When I look at the economics of frontier artificial intelligence, I see a question that extends far beyond model architecture: where, exactly, is all the capital going? If a comparatively recognizable transformer-based system can approach the performance of the most advanced frontier models while competing impressively on the cost frontier, then the central economic problem is no longer simply whether AI laboratories can invent a fundamentally new architecture. It is whether enormous expenditures are producing sufficiently large economic advantages to justify their scale.
That distinction matters because technological revolutions are ultimately governed by economics. A technically impressive system is not automatically a commercially superior system. The decisive variables include performance, inference cost, training cost, capital intensity, energy consumption, latency, reliability, scalability, availability of compute, and the value generated by each incremental improvement. AI is therefore becoming an increasingly useful case study in the economics of innovation: the underlying technology may remain relatively recognizable even as the industrial system surrounding it becomes extraordinarily sophisticated and capital intensive.
The Persistence of the Transformer Is an Economic Signal
I find the continued relevance of the transformer particularly important because it challenges a common assumption about technological progress: that each generation of superior systems must be built upon an entirely new conceptual foundation. That is not necessarily how technological revolutions work. Mature architectures can continue improving through incremental innovations in efficiency, routing, attention mechanisms, training methods, data utilization, and systems engineering.
A transformer-like architecture remains economically significant precisely because it is already supported by an enormous ecosystem of knowledge, software, hardware, research, engineering talent, and accumulated experimentation. That ecosystem creates path dependence. Once an architecture becomes sufficiently useful, every improvement to the surrounding stack increases the value of continuing to exploit it.
From an investment perspective, this is analogous to the way established industrial technologies can remain dominant long after researchers have identified theoretically different alternatives. The incumbent technology possesses complementary assets. It has supply chains, specialized equipment, experienced labor, standardized processes, intellectual capital, and enormous installed infrastructure. Replacing the core technology requires more than demonstrating that an alternative works. The alternative must overcome the economic value of the existing ecosystem.
That is why I would resist interpreting architectural familiarity as technological stagnation. A recognizable architecture can support extraordinary innovation if the optimization space around it remains large.
Performance Alone Is the Wrong Metric
The most important analytical shift in AI economics is to stop treating benchmark performance as the complete measure of technological superiority. I care much more about the relationship between performance and cost.
Suppose two models produce broadly comparable results. If one requires dramatically more computation, energy, specialized hardware, or capital expenditure to deliver those results, its economic position may be weaker even if it wins a benchmark by a modest margin. Conversely, a model that achieves nearly equivalent capability at substantially lower cost can be enormously disruptive.
This is the logic of the cost frontier. Innovation becomes economically powerful when it changes the amount of output obtainable from a given quantity of capital and compute. In that sense, AI development increasingly resembles a productivity problem. The relevant question becomes: how much intelligence can I purchase with a dollar of compute?
That metric has profound implications. If model quality continues to improve while unit costs decline, demand can expand rapidly because previously uneconomic applications become viable. The resulting effect is not merely technological improvement; it is a potential expansion of the addressable market.
This is the same broad mechanism that has driven technological diffusion throughout economic history. Falling costs transform luxuries into commodities, commodities into infrastructure, and infrastructure into platforms upon which entirely new industries are built.
Why Spend Billions If the Architecture Is Familiar?
This is where the economics becomes most interesting. If the core architecture is recognizable, enormous AI expenditures must be explained by everything surrounding that architecture.
I would break the investment problem into several broad categories: compute infrastructure, semiconductor capacity, data-center construction, networking, energy, model training, experimentation, talent, software systems, inference infrastructure, and the organizational capacity required to operate all of these components simultaneously.
The architecture may be relatively familiar while the industrial system required to push it toward the frontier is anything but simple.
There is an important distinction between inventing an algorithm and industrializing an algorithm. Once an architecture becomes powerful enough, the bottleneck can migrate from theoretical computer science toward systems engineering and capital deployment. Building larger and more reliable clusters, coordinating vast quantities of computation, reducing communication overhead, improving utilization, managing power constraints, and translating research improvements into production systems can require extraordinary resources.
This helps explain why two organizations can use fundamentally similar architectural concepts while exhibiting radically different costs and capabilities. The competitive advantage may reside less in the abstract architecture than in execution across the entire technology stack.
Capital Intensity Changes the Competitive Structure of AI
AI is becoming an unusually important example of a technology in which software economics and industrial economics collide. Traditional software businesses often benefit from extremely low marginal costs and relatively modest physical capital requirements. Frontier AI introduces a much heavier infrastructure component.
Training and operating advanced models can require enormous concentrations of compute. Compute requires semiconductors. Semiconductors require fabrication capacity and advanced packaging. Data centers require land, cooling, electrical infrastructure, networking, and power generation. All of this creates a capital-intensive production function behind what ultimately appears to the customer as software.
That changes competitive dynamics.
High fixed costs can create barriers to entry and favor companies capable of raising and deploying large amounts of capital. But high fixed costs do not automatically produce durable monopolies. If technological progress rapidly improves efficiency, a smaller competitor can potentially challenge an incumbent by achieving similar output with dramatically less capital or compute.
This creates a fascinating tension between economies of scale and technological substitution. Scale can be an advantage when larger systems produce better models, but efficiency can undermine that advantage when smaller systems approach the same performance at lower cost.
Efficiency Can Be More Disruptive Than Raw Scale
I see one of the most important investment lessons in this dynamic: technological disruption does not always come from the company spending the most money. Sometimes it comes from the company discovering how to obtain comparable economic output with less money.
That is a classic productivity story.
If one generation of AI requires massive resources to achieve a particular level of capability, while a subsequent approach reaches nearly the same level using substantially fewer resources, the economic value of the underlying technology has changed. The innovation effectively lowers the price of intelligence.
Lower prices can have enormous second-order effects. Businesses that previously could not justify AI automation may adopt it. Developers may embed models into more applications. Consumers may use AI for tasks that were previously too expensive. Enterprises may redesign workflows around machine intelligence rather than simply adding AI as an incremental feature.
In other words, efficiency can expand demand enough to compensate for lower revenue per unit. This is one reason I would pay close attention to inference economics rather than focusing exclusively on the prestige of frontier training runs.
The Economics of Diminishing Returns
There is another issue I cannot ignore: the possibility of diminishing returns to additional expenditure.
When a technology is immature, relatively small investments can generate enormous breakthroughs. As the frontier advances, however, extracting another increment of capability may require exponentially more resources. The economic question then becomes whether the marginal value of additional intelligence rises quickly enough to justify the marginal cost.
This is fundamentally a marginal analysis problem. A laboratory should not be evaluated simply by the absolute amount it spends. I need to ask what incremental capability that expenditure buys, how much customers value that capability, and whether competitors can obtain similar results through more efficient methods.
If the answer to the last question increasingly becomes yes, the strategic importance of capital efficiency rises sharply.
That does not mean frontier spending becomes irrational. A tiny performance improvement can have enormous economic value if it unlocks a new category of applications, improves autonomous systems, strengthens enterprise adoption, or produces a durable technological lead. But the burden of proof becomes economic rather than purely technical.
AI May Be Entering an Industrial Optimization Phase
I think the most useful way to interpret this environment is as a transition from pure architectural discovery toward industrial optimization.
The early phase of a technology is dominated by questions such as: What works? What architecture is viable? Can the system scale? Can the underlying problem be solved at all?
Once those questions have largely been answered, the questions change. How cheaply can it work? How reliably? How quickly? With how little energy? Using how little memory? At what latency? With what utilization rate? How effectively can hardware and software be co-designed?
Those questions may sound less glamorous than discovering an entirely new architecture, but they are often where enormous economic value is created.
The history of computing repeatedly demonstrates this pattern. Commercial dominance frequently emerges from incremental improvements in manufacturing, efficiency, standardization, integration, distribution, and cost rather than from a single revolutionary conceptual breakthrough.
What This Means for AI Infrastructure
If model architectures converge toward broadly recognizable designs, infrastructure can become an even more important competitive battleground.
I would therefore think about AI not simply as a software industry but as an integrated stack. At the bottom are energy and physical infrastructure. Above that sit semiconductor manufacturing, accelerators, memory, networking, and data centers. Above those are distributed computing systems and model-training infrastructure. Above that sit model architectures, post-training techniques, applications, and distribution.
Value can migrate between these layers.
If models become commoditized, infrastructure providers may capture more value. If compute becomes abundant and inexpensive, model developers may capture more value. If model capabilities converge, distribution and proprietary data may become more important. If inference costs collapse, application-layer businesses may capture the largest economic gains because they can incorporate intelligence into products at previously impossible price points.
For investors, this means I would avoid assuming that the most technologically visible layer will necessarily capture the greatest long-term economic surplus.
The Investment Lesson: Follow the Production Function
When I evaluate an emerging technology, I want to understand its production function. What inputs generate the output? Which inputs are scarce? Which are becoming cheaper? Where are economies of scale strongest? Where can substitution occur?
For AI, this means monitoring the relationship among compute, capital, energy, labor, software efficiency, and model capability.
If capability improves mainly by adding more compute, capital intensity remains central. If capability increasingly improves through algorithmic efficiency, the competitive landscape becomes more fluid. If inference costs fall dramatically, application adoption could accelerate. If energy becomes the binding constraint, access to power becomes a strategic asset. If semiconductor supply becomes the bottleneck, hardware availability can determine the pace of expansion.
This framework is more useful to me than simply asking which model is currently ranked first. Rankings describe the frontier at a point in time. Production economics helps explain who can profit from moving the frontier.
The Broader Economic Principle: Innovation Is Often Combinatorial
One of the deepest lessons here is that innovation does not require abandoning the past. Technological progress is often cumulative and combinatorial. Researchers can take an established architecture and improve individual components, remove inefficiencies, alter routing, optimize computation, redesign attention mechanisms, and integrate better hardware and software.
The resulting system can be dramatically more capable without looking revolutionary from a distance.
This is economically significant because cumulative innovation tends to reward ecosystems. Every improvement increases the return on previous investments in tooling, expertise, infrastructure, and knowledge. That creates increasing returns to accumulated technological capability.
It also means that investors should be cautious about narratives built around technological discontinuity. A new architecture may ultimately win, but the mere existence of a novel technical concept does not establish commercial superiority. The economic hurdle is much higher: the new approach must outperform an incumbent ecosystem that has years of accumulated optimization behind it.
Why This Matters for the Future of AI Valuations
The distinction between capability and economics becomes particularly important when evaluating valuations.
High expectations can be rational if AI produces extraordinary productivity gains across the economy. But those expectations require assumptions about the distribution of economic value. The existence of powerful models does not tell me which companies will capture the resulting surplus.
I need to distinguish between technological value and shareholder value.
A technology can be enormously valuable while generating mediocre returns for investors if competition drives prices toward marginal cost. Conversely, a company can create substantial shareholder value by controlling a scarce complementary asset even if its underlying technology becomes increasingly standardized.
This is a familiar principle in economics. Consumer surplus and producer surplus are not the same thing. A transformative technology can deliver enormous benefits to consumers while the firms supplying it compete those benefits away.
For AI investors, therefore, the crucial questions include pricing power, switching costs, proprietary distribution, scarcity of infrastructure, capital efficiency, utilization rates, and the durability of technological advantages.
A Framework for Thinking About the AI Capital Cycle
I would frame the emerging AI capital cycle around four questions.
First, how much incremental capability does each additional dollar of investment purchase?
Second, how quickly are the costs of producing and serving intelligence falling?
Third, where in the AI value chain is scarcity actually occurring?
Fourth, who possesses durable pricing power once the technology becomes widely available?
These questions allow me to separate a genuine productivity revolution from a simple capital-spending race.
A spending race can create spectacular infrastructure buildouts without guaranteeing equivalent returns. A productivity revolution is different: it lowers the cost of producing valuable economic output and allows businesses throughout the economy to reorganize around that lower cost.
The ultimate test of AI investment, in my view, will therefore be measured not by the size of training clusters or the magnitude of capital announcements, but by the economic output generated per unit of capital, compute, and energy.
The Strategic Bottom Line
I take considerable significance from the fact that a recognizable transformer-based approach can remain highly competitive near the frontier. It suggests that the future of AI may be less about waiting for a single miraculous architectural replacement and more about relentlessly improving the economics of an already powerful computational paradigm.
That changes how I think about the industry.
I see a contest between scale and efficiency, between capital intensity and technological productivity, and between frontier capability and commercial usefulness. The largest laboratories may continue spending extraordinary sums because frontier development requires enormous infrastructure and because even small capability advantages can have strategic value. But their spending should not be interpreted as proof that comparable outcomes necessarily require comparable resources.
The more important possibility is that AI is entering an era in which efficiency itself becomes a source of disruption. If increasingly capable intelligence can be delivered with progressively less computation and capital, the economic consequences could be much larger than the model leaderboard suggests.
For me, that is the central investment thesis. The defining question of the next stage of artificial intelligence may not be who can spend the most to build the largest system. It may be who can convert capital, compute, energy, and engineering effort into useful intelligence most efficiently.
And if the underlying architecture remains recognizable while the economics improve dramatically, that would not be evidence that AI innovation has stalled. It could be evidence that the technology is entering the more consequential phase of technological development: the phase in which engineering optimization turns an extraordinary invention into economically ubiquitous infrastructure.