Cheap AI is Coming—But Who Owns the Compute Stack?
By Maya Ellison ·
While intelligence may become a utility, the true value lies in owning the persistent infrastructure that runs it.
The idea that machine intelligence is set to become a commodity—a utility delivered at "unbeatable market prices"—is breathtaking, almost utopian. On the podcast Invest Like the Best, an ex-NVIDIA engineer detailed how AI could get 1000x cheaper, moving from expensive consultation services toward true abundance. But when I listen closely beneath the technical jargon of tokens and throughput, what I hear is not a promise of universal benefit, but a blueprint for concentrating unprecedented power in the hands of those who own the compute stacks.
The Shift From Speed to Scope
The core argument presented was that AI’s future isn't about being fast; it's about persistence. Current chatbots prioritize low latency—spitting out answers instantly. But the industry, according to the interviewee, is shifting toward background processing and "proactive intelligence." This means agents running for hours or days to complete long-horizon tasks like deep research (analyzing 10,000+ sources) or building authoritative indexes. The market share prediction—a shift to 90/10 in favor of background work—is a technical observation with profound economic weight: the most valuable AI will be invisible labor, running constantly and autonomously.
This isn't just an architectural change; it’s a fundamental redefinition of work itself. Instead of paying for human hours or even quick API calls, companies are building systems that require continuous compute time—the "long lens view"—to solve verifiable problems like formal math proofs. The goal is to predict the user’s next action and surface it before they ask for it. This sounds efficient, but when intelligence becomes a utility running in the background, who owns the data stream? And more critically, what happens to the human labor that used to structure, verify, or even prompt those tasks?
The New Industrial Landscape of Compute Power
The discussion quickly devolved into the mechanics of power: chips, memory, and infrastructure. We hear about specialized units like Tensor Cores and architectural battles between companies using massive amounts of high-density memory (like Cerebras utilizing an entire wafer for SRAM capacity). This is where the progressive alarm bells should ring loudest. The true bottleneck isn't computation; it’s power and geography.
The industry consensus suggests that building these systems requires coupling chips into small, distributed "mini data centers" to bypass the need for massive, centralized facilities—a strategy the speaker calls the "scavenger approach." This focus on decentralized compute is marketed as resilience, but its economic reality is a race for power access and arbitrage. The ultimate winners will be those who can amass aggregate compute supply from overlooked sources, ensuring that the most valuable resource remains electricity, not silicon.
Abundance or Automation?
The strongest case made for this technology is that it delivers "abundance"—an intelligence commodity at sustainable cost for all industries. This is the seductive narrative: technological progress solves scarcity. But history shows that when a new form of abundance emerges—whether it was steam power or electricity—the immediate effect is not universal prosperity, but massive capital consolidation.
The historical pattern predicts that any dramatic deflationary force in the cost of labor (even digital labor) will be absorbed by capital owners who own the infrastructure of that labor. The focus on solving "verifiable problems" and away from "human taste" confirms this: AI is optimized for measurable, scalable tasks—the backbone of corporate efficiency—not for human flourishing or complex cultural exchange. We are being sold a machine designed to automate not just manual work, but cognitive overhead itself.
The promise of cheap intelligence does not equate to equitable wealth distribution. It simply means that the marginal cost of advanced analysis will approach zero, creating enormous pressure on wages and job security across every white-collar sector. The true economic policy question is not how cheaply we can run these agents, but who gets to govern the output of their tireless operation.