusual part.
Most technology booms begin with a promise. The infrastructure is built in anticipation of future demand. Sometimes the demand arrives. Sometimes it does not. That is how bubbles form.
AI is different because, at least for now, demand appears to be real, immediate, and supply-constrained.
The GPUs are not dark.
The data centers are not empty.
The frontier models are not unused.
The bottleneck is not a lack of imagination.
The bottleneck is the physical ability to produce enough intelligence.
That changes the nature of the boom.
It does not mean there cannot be a bubble. There probably will be excess somewhere. History suggests that foundational technologies attract too much capital, too much certainty, and eventually too much supply.
But this AI cycle begins from a strange place: not too much capacity, but not nearly enough.
The internet made information abundant.
AI is making intelligence abundant.
But intelligence is not free. It has a cost structure. Every answer, agent action, code generation, simulation, research pass, image, video, protein model, or robotics policy has a physical footprint somewhere in the system.
The token looks weightless to the user.
It is not.
Behind the token is a stack: power generation, grid connections, data centers, GPUs, memory, networking, semiconductor manufacturing, advanced packaging, cooling systems, and capital markets.
This is why watts and wafers matter. They are not side issues. They are the foundation.
If AI demand keeps growing, the economy will be forced to reorganize around compute the same way previous eras reorganized around railroads, electricity, oil, highways, and the internet.
The companies that control scarce inputs will capture enormous value. The companies that relieve bottlenecks will become strategically important. The countries that can produce energy, build infrastructure, and secure chip supply will gain geopolitical leverage.
AI is often described as a software revolution.
It may be more accurate to describe it as the largest industrial mobilization since the internet, and possibly since electrification.
The first constraint is power.
AI needs enormous amounts of electricity. Training frontier models is energy-intensive. Inference at scale may become even more important as billions of people and millions of companies begin using agents constantly.
Baker’s view is that capitalism is good at solving this kind of problem over time. If the world needs more energy, capital will move toward turbines, gas, nuclear, solar, batteries, grid infrastructure, and new forms of generation.
The harder constraint may not be the existence of energy sources, but the social and political ability to build.
Permitting.
Zoning.
Grid connection queues.
Environmental opposition.
Local politics.
Transmission constraints.
Regulation.
In other words, the energy bottleneck is partly physical, but partly institutional.
This is a recurring theme in the age of AI: the limits are not only technical. They are civilizational. Can societies build fast enough? Can democratic systems approve infrastructure quickly enough? Can countries align energy policy with AI competitiveness? Can local concerns be reconciled with national strategic needs?
AI turns these questions from boring infrastructure debates into central economic questions.
The country that can build power fastest may gain an AI advantage.
One of the most fascinating parts of Baker’s argument is orbital compute.
The phrase “data centers in space” sounds absurd at first. It evokes giant floating buildings orbiting Earth. But Baker reframes it differently: not data centers as buildings, but compute racks in space, connected through lasers, powered by continuous solar exposure, and launched economically through reusable rockets.
Whether or not this happens exactly as imagined, the thought experiment is important.
If AI demand becomes large enough, the economy will search for compute wherever the constraints are weakest.
On Earth, data centers face power limits, land limits, cooling limits, water limits, regulatory limits, and political limits.
In space, some constraints change. Solar exposure can be abundant. Cooling is different. Land use disappears. Connectivity can be reimagined. Launch cost becomes the gating factor.
The fact that serious investors and operators are even discussing orbital compute tells us something about the scale of the AI build-out.
When demand is small, you optimize software.
When demand is large, you build data centers.
When demand becomes civilization-scale, you start asking whether compute should leave the planet.
That is the real signal.
AI is expanding the frontier of infrastructure imagination.
Power may be solvable by capital over time. Wafers are harder.
Advanced AI chips depend on semiconductor manufacturing at the very edge of human industrial capability. This is not a market where capacity appears instantly because money wants it to. It requires fabs, extreme ultraviolet lithography, advanced process knowledge, supply chains, equipment, chemicals, packaging, talent, and years of accumulated know-how.
This is why TSMC matters so much.
The world’s AI ambitions run through a small number of extremely complex manufacturing processes. The chip supply chain is global, but its most advanced node is highly concentrated. That makes wafers not just an economic bottleneck, but a geopolitical one.
If watts are about energy abundance, wafers are about precision, patience, and industrial mastery.
A data center can be financed.
A turbine can be ordered.
A fab cannot be improvised.
This is the “Silicon Shield” logic around Taiwan: the world depends on Taiwan’s semiconductor ecosystem so deeply that it becomes geopolitically central.
In the age of AI, this dependency grows even more consequential. AI capability is downstream of chips. Chips are downstream of fabs. Fabs are downstream of geopolitical stability.
The frontier of intelligence now depends on a fragile physical geography.
One of Baker’s most interesting claims is that wafer scarcity might actually help prevent an AI bubble.
Historically, major technology waves often produce overbuilding. Railroads, canals, telecom fiber, and internet infrastructure all attracted huge amounts of capital. The market correctly understood that the technology mattered, but then overestimated how quickly demand would absorb supply.
AI could follow that pattern.
But the current cycle has one difference: supply is constrained.
If the world cannot produce enough advanced chips fast enough, then capacity cannot run too far ahead of demand. The shortage acts as a brake. It limits how irrational the build-out can become.
That does not eliminate risk. It changes where risk lives.
If TSMC expands too slowly, AI progress is constrained.
If TSMC expands at the right pace, the market stays hungry but not flooded.
If alternative suppliers flood the market, a classic overcapacity cycle becomes more likely.
In this sense, semiconductor capacity decisions may become one of the most important macro variables in the world.
Not interest rates.
Not just oil.
Not just cloud budgets.
Wafer capacity.
The pace at which the world can manufacture advanced chips may determine whether AI remains a high-return infrastructure boom or becomes an overbuilt bubble.
The software side of the conversation is equally important.
A major question is whether frontier tokens will continue to capture most of the economic value at the model layer.
In theory, intelligence should commoditize. Models should get cheaper. Open-source systems should catch up. Companies should prototype with frontier models and then move production workloads to cheaper alternatives.
Some of that will happen.
But Baker points to a surprising reality: the frontier still matters enormously. The best models are not just marginally better. They can be meaningfully more useful, more reliable, more agentic, and more economically valuable.
That creates a strange market structure.
The world wants cheap intelligence.
But the most valuable work often still wants the best intelligence.
And the best intelligence remains expensive, scarce, and compute-constrained.
This helps explain the shift toward usage-based pricing.
The age of flat-rate AI subscriptions may be a temporary phase. If one person can run dozens or hundreds of agents, “all you can eat” pricing breaks. Intelligence becomes metered. Tokens become more like electricity, cloud compute, or mobile minutes.
That has a democratic downside.
If the best AI is available only to those who can afford usage-based frontier access, then intelligence becomes stratified. The rich, the large companies, and the best-capitalized startups get the strongest cognitive leverage. Everyone else receives a throttled version.
That is a major social question for the age of AI.
The future may not be divided between people who use AI and people who do not.
It may be divided between those with access to frontier intelligence and those stuck below the frontier.
The most uncomfortable part of the AI boom is the application layer.
The obvious assumption was that AI would create thousands of valuable software companies on top of the models. That will still happen. But the early reality has been more difficult.
Much of the value has accrued to the infrastructure layers: energy, data centers, chips, model companies, and cloud platforms.
Application companies face a brutal question: what is defensible?
If a product is obvious and easy, the model labs can build it.
If it is narrow, the market may be too small.
If it depends on proprietary data, the data must be valuable enough to beat frontier general intelligence.
If it is a workflow wrapper, it may be absorbed by a platform.
If it is not in the “token path,” it may struggle to capture economics.
This is why coding has been such a strong early category. Code is both a high-value workflow and a path toward more AI capability. If a model can code, it can increasingly build tools for itself and for users. Coding is not just another vertical. It is a meta-vertical.
For other application founders, the bar is rising.
The question is no longer: can you add AI to a workflow?
The question is: can you build something different, hard, and strategically positioned before the frontier models or platforms reach your market?
This is a very different startup environment.
AI lowers the cost of building, but raises the standard for defensibility.
Baker uses a useful lens from chip design that applies far beyond chips: different and hard.
If you are building a chip startup, do not simply try to build a better GPU. Nvidia has scale, customer relationships, supply chain access, software, and years of accumulated optimization. A small company trying to beat Nvidia at Nvidia’s own game is unlikely to win.
Instead, a startup must do something different — and that difference must be hard enough that incumbents cannot copy it immediately.
The same is true for AI applications.
A founder needs to ask:
Is this obvious?
Can a frontier lab add it as a feature?
Can a cloud platform bundle it?
Can an incumbent with distribution copy it?
Can open source erode it?
Can the data moat actually compound?
Is the workflow valuable enough to support a venture-scale company?
Is the company in the token path?
Is the hard part technical, operational, regulatory, distributional, or cultural?
In the age of AI, “we use AI for X” is not enough.
Everyone can use AI for X.
The opportunity is in doing something that becomes more valuable as AI gets better, not less.
The big technology companies are no longer just product companies. They are compute allocation machines.
Google, Microsoft, Amazon, Meta, Nvidia, and others must decide not only what products to build, but where to allocate scarce compute.
Should GPUs serve external customers?
Should they train internal models?
Should they improve existing products?
Should they support startups?
Should they be reserved for frontier research?
Should they be monetized today or used to defend strategic position tomorrow?
Compute has become strategy.
This is one of the defining changes of the AI era. In the cloud era, software companies competed through product, distribution, and infrastructure scale. In the AI era, the largest companies also compete through access to intelligence production.
A company with more compute can train stronger models, serve more users, run more agents, improve products faster, and experiment at greater scale.
The installed base matters.
The chip roadmap matters.
The cloud relationship matters.
The model strategy matters.
The capital budget matters.
Big Tech is no longer simply competing in AI.
Big Tech is becoming the AI supply chain.
The final layer is geopolitics.
AI will not remain a business story. It is already becoming a national power story.
The countries with the best models, chips, drones, cyber capabilities, manufacturing base, energy supply, and autonomous systems will gain military and economic advantages. The battlefield is becoming more intelligent. Cyberattacks will become more sophisticated. Deepfakes and social engineering will become more convincing. Personal safety, corporate security, and national security will increasingly overlap.
This creates a paradox.
AI may be one of the most positive technologies ever created. It could accelerate drug discovery, improve education, increase productivity, help cure rare diseases, and extend human life.
But the same technology also destabilizes power balances.
The more powerful AI becomes, the more it matters who controls it, who can access it, who can defend against it, and how quickly societies adapt.
This is why the age of AI cannot be understood only through markets.
It is a civilizational transition.
The biggest mistake is to think AI is weightless.
Because the interface feels magical, we forget the machine underneath.
A user types a prompt.
A model answers.
An agent acts.
A company becomes more productive.
A founder builds faster.
A scientist discovers a molecule.
A student gets a tutor.
A designer creates a world.
A developer ships a product.
It feels like pure software.
But behind it is an industrial stack of almost absurd complexity.
Power plants.
Gas turbines.
Solar fields.
Nuclear debates.
Transmission lines.
Data centers.
Cooling systems.
Advanced packaging.
HBM.
EUV machines.
Taiwanese fabs.
American export controls.
Cloud capex.
Private credit.
Reusable rockets.
Lasers in orbit.
National security policy.
The future of intelligence is being built out of atoms.
That is the real lesson of watts and wafers.
The AI era will not belong only to the best software companies.
It will belong to the companies, countries, and founders that understand the full stack.
The model matters.
The chip matters.
The watt matters.
The wafer matters.
The data center matters.
The capital structure matters.
The regulatory environment matters.
The supply chain matters.
The application layer matters.
The geopolitical context matters.
AI is forcing the digital world and the physical world back together.
For years, software seemed to float above the economy. It scaled with near-zero marginal cost. It turned atoms into bits. It rewarded asset-light models, network effects, and distribution.
AI reverses part of that logic.
The most important software in the world now demands some of the most expensive physical infrastructure ever built.
The next internet may require power plants.
The next platform may require fabs.
The next productivity boom may require orbital compute.
The next startup wave may depend on access to frontier tokens.
The next geopolitical order may depend on who controls the compute stack.
This is why the age of AI is not merely a software age.
It is an industrial age.
And the central question is no longer only: what can intelligence do?
It is also: can we build enough of it?
