Follow the Megawatts: AI Load and the Return of Electricity-Driven Industrial Location

01 An Old Idea Returns to the Siting Table

For roughly half a century, the standard answer to "where should this factory go?" was some weighted blend of labor cost, market access, tax treatment, and port proximity. Electricity was a line item, not a determinant. There is a well-known exception embedded in industrial history: aluminium smelting, the process of extracting aluminium from its oxide, alumina, generally via the Hall-Héroult process, with alumina extracted from bauxite at an alumina refinery (source: Wikipedia summary — Aluminium smelting). Electrolysis is so electricity-intensive that smelters have always been built where power is cheap and plentiful — hydro-rich regions — rather than where bauxite is mined or where customers live. The industry's geography is, in effect, a map of cheap electrons: the great smelting clusters sit in Quebec, Norway, Iceland, Russia's Siberian hydro belt, and the Pacific Northwest of the United States. Bauxite travels; smelters stay near the dam.

The seed question for this analysis is whether AI data centers, and the energy-hungry manufacturing that may follow them — semiconductor fabs, battery plants, hydrogen electrolysis — are re-creating that pattern at scale. The proposition has a strong a priori logic: an AI training campus is, in energy terms, a smelter. It imports a physical input (electricity, continuously and at high capacity factor) and exports an immaterial product whose transport cost is nearly zero. Just as the smelter's location was solved by moving alumina instead of the plant, the data center's location is solved by moving the work to the power rather than the power to the work — because moving electricity over long distances is expensive and lossy, and interregional transmission capacity is the scarcest commodity in the North American and European grids today.

Journalistic coverage has already begun tracking the community-level texture of this shift — one widely viewed report examines a large data center campus in an American small-town setting through the experience of nearby residents, documenting how abruptly a rural locality's land, noise, and fiscal profile changes when a hyperscaler arrives (source video: More Perfect Union, "I Live 400 Yards From Mark Zuckerberg's Massive Data Center"). That texture matters for politics, but the analytical core is economic: the location decision itself, and what it implies about regional divergence.

02 The Location Decision, Properly Specified

Model the firm's problem. A hyperscale operator chooses a site to minimize the levelized cost of delivered compute, subject to constraints. The components are: the all-in electricity price (energy plus capacity plus transmission charges) over a 15-to-25-year horizon; the speed at which interconnection can be secured; latency to users and peering points (binding for inference workloads, nearly unbinding for large training runs); land and cooling resources; state and local tax treatment; and construction labor availability. Two facts about the current environment transform this objective function. First, interconnection wait times in major markets have lengthened to the point where they, not construction, define project timelines; a queue position is worth real money. Second, large loads can now negotiate directly with generators through power purchase agreements (PPAs), so the effective electricity price is set by bilateral deals, and regions with surplus generation — notably those with large renewable buildouts whose output exceeds local demand — can offer prices and availability that load-constrained regions cannot.

The result is a reversal of the historical relationship between generation and load siting. During most of the twentieth century, generators were built to follow load: utilities sited plants where demand existed or was forecast. In the renewable era the resource map flipped — the best wind and solar is where the people are not — and for twenty years that mismatch produced curtailment and under-built transmission. AI load is the first industrial customer class large and locationally flexible enough to follow the generation map instead. This is why the announced campuses cluster in exactly the regions one would predict from resource abundance and cheap land: the U.S. Midwest and Plains wind belts, the Ohio Valley gas corridor, the desert Southwest, Quebec and the Nordic hydro countries, and, prospectively, oil-producing regions offering stranded gas for behind-the-meter generation.

03 The PPA and the Queue: Mechanisms That Make Location Sticky

Two institutional mechanisms convert the geography of generation into the geography of compute, and both deserve precise treatment. The first is the power purchase agreement. A hyperscaler signs a long-term contract with a wind, solar, hydro, or nuclear generator in a resource-abundant region; the generator's revenue certainty finances new capacity, and the buyer secures a known price and a green attribute. Crucially, the PPA is also a financing instrument for the region itself: it allows developers to build generation ahead of organic local demand, effectively importing demand through the contract. The second is the interconnection queue. A new load or generator seeking to connect to the bulk system in a congested market may wait years for studies and upgrades, and the wait varies enormously by region. For a location decision with a two-to-four-year construction horizon, queue length is not a detail — it is often the binding constraint, and it systematically favors regions with surplus capacity and cooperative transmission planning over constrained coastal markets.

Each mechanism deserves a note on its political economy, because both create constituencies that entrench the pattern once it starts. A PPA binds a global firm and a local generator into a multi-decade relationship; the generator gains an anchor tenant, local governments gain a ratepayer and construction payroll, and the contract's term makes the arrangement hard to unwind even if relative prices shift. The queue, meanwhile, rewards incumbency: the first large loads into a region consume its scarce interconnection headroom, so being early is itself a competitive asset — which explains the observed land-rush behavior around queue positions in favored regions and why established hyperscalers hold options on capacity years before they need it. Together these mechanisms mean the follow-the-power pattern, once established, is sticky: contracts, queue positions, and local political relationships all raise the cost of relocating again, exactly as decades-long hydro contracts kept smelters tied to their watersheds through commodity cycles good and bad.

Together these mechanisms generate a paradoxical effect: the harder it is to build transmission into a demand center, the more valuable it is to site the load where the power already is. In the smelting era, the equivalent logic was explicit — smelters signed decades-long hydro contracts and even invested in the dams — and the contracts outlived several commodity cycles, tying whole towns to the price of one metal. The analogy carries a warning the boosters rarely mention: smelting towns have lived through brutal boom-bust cycles precisely because they concentrated on a single electricity-priced industry. Regions recruiting AI campuses on cheap-power grounds are buying exposure to a different single demand driver — model-training economics — whose cycle length nobody yet knows.

04 Will Manufacturing Follow? The Follower-Industry Question

The seed question extends beyond data centers: will semiconductor fabs, battery gigafactories, and other energy-intensive manufacturing also migrate toward abundant generation? Here the evidence argues for a qualified yes with a strong qualifier — the qualification being that manufacturing location decisions bundle energy price with workforce, logistics, and industrial-policy subsidies, and the weight on energy rises with the process's electricity intensity. Semiconductor fabrication is a strong candidate: fabs run continuous process loads at high power quality requirements, energy is a meaningful share of operating cost, and the political economy of fab subsidies already interacts with power-supply guarantees. Hydrogen electrolysis is the purest case — it is literally a mechanism for converting electricity into a storable commodity, and its economics are almost entirely a function of the electricity price, so green-hydrogen projects behave like the smelters of the coming decades. Batteries and aluminum sit in between.

The plausible equilibrium is not a wholesale relocation of manufacturing but a bifurcation: electricity-intensive, automation-heavy, low-labor-content processes migrate toward power abundance, while labor-intensive assembly and customer-proximate manufacturing stay put. If that is right, the regional consequence is a new division of industrial geography — energy-rich peripheries capture the capital-intensive, high-value-per-worker plants, while legacy industrial regions retain what is left of the labor-weighted segments. That is a recipe for regional divergence sharper than the smokestack-era pattern, because the migrating plants employ fewer people at higher productivity, meaning the receiving regions get tax base and fiscal capacity without proportional population inflows — a different social contract than the mill-town model, as the small-town reporting around new campuses already suggests (source video: More Perfect Union).

A third pattern deserves inclusion because it complicates the clean bifurcation: co-location driven by the data center's own supply chain. Large campuses require continuous supplies of electrical equipment, cooling systems, and construction services over their operating life, not merely at commissioning — and the suppliers of those services face the same energy-and-wage logic in their own location decisions that the campuses do. Where several campuses concentrate, a supplier ecosystem can form around them, adding manufacturing employment of a kind the pure bifurcation model assigns to neither camp. The smelting precedent again instructs: smelting regions did not remain pure metal exporters; alumina refining, carbon-anode production, and fabrication industries clustered around the smelters over decades, converting an energy-driven location decision into a diversified industrial base — though diversification arrived on a timescale longer than any single recruiting government's term, which is a caution against expecting supplier ecosystems on a subsidy-cycle timetable.

Against these pull factors, an honest analysis must register the resistances. Semiconductor fabs chase not only power but ultra-pure water, seismically stable geology, and talent pools that only a few metropolitan regions can supply, so the fab-follows-power effect is bounded. Battery plants sit close to automotive assembly for logistics reasons. And industrial policy — the semiconductor-subsidy programs of the United States, Europe, and Japan — deliberately steers fab location toward national-security priorities rather than energy abundance, injecting a political override into what this article otherwise models as economics. The correct synthesis is that electricity abundance sets the feasible region within which other determinants pick the exact site: it narrows the choice set rather than dictating the choice, but over dozens of projects, a narrowed choice set produces the geographic pattern all the same.

05 Second-Order Effects: Grids, Rates, and Politics

Second-order effects concentrate in three places. First, retail rates and the allocation of system costs. When a large industrial load arrives, the local utility builds infrastructure whose costs are recovered through rates; whether residents benefit or pay depends on rate design and on whether the load's contribution covers its system costs. The community-level reporting around new campuses documents exactly this negotiation — fiscal windfalls, strained local services, and contested cost allocation — playing out town by town (source video: More Perfect Union). Second, generation investment: sustained AI-class demand is now cited by developers and utilities as justification for extending the lives of existing thermal plants, uprating nuclear units, and accelerating new build in host regions, which alters the emissions trajectory of those regions in direction depending on the resource mix. Third, transmission politics: the more load self-sites near generation, the more the national transmission expansion case weakens in some regions while strengthening in others, creating a patchwork that entrenches regional divergence.

Third-order effects are more speculative and should be labeled as scenarios rather than expectations. A region that wins a large share of AI load could see its wholesale prices firm up, eroding the very abundance that attracted the load — a self-limiting dynamic with a smelting precedent, since decades of smelter demand raised the opportunity cost of dedicated hydro in several systems. Alternatively, abundance-chasing loads could drive genuine electrification-led regional booms, with local educational institutions, supplier ecosystems, and heat-reuse industries clustering around campuses, moving these regions from resource peripheries to industrial centers. Which path dominates depends on whether generation capacity expands ahead of the load — that is, on the very interconnection and permitting systems discussed in this series' bottleneck analysis.

06 Counterfactual and Competing Explanations

Discipline requires the counterfactual. Absent AI load growth, would the follow-the-power pattern have emerged from renewables economics alone? Probably in attenuated form — green hydrogen and some smelter-class processes were already moving toward resource abundance before AI — but the pace and scale would be far smaller: AI data centers are unique among candidate loads in both magnitude and location flexibility. So the observed siting wave is not purely an AI phenomenon, but AI is the accelerant, and removing it would slow the regional reallocation substantially. Competing explanations for the siting pattern: Hypothesis one, the pure electricity-price hypothesis — sites follow the cheapest PPA; supported by the observable clustering in wind-rich, gas-rich, and hydro-rich regions. Hypothesis two, the speed hypothesis — sites follow the shortest interconnection queue, and the correlation with abundant regions is incidental (they simply have queue headroom); discriminating evidence would be data centers paying premium prices in fast-queue regions, which the record partly supports. Hypothesis three, the incentives hypothesis — siting follows state and local subsidy packages, and power abundance is coincidental; discriminating evidence would be campuses in expensive-power, high-incentive markets, which do exist. The most defensible synthesis is that all three operate, with weights varying by project: training-dominant campuses weight power and queue; inference-dominant facilities weight latency and thus stay near demand centers; and the subsidy layer distorts at the margin. This synthesis predicts a durable geographic split between training and inference capacity — itself an observable pattern going forward.

07 Scenarios, Indicators, and the Bottom Line

Scenario A — convergence: transmission buildout accelerates and interconnection queues shorten, power-price differentials between regions narrow, and location flexibility loses value; siting reverts toward demand centers. Trigger: measurable queue-time reduction in major markets plus completed long-haul lines into load pockets. Scenario B — persistence: queues stay long, PPAs keep pulling load to abundance, and the training-inference geographic split hardens; regional divergence continues on current trend. Trigger: continued campus announcements in the now-familiar resource regions with unchanged queue statistics. Scenario C — structural divergence: regions with abundant, fast-interconnected generation assemble complete industrial stacks — campuses plus fabs plus electrolysis plus supplier ecosystems — while queue-constrained regions deindustrialize further at the energy-intensive margin, transforming the map of regional fiscal capacity. Trigger: a marquee fab or electrolysis plant siting that is unexplainable except by power abundance, followed by supplier clustering.

Indicators to watch: the geographic distribution of new large-load interconnection agreements by region; average and median interconnection study durations in major markets; the share of announced AI capacity co-located with contracted generation versus relying on market purchases; PPA price spreads between surplus and constrained regions over time; state incentive packages that include power-supply guarantees as a formal term; the number of announced electrolysis and fab projects citing contracted power as a primary siting factor; retail-rate outcomes in host communities after campus energization; and transmission projects completed versus announced into the major AI corridors.

What we know: aluminum smelting demonstrates that sustained, electricity-intensive industrial location follows power abundance, and AI data centers are the first major new load class since then with comparable energy intensity and far greater locational flexibility; interconnection queue length and PPA pricing are now first-order siting variables; and campuses are observable today in resource-abundant regions. What we think we know: the training-inference split is producing a durable two-track geography, and electricity-intensive manufacturing will follow campuses at least partially. What we do not know: whether transmission buildout will erode the regional price differentials driving the pattern, and whether host regions are importing stability or a new single-industry boom-bust cycle. The bottom line: the smart-money phrase of the smokestack era was "follow the water." The emerging equivalent is "follow the megawatts" — and the smelting precedent suggests both how real that pull can be, and how much a region sacrifices when it becomes a company town for one kind of demand.