Few Labs Poised to Control Global Compute and Economic Future
By Josie Calloway ·
AI compute capacity is rapidly centralizing, creating a new economic structure dominated by handful of frontier labs.
The most striking takeaway from the recent episode of Dwarkesh Patel is not the sheer scale of the money involved—the projected $10 trillion in infrastructure CAPEX alone—but the degree to which this wealth and power are predicted to concentrate within just a handful of private labs. The conversation painted a picture of an economic future where global compute capacity, and by extension, much of human labor's value capture, will be dominated by OpenAI and Anthropic.
On the podcast, Dylan Patel argued that the global economy is rapidly becoming a function of "lab economics" and compute capacity. He cited projections showing CAPEX spending ballooning to over $2 trillion by 2028, with labs’ spending moving toward potentially trillions annually. The financial shift supporting this growth is equally alarming: major labs are transitioning from venture-funded losses to generating revenue, which has allowed margins to "skyrocket." For instance, Anthropic's revenue was noted as high as $50 million per megawatt, vastly exceeding the base cost of compute.
The core mechanism driving this centralization, according to the discussion, is the exponential growth rate of frontier labs’ compute, projected to triple every year. This trajectory suggests that by 2028, these few players could control most usable flops in the world. While the market structure appears profitable—with revenue per gigawatt generating massive returns far exceeding initial fab CAPEX—the underlying tension is one of resource hoarding and systemic risk.
The Illusion of Market Efficiency
The prevailing narrative presented on Dwarkesh Patel is that this concentration is simply a function of superior profit margins, arguing that the huge discrepancy between the cost to build compute (massive CAPEX) and its eventual revenue generation creates an irresistible market incentive for the dominant players. They argue that value capture shifts away from model developers toward end-users like Meta or Jane Street who monetize the models.
This is where the argument falters in a critical way. It assumes that capital, once mobilized by private labs, will flow smoothly and sustainably through global infrastructure without major friction. The discussion rightly pointed out that the primary physical bottleneck isn't capacity itself, but specialized components like ASML mirrors—a choke point controlled by few suppliers. Furthermore, it highlighted the massive macroeconomic constraint: this required $11T CAPEX must be financed either by debt or consumer spending, leading to a "crowding out effect" and dramatically increased interest rates across all sectors. The assumption that limitless capital will meet exponential demand ignores the historical pattern of resource scarcity and economic contraction that accompanies such rapid asset inflation.
When Infrastructure Becomes the Public Health Crisis
When I hear about massive, centralized infrastructure build-outs—whether it’s a new hospital wing or an AI data center cluster—I always see the same failure pattern: the system is designed for profit extraction before human need. The podcast touched on this when discussing regulatory risks, noting that government action could slow down labs more than open-source models. This isn't merely a technical detail; it’s a profound warning about who controls essential services.
In public health, we know what happens when critical infrastructure—the supply of skilled nurses, the capacity of ICU beds, or reliable power to life support systems—becomes centralized and profit-driven. The current AI model mirrors this vulnerability. By concentrating compute at the hands of a few hyperscalers, the system creates an oligopoly over a resource that will soon underpin everything from medical diagnostics to educational tools. This centralization is not merely economic; it is structural power consolidation, making the entire public sphere dependent on the goodwill and internal R&D priorities of private corporations.
The Long Shadow of Debt and Centralized Power
The deepest systemic threat revealed by these projections ties back to a fundamental truth: every technological boom creates an accompanying debt bubble that favors those who own the foundational assets. Historically, when new energy sources or communication methods were introduced (the telegraph, electricity, oil), the infrastructure required was so vast it necessitated massive government borrowing and corporate consolidation.
The current model—where internal value generated from AI research is deemed far more profitable than external revenue generation—is simply the latest iteration of this pattern: private entities are building essential global infrastructure under a shell of "market efficiency," while simultaneously accumulating colossal debt that will inevitably impact every consumer, every small business, and ultimately, the public purse. The sheer scale of required investment means that when the inevitable economic correction hits, the burden of servicing trillions in corporate debt will fall disproportionately on the general taxpayer.
The current trajectory is not a market revolution; it is a resource transfer. The immense concentration of compute power within a few labs guarantees that any future advancement—be it medical or educational—will first be filtered through and priced by these private monopolies, solidifying an unprecedented level of corporate control over foundational human capability.