A.I. Data Centers and the Future of Distributed Compute
Zoning hearings in Tucson, Reno and smaller towns across the American West now fill up with residents holding groundwater charts and utility bills. At least 75 U.S. projects worth roughly $130 billion were blocked or delayed in the first quarter of 2026, a total nearly equal to the whole of the previous year. Active opposition groups climbed from 396 to 833 across 49 states over those same three months.
Reporting on the revolt usually settles on the physical inputs like megawatts, acre-feet, transmission corridors and generator noise. Most of those objections survive scrutiny. Sitting underneath them is a question about ownership that gets much less airtime. Who owns the compute, and who gets to participate in the economy it creates?
Costs land inside county lines
Every physical burden an A.I. campus imposes can be mapped inside a county line, starting with
Electricity impacts are wider, yet still bounded by the reach of the regional grid operator. Electricity bills for households hundreds of miles from any server hall carry a similar imprint. Data center demand drove 63 percent of one year’s capacity price increase across the 13-state PJM grid region, or about $9.3 billion recovered from ratepayers, including households and small businesses.
Whatever these buildings produce departs the county almost immediately after it is computed. Compute belongs to a handful of firms; the weights trained on it stay proprietary; and the revenue books in Seattle, Redmond or Menlo Park. A county keeps the substation, the truck traffic and a property-tax line frequently discounted in advance by state incentives.
Loosening the co-location rule
The physical burdens stay unusually local. A county 30 miles up the road feels very little of them because the aquifer draw, the substation and the truck traffic all land on whoever sits closest. Piling more load onto those same few spots looks perverse until you examine what forces the concentration in the first place.
These problems are not true of data centers as such. They become serious at a certain scale. A modest site rarely drains an aquifer, moves a regional power market or fills a zoning hall. So why squeeze as many chips as possible into the same spot, rather than spread them across smaller sites?
Several forces pull operators toward enormous single sites, from power contracts to construction economics. But those constraints can be overcome. They mainly raise costs, and paying more to spread the chips can look like a reasonable trade once local objections, grid limits and permitting are counted. The binding constraint on the largest sites is the training of new models. That one is closer to binary: either there is enough capacity to train a model at the next scale, or there isn’t.
Training a frontier model has typically required thousands of high-end GPUs packed into a single facility. The algorithms that train these models depend on constant synchronization, with chips exchanging updates at every training step. That traffic demands ultra-low latency and massive bandwidth, which traditionally exist only inside one tightly networked building. That is why A.I. data centers have grown so large, and why they are now pressing against the physical limits of how many chips can sit in one place.
That requirement is starting to loosen. In 2023, Google DeepMind researchers showed that machines in separate locations could train a model just as well while exchanging roughly 500 times less information with each other.
The demonstrations followed quickly. In 2024, a startup called Prime Intellect trained a model across three continents at once. This March, a project called Templar went considerably further, training a 72-billion-parameter model over ordinary internet connections, with more than 70 independent hardware contributors and no central cluster.
On a standard suite of benchmarks, it came close to Meta’s Llama 2 70B, a centrally trained baseline of similar size. That is evidence that aggressive communication reduction does not have to come at the expense of model quality. Coordination survived peers joining and dropping out mid-run, while gradient compression kept the hardware busy instead of idling on bandwidth.
Work has continued since then, including an 8.6-billion-parameter model trained across 330 contributors on consumer-grade cards. Every one of these runs lands in public with weights and evaluations attached, making the rate of progress something outsiders can verify for themselves.
Inference has always been distributed with far less difficulty, and it accounts for a growing share of total compute demand. The asymmetry is severe. Meta’s 405-billion-parameter Llama 3.1 can be served comfortably from a single eight-GPU server, while building the model took upward of 16,000 of the same chips wired together. Those of us working on distributed training keep pushing outward the boundary of what genuinely needs co-location.
The hyperscalers themselves have reached much the same conclusion about the limits of a single site. Nvidia introduced Spectrum-XGS Ethernet in August 2025 with a candid rationale: individual facilities are running into the ceiling of available power and capacity. Its answer is to stitch sites together across long distances. The physical limits of concentration, in other words, are becoming something operators can design around.
Suppliers beyond the hyperscalers
Distributed methods change who can sell compute into a serious training run. Regional operators, university clusters, mid-sized co-location providers and independent hardware owners all become plausible participants. Researchers at smaller labs can gain access without raising billions first.
For a county in northern Nevada or eastern Oregon, a 30-megawatt facility under local or regional ownership carries a different economic profile from a 500-megawatt campus whose entire output is committed elsewhere before the concrete cures. Local ownership gives a county a stake in the market it supplies. Absentee ownership can leave it hosting a substation for someone else’s business. As compute becomes economic infrastructure, its physical location is only part of the equation. The larger questions are who supplies it, who controls access and who ultimately captures the value built on top of it.
Hybrid is the honest forecast
Hyperscale sites will keep their place. The tightest synchronous workloads still reward co-location, and Nvidia is right about the underlying physics. Everything without that requirement has good reason to spread across hardware that already sits nationwide, and the size of that category grows with every publication cycle.
Treating hyperscale as the sole viable design imposes an unnecessary trade-off on communities from Tucson to Tooele. Some of the concentration reflects genuine engineering constraints. A great deal of it now reflects habit, capital structure and procurement patterns set before the research caught up.
What policy can actually do
Sensible policy would start by making large loads pay for the grid upgrades they require, alongside minimum contract terms and exit fees so that ratepayers are not left holding assets built for a tenant who departed. Oregon has already created a dedicated data center rate class, and the mechanics for other states are well documented.
Procurement can pull in the same direction. Public agencies and enterprises buying A.I. capacity can favor distributed and locally owned supply where the workload permits it, which turns a technical option into an economic one.
Spreading supply across many providers widens who can access compute, alongside who can sell it. Distributed infrastructure and open access are separate things, though the first loosens the grip of whoever currently stands between a research group and a working cluster.
Open access to compute shapes what A.I. eventually delivers in medicine, materials science and public administration. Those gains reach the public unevenly if the machinery producing them sits inside a handful of companies, and that concern belongs in land-use debates as much as
The people filling council chambers in Tucson and Reno have priced the offer in front of them and judged it poor. Nothing in this verdict is anti-technology, and a better-designed build-out would put a different offer on the table.
