One rack. One hundred and thirty kilowatts.
Four hundred of those racks and you are running a campus that pulls more electricity than the city it borders, employs fewer people than a mid-sized car dealership, and pays property tax against an assessed value the county assessor is still learning how to calculate. The water bill is metered in millions of gallons. The noise is a constant low-frequency hum that carries further at night. The permanent jobs number โ the one every economic development office puts in the press release โ rounds to something you can count on two hands.
That is the machine Trump pointed at when he floated the idea that data centers should pay a dividend back to their host communities.
Here is everything the report actually contained. One: the proposal exists as a rhetorical position. Two: the word used was dividend. Three: adoption was framed as contingent on companies. Four: the sourcing was a crypto outlet reprinting a policy item with no first-party interview. No publication date. No direct quotation. No mechanism. No operator response.
Four information points. One fact. When a document has more adjectives than variables, you are not reading a policy. You are reading a trial balloon with a headline attached.
I have run this audit before. In late 2017 I spent three weeks inside the Tezos smart contracts while my classmates bought the token on narrative. I found a delegation-logic race condition, filed a GitHub issue with the reproduction path, and sold my allocation on mainnet launch for a $4,200 gain while the people who read the whitepaper held through the unwind. The lesson was not that I was smart. The lesson was that the gap between what a document says and what a document specifies is where all the money lives. A dividend with no payer, no base, and no collection authority is not a policy. It is a headline searching for a legal theory.
The ledger does not forgive emotion, only math. So let us do the math on what a community dividend would actually price.
The Constraint Moved, and the Politics Moved With It
The AI buildout stopped being a semiconductor story somewhere in the last eighteen months and became a power story. That transition is the entire context for this proposal, and the report I am working from never says it once.
Three facts define the regime. First, rack density. A 2015-era enterprise rack drew five to ten kilowatts and was air-cooled. An H100-class rack draws forty to sixty. A GB200 NVL72-class rack draws somewhere in the hundred-and-twenty to hundred-and-forty kilowatt band and requires liquid cooling because air cannot move the heat. Power density per square foot has gone up by roughly an order of magnitude in under a decade. Second, the bottleneck migrated. The binding constraint on deployment is no longer GPU allocation, it is energized land โ a site with a substation, a transmission path, water rights, and a permit. Third, the queue. Interconnection requests in the largest US grid operator run into multi-year study windows, and the large-load queue in Texas has at times held requests measured in hundreds of gigawatts against a system peak that is itself only around eighty-five to ninety gigawatts. Most of those requests will never energize. The queue is not a waiting list. It is a speculative asset class.
When capacity was cheap and load growth was flat, a data center was a tax base. Counties competed for them. Now that load growth has flipped positive for the first time in two decades and capacity auctions are clearing at the administrative cap in large parts of the footprint, a data center is a ratepayer liability. The political sign flipped before the economics did.
The geography of the backlash is not random. It clusters where the density of load meets the density of people. Loudoun and Prince William in Virginia. Fayette and Newton County in Georgia. Maricopa in Arizona. New Albany and the Columbus ring in Ohio. These are places where residents opened an electricity bill, saw a double-digit increase, read that the local utility had signed a special contract with a hyperscaler, and connected two dots that were in fact connected.
The instruments already exist, which is the part the report missed entirely. Local governments have been extracting value from large industrial load for decades. Payment in lieu of taxes. Community benefit agreements. Special rate schedules. Utility demand-response contracts. Host community funds. What the proposal calls a dividend is, in almost every plausible construction, a rebranding of one of those four things. The question is not whether data centers pay. They already pay. The question is which instrument gets used, and therefore who bears the cost of capital.
And there is a precedent nobody in the AI conversation wants to cite, because it comes from the industry they treat as a punchline. Bitcoin miners have been paying this bill for years. In Texas they sell curtailment back to the grid operator and call it revenue. In upstate New York and the Pacific Northwest they sign host community agreements before they break ground. The AI data center industry is being asked to pay a price the mining industry already negotiated. That is not a coincidence. It is the same physical asset with a different revenue line on top.
What a Dividend Actually Is, Legally
The word is doing an enormous amount of work and almost none of it is defined. There are three distinct legal paths a community dividend can take, and they have three completely different costs of capital. Conflating them is how people lose money.
Path one: a tax. The county raises the assessed value of the industrial property and the millage rate does the rest. This is the least disruptive path and the most likely one. It requires no new federal authority, no new statute, and no corporate consent. It also has almost no marginal impact, because data centers already pay property tax and reassessment is a slow, litigated, backward-looking process. If the dividend resolves into a tax, the headline was noise and the operators barely notice.
Path two: a community benefit agreement. A negotiated contract between the operator and a local government or community organization, typically covering infrastructure contributions, workforce funds, or fixed annual payments. This is a CAPEX-adjacent instrument. It is paid at or near construction, it can be amortized over the asset life, and it is a one-time negotiation. Cost impact: low single-digit percentage on project capital. Annoying, not existential.
Path three: an ongoing revenue or consumption levy. A per-megawatt-hour charge, a percentage of revenue, or a per-rack fee, collected continuously. This is the dangerous version. It is an operating expense that never amortizes, it compounds against every future capacity expansion, and it cannot be renegotiated without renegotiating the site. Cost impact: permanent compression of the internal rate of return, and a direct increase in the cost of every token the campus will ever produce.
The report I am working from gestures at all three and commits to none. That is the single most important defect in the source material, and it is the reason anyone who tells you they know what this proposal costs is selling you something. The difference between a one-time community payment and a per-megawatt-hour levy is the difference between a rounding error and a credit downgrade.
There is a fourth path nobody lists because it is not a policy at all: nothing happens. The statement is made, the cycle moves on, no bill is filed, no agreement is signed. In my experience auditing announcements against outcomes, this path has the highest base rate of any of them. Political statements that lack a legislative vehicle and a named sponsor die quietly. The absence of a date in the source material is itself a signal about how seriously to take the timing.
The Unit Economics of a Dividend
I am going to build the cost model from the bottom, because the public conversation keeps treating this as a moral question when it is a spreadsheet question.
Start with the capital stack of a modern AI campus. Shell, electrical, cooling, and land run somewhere in the nine to thirteen million dollar range per megawatt of critical IT load, depending on site conditions and how much substation work the operator has to fund itself. Add the compute. A megawatt loaded with current-generation accelerators is a different order of magnitude entirely. Then add the thing that never appears in the pitch deck: the interconnection deposit, the network upgrade cost allocation, and the eighteen to thirty-six months of carrying cost while the site waits for energization.
Now the operating stack. For a colocation provider, electricity is typically thirty to fifty percent of operating cost depending on the market and the contract structure. For a self-mining operation it is sixty to eighty percent of cash cost. Everything else โ labor, maintenance, depreciation, bandwidth, insurance โ sits on top of a number that is set by the grid.
Against that, layer a dividend. Run the three scenarios.
| Scenario | Mechanism | Operator impact | Tenant impact | Probability | |---|---|---|---|---| | One-time community investment | Paid at construction, amortized | CAPEX up 3 to 8 percent | Negligible | Medium | | Ongoing revenue or consumption share | 1 to 5 percent of revenue or per-MWh levy | OPEX permanently higher, IRR down 100 to 300 basis points | Rents up 2 to 5 percent | Medium | | Property tax reassessment | Assessor raises valuation | Marginal, overlaps existing regime | Negligible | High |
A hundred to three hundred basis points of IRR sounds survivable. It is not, at the margin where these assets are financed. A data center development is a levered, long-duration, single-tenant-credit bet. If the underwritten return is fifteen percent and the dividend takes it to twelve, the project does not get marginally worse โ it fails the investment committee threshold and does not get built. At the margin, policy does not shave returns. It decides which projects exist.
The second-order effect is the one that actually matters for anyone holding tokens. If the dividend is levied on consumption rather than revenue, it is a tax on efficiency. An operator who improves power usage effectiveness pays the same per megawatt-hour and therefore pays more relative to output. That inverts the incentive the entire industry has been optimized around for fifteen years. Efficiency is just another word for fragility when the pricing mechanism punishes it.
The third-order effect is contract structure. Existing power purchase agreements and colocation contracts are a mix of fixed-price and floating. An operator who signed a seven-year fixed-price contract in 2024 cannot pass through a 2027 levy. They absorb it. That is a pure margin transfer from shareholders to a county government, executed retroactively, with no ability to reprice. I have watched this exact mechanism destroy a yield product before. In 2020 I had fifteen thousand dollars in a newly launched automated market maker and a Python script watching gas and slippage in real time. When the oracle got manipulated and the pool started bleeding, my script exited in forty-five seconds and I recovered ninety-two percent of principal. The people who stayed were not stupid. They were locked into a position whose risk had been repriced by someone else's decision. Retroactive repricing is the most expensive kind, because you cannot hedge what you already signed.
The Asymmetry Nobody Has Priced
Here is the part of this story that no headline has touched, and it is the only part I would actually trade.
A community dividend is a fixed cost. Fixed costs are regressive. They fall hardest on the smallest balance sheet in the room. That means the proposal, if it becomes real, does not tax the data center industry uniformly. It redistributes market share inside it.
| Operator class | Project scale | Dividend as share of economics | Net position | |---|---|---|---| | Hyperscaler | $1bn to $10bn per campus | Well under 1 percent | Strongly positive | | Mid-cap colocation | $200m to $800m per site | 2 to 5 percent | Neutral to negative | | Emerging neocloud | $500m to $3bn, heavily levered | 4 to 10 percent plus delay risk | Negative |
The hyperscaler calculus is not close. A commitment worth tens of millions against a ten-billion-dollar campus buys three things that are worth far more than the check: approval velocity, political cover, and a narrative asset that the sustainability team can put in front of institutional investors. If the choice is nine months of county hearings or a funded community center and a workforce pipeline, the arithmetic answers itself. The hyperscalers will not resist this proposal. They will co-opt it, and then they will use it.
For the mid-cap colocation operators, the dividend is a margin event. Their tenants are price-sensitive enterprises on multi-year contracts, and their cost of capital is already higher than the hyperscaler's. They can pass through some of it in a seller's market. In a buyer's market they eat it.
For the neoclouds, this is not a margin event. It is a solvency event. The emerging GPU-cloud cohort is built on a specific structure: five-to-seven-year take-or-pay contracts secured against GPU assets depreciating on a much shorter schedule, financed with debt that assumes energization on schedule. Anything that adds eighteen months to a permitting timeline does not reduce their return โ it triggers a covenant. The dividend itself might cost them two percent. The delay the dividend is designed to prevent might cost them the company.
Which is the paradox at the center of this whole story. The proposal is framed as a cost imposed on data centers. In practice it is a pricing mechanism for speed, and speed is the only asset the leveraged players actually need. The operators who can afford to pay it will pay it and accelerate. The operators who cannot afford it will be stuck in the queue, watching the ones who paid move ahead of them. Structure survives the storm; chaos drowns it.
The Interconnection Queue Is the Real Balance Sheet
If you want to understand why a dividend is cheap, you have to price the alternative, and the alternative is a delay.
A queue position is an option. It has a strike price โ the deposit, the study costs, the network upgrades โ and it has a payoff that is a function of time to energization. Options decay. Every month a campus sits unenergized, the operator carries land cost, capitalized interest, pre-paid equipment, and the opportunity cost of compute that could be earning. On a large campus, carrying cost runs into seven figures per month before a single accelerator is installed.
Now put a number on the political risk. A county board can delay a project by refusing a rezoning, contesting a water permit, or intervening in a utility rate case. Each of those actions adds months, not weeks. Each month costs more than a community benefit agreement would have cost in total.
So the rational operator does not negotiate the dividend. They pre-empt it. They walk into the first hearing with a community fund already sized, a workforce commitment already drafted, and a local hire target already agreed with the technical college. This is not charity. This is buying optionality on the queue, and the price is trivial relative to the value of the option.
I built a version of this trade in 2024. When the spot ETF approvals forced institutional reporting standards onto a market that had none, I led four analysts through a rebuild of our flow templates and cut report generation from four hours to forty-five minutes by automating the terminal extraction. The efficiency gain did not just save labor. It let us rebalance ahead of the crowd and identify a $2.3 billion inflow trend before it hit the financial press. The lesson generalizes: the entity that shortens the reporting cycle captures the alpha, and the entity that shortens the permitting cycle captures the capacity. A community dividend is a permitting-cycle accelerator purchased at a discount.
Liquidity is a ghost; it vanishes when you blink. Queue capacity behaves the same way. It looks abundant in the aggregate โ hundreds of gigawatts of requests โ and it is almost nonexistent in the specific. The operators who understand this are already signing host community agreements in counties where no one has filed yet. The operators who do not are still waiting for the interconnection study to come back.
From Dividend to Token Price: The Transmission Chain
Every analysis I have read on this topic stops at the operator. That is a mistake, because a data center is not a terminal node. It is the physical floor of the AI stack, and a cost imposed on the floor propagates upward.
The chain runs like this. A per-megawatt-hour levy raises the operating cost of the campus. The campus passes a portion to its tenants through colocation rent or reserved-capacity pricing. The cloud provider passes a portion to its customers through instance pricing, and eats the rest in gross margin. The inference provider passes a portion to the application layer through per-token pricing. The application layer either raises consumer prices, degrades model quality, or dies. There is no fifth option.
The magnitude matters. Inference is the most power-intensive part of the AI value chain per unit of revenue, because it runs continuously against a marginal-cost curve dominated by electricity. A one percent increase in power cost is not a one percent increase in inference cost โ it is a larger number applied to a business that is already being priced on the assumption of continuous deflation. Every AI application business plan written in the last two years assumes inference cost per token falls by an order of magnitude. A policy that puts a floor under power cost is a policy that puts a floor under token price.
And the crypto leg is not exempt. The same substations serve both industries. When a county raises the cost of industrial load, mining economics compress alongside AI hosting economics, because both are competing for the same energized megawatt. In a bear market where hashprice is already compressed against post-halving issuance, an incremental per-megawatt-hour charge is not a haircut. It is the difference between a machine that pays back and a machine that gets unplugged.
This is the transmission mechanism the report completely missed. It treated the dividend as a discrete policy event affecting a discrete industry. It is actually a repricing of the marginal cost of computation, and computation is the input to everything this market is currently valued on. Numbers do not lie, but narratives do โ and the narrative here is that this is a local zoning story. It is not. It is an input-cost story with a two-year lag before it shows up in someone's gross margin line.
The Miners Already Paid This Bill
Here is the contrarian read, and it is the one I would stake capital on.
Bitcoin mining is the only industry in the world that has already fully solved the community externality problem, and it did so without legislation, because the grid made it profitable to solve. In Texas, a curtailment program pays a large flexible load to stop consuming when the system is stressed. That payment is not a subsidy. It is the market price of the option the miner sold to the grid operator. The community gets reliability. The miner gets revenue. The transaction is voluntary, priced, and settled in real time.
Compare that to a dividend, which is a fixed obligation with no optionality. The miner's structure is strictly superior from a risk standpoint because it converts a political cost into a tradable one. When the system is stressed, the miner earns. When the system is calm, the miner mines. There is no scenario in which the miner writes a check to a county and receives nothing in return.
Now apply that lens to the AI campus. A campus that can modulate load โ batching inference, deferring training, shifting cooling โ has a demand-response asset that it is currently leaving on the table. The dividend conversation is happening because operators are treating themselves as passive consumers of power instead of active participants in a market. The moment they start selling flexibility back to the grid, the political problem changes shape. They stop being the thing that raised the electricity bill and start being the thing that keeps the lights on during the heat wave.
The miners figured this out because they had no choice. Their margins were thin enough that every ancillary revenue stream mattered. The AI operators have had fat enough margins that they never had to learn. A bear market fixes that. When capital gets expensive, everyone suddenly discovers operational discipline. I have watched this exact pattern play out in the DeFi liquidity mining cycle. When emissions were paying, projects reported TVL that existed only because the subsidy existed, and nobody asked what the number would look like if the incentive stopped. When it stopped, it stopped in a week. Subsidy-dependent metrics are always a liability dressed as an asset. A data center that only works because power is cheap and politics is quiet is the same structure.
The Bear Market Read: Who Bleeds
I do not write about this as an abstraction. I write about it because there are balance sheets on the other side of it, and in a down market the difference between a good operator and a bad one is not returns. It is survival.
Four exposures, ranked by how fast they bite.
Refinancing risk. Any operator with debt maturing inside twenty-four months and a project that has not energized is exposed to both a higher rate environment and a permitting delay. The dividend is the smaller problem. The covenant is the bigger one.
Take-or-pay mismatch. Contracts signed at the top of the cycle assume a compute price that a down market will not sustain. If the tenant reprices or restructures, the operator is left holding power contracts and depreciation schedules sized for revenue that no longer exists.
Depreciation curve. Accelerators lose economic value faster than any industrial asset in history. A three-to-four-year useful life against a seven-year contract is a duration mismatch that only works if utilization stays near capacity. Utilization is the first thing to fall in a bear market.
Political cost. This is the newest and least priced. It does not show up in a filing until it shows up as a delay, and a delay shows up as a missed covenant, and a missed covenant shows up as a restructuring.
Rank them honestly and the dividend lands fourth. That is the point. The proposal is not dangerous because of the money. It is dangerous because of the uncertainty, and uncertainty is the only input that cannot be modeled into a discounted cash flow. You cannot underwrite a cost whose base, rate, and collector are all undefined. So the market does what markets do with undefined variables: it applies a haircut to everything in the category, indiscriminately, and lets the strong names buy the weak ones at a discount.
The Blind Spot
Everyone reading this story has the same reaction: this is a tax on technology, and technology will fight it.
That is backwards.
A community dividend is the cheapest competitive weapon a hyperscaler could ask for. It is a fixed cost that the largest balance sheets can absorb at a rounding error and the smallest cannot absorb at all. It converts a reputational liability into a procurement advantage. It gives the incumbents a publicly defensible reason to demand that their competitors meet the same standard, and the standard is one only they can meet. Every regulatory regime in history has been shaped by the incumbents who could afford to comply with it, and this one will be no different.
The second blind spot is the assumption that the proposal is about fairness. It is about throughput. A county does not want a dividend. A county wants the data center to exist without costing the supervisor an election. The dividend is the price of removing an objection. Framed that way, the operators who pay it fastest are not the ones being punished. They are the ones buying the lane.
The third blind spot is the assumption that this is an American story. It is not. If the United States establishes a norm that large compute load owes a host community payment, the European Union's existing energy and data-sovereignty frameworks will adopt the language within a cycle, and the sovereign AI programs in the Gulf and Southeast Asia will build it into their incentive packages from the start. The result is not a US cost. It is a global cost of compute, applied unevenly, with jurisdictions that move fastest capturing the deployment. Anchor pegs break before trust does, and a jurisdiction's promise of cheap power breaks the moment the queue gets long.
Takeaway
Do not trade the headline. Trade the mechanism, and the mechanism is not defined yet.
Three things I would actually watch, in order. First, state-level legislation. If three or more states introduce data center community payment statutes within a year, the norm is forming and the cost is real. Until then it is a press release. Second, hyperscaler behavior. If two or more of the largest operators publicly commit to a community dividend framework voluntarily, the standard has been set by the industry itself and the mid-caps have eighteen months to adapt. Third, capacity auction clearing prices in the largest grids. If they clear at the cap again, the political pressure that produced this proposal does not go away โ it intensifies, and the next version will have a sponsor and a bill number.
My position is not complicated. The dividend is small, the asymmetry is large, and the only operators who should be worried are the ones who cannot write the check and cannot afford the wait. The rest of this industry will pay the fee, call it stewardship, and use it to take market share from the people who called it a tax.
I audit the code, not the promises. Right now there is no code โ only a sentence, no date, no base, no collector. The question worth asking is not whether data centers will pay their communities. It is which ones will be able to afford to, and what happens to the ones that cannot.