The Unit of Computing Has Changed
I believe one of the most important economic transformations of our time is occurring beneath the surface of the AI boom. We are not simply building faster computers. We are changing what a computer is.
For most of the digital era, I could think of a computer primarily as a retrieval machine. Information was created in advance, stored somewhere, and subsequently retrieved. Search engines, databases, operating systems, websites, cloud storage, and recommendation systems all operated largely within that paradigm. The fundamental economic value of computing came from storing, organizing, retrieving, transmitting, and processing information that already existed.
Artificial intelligence changes the economics of that model because the machine increasingly generates information in real time.
Instead of merely retrieving an answer, an AI system can reason through a problem, generate code, conduct research, query databases, use software tools, manipulate files, interact with other systems, and create an output that did not previously exist. Computing therefore moves from a warehouse model toward a factory model.
That distinction is enormously important.
A warehouse stores inventory. A factory produces something.
The more useful AI becomes, the more computing itself becomes a productive asset. An AI data center is not simply infrastructure supporting applications. It is increasingly part of the production process. Its output can be measured in tokens, decisions, software, discoveries, designs, analyses, customer interactions, and eventually physical-world actions.
This changes the capital-allocation equation.
If computation produces economically valuable output, then demand for computation is no longer determined simply by how much information society wants to store. It is determined by how much intelligence society is willing to purchase.
That could ultimately be a much larger market.
From Chips to Systems
I see another important transition taking place simultaneously: the unit of competition is moving upward.
The semiconductor industry once competed primarily around individual chips. Then the relevant unit became the computer. Then the cluster. Now increasingly it is the rack, the pod, the data center, and ultimately the AI factory.
This happens because artificial intelligence has become a distributed-computing problem.
A sufficiently large model cannot simply be accelerated by making one GPU faster. Once computation is distributed across thousands or tens of thousands of processors, the performance of the entire system becomes constrained by everything surrounding the processors: memory, networking, switching, storage, software, power delivery, cooling, communication latency, and the algorithms themselves.
This is fundamentally an Amdahl's Law problem.
If only half of a workload benefits from an enormous improvement in computation, making that computational component infinitely fast cannot make the entire system infinitely fast. The remaining bottlenecks become increasingly important.
That means the economic value of the individual component cannot be evaluated independently of the system.
The fastest GPU in the world is not necessarily the most valuable system. The most valuable system is the one that produces the greatest useful computation at the lowest effective cost while satisfying the constraints of power, bandwidth, latency, reliability, software compatibility, and capital expenditure.
This is why extreme co-design matters.
I think of extreme co-design as the optimization of the entire technology stack simultaneously: algorithms, models, software, CPUs, GPUs, memory, networking, storage, power, cooling, racks, data centers, and ultimately applications.
The implication for investors is straightforward: the AI economy cannot be understood by looking exclusively at AI model companies or GPU manufacturers. Value is distributed across an increasingly elaborate industrial ecosystem.
Memory manufacturers matter.
Advanced semiconductor manufacturing matters.
Packaging matters.
Networking matters.
Optics matter.
Power infrastructure matters.
Cooling matters.
Construction matters.
Utilities matter.
Data-center operators matter.
Software matters.
And eventually the companies using AI to increase their own productivity matter most of all.
The AI economy is therefore becoming an industrial ecosystem rather than merely a software sector.
The Real Moat Is Not Always the Best Technology
One of the most important principles I take from the evolution of computing is that technological elegance alone does not determine which architecture wins.
Install base matters.
Developers build for platforms that give them access to users. Users adopt platforms that have applications. Applications attract developers. Developers create more applications. More applications attract more users.
That is a network effect.
Once enough software, developers, knowledge, tools, libraries, educational material, infrastructure, and institutional expertise accumulate around an architecture, replacing that architecture becomes extraordinarily difficult.
This is why an architecture that is technically imperfect can defeat one that is theoretically superior.
The installed base becomes an economic asset.
In the AI era, software ecosystems may become even more powerful because the switching costs can extend beyond the software itself. Developers invest years in learning an architecture. Companies invest in codebases. Universities teach it. Cloud providers support it. Researchers optimize algorithms for it. Businesses build infrastructure around it.
The resulting moat is not merely intellectual property.
It is accumulated coordination.
That is a much harder moat to displace.
I therefore look at AI companies through a broader framework than technological capability. I ask:
How many developers are committed to the platform?
How much software has been built on it?
How many clouds support it?
How many enterprises depend on it?
How much institutional knowledge exists around it?
How quickly is the platform improving?
How much trust exists that the platform will still be supported years from now?
The combination of install base, ecosystem, execution velocity, and trust can be considerably more durable than a temporary technological lead.
The Economics of Strategic Sacrifice
The history of technology also teaches me that some of the most important investments initially look economically irrational.
A company can spend enormous amounts of money developing infrastructure that customers are not yet demanding. It can sacrifice current margins to establish capabilities whose economic payoff may not arrive for a decade.
From a conventional quarterly-management perspective, this can look irresponsible.
From a strategic perspective, it can be exactly the right decision.
The critical distinction is between an expense and an investment in future optionality.
If a company spends money merely to defend its existing position, the expenditure may produce incremental returns. But if the spending creates an entirely new platform upon which future markets can develop, the payoff can be nonlinear.
That is why I think investors need to distinguish between short-term profitability and strategic capital formation.
A business that deliberately sacrifices present margins to establish a future ecosystem may look weaker before it looks stronger.
The difficulty is that most strategic investments fail.
That is why conviction alone is insufficient.
The investment must be based on a coherent causal model of the future: What changes? Why does it change? What capabilities become necessary? Why is this company positioned to capture the resulting economics? What must be true for the thesis to work?
The deeper lesson is that extraordinary returns often emerge when a company invests ahead of the market rather than merely responding to current demand.
The Four Scaling Engines of AI
I see AI's development as a sequence of increasingly powerful scaling mechanisms.
The first is pre-training.
More computation applied to larger quantities of data historically produced more capable models. But the apparent limitation of finite human-generated data does not necessarily represent a terminal constraint.
Synthetic data changes the equation.
AI systems can use existing ground truth to generate additional training examples, augment information, create new scenarios, and produce experiences that can subsequently become training material.
This creates a feedback loop.
The second scaling engine is post-training. Models can be refined using increasingly sophisticated generated experiences and targeted training.
The third is test-time scaling.
Inference is not necessarily a trivial stage after training. If an AI system is actually reasoning—searching, planning, decomposing problems, exploring alternatives, using tools, and evaluating potential solutions—then inference itself can consume substantial computation.
In other words, AI is moving from merely remembering patterns toward spending computation to think.
The fourth scaling mechanism is agentic scaling.
One AI system can create or coordinate many sub-agents. Instead of one digital worker performing one sequence of operations, a system can deploy a team of specialized digital workers simultaneously.
That creates a powerful economic possibility.
Human organizations are constrained by the number of people who can be hired, trained, coordinated, and managed. Software agents can potentially be replicated at extremely low marginal cost relative to human labor.
The implication is profound: intelligence becomes increasingly scalable.
And when intelligence becomes scalable, the primary constraint shifts toward compute, energy, data, software, capital, and organizational integration.
AI Agents Will Not Eliminate Software
I reject the simplistic idea that increasingly capable AI means software and tools disappear.
I see almost the opposite.
The more capable an AI agent becomes, the more useful external tools become.
A digital worker needs access to ground truth. It needs files. It needs databases. It needs the internet. It needs specialized applications. It needs APIs. It needs enterprise systems. It needs communication channels.
A sufficiently capable agent does not need to reinvent every tool.
It needs to know how to use the tools that already exist.
This produces a new layer of software economics.
The interface between humans and software may increasingly become the AI agent itself, but the underlying software infrastructure does not necessarily disappear. Instead, the agent becomes the orchestration layer connecting intelligence to existing systems.
That could make APIs, data infrastructure, enterprise software, security systems, and specialized applications more important rather than less important.
The computer is effectively being reinvented as an autonomous system that can perceive, reason, retrieve information, use tools, execute actions, and learn from the results.
The Agentic Economy Requires a New Security Model
There is, however, a fundamental problem.
An AI agent becomes economically useful precisely because it has access to things.
The more access it has, the more useful it becomes.
But the more access it has, the greater the potential damage if it is compromised, manipulated, or simply makes an incorrect decision.
Three categories of capability are particularly important:
Access to sensitive information.
The ability to execute code or actions.
The ability to communicate externally.
Giving an agent all three simultaneously creates a much larger security surface.
A safer architecture can constrain the combination of permissions. An agent might have access to information and execute certain actions but lack unrestricted external communication. Or it might communicate externally while having tightly constrained access to sensitive systems.
This suggests that agentic computing will generate a substantial new market for identity, authorization, policy engines, sandboxing, monitoring, auditability, and secure execution.
Security is therefore not an accessory to the AI economy.
It is part of the infrastructure.
Power Is Becoming an Economic Input to Intelligence
The most obvious physical constraint on AI expansion is energy.
But I think the deeper issue is not simply how much electricity AI consumes. It is how efficiently electricity can be converted into useful intelligence.
The relevant metric increasingly becomes performance per watt.
If hardware becomes more expensive while its computational productivity increases much faster, the cost of generating useful AI output can still fall dramatically.
This creates a familiar technological pattern.
The absolute quantity of infrastructure increases while the unit cost of the output declines.
That dynamic can stimulate demand rather than destroy it.
If tokens become cheaper, society may simply consume vastly more tokens.
This is analogous to many technological revolutions in which falling unit costs create entirely new categories of demand.
The result could be an enormous expansion in computational consumption even while the cost per unit of intelligence falls.
That is why I do not think the energy problem should be viewed simply as a ceiling.
It is simultaneously a constraint and an investment opportunity.
The Grid May Contain Hidden Capacity
There is another potentially important economic insight in the energy discussion.
Electric grids are engineered for peak demand rather than average demand.
That means substantial capacity may remain unused for much of the year.
If AI data centers can become flexible electricity consumers, they could potentially absorb some of this otherwise idle capacity while reducing consumption during periods of extreme grid stress.
The key concept is graceful degradation.
An AI system does not necessarily need to operate at maximum performance every second.
Some workloads can be delayed.
Some can be moved geographically.
Some can operate at lower throughput.
Some latency-sensitive workloads can be shifted to other facilities.
If data centers were designed around variable power availability rather than absolute 100% power guarantees, the economics of the grid could change.
The data center becomes not merely an electricity consumer but a flexible load.
That could reduce pressure to build enormous amounts of new generating capacity solely to satisfy occasional peaks.
It also illustrates a broader principle: infrastructure constraints are often partly architectural problems.
Instead of asking only, "How do we build more power?" I should also ask, "How can I redesign the system so that existing power is used more intelligently?"
AI Is Creating a New Industrial Supply Chain
The semiconductor supply chain is becoming one of the most strategically important industrial systems on Earth.
Advanced lithography, semiconductor fabrication, advanced packaging, high-bandwidth memory, networking, optical systems, power equipment, cooling systems, construction, and data-center integration are increasingly interconnected.
The complexity is extraordinary.
A modern AI rack can contain an enormous number of components supplied by hundreds of companies.
This creates both opportunity and fragility.
A bottleneck in one part of the system can constrain the economics of the entire system.
But it also creates an unusual form of industrial leverage.
A company that can coordinate this ecosystem effectively can influence capital investment far beyond its own balance sheet.
If a dominant platform provider convinces suppliers that demand will grow dramatically, suppliers may invest billions of dollars in capacity before that demand fully materializes.
The platform company is therefore helping create the supply curve for its own future market.
That is an important form of economic power.
The New Industrialization of Computing
The AI revolution is increasingly blurring the line between software and heavy industry.
Building AI infrastructure requires enormous quantities of physical capital.
Semiconductor fabs require advanced manufacturing.
Data centers require land, power, cooling, networking, construction, and specialized equipment.
The result is a capital-intensive technology cycle.
This matters for macroeconomics.
An AI investment boom can generate demand across multiple layers of the economy simultaneously: semiconductors, construction, electrical equipment, utilities, industrial machinery, engineering, telecommunications, manufacturing, and eventually labor productivity.
The multiplier effects could therefore extend considerably beyond technology companies.
This also helps explain why AI could become a macroeconomic phenomenon rather than merely a sectoral one.
If AI increases productivity, the consequences can propagate through corporate margins, wages, output, capital expenditure, tax revenues, interest rates, and asset valuations.
The question is no longer simply whether AI companies will grow.
The question is how much of the economy will eventually be reorganized around machine intelligence.
The Productivity Question
Ultimately, the economic justification for enormous AI investment rests on productivity.
If AI merely produces impressive demonstrations, the capital cycle eventually breaks.
If AI materially increases output per worker, accelerates scientific discovery, reduces administrative costs, improves manufacturing, develops new products, discovers drugs, optimizes logistics, and creates entirely new services, then the economic payoff can be enormous.
Productivity growth is one of the most powerful forces in economics because it increases the amount of output society can produce from a given quantity of resources.
AI potentially attacks productivity at the level of cognitive labor.
That is particularly important because advanced economies have accumulated enormous quantities of physical capital while much of the remaining economic bottleneck consists of human decision-making, knowledge work, research, coordination, and specialized expertise.
If those activities become partially automatable, the effective supply of intelligence increases.
And unlike human intelligence, digital intelligence can potentially be replicated.
That is the fundamental economic asymmetry.
Intelligence May Become a Commodity
I think one of the most consequential implications of AI is that intelligence itself may become increasingly commoditized.
Historically, high-quality intelligence was scarce because highly capable humans were scarce.
A company could employ only so many brilliant engineers, scientists, analysts, programmers, lawyers, researchers, or strategists.
AI changes the marginal economics.
Once a capable model exists, additional instances can potentially be created at dramatically lower marginal cost.
The economic value therefore migrates.
If intelligence becomes abundant, possessing intelligence alone becomes less of a competitive advantage.
The scarce resources may instead become:
Capital.
Energy.
Data.
Distribution.
Trust.
Physical infrastructure.
Proprietary information.
Customer relationships.
Brand.
Execution.
Judgment.
Creativity.
And perhaps most importantly, the ability to coordinate intelligence toward valuable economic outcomes.
This could have major implications for education and labor markets.
The ability to memorize information becomes less economically valuable when machines can retrieve and synthesize information instantly.
The ability to ask good questions, define problems, exercise judgment, communicate, build relationships, manage uncertainty, and take responsibility may become relatively more valuable.
AI may therefore reduce the economic scarcity of intelligence while increasing the relative importance of other human characteristics.
The Labor Market Will Be Disrupted at the Task Level
I do not think it is useful to frame AI simply as "jobs versus no jobs."
A job is a bundle of tasks.
Some tasks can be automated.
Some can be augmented.
Some require human judgment.
Some require physical presence.
Some require trust or interpersonal interaction.
Some require accountability.
The degree of disruption will therefore depend on how much of a person's economic value consists of automatable tasks.
This suggests a more practical strategy for workers and businesses.
Instead of asking whether AI will eliminate a profession, I should ask which tasks within the profession AI can perform better, faster, or more cheaply.
A worker who learns to use AI to automate the low-value components of his job may become substantially more productive.
The greatest risk may not be that AI replaces every worker.
It may be that AI-enabled workers replace workers who refuse to adapt.
That distinction matters.
Technology historically tends to reorganize tasks before it completely reorganizes occupations.
China Demonstrates the Importance of Industrial Ecosystems
Another important structural lesson comes from the development of China's technology sector.
Technological leadership does not emerge from isolated companies.
It emerges from ecosystems.
Large populations of technically trained workers, strong educational systems, intense competition, regional experimentation, manufacturing capacity, software expertise, open-source participation, and dense networks of engineers can reinforce one another.
Competition can itself become a source of innovation.
When many companies simultaneously pursue the same opportunity, most will fail or remain mediocre. But the competitive process can also generate exceptional survivors.
This creates a selection mechanism.
The broader lesson is that technological competitiveness depends on the environment in which companies operate, not simply on individual corporate strategy.
Countries that produce engineers, researchers, manufacturers, entrepreneurs, capital, infrastructure, and competitive markets can create reinforcing technological ecosystems.
The industrial policy implications are therefore substantial.
Open Source Can Accelerate Technological Diffusion
I also see open-source AI as an important economic mechanism.
Proprietary systems can create enormous commercial value, but closed technology can also slow diffusion.
Open models allow researchers, businesses, students, governments, and developers to experiment without having to negotiate access to a proprietary platform.
That can accelerate innovation because thousands of independent actors can build on the same foundation.
Open source effectively turns innovation into a distributed research process.
For companies operating inside the ecosystem, this can be strategically valuable even when they do not directly monetize every open-source artifact.
If open models increase AI adoption, they expand the total addressable market for the infrastructure required to run AI.
The economics therefore need to be evaluated at the ecosystem level rather than simply at the product level.
TSMC Illustrates the Power of Trust and Execution
Another lesson I take from advanced semiconductor manufacturing is that technological superiority is only one component of competitive advantage.
Operational excellence matters enormously.
A manufacturer serving hundreds of technologically demanding customers must constantly adjust production schedules, wafer starts, capacity allocation, packaging, yields, delivery timing, and customer requirements.
The extraordinary capability is not simply making advanced chips.
It is orchestrating a massive system reliably.
That produces something difficult to quantify on a balance sheet: trust.
Trust becomes economically valuable when companies organize their businesses around another company's promises.
If a major technology company can confidently plan product launches, revenue forecasts, inventory, and capital expenditures around a supplier's ability to deliver, that supplier becomes deeply embedded in the customer's economic system.
The switching cost is not merely technical.
It is operational.
First Principles Versus Incrementalism
One of the most useful management frameworks I take from this material is the distinction between continuous improvement and first-principles reconstruction.
Continuous improvement asks:
How can I make the current process slightly better?
First-principles thinking asks:
Why does the process exist in its current form at all?
Suppose something takes seventy-four days.
The incremental approach might attempt to reduce that to seventy-two.
The first-principles approach asks why it takes seventy-four days and what would happen if I designed the entire process from scratch.
Perhaps the physical constraints imply that six days is achievable.
The forty percent, fifty percent, or ninety percent improvement becomes conceivable only after I abandon the assumption that the existing process represents the natural limit.
This is especially powerful in industries with accumulated bureaucracy.
Processes often contain historical compromises that made sense when they were created but no longer represent physical or economic necessity.
First-principles analysis exposes those hidden assumptions.
The Speed-of-Light Framework
I find another useful way to think about optimization: measure every system against its physical or economic limit.
How fast could memory theoretically operate?
How quickly could a product be manufactured?
How much throughput is physically possible?
What is the minimum latency?
What is the minimum energy consumption?
What is the minimum cost?
How many people are actually necessary?
This creates a benchmark against which current performance can be measured.
The purpose is not to demand the physically impossible.
The purpose is to distinguish genuine physical constraints from inherited organizational constraints.
That distinction is enormously valuable.
Physics is a constraint.
Tradition is not.
A regulation may be a constraint.
A habit is not.
A supply-chain convention may be a constraint.
An organizational preference is not.
The entrepreneur's job is often to discover which is which.
The Investment Implications
For investors, I think the most important lesson is that AI should not be analyzed merely as a collection of rapidly growing technology stocks.
It is a capital cycle.
I should ask where capital is flowing, what bottlenecks exist, which suppliers capture pricing power, where capacity is expanding, where margins are rising, and where new infrastructure is being built.
I should also distinguish between revenue growth caused by temporary scarcity and revenue growth caused by structural expansion of the market.
A company benefiting from a temporary bottleneck can generate extraordinary profits until capacity catches up.
A company controlling a platform with network effects can potentially maintain superior economics much longer.
The distinction between the two is crucial for valuation.
High growth does not automatically justify a high valuation.
I need to consider the duration of that growth, the marginal capital required to sustain it, competitive threats, technological substitution, pricing power, operating leverage, and the eventual steady-state return on invested capital.
At the same time, traditional valuation frameworks can struggle when the addressable market itself is expanding rapidly.
If AI creates entirely new categories of computational demand, today's market-share calculations may understate the future opportunity.
The challenge for investors is therefore two-sided.
I must avoid underestimating genuine structural change.
But I must also avoid paying any price simply because the future is exciting.
The more transformational the technology, the more important valuation discipline becomes.
The Energy Trade May Become an AI Trade
The investment implications extend beyond semiconductors.
If computational demand grows dramatically, electricity demand becomes an economic input to the AI industry.
That creates potential opportunities in power generation, transmission, grid modernization, nuclear energy, natural gas, renewable generation, electrical equipment, transformers, cooling, and energy-management technologies.
But the relationship is not one-directional.
More power creates more compute.
More efficient compute reduces the cost of AI.
Lower AI costs stimulate demand.
Greater demand requires more compute.
More compute requires more power.
This creates a feedback loop.
The critical variable is therefore not simply total energy consumption. It is the relationship between energy consumption, computational efficiency, and the economic value generated by computation.
The New Scarcity May Be Coordination
As technology becomes increasingly abundant, I think coordination itself becomes a scarce economic resource.
It is difficult enough to design a GPU.
It is considerably harder to coordinate GPUs, CPUs, memory, networking, software, power, cooling, suppliers, manufacturing, customers, developers, and capital into a coherent system.
This principle applies far beyond AI.
Complexity creates opportunities for companies that can integrate fragmented systems.
The winner is not necessarily the company with the best individual component.
It may be the company capable of orchestrating the greatest number of components into a superior economic machine.
That is why ecosystems matter.
That is why platforms matter.
That is why supply-chain relationships matter.
And that is why trust matters.
Technology Changes What Capital Means
Historically, capital meant factories, machinery, land, transportation networks, buildings, and financial assets.
Increasingly, capital includes computation.
A company can deploy billions of dollars into data centers and use those facilities to generate economically valuable intelligence.
This means compute itself becomes a form of productive capital.
The distinction between information technology and industrial infrastructure therefore becomes less meaningful.
The data center becomes a factory.
The GPU becomes industrial equipment.
Electricity becomes a direct production input.
Software becomes the control system.
Models become production technology.
Tokens become an output.
Agents become digital labor.
And the resulting system becomes a new form of industrial capital.
That is perhaps the most important conceptual shift of all.
The Long-Term Economic Possibility
If I extend these trends far enough, the implications become extraordinary.
Suppose AI systems continue improving.
Suppose inference becomes increasingly capable.
Suppose agents can perform increasingly complex sequences of tasks.
Suppose the cost per unit of computation continues falling.
Suppose energy infrastructure expands.
Suppose AI becomes integrated into medicine, engineering, manufacturing, finance, science, agriculture, transportation, logistics, education, and government.
The resulting economy could experience a significant acceleration in productivity.
New medicines could be discovered faster.
Engineering cycles could shrink.
Software could become cheaper to produce.
Businesses could automate enormous quantities of administrative work.
Scientific research could accelerate.
Physical production could become more autonomous.
Entire categories of products and services could emerge that are difficult to imagine today.
The ultimate economic question becomes:
How much intelligence can society economically deploy?
That question is much larger than asking how many computers society needs.
Humanity May Become More Valuable as Intelligence Becomes Cheaper
There is an important paradox here.
If intelligence becomes abundant, humanity may become more economically and philosophically important rather than less.
Intelligence is a capability.
Humanity encompasses considerably more than intelligence.
Judgment, character, compassion, courage, trust, responsibility, empathy, determination, creativity, and the ability to establish meaningful relationships do not automatically disappear when machines become better at reasoning.
In fact, their relative importance may increase.
If machines can perform more of the computational work, humans may increasingly be judged by what they choose to do with that capability.
That could be one of the most important consequences of AI.
The technology may not simply make humans obsolete.
It may separate human value from the narrow category of information processing.
The Ultimate Competitive Advantage: The Ability to Imagine
The most difficult investment and business problem in periods of technological transition is not calculating what exists today.
It is imagining what becomes possible tomorrow.
Markets are generally much better at extrapolating existing categories than imagining new ones.
That creates opportunities.
The companies that build the infrastructure for an emerging market must often invest before the market is obvious.
The investors who understand the emerging market early can potentially capture enormous returns.
But the same characteristic creates enormous risk.
An imagined future is not guaranteed to arrive.
The discipline is therefore to reason from first principles.
What physical constraints exist?
What technological constraints exist?
What economic incentives exist?
What becomes cheaper?
What becomes more expensive?
What becomes scalable?
What new bottlenecks appear?
What industries become larger?
What industries become smaller?
Where does value migrate?
Where does capital flow?
And, ultimately, who captures the economics?
Those are the questions I want to ask when I analyze AI.
The AI Factory Is Only the Beginning
I increasingly think of the AI revolution not as the invention of a better software application but as the construction of a new industrial system.
The computer is becoming a factory.
The factory produces intelligence.
The intelligence can perform work.
The work can generate economic output.
The output creates demand for more intelligence.
More demand creates demand for more computation.
More computation creates demand for more semiconductors, memory, networking, energy, cooling, data centers, and capital.
Lower computational costs then expand the market further.
That is a potentially self-reinforcing economic cycle.
The ultimate limiting factors may therefore move continuously.
First it was data.
Then compute.
Then inference.
Then agents.
Then energy.
Then manufacturing capacity.
Then software integration.
Then security.
Then perhaps something we have not yet identified.
That is how technological revolutions tend to work. Solving one bottleneck creates another.
The opportunity lies in identifying the next bottleneck before the market fully recognizes it.
That is the framework I find most useful for thinking about the AI economy.
I do not want to think merely about which company has the fastest chip, the largest model, or the most popular application.
I want to understand the entire machine.
I want to understand the flow of capital, energy, computation, software, talent, manufacturing capacity, and economic value through the system.
Because if AI truly becomes a new general-purpose production technology, the ultimate opportunity is not merely to sell computers.
It is to participate in the economic transformation created by machines that can increasingly perform cognitive work.
And that transformation could become one of the largest capital-allocation events in modern economic history.
01 Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution