Follow the Money: Who Profits When America Stops Building AI Infrastructure

Edge Computing Won't Save You From China, But It Will Sell Software

There is an old rule in epistemology that deserves wider use in technology policy: when a man tells you a category of thing is obsolete, ask what he sells. The answer does not settle whether he is right. Salesmen are sometimes right. But it tells you where to apply pressure, because an argument that conveniently terminates in the arguer’s invoice should be examined at the joints rather than swallowed whole.

A genre of essay now circulates widely in conservative media making roughly the following claim. Data centers, the argument runs, are relics of an older computing era, built to run sluggish corporate software for batch reports and billing systems. Artificial intelligence, by contrast, is about fast decisions made where the data lives, on the battlefield, on the oil rig, in the vehicle. Centralized facilities can never deliver that speed. Therefore the hundreds of billions of dollars now flowing into American AI infrastructure are being poured into stranded assets, and the citizens fighting data center construction in their counties are not NIMBYs but prophets. The essays urging this view are frequently written by people whose companies sell the alternative, distributed edge software that, by remarkable coincidence, renders the entire data center industry unnecessary.

I want to take this argument seriously, because it contains a grain of truth, and arguments with grains of truth are more dangerous than arguments with none. The grain is this: a real and growing class of AI work genuinely belongs at the edge. When a drone must decide in milliseconds whether the object below is a tank or a tractor, it cannot wait for a round trip to a facility in Virginia. The industry knows this. Edge inference is a thriving field precisely because everyone serious already concedes the point. So far, the doomsayers and the builders agree.

The error enters through an equivocation, and it is worth walking through slowly, because once you see it the whole edifice collapses. The word “AI” is doing double duty. It names two profoundly different activities. The first is training, the process by which a frontier model is built. The second is inference, the process by which a finished model is used. The obsolescence argument quietly assumes that because inference can be distributed, training can be too. It cannot, and the reason is not a matter of opinion or vendor preference but of physics and arithmetic.

Training a frontier model requires tens of thousands of specialized chips wired together with high-bandwidth interconnects, computing synchronously on a single enormous problem. The chips must exchange results constantly, billions of times per second, which is why they must sit meters apart rather than miles. You cannot train a frontier model across a mesh of phones and desktop boxes any more than you can build an aircraft carrier in 10,000 backyard workshops and weld the pieces together at sea. The thing being built is unitary. Its construction is unitary. And here is the part the obsolescence school never mentions: every smart edge device runs a model that was trained somewhere first. The slick inference engine on the drone is downstream of a training cluster. Edge computing does not replace data centers. It depends on them, the way a paperback depends on a printing press. The relationship was never edge versus data centers. It was always edge plus data centers, a division of labor as natural as the one between the factory and the showroom.

Consider the favorite example of the obsolescence school, the self-driving car. The story goes that autonomous vehicles are failing because they must stream sensor data to a central hub and wait for instructions, a latency disaster. But this describes no autonomous vehicle actually on American roads. Waymo’s cars do their perception and planning onboard, at the edge, in real time, precisely because the engineers understood latency from the start. And the models making those onboard decisions were trained on vast centralized clusters, refined, and then deployed to the fleet. The example meant to prove that data centers are obsolete is in fact a textbook illustration of the complementary architecture, centralized training feeding distributed inference. When your marquee case proves your opponent’s thesis, the problem is not the marketing. The problem is the thesis.

The speed claims deserve the same scrutiny. The obsolescence essayists boast of software that runs 1,000 or even 1,000,000 times faster by eliminating input/output bottlenecks, the wait states that leave a conventional server’s processor idle 95% of the time. Grant every word of it. The claim is about I/O-bound workloads, database queries, billing runs, record searches, the very legacy corporate computing the essayists correctly describe as the old world. But training a neural network is not I/O-bound. It is compute-bound, limited by raw mathematical operations, trillions upon trillions of them. No amount of I/O optimization conjures the floating-point operations needed to train a GPT-class model, just as no amount of streamlining the loading dock makes the factory floor produce more steel. The advertised speedup, however real, applies to a workload class that is not the one filling an AI data center. The argument refutes a machine that the buildout was never constructing.

Now ask the question I posed at the outset. Who benefits from the claim, framed as it is? The vendor of edge software benefits, obviously, since every county commission that blocks a data center is a future customer persuaded that the alternative is not merely viable but inevitable. But the beneficiary class is larger than any one firm. An entire ecosystem now profits from data center fear. Activist organizations raise money on it. Foreign-funded NGOs, which I have documented elsewhere, organize local opposition with it. Writers build subscriber bases on it. Consultants bill hours explaining it. None of this makes every claim in the genre false. Data centers do use water. They do draw power. But a claim can be locally true and globally misleading, and the framing here, that these facts add up to obsolescence and that opposing construction is therefore costless, is wildly misleading. The water story is the cleanest illustration: modern facilities increasingly use closed-loop or reclaimed-water cooling, and the aggregate draw is a rounding error beside agriculture. The Chinese have even sunk a seawater-cooled facility off Shanghai using zero fresh water, demonstrating that cooling is a solvable engineering problem rather than a civilizational limit. The honest framing is that siting matters. The dishonest framing is that the technology is inherently extractive.

The electricity story is, if anything, the reverse of the popular telling. The fear is that data centers will raise your power bill. But consider how utility economics actually work. The grid is a fixed-cost machine. The poles, wires, substations, and plants must be paid for whether demand grows or shrinks, and those costs are spread across every kilowatt-hour sold. After 2 decades of flat demand, in which ratepayers shouldered the full fixed cost of an aging system, large new industrial customers arrive willing to buy enormous quantities of power around the clock. More kilowatt-hours sold against the same fixed base means a lower cost per unit for everyone, which is why analysts at conservative institutions like the Heritage Foundation have long argued that demand growth, properly priced, is a ratepayer’s friend rather than his enemy. The economics resemble a church that has been splitting the building’s mortgage among 40 families and suddenly welcomes 100 more. The mortgage does not grow. The shares shrink. Moreover, the hyperscalers are not even waiting for the grid. Most now plan to build or contract their own generation, gas turbines, small modular nuclear, dedicated solar, behind-the-meter plants that never touch the residential system at all. A customer who brings his own power plant is not a burden on your bill. He is a co-investor in American generation capacity that the country needs regardless.

And the country does need it regardless, which brings us to the stakes the obsolescence school never weighs. The race that matters is not between two architectures. It is between two nations. China is building training compute at ferocious speed, backed by an electricity buildout that already dwarfs ours, and the layer of the AI stack that decides who leads, frontier model training, cannot be won at the edge by anyone, American or Chinese. It can only be won in large centralized clusters. If the United States internalized the doctrine that data centers are obsolete, if county by county we blocked construction on the advice of essayists selling the alternative, we would not be choosing a cleverer architecture. We would be unilaterally conceding the decisive layer of the most important technological competition since the Manhattan Project, while telling ourselves we had outsmarted everyone. Beijing could not purchase better propaganda. It is getting it free, dressed in the language of property values and farmland.

So let us keep two thoughts in our heads at once, since the truth requires both. Edge computing is real, valuable, and growing, and the men who sell it deserve their market. And the claim that its rise renders centralized AI infrastructure obsolete is false, refuted by the dependence of every edge device on centrally trained models, by the compute-bound nature of training itself, and by the capital now flooding into the very facilities the theory declared stranded. When someone tells you America should stop building the infrastructure of the AI age, check what he sells, check who funds the fear, and then check the only scoreboard that matters, the one with China on it.

If you enjoy my work, please subscribe https://x.com/amuse/creator-subscriptions/subscribe

---

Alexander Muse is a Fellow at the John Milton Freedom Foundation and publishes daily political analysis at amuseonx.com. Primary sources cited in this piece are linked inline; campaign finance figures are drawn from FEC filings, polling data from publicly released crosstabs, and legal claims from filed pleadings. Corrections are posted to the original URL with a dated changelog. Readers who identify errors are invited to contact the author directly.