AI Is Becoming Productive Infrastructure
I increasingly view artificial intelligence not as another consumer technology cycle, but as the emergence of a new category of productive infrastructure. The important transition is from computing as a tool used by individuals and businesses toward computing as an industrial system that converts energy and capital directly into economically valuable intelligence.
The analogy I find most useful is the historical evolution of infrastructure. Electricity grids, roads, water systems, telecommunications networks, and the internet created platforms upon which entire economies could build additional productive capacity. AI infrastructure is beginning to occupy a similar position. Electricity provides the raw energy, semiconductors transform that energy into computation, data centers organize the computation at industrial scale, models convert computation into increasingly capable reasoning, and applications turn that capability into measurable economic output.
This changes the economic unit I care about. In the energy economy, a fundamental unit of monetization can be dollars per kilowatt-hour. In an AI economy, an analogous operational metric is dollars per million tokens. The token is not merely a technical abstraction; it represents the monetizable output of computational processes. The economic question therefore becomes increasingly straightforward: how much intelligence can I produce from a given quantity of energy, hardware, capital, data, and engineering effort, and what economic value can that intelligence ultimately generate?
That framework makes AI capital expenditure easier to understand. A data center is not simply a collection of servers. It is a productive asset that combines electricity, semiconductor capacity, buildings, networking, cooling, software, and models to manufacture computational output. The resulting output can be sold directly as compute or embedded into higher-value products and services.
The Five-Layer Economy of Artificial Intelligence
I find it useful to analyze the AI economy as a five-layer stack: energy, semiconductor hardware, physical infrastructure, models, and applications. Each layer represents a different form of capital and a different potential source of competitive advantage.
The first layer is energy. Without abundant, reliable electricity, there is no large-scale AI economy. Computation ultimately requires physical energy, so the expansion of intelligence becomes partly an energy-infrastructure problem. This creates an important economic connection between AI investment and power generation, transmission, grid capacity, and industrial construction.
The second layer is semiconductor hardware. Highly parallel processors transform electrical energy into enormous quantities of arithmetic operations. This layer sits at the intersection of semiconductor economics and AI economics, making chip architecture, manufacturing capacity, supply chains, and software compatibility strategically important.
The third layer is physical infrastructure. Gigawatt-scale computing requires land, electrical interconnection, thermal management, networking, buildings, and specialized equipment. The economics therefore extend well beyond the semiconductor industry. Construction companies, utilities, power developers, networking providers, equipment manufacturers, and infrastructure financiers can all participate in the expansion of the AI capital cycle.
The fourth layer is the model layer. Models transform raw computational capacity into increasingly useful representations of language, biology, chemistry, physics, mathematics, spatial relationships, and other structured information. Large language models are only one manifestation of this broader capability. The economically significant development is the ability to convert computational resources into reusable intelligence across many domains.
The fifth layer is applications and data. This is where infrastructure becomes embedded in actual economic activity. Education, healthcare, manufacturing, logistics, scientific research, software development, and other industries can incorporate AI into their existing workflows. The ultimate economic return on the infrastructure therefore depends not only on how efficiently computation is produced, but on how effectively it diffuses through the real economy.
This distinction is crucial. Building compute capacity creates potential productive power. Deploying that capacity into businesses creates actual productivity gains. I therefore see AI diffusion as at least as important as AI invention. A country or company does not necessarily need to dominate every layer of the stack to capture economic benefits. It can specialize in particular layers while aggressively integrating AI into its existing industries.
Capital Expenditure Is Moving Toward Infrastructure Scale
The scale of investment implied by this transformation is one of its most consequential economic features. The relevant comparison is no longer merely between software companies and traditional technology companies. AI infrastructure increasingly resembles heavy industrial investment.
A state-of-the-art semiconductor fabrication facility has historically represented an investment on the order of tens of billions of dollars. A one-gigawatt AI data-center environment can require roughly $50 billion to $60 billion when power, compute, facilities, and networking are considered together. A global buildout on the order of 100 gigawatts would therefore imply several trillion dollars of cumulative capital deployment.
I interpret these figures less as precise forecasts than as an indication of the economic category into which AI is moving. Once infrastructure requirements reach this magnitude, AI becomes deeply connected to interest rates, credit conditions, utility investment, construction cycles, government policy, industrial capacity, and capital-market financing.
This also creates a classic capital-allocation problem: enormous amounts of capital are being committed today in anticipation of future computational demand. The critical question is not simply whether AI is useful. It is whether the future cash flows generated by AI-enabled economic activity will justify the infrastructure investment required to produce them.
That distinction matters for investors. A technological revolution can be economically transformative while still producing periods of excessive valuation, overinvestment, or disappointing returns on individual projects. Technology adoption and investment returns are related, but they are not identical.
Why Hardware Fungibility Matters to Capital Allocators
I place particular importance on hardware flexibility because rapid technological change creates depreciation risk. When infrastructure costs tens of billions of dollars, an architecture that becomes obsolete quickly can destroy capital returns even if demand for AI remains extremely strong.
General-purpose parallel computing addresses part of this problem by allowing the same underlying hardware architecture to serve multiple computational workloads. The hardware can support different models, numerical formats, scientific simulations, graphics workloads, biological applications, physical simulations, and increasingly embodied AI systems.
This creates a form of technological fungibility. I can think of the investment as purchasing a flexible computational platform rather than a machine designed for only one narrow algorithmic regime.
The economic advantage compounds when hardware is supported by a broad software ecosystem. If software optimization can extend the useful applications of installed hardware as models and computational techniques change, the effective economic life of the capital asset can be longer than its initial use case suggests.
This is analogous to an important principle in capital-intensive industries: flexibility has economic value. When uncertainty about future technology is high, an asset capable of serving multiple markets can carry a significant advantage over a narrowly optimized asset.
AI Turns Natural Language Into a New Interface for Capital
The most profound change may not be the model itself, but the interface between humans and computation. For decades, direct control of computers required specialized technical knowledge. Programming languages, hardware architectures, and software engineering created a relatively narrow gateway through which people could access digital leverage.
Natural-language interfaces change that relationship. If I can express a desired computational outcome in ordinary language and have an AI system translate that intent into executable operations, the effective population capable of directing software expands dramatically.
I see this as an economic democratization of computational capital. The scarcity is no longer exclusively the ability to write code. It increasingly becomes the ability to define objectives, provide context, evaluate outputs, and integrate computational capabilities into productive workflows.
This distinction has major labor-market implications. AI can automate individual tasks without necessarily eliminating the broader occupational function that contains those tasks. Syntax generation, documentation, translation, routine analysis, and repetitive communication can become increasingly automated while judgment, objective-setting, domain knowledge, organizational context, and accountability remain valuable.
That suggests a more nuanced framework for understanding AI and employment. The relevant unit of automation is often the task rather than the job. Productivity can therefore rise even when the number of occupations remains relatively stable, because each worker can accomplish more with a given amount of time and capital.
From Language Models to Economic Agents
The next stage of AI development is not simply making models larger. It is making them operational.
A standalone language model can generate information, but an economically useful agent requires an architecture around the model. I think of this architecture as an operational harness consisting of retrieval, working memory, tool use, and collaboration.
Retrieval connects the model to external information. Working memory allows it to maintain state across longer tasks. Tool use lets it interact with spreadsheets, software, browsers, databases, code environments, and enterprise systems. Multi-agent collaboration allows complex problems to be decomposed into specialized tasks.
This distinction explains why achieving high cognitive capability does not automatically translate into immediate enterprise productivity. A highly capable intelligence still needs a purpose, context, permissions, data boundaries, workflows, and measurable objectives.
The economic bottleneck therefore shifts. As the marginal capability of foundation models rises, organizational design becomes increasingly important. Companies will need to determine which decisions can be delegated, what information an agent can access, how its work is verified, and where human accountability remains necessary.
I expect a substantial portion of the economic value of AI to come from this integration layer. The model may be generalized, but the productive implementation can be highly specific to an enterprise's data, processes, customers, intellectual property, and operational constraints.
The Enterprise AI Moat Will Often Be Context
This leads to an important strategic implication. I do not assume that owning access to a powerful general model automatically creates a durable competitive advantage. If many companies can purchase similar foundation-model capabilities, differentiation increasingly depends on what each organization places around those models.
Proprietary data, institutional knowledge, workflow integration, security architecture, customer relationships, specialized evaluation systems, and domain-specific applications can become sources of economic advantage.
I therefore distinguish between generalized intelligence and organizational intelligence. Generalized intelligence can increasingly be purchased as infrastructure. Organizational intelligence must be constructed.
This also creates a tension between outsourcing and sovereignty. Companies can rely on external AI providers for generalized capabilities while retaining sensitive institutional knowledge and proprietary intelligence layers internally. The resulting architecture resembles a hybrid capital model: generalized computational infrastructure is rented or purchased from the broader ecosystem, while organization-specific intelligence remains strategically controlled.
Embodied AI Extends the Productivity Frontier
The transition from digital agents to physical agents expands the economic opportunity further. Once AI systems can control actuators, sensors, vehicles, robotic arms, laboratory equipment, and other machines, computation begins to influence physical production directly.
Autonomous vehicles can perform mobility and logistics functions. Robotic manipulators can execute manufacturing operations. Automated delivery systems can move goods. Surgical systems can assist with highly precise procedures. Laboratory robotics can automate repetitive experimental processes in scientific research.
This matters because software historically had a particularly powerful economic characteristic: it could be reproduced at extremely low marginal cost. Embodied AI introduces a different dynamic. Intelligence can increasingly be replicated into physical capital, allowing machines to perform tasks that previously required human labor.
The long-term implication is a potential increase in the productivity of both labor and physical capital. A worker equipped with intelligent software can accomplish more. A factory equipped with intelligent machines can operate differently. A research laboratory equipped with autonomous experimentation can potentially increase the number and speed of empirical cycles.
The economic value therefore emerges from the interaction between intelligence and capital rather than from AI in isolation.
AI Can Compress the Time Dimension of Economic Development
One of the most consequential possibilities is the compression of development cycles. If AI can reduce a process that once required years to something that can be executed in months, and eventually compress months into weeks, the effective rate of capital turnover and innovation can increase substantially.
Time is an economic variable. Shorter development cycles allow companies to test more ideas with the same calendar horizon, shorten feedback loops, bring products to market faster, and potentially generate returns on invested capital sooner.
The effect can become nonlinear when improvements compound. Faster research creates faster commercial deployment; faster deployment creates more data; more data improves subsequent systems; improved systems accelerate the next research cycle.
I see this feedback loop as one of the strongest mechanisms through which AI could influence productivity growth. The key question is not merely whether AI can perform a task, but whether it accelerates the entire innovation system surrounding that task.
AGI Does Not Eliminate the Economics of Management
I also reject the simplistic assumption that achieving broadly capable artificial intelligence automatically produces proportional economic output. Capability is a necessary input, but organizational structure determines how much of that capability becomes productive.
A highly capable cognitive system without objectives, constraints, data access, feedback mechanisms, or organizational context is comparable to highly educated human talent without a clearly defined role. The economic value emerges only after the capability is connected to a production function.
This means management itself becomes part of the AI productivity equation. Companies will need to redesign workflows, define decision rights, establish verification systems, determine where humans remain in the loop, and allocate computational resources toward the highest-return activities.
The result could be a shift in the composition of corporate capital. Traditional firms already allocate capital among factories, equipment, software, employees, and financial assets. Increasingly, they will also allocate capital toward models, data infrastructure, agents, inference capacity, and AI-enabled workflows.
National Competitiveness Depends on Diffusion, Not Just Invention
At the national level, I see AI competitiveness as a diffusion problem as much as an invention problem. A country can possess world-class models and still fail to capture their full economic value if those capabilities do not spread into domestic industries.
The five-layer framework provides a useful policy lens. A nation with abundant energy resources may have an advantage at the power layer. Another may possess semiconductor expertise. Another may excel at data-center construction. Another may specialize in foundational models. Another may capture value by integrating AI into manufacturing, healthcare, education, logistics, or scientific research.
The critical strategic objective is therefore not necessarily vertical dominance across every layer. It is the ability to convert available advantages into widespread domestic productivity.
This is particularly significant for developing economies. If AI lowers the cost of accessing advanced knowledge and technical capabilities, countries with large populations can potentially improve educational, industrial, scientific, and entrepreneurial capacity without reproducing every stage of technological development followed by earlier industrial economies.
That creates the possibility of technological leapfrogging. The same phenomenon has appeared repeatedly in economic history when emerging technologies allowed countries to bypass legacy infrastructure. AI could extend that pattern into knowledge-intensive production.
Regulation Must Manage Real Risk Without Freezing Technological Diffusion
I see regulation as a capital-allocation question as well as a safety question. Excessive restrictions can raise the cost of deployment, slow experimentation, and divert investment toward jurisdictions with more favorable conditions. Insufficient oversight can create real operational, financial, or social harms.
The most economically coherent approach is to distinguish between hypothetical risks and demonstrated risks. Regulation should increasingly focus on measurable harms arising from actual deployments while using established domain-specific regulatory institutions wherever possible.
Autonomous vehicles, for example, can be evaluated within transportation-safety frameworks. AI-assisted drug discovery can be incorporated into existing pharmaceutical and medical regulatory processes. This approach avoids treating artificial intelligence as an entirely separate category of economic activity when established institutions already possess relevant expertise.
There is also a deeper argument about technological progress and safety. More capable systems can potentially reduce certain failure modes through better grounding, verification, reasoning, and reliability. That does not imply that acceleration automatically makes every AI system safe. It does imply that freezing technology at an earlier capability level can preserve the weaknesses of earlier systems rather than eliminate them.
The policy challenge is therefore dynamic. I would not frame the choice as regulation versus innovation. I would frame it as determining which regulatory mechanisms reduce genuine risk while allowing technological learning to continue.
The Monetary and Financial Consequences of the AI Capital Cycle
The magnitude of AI infrastructure investment creates an important connection to macroeconomics. Trillion-dollar capital programs can influence demand for construction, power, semiconductors, networking equipment, engineering services, and financial capital.
That means AI investment can operate through several channels simultaneously. It can increase capital expenditure, create employment, stimulate industrial demand, expand electricity requirements, and alter corporate investment plans. If sustained, these effects can influence productivity growth and the composition of aggregate investment.
At the same time, large infrastructure programs require financing. Interest rates therefore matter. Higher discount rates raise the cost of capital and reduce the present value of long-duration expected cash flows. Companies building infrastructure today against uncertain future demand are particularly sensitive to the relationship between financing costs and expected utilization.
Liquidity and credit conditions can consequently affect the pace of AI deployment even when the underlying technology remains attractive. A technology cycle can be structurally powerful while still moving through conventional financial cycles of expansion, tightening, overinvestment, and consolidation.
This is where I separate the technological thesis from the investment thesis. I can believe that AI will transform productivity while remaining cautious about paying any price for companies exposed to the theme. Strong secular growth does not eliminate valuation risk.
Investment Returns Depend on the Distribution of Economic Value
The central investment question is ultimately not whether AI creates value. It is who captures that value.
Infrastructure providers may benefit from rising demand for compute. Semiconductor companies may capture value through hardware and software ecosystems. Data-center developers may monetize physical capacity. Energy producers and grid infrastructure providers may benefit from rising power requirements. Cloud platforms may monetize compute and distribution. Application companies may capture value by embedding intelligence into customer workflows.
But these layers do not necessarily earn identical returns. Competition can push economic rents from one layer to another. High margins attract capital, and capital attracts competition. Over time, the extraordinary returns of one segment can become the ordinary returns of another.
This is why I focus on marginal returns on invested capital rather than revenue growth alone. The strongest businesses should be able to convert technological demand into durable cash generation without requiring disproportionate amounts of new capital merely to maintain their position.
Capital intensity also changes the valuation framework. Businesses with enormous infrastructure requirements may have substantial growth opportunities but also significant depreciation, financing, and utilization risks. Investors therefore need to distinguish between accounting growth, physical capacity growth, cash-flow growth, and economic returns on capital.
The Difference Between a Technology Revolution and an Investment Bubble
History teaches me that transformative technologies and financial bubbles can coexist. Railroads, telecommunications, electricity, automobiles, semiconductors, and the internet all generated enormous economic changes, while their associated investment cycles also experienced periods of speculation and overcapacity.
The lesson is not to dismiss the technology because valuations can become excessive. The lesson is to analyze the capital cycle separately from the technological cycle.
When expected returns become very high, capital floods into the sector. Infrastructure expands. Competitors emerge. Supply increases. Pricing can eventually decline. The technology may become more economically important even as individual investments generate lower returns.
This distinction is particularly important for AI because falling compute costs can be simultaneously positive for the economy and challenging for certain infrastructure providers. Lower prices can expand demand dramatically while reducing the unit economics of existing capacity. The winners are therefore not necessarily the companies with the most capacity; they are the companies that can preserve attractive returns as capacity expands and prices evolve.
Productivity Is the Ultimate Test
For all the excitement surrounding models, chips, data centers, and agents, I ultimately judge the AI transformation by productivity. The economic case becomes durable when intelligence allows the economy to produce more output with the same or fewer inputs.
Productivity gains can appear in several forms. Workers can complete more tasks per hour. Businesses can reduce development times. Researchers can conduct more experiments. Factories can increase throughput. Logistics networks can optimize routes and inventories. Software teams can build systems faster. Small enterprises can access capabilities previously available only to large organizations.
The aggregate result could be an expansion of potential output. If AI raises productivity across enough industries, the economic effect extends beyond the technology sector itself.
This is why I see the largest opportunity not in selling AI as an isolated product, but in embedding intelligence throughout the existing economy. The addressable market is not merely the software industry. It is the entire production system.
The Strategic Shift From Scarcity of Intelligence to Abundance
The deepest structural implication is that intelligence itself may become increasingly abundant relative to its historical scarcity.
For centuries, societies expanded their productive capacity partly by expanding access to education and skilled human capital. Institutions such as universities effectively scaled the production and transmission of specialized knowledge. AI represents another mechanism for scaling cognitive capability, this time through computational infrastructure.
If intelligence becomes cheaper and more widely accessible, the economic bottleneck can migrate toward other scarce resources: energy, physical infrastructure, high-quality data, organizational attention, trusted institutions, specialized physical assets, and capital.
This is a recurring pattern in economics. When technology reduces the scarcity of one input, the relative importance of complementary scarce inputs increases. Cheap computation can make energy more valuable. Abundant intelligence can make high-quality data more valuable. Automated execution can make strategic judgment more valuable. Rapid experimentation can make access to physical production capacity more valuable.
Understanding these substitutions is essential for both corporate strategy and investing.
What I Watch From Here
I see several variables determining whether the AI infrastructure cycle develops into a durable productivity revolution.
First, I watch the cost and availability of energy. AI demand ultimately depends on the physical ability to supply computation.
Second, I watch utilization and returns on infrastructure. Massive capital expenditure must eventually translate into sufficient economic activity to justify the investment.
Third, I watch the pace of model improvement and hardware efficiency. Better algorithms can increase the amount of useful intelligence produced from the same physical resources.
Fourth, I watch AI diffusion across traditional industries. The strongest evidence of a structural productivity revolution will come from deployment outside the technology sector itself.
Fifth, I watch organizational adaptation. Businesses must redesign processes around AI rather than simply adding AI tools to legacy workflows.
Sixth, I watch labor productivity rather than headline automation. The most important economic outcome may be the increase in output generated by workers whose capabilities are augmented by increasingly capable digital agents.
Finally, I watch capital discipline. Even the most important technological transformation can produce poor investments when expectations outrun cash flows, when capital is allocated indiscriminately, or when technological depreciation occurs faster than anticipated.
The Larger Economic Thesis
I see AI as the beginning of a transition from an economy that primarily consumes computation to one that manufactures intelligence as a productive input. Energy becomes computation; computation becomes tokens; models transform tokens into reasoning and specialized capability; applications convert that capability into economic output.
The significance of this process extends far beyond software. It connects energy infrastructure to semiconductors, semiconductors to data centers, data centers to models, models to agents, agents to businesses, and businesses to aggregate productivity.
The most important economic variable is therefore not the intelligence of an individual model. It is the rate at which increasingly capable intelligence diffuses through the capital stock and labor force.
If that diffusion is rapid and capital is allocated productively, AI can become a general-purpose technology capable of increasing the economy's productive frontier. If deployment is slow, organizational adaptation fails, infrastructure is overbuilt, or valuations become detached from cash flows, the financial outcome can be much less impressive than the technological outcome.
That distinction gives me the framework I need to think about the next cycle. I do not need to choose between believing in AI and worrying about valuations. I can hold both views simultaneously. The technology can be transformative, infrastructure investment can be enormous, productivity can accelerate, and individual securities can still be overpriced.
Ultimately, I view the AI transition as a new capital-allocation regime. The scarce resource is shifting from access to computation toward access to energy, infrastructure, organizational context, proprietary data, and productive applications. The businesses that matter most will be those that convert these inputs into durable economic output. The investors who matter most will be those who distinguish technological progress from financial returns and who understand where, across the five-layer stack, the economic rents are likely to accumulate.
That is the central transformation I see: AI is moving from being a category of software to becoming an industrial system for producing intelligence. Once intelligence becomes infrastructure, the consequences are no longer confined to technology markets. They reach into productivity, capital formation, labor economics, corporate strategy, national competitiveness, financial markets, and the long-run capacity of the global economy to create wealth.