AI’s Real Moat Is Infrastructure, Not Algorithms

The Central Investment Thesis

I view the artificial-intelligence economy as having moved beyond a conventional software cycle and into something closer to an industrial infrastructure cycle. The most important distinction is that AI is no longer merely a contest over who can produce the cleverest model. It is increasingly a contest over who can mobilize sufficient compute, electricity, networking, capital, manufacturing capacity, software, data and distribution to turn intelligence into economically valuable production.

That distinction changes the investment framework. If AI were primarily a software phenomenon, the dominant question would be which model developer captures the most users and monetizes them at the highest margin. If AI is an infrastructure transformation, however, the economic opportunity extends much further down the stack. Semiconductor manufacturing, advanced packaging, memory, networking, cooling, transformers, electricity generation, transmission, data-center construction, cloud capacity and specialized software all become components of the same capital cycle.

The financial evidence increasingly supports this interpretation. NVIDIA generated $215.9 billion of revenue in fiscal 2026, up 65% year over year, while Data Center revenue reached $193.7 billion, up 68%. By the second quarter of fiscal 2027, quarterly revenue had reached $96.2 billion and Data Center revenue $89.0 billion, representing year-over-year growth of 106% and 117%, respectively. r1

Those numbers do not prove that every AI investment will succeed. They do demonstrate that the leading infrastructure supplier is experiencing a scale of demand that is economically different from the early stages of a typical software category. Our analysis therefore starts with the physical economy rather than the narrative economy. The key question is not whether AI is impressive. It is whether the world can build enough productive capacity to exploit it.

From Model Race to Industrial Revolution

I think investors make a conceptual error when they treat AI as a single product category. The more useful analogy is electricity or the internet: foundational infrastructure creates an enormous number of downstream businesses whose eventual economic value can exceed that of the original infrastructure itself.

That is why the distinction between closed models, open-weight models and application-layer systems matters less than it initially appears. Different architectures can coexist because customers have different requirements. Some applications will prioritize maximum frontier capability and therefore favor managed proprietary models. Others will require privacy, sovereignty, customization, predictable economics or local deployment and therefore favor open-weight systems.

The commercial consequence is significant. AI does not require one model to win. It requires a growing number of economically useful workloads to emerge. If developers can reduce the cost of inference, increase reliability, improve reasoning, automate workflows and embed AI into physical machines, the addressable market expands even when individual model prices decline.

This is the same economic mechanism that repeatedly appears in technology transitions: technological progress lowers the cost of a capability, and lower cost expands the number of economically viable uses. The result can be rising total demand despite falling unit economics. I therefore do not regard declining token prices as inherently bearish for the infrastructure layer. If lower inference costs stimulate exponentially greater utilization, infrastructure demand can continue expanding.

Compute Has Become a Production Asset

The most consequential shift is the transformation of compute from a technical input into a productive asset. Historically, investors thought about computers primarily as capital equipment used by employees. AI reverses that relationship. The computer increasingly performs the cognitive work itself.

That means the economics of an AI data center resemble a factory more than an office building. The facility consumes electricity, networking equipment, accelerators, storage and cooling, then produces something economically measurable: predictions, code, images, scientific discoveries, autonomous decisions, customer-service interactions or other forms of machine-generated work.

Once compute becomes a production asset, utilization and return on invested capital become central variables. A GPU sitting idle is depreciating capital. A GPU running productive workloads can generate recurring revenue. The strategic objective therefore becomes maximizing the utilization, throughput and economic value of every unit of installed compute.

This is why the infrastructure architecture matters so much. The accelerator is only one component. Networking determines how effectively thousands of accelerators operate together. Memory determines how efficiently models can be served. Cooling determines usable density. Software determines how effectively hardware is utilized. Power availability determines whether the entire facility can operate at planned capacity.

In my framework, this creates an unusual form of vertical economic interdependence. The value of the chip is partly determined by the data center around it, while the value of the data center is partly determined by the software and models running on its chips. The economic system is therefore increasingly optimized as an integrated platform rather than a collection of independent components.

The Power Constraint Is Becoming an Investment Constraint

The physical bottleneck that I consider most underappreciated is electricity. AI models are digital, but their deployment is profoundly physical. Every inference ultimately requires electrical power, and the rapid scaling of AI data centers is beginning to reshape regional power markets.

The International Energy Agency estimates that global data-center electricity consumption was approximately 485 TWh in 2025 and projects roughly 950 TWh by 2030 in its base case. AI-focused data centers are growing faster than the broader data-center market. The IEA also reports that data-center electricity demand increased 17% in 2025, while electricity use at AI-focused facilities rose approximately 50%. r2

The geographic concentration matters as much as the global total. A 100-megawatt AI facility is not merely a large customer on a national grid; it can be a material load for a regional utility. The IEA notes that hyperscale AI facilities can exceed 100 MW, while the largest facilities under construction or planned can reach into the gigawatt range. r3

This creates a new hierarchy of scarce resources. Historically, the semiconductor industry obsessed over wafer capacity. Today, investors must think about land, power, transmission, transformers, cooling and construction schedules with comparable seriousness.

The investment implication is straightforward: the AI infrastructure cycle cannot be evaluated solely through semiconductor shipment forecasts. A chip can be manufactured faster than a power plant can be permitted, a transmission line constructed or a large data center connected to the grid. The physical bottleneck can therefore migrate from silicon to electricity.

The IEA explicitly identifies transformers, grid connections, advanced chips and other components as potential constraints on the pace of data-center expansion. It also estimates that data-center investment by five major technology companies exceeded $400 billion in 2025 and could increase by another 75% in 2026. r4

The AI Capital Cycle Is Broader Than NVIDIA

NVIDIA is the clearest public-market expression of the AI infrastructure cycle, but I would resist the temptation to equate NVIDIA with the entire opportunity. Its extraordinary financial performance demonstrates the scale of demand flowing through the accelerator layer, yet that very success should encourage investors to search for the constraints surrounding it.

The relevant question becomes: what must exist for another dollar of AI compute to be deployed?

Our investment interpretation is that economic rents should migrate over time. During the initial scarcity phase, suppliers of critical hardware can capture extraordinary margins. As capacity expands, competition tends to move upward toward software, applications and distribution. But this migration is neither instantaneous nor guaranteed. If demand continues to grow faster than supply, infrastructure providers can retain pricing power much longer than conventional technology-cycle models would suggest.

Why the Data-Center Boom Is Not Simply a Technology Story

The data center has become a bridge between technology and macroeconomics. It requires debt financing, equity capital, utility investment, construction labor, semiconductor capacity and energy infrastructure. The AI boom therefore has multiplier effects that reach well beyond technology companies.

The IEA estimates that data centers accounted for approximately 1.5% of global electricity consumption in 2024 and projects their share to approach 3% by 2030 in its base case. In the United States, data centers are expected to account for a substantial portion of electricity-demand growth through the end of the decade. r5

This creates an important macroeconomic feedback loop. AI increases demand for data centers. Data centers increase demand for electricity. Electricity demand increases investment in generation and grids. Construction requires financing and labor. Semiconductor demand supports manufacturing and equipment suppliers. Rising compute availability then stimulates additional AI adoption.

The feedback loop can become self-reinforcing until the marginal return on infrastructure investment falls. That is the critical transition I would monitor. The bear case is not simply that AI stops working. It is that capital expenditure grows faster than monetizable workloads, causing utilization, returns and financing conditions to deteriorate.

That distinction is crucial. An AI infrastructure bust would not require AI to be a technological failure. It would only require investors to build capacity ahead of sustainable economic demand. History provides numerous examples of transformative technologies that generated enormous long-term value while still producing severe intermediate capital losses.

Recursive Improvement Is Better Understood as Engineering

I am similarly cautious about treating recursive self-improvement as a single discontinuous event. The economically relevant process is more incremental. AI systems can generate synthetic data, write software, perform evaluations, search solution spaces, assist researchers and optimize components of future AI systems. Each capability can increase the productivity of the development process without implying an uncontrolled technological singularity.

For investors, the distinction is important because productivity improvements can accelerate both supply and demand. If AI helps engineers build better AI systems, the cost of technological progress declines. That can compress the price of intelligence while simultaneously expanding its consumption.

The same dynamic exists inside applications. Better coding agents may reduce the cost of software development. Better scientific models may reduce the cost of drug discovery. Better autonomous systems may reduce the cost of transportation and industrial labor. The immediate beneficiary may therefore not be the model developer. It may be the company capable of reorganizing its production process around cheaper intelligence.

This is why I believe the largest economic consequences of AI may emerge outside the technology sector. AI should ultimately be judged not by how many benchmark points a model gains, but by how much human and machine productivity it unlocks across the economy.

Open Models Create a Strategic Multiplier

Open-weight AI is economically significant because it reduces the cost of experimentation. A startup does not necessarily need to finance a frontier model from scratch to build a differentiated product. It can adapt an existing model, combine it with proprietary data, optimize inference and construct an application-specific system.

This broadens the innovation base. A closed model can be extraordinarily capable, but an ecosystem of accessible models can generate thousands of specialized experiments simultaneously. From an economic perspective, that resembles a distributed research network.

The strategic implication is that the AI race should not be measured exclusively by the number of frontier laboratories. A country or region can benefit enormously if its companies, universities and startups become exceptionally good at applying available AI technology.

That is why I distinguish between invention and exploitation. The original technology does not have to be invented in the country that ultimately captures the greatest economic benefit. The decisive variable can be how effectively a society commercializes, industrializes and distributes the technology.

China and the Semiconductor Geopolitical Constraint

Technology competition is also becoming inseparable from semiconductor geopolitics. Export controls have already demonstrated that restrictions can alter market access, inventory economics and competitive dynamics.

NVIDIA's filings show the financial consequences directly. The company recorded a $4.5 billion charge in fiscal 2026 related to H20 inventory and purchase obligations after U.S. export restrictions reduced demand. Its filings also state that, under the prevailing regulatory environment, it was effectively prevented from competing broadly in China's data-center computing market, while competitors gained opportunities to build developer and customer ecosystems. r6

I therefore treat semiconductor policy as an operating variable rather than a peripheral political issue. Restrictions can influence product design, inventory planning, geographic revenue exposure, customer relationships and the rate at which competing ecosystems mature.

The strategic problem is asymmetric. Limiting access to advanced compute may constrain a competitor in the short term, but it can simultaneously create incentives to develop alternative domestic supply chains. Over a decade, the competitive effect depends on whether restrictions slow technological diffusion faster than they accelerate substitution.

The Next Scarcity May Be Applications, Not Chips

There is a natural tendency to assume that infrastructure scarcity guarantees infrastructure profitability indefinitely. I do not believe that. Every successful infrastructure boom eventually stimulates capacity expansion, and capacity expansion changes bargaining power.

The most important transition will occur when compute becomes sufficiently abundant that customers stop asking whether they can access GPUs and start asking what economically valuable work they can perform with them.

At that point, application-layer differentiation becomes much more important. The scarce asset becomes proprietary workflow integration, domain-specific data, distribution, trust, customer relationships and organizational capability. The winners may be companies that transform AI from a feature into a new operating model.

That is where I would expect economic rents to migrate. The infrastructure layer can generate extraordinary profits while capacity is scarce, but applications can potentially capture durable value if they fundamentally alter productivity.

The Investment Framework I Would Use

My framework is therefore built around five questions. First, is AI compute demand growing faster than supply? Second, are infrastructure bottlenecks moving downstream from chips into power, networking and construction? Third, are AI applications producing measurable economic returns for customers? Fourth, is capital expenditure being financed against credible future cash flows rather than purely optimistic assumptions? Fifth, is technological progress lowering the cost of intelligence fast enough to expand utilization faster than it destroys pricing?

The fifth question is particularly important. A mature AI economy could simultaneously exhibit falling model prices, falling inference costs and rising total AI expenditure. That is not contradictory. It is precisely what happens when a technology becomes more economically useful.

The current evidence suggests that the infrastructure phase remains powerful. NVIDIA's latest reported quarter produced $89.0 billion of Data Center revenue, while its first-half fiscal 2027 Data Center revenue reached $164.3 billion. The company forecast $108 billion of total revenue for the following quarter, while explicitly assuming no Data Center compute revenue from China in that outlook. r7

Those figures indicate that the demand engine remains exceptionally strong. But strong demand does not eliminate valuation risk, capacity risk or capital-cycle risk. The more AI becomes an industrial system, the more investors must analyze it using industrial disciplines: utilization, depreciation, financing costs, supply-chain resilience, power availability, return on invested capital and replacement cycles.

The Bottom Line for Capital Allocation

I would frame the AI opportunity as a multi-stage economic transformation rather than a single technology trade. The first stage is infrastructure scarcity, where accelerators, networking, memory and data-center capacity command extraordinary demand. The second stage is infrastructure proliferation, where power, construction, financing and regional cloud capacity become increasingly important. The third stage is application diffusion, where enterprises reorganize workflows around abundant machine intelligence. The fourth stage is productivity realization, where the macroeconomic benefits become visible through higher output per worker and lower production costs.

We are already deep into the first stage and rapidly entering the second. The scale of current data-center investment makes that unavoidable. The IEA's latest analysis shows both the magnitude of capital deployment and the physical bottlenecks that could constrain it. r8

My core conclusion is consequently neither that AI is a speculative illusion nor that every AI-related asset deserves a premium. The correct conclusion is more demanding: AI is becoming an industrial economy, and industrial economies produce both enormous fortunes and enormous capital misallocations.

The durable winners will be determined by control of scarce inputs, technological execution, capital discipline and the ability to convert falling computational costs into rising economic output. Chips matter. Models matter. Open ecosystems matter. But the deepest moat is the ability to coordinate the entire system.

That is why I see the AI investment thesis increasingly shifting from “Who has the smartest model?” to “Who can build, finance, power, utilize and monetize the largest productive intelligence system?” That is a far larger question, and it is the question that will ultimately determine whether today's extraordinary AI capital expenditure becomes a durable productivity revolution or merely another spectacular technology investment cycle.

01 Jensen Huang: The Doomer Hoax, Superintelligence is Here, and The Future of AI (ft. President Trump)