From Heat Sink
to Harvest.

The product is permission, not produce.

Artificial intelligence is being built on a trillion-dollar wager, and that wager is now stalling—not in capital markets, but in town halls. This brief argues that the scarce resource in the AI build-out has quietly changed. It is no longer money, chips, land, or even power. It is community permission. And it argues that the most credible instrument yet found for manufacturing that permission is also the most literal: capture the waste heat a data center is already throwing away, and grow food with it, next door, for the people whose consent the project requires.

The pages that follow make that case as a thesis—problem, solution, proof, product, and close—and hold themselves to an evidentiary standard, because the argument only works if it is true. Where a claim is independently established, it is cited to its primary source; where a number is our own engineering target, it is labeled as such and never dressed up as a finding. That discipline is not incidental to the thesis. It is the thesis: a project whose entire value proposition is verifiable benefit must be the first to submit to verification.

Three claims carry the argument, and the document is organized to defend each in turn. First, the problem: opposition has become the rate-limiting step in data-center development, measurable in the hundreds of billions of dollars. Second, the solution and its proof: the cooling transition the industry is already making turns waste heat into a capturable, greenhouse-ready resource, and operators on four continents are already reusing it. Third, the product: a co-located greenhouse converts that heat into a benefit communities can see, taste, and verify—answering the objections on water, grid, jobs, and trust at once. The numbers below preview the case; the parts that follow build it.

^† Intelligent Harvest design target / internal analysis; all other figures are externally sourced and cited in References.

The permission crisis.

For two decades, the limiting factors on data-center growth were tangible and familiar: land, fiber, power, and capital. Those constraints have not disappeared, but they have been decisively eclipsed. To see why, begin with the one number that is not a constraint. In May 2026, Moody’s raised its capital-spending projections for the six largest U.S. hyperscalers to roughly $785 billion for 2026 and nearly $1 trillion for 2027.¹⁷ Whatever is slowing this industry, it is not a shortage of money.

What is increasingly missing is permission—the practical consent of the communities asked to host the buildings—and unlike capital, it is becoming scarcer by the quarter. According to analysis maintained by Data Center Watch and reported by Fortune, at least 48 projects representing $156 billion in investment were blocked or stalled by local opposition in 2025 alone.¹ This is not a one-year anomaly; it is an accelerating trend. Annual project cancellations climbed from a small handful in 2023 to roughly two dozen in 2024 and 25 in 2025—and then, in the first quarter of 2026 alone, more than 20 additional projects were killed, a record quarterly pace.¹ The opposition has also organized: there are now 188 distinct opposition groups operating across 40 states, and by Data Center Watch’s accounting, the projects they contest are blocked or delayed roughly two times in three.^1, 2

Two features of this opposition make it strategically important rather than merely noisy. The first is that it is bipartisan and cross-regional—blue and red states alike are tightening rules and rethinking subsidies—which means it cannot be dismissed as a partisan phase that will pass.² The second is that it has begun to win even where the developer holds every conventional advantage. The reversal is sharpest exactly where the industry is most entrenched.

Consider Virginia, the global capital of data centers. A Washington Post–Schar School poll conducted in late March 2026 found that the share of voters comfortable with a new data center in their community had fallen to 35 percent, down from 69 percent in 2023—a 34-point collapse in three years.³ Support for the tax incentives that built the industry fell from 61 percent to 37 percent over the same window, and 67 percent of voters now say the state should end the incentive outright.³ Days after the poll, Prince William County abandoned one of the largest proposed data-center projects in the country.³ If the most experienced data-center community in the world has turned this sharply, the trajectory elsewhere is unlikely to be gentler.

This is the problem, stated plainly: the industry must build at a once-in-a-generation scale, and its license to do so is evaporating in precisely the places it most needs to expand. Every blocked project strands enormous sunk cost in land, design, and interconnection studies; every contested one adds delay, legal exposure, and political risk. A developer who can credibly de-risk permission—not through louder lobbying, but by delivering a benefit the community can see and value—holds something rarer and more durable than another feasibility study. To design that benefit well, though, we first have to understand what the public is actually asking for. That is the subject of the next section.

What the public actually wants.

It would be easy—and wrong—to read the backlash as blanket hostility to technology. Read closely, the public is not saying “no.” It is saying “not like this.” The opposition is a demand for reciprocity, and that distinction is the entire opening for this project.

Start with the shape of the sentiment. The collapse in Virginia is not an outlier but the leading edge of a national mood: a January 2026 Marquette Law poll found that a clear majority of Americans believe data centers’ costs outweigh their benefits.¹³ Yet the same public is strikingly supportive in the abstract. A 2025 survey captured the paradox in a single line—overwhelming support for data-center development as a national priority, collapsing the moment the proposed site is nearby.¹¹ Support in principle; refusal in practice. The gap between those two opens at the property line.

Why does the gap open there? Because at the property line, the ledger goes lopsided. The costs of a data center are local and concrete—noise, water draw, traffic, transmission lines, the industrial transformation of a familiar place—while the benefits are diffuse and distant: national competitiveness, corporate revenue, a tax base that residents suspect they subsidize. People are not refusing the digital economy. They are refusing a bargain in which they absorb the costs and someone else collects the benefits.

The evidence that this is the real grievance—rather than reflexive opposition—is that the public will say yes when the bargain is rebalanced. Offered a third path between “build freely” and “stop building,” one that builds data centers but requires companies to protect consumers and the environment and pay their fair share, roughly 72 percent of national respondents choose it.¹² The public is not anti-compute; it is anti-extraction. There is even evidence that some specific fears are misplaced: a George Mason University analysis found that in Virginia, residential property values tended to rise the closer a home sat to a data center, not fall.¹⁶ But in a permission fight, facts rarely carry the room on their own. What carries the room is a benefit a skeptic can see, name, and—ideally—taste.

What Intelligent Harvest does.

We co-locate a controlled-environment greenhouse on a data center’s cooling loop, so the heat the facility is already paying to throw away instead grows food for the town next door. The design goal is not novelty. It is reliability—and, above all, honesty about what the system does and does not require.

Begin with a fact that sounds like a provocation but is simply thermodynamics: a large data center is a furnace that happens to compute. Essentially all of the electricity a server draws leaves the building as heat.⁵ The only question a facility ever faces is where that heat goes. Today, the answer is “away”—vented to the sky or evaporated into cooling towers. Our proposal is to give a fraction of it a second destination, without changing anything about the first.

The mechanism is deliberately unglamorous, because reliability is the point—and it works no matter how the servers are cooled. Where a facility is liquid-cooled, we tap its existing cooling loop directly; where it is air-cooled, we add an air-to-water recovery coil on the contained hot-aisle exhaust. Either way, the captured warmth flows into a separate, clean loop and a buffer tank that stores it around the clock, then into low-temperature under-bench and floor loops in an adjacent greenhouse.⁵ A heat pump is available to lift the grade, but it is a small, optional step—needed mainly for the lower-grade heat of air-cooled sites or the coldest winter nights, not for everyday operation. Cooled water returns to the data center, which runs its own cooling exactly as before. We never replace the facility’s cooling—the compute load must always be protected—we give a beneficial second life to heat it already rejects.

Two design choices in that diagram are worth dwelling on, because they are where honesty does real work. The first is that the system is cooling-agnostic. A fair skeptic asks whether this depends on the newest liquid-cooled hardware; it does not. The densest AI is indeed going liquid—and that favors us—but air cooling is not disappearing from storage, networking, and much of the enterprise market, so a durable concept has to work with both. It does, by changing only the first step: “tap a water loop” becomes “recover heat from the exhaust stream,” and everything downstream is identical.⁵ The second choice is that the heat pump is intentionally drawn small. A great deal of waste-heat marketing leans on heat pumps as the hero; we present ours as an occasional helper, because over-promising the easy part is exactly how a project loses the trust it is trying to build.

And so two outputs leave the greenhouse. The first is food—fresh produce grown year-round, in regions that are frequently food-insecure and far from fresh supply.⁵ The second is the output that matters to the developer: a concrete, visible, locally owned benefit that changes the answer a community gives when asked whether the next facility should be built. The greenhouse is the instrument; permission is the product.

The physics: why water, not air.

You cannot harvest heat you cannot catch. The viability of this entire model rests on a physical transition the industry is already making for its own reasons—from cooling with air to cooling with liquid—because air scatters heat and water concentrates it.

For most of the data-center era, servers were cooled by blowing cold air across them and rejecting the warmed air outdoors, frequently through evaporative cooling towers that consume large volumes of water.⁵ Air, however, is a poor heat carrier: it holds very little energy per unit volume, so capturing usable heat from a diffuse, lukewarm air stream is inefficient and rarely worth the trouble. This is the real reason waste-heat reuse has lagged in the United States. The heat was never absent—air-based systems simply made it uneconomical to catch.⁵

Water changes the equation entirely. By volume, water carries on the order of several thousand times more heat than air—a direct consequence of its far greater density and specific heat. Direct-to-chip cold plates and immersion cooling, which the largest operators are now adopting for the densest AI workloads, move server heat into a closed liquid loop that delivers heat which is hotter, more concentrated, and easy to pipe.^4, 5 The single transition the industry is making for its own efficiency is precisely the transition that makes a neighboring greenhouse feasible. We are not asking the industry to change course; we are building on the course it is already on.

The environmental case rides on the very same shift, which is what makes it so durable. A two-year life-cycle assessment led by Microsoft researchers and published in Nature in 2025 found that cold-plate and immersion cooling reduce greenhouse-gas emissions by 15–21%, energy demand by 15–20%, and blue-water consumption by 31–52% across the full life cycle, compared with air cooling.⁴ Read that against the public’s top concerns from Part I: the move that makes heat capturable is the same move that independently cuts the water and carbon footprint communities object to most. Solving the water fight and enabling reuse are, physically, one decision.

The grade of the heat—and why it is already enough

The standard objection to data-center heat is that it is “low-grade”—warm, but cooler than many uses require. The objection is real, and it is why district heating struggles: hot-water networks for buildings typically want 60–70 °C or more, which air-cooled exhaust strains to deliver.⁵ An ACEEE analysis puts typical reject temperatures at roughly 25–35°C for air-cooled systems, 50–60°C for water-cooled systems, and up to 90°C for two-phase refrigerant systems.⁶ For most offtakers, the low end is a problem. For a greenhouse, it is the whole point.

A greenhouse is the rare customer that wants exactly what data centers most easily give. Comfortable growing air sits around 18–27 °C, and root-zone and bench heating need only mild warmth delivered through low-temperature under-bench and floor loops.⁵ As the ladder below shows, even the low end of water-cooled reject heat clears the greenhouse’s comfort line before any lift at all—and the optional heat pump exists only to add flexibility for the lower-grade air-cooled case or the coldest weeks of winter. The low grade of this heat, a deal-breaker for almost everyone else, is precisely what a greenhouse can use.

Why a greenhouse.

Low-grade heat does not travel. The hardest problem in waste-heat reuse is not capturing the heat—it is finding an anchor offtaker with continuous, year-round demand, sited close enough to use the heat before it is lost. For a rural data center, a greenhouse is the near-perfect answer, and it is worth seeing exactly why.

The economics of reuse live and die on distance and demand. Hot-water heat networks are efficient over only a few miles, and costs climb steeply with pipe length and diameter.⁵ District heating—the classic reuse pathway—therefore needs dense housing or campuses nearby. But most new AI data centers are sited in rural areas precisely because that is where land and power are; the dense neighborhoods district heating requires usually are not there.⁵ A greenhouse dissolves the distance problem by sitting on the fence line: zero transmission distance, and an appetite for heat that runs 24 hours a day, every day of the year.

The supply, meanwhile, is more than ample. A Microsoft analysis estimates that between 0.69 MWh (winter) and 0.86 MWh (summer) of reusable heat is available for every 1 MWh of electricity a data center consumes.⁵ To put that at community scale: a 24-MW facility running through an Ohio winter could in principle supply much of the seasonal space heat for 15,000–20,000 homes—if the dense housing and district piping existed to use it.⁵ A co-located greenhouse captures a slice of that abundance immediately, without waiting years for a pipe network to be built. The data center rejects far more heat than the greenhouse needs, which is exactly why the recovery gear can stay small and the project can start now.

Zone 1
Strawberries

Zone 2
Ginger & turmeric

Zone 3
Gourmet mushrooms

Zone 4
Citrus orangery

Four climate zones on one heat loop, each tuned to a high-value crop that thrives on steady, low-grade warmth—the part of the system the public can taste. Images: Intelligent Harvest.

But the deepest reason for a greenhouse is the one that connects back to Part I. Recall the brief any solution had to meet: local, visible, durable. A greenhouse satisfies all three in a form a skeptic can physically stand inside. Many of the rural communities now weighing data centers are food-insecure, with limited access to fresh produce; a greenhouse converts an invisible, abstract by-product—warm water—into something a resident can see from the road and buy at the market.⁵ Of everything a data center could do with its waste heat, growing food produces the benefit that is most legible to exactly the people whose permission the developer needs. That legibility is not a side effect of the strategy. It is the strategy.

It already works.

The most powerful answer to “will this work?” is “it already does.” Data-center heat warms greenhouses, homes, swimming pools, and fish farms across four continents today—and a wave of U.S. projects is now in development, several converging on precisely the Intelligent Harvest pattern.

Europe is roughly a decade ahead, in part because several countries now require reuse. The clearest greenhouse analogue runs north of almost everywhere: in Boden, Sweden, Hive Digital’s 32-MW data center heats an approximately 90,000-square-foot greenhouse, enabling vegetable production near the Arctic Circle.⁹ Across the continent, hyperscale and colocation operators feed municipal heat networks—Meta in Odense, Denmark; Yandex in Mäntsälä, Finland; atNorth in Kista, Sweden—while others drive uses from wood-pellet drying to commercial trout and lobster farming.⁷ The technology is not experimental abroad. It is infrastructure.

The American pipeline is the more telling signal, because it is converging independently on the exact model this brief proposes. In Ohio, SAIHEAT’s Marietta facility already demonstrates liquid-cooled compute heating a local greenhouse.⁵ In West Virginia, Fidelis New Energy’s proposed Monarch Cloud Campus would co-locate up to 1,000 MW of data centers with controlled-environment agriculture, using waste heat and captured CO₂ to supply adjacent greenhouses.⁵ In Mansfield, Ohio, EnergiAcres is pursuing a “gas → data → food” campus that routes waste heat into greenhouses and cold-chain logistics.⁵ When multiple independent teams arrive at the same design, that is not coincidence—it is convergence on what the constraints allow. The open question is no longer whether the pattern works, but who executes it cleanly, at community scale, with the brand and benefit framing that actually earns permission.

The weight of the evidence.

A thesis is only as strong as the numbers under it. Here we assemble the load-bearing figures in one place, separate what is independently established from what is our own projection, and show how the pieces interlock into a single, coherent opportunity.

Read in isolation, each figure is merely interesting. Read together, they describe one opportunity with an unusual property: the pieces were not designed to fit, and they fit anyway. A trillion-dollar build-out¹⁷ is colliding with a permission crisis measured in the hundreds of billions of blocked dollars.¹ The cooling transition that resolves the public’s water objection⁴ is the very transition that yields heat already warm enough to grow food.⁶ And that heat is available at a generous ratio to the power drawn—up to 0.86 MWh of usable warmth per MWh consumed.⁵ A solution does not usually arrive pre-assembled from four independent directions. This one effectively does.

Two cautions keep this honest. First, organized opposition is not the sole cause of every delay—permitting, power availability, and economics all play roles, as Data Center Watch itself notes; the $156 billion figure marks projects opposition touched, not projects it single-handedly killed.^1, 2 The trend and direction, however, are unambiguous. Second, the figures below are ours, and we mark them as such so they can never be mistaken for third-party findings.

† Internal Intelligent Harvest projections; presented as design targets, not measured results.

That final ratio is the thesis compressed into arithmetic. A roughly $16 million greenhouse—about one one-hundredth of one percent of the pipeline opposition has stalled—is positioned not to pay for itself in produce, but to protect a multi-billion-dollar development program by changing the answer a community gives. With the problem documented, the mechanism explained, and the evidence assembled, what remains is to show that the result is a real, defensible product rather than a hopeful gesture. That is Part IV.

Permission, not produce.

The reason this is a product and not a publicity stunt is that a single intervention pushes on four of the industry’s hardest constraints at once—and each gain reinforces the next. Understood correctly, the greenhouse is a flywheel for manufacturing consent.

The logic is circular by design. Captured heat grows food. Food and the jobs around it create a visible local benefit. That benefit earns community trust. Trust converts into permission—approvals, rezoning, the absence of a lawsuit. Permission lets more compute get built. And more compute produces more heat to capture. Each revolution lowers the resistance to the next, which is what separates a flywheel from a one-time favor.

Trace the flywheel against the four objections that actually sink projects, and the product comes into focus.

1 · The permission problem

This is the binding constraint, and the one the industry under-prices. With roughly $156 billion in projects already blocked or delayed and a record number killed in a single recent quarter,^1, 2 the scarce resource is consent. A greenhouse is the most legible consent-builder available, because it converts an abstract corporate promise into a building residents can walk into—satisfying, at last, the local-visible-durable brief from Part I.

2 · The water problem

Water is the concern that most reliably mobilizes opposition, and the move to liquid cooling answers it directly: advanced cooling can cut data-center water consumption by 31–52% versus air-based systems.⁴ That same closed loop is what makes the heat capturable in the first place. Intelligent Harvest’s own design then targets roughly 18% less evaporative water than a comparable conventional greenhouse, by drawing on that loop rather than fresh make-up water.^†

3 · The grid problem

Interconnection queues now stretch for years, and new load is often exactly what neighbors fear. Counter-intuitively, reuse can help. A 2025 National Renewable Energy Laboratory analysis showed that when nearby buildings draw on data-center waste heat instead of their own electric or fossil heating, their demand drops—freeing “headroom” on the same feeders and substations and letting the grid absorb new 15–30 MW loads it otherwise could not.¹⁵ The heat the project gives away is also what helps it plug in faster.

4 · The jobs problem

Data centers are famously capital-rich and labor-poor. An Ohio River Valley Institute analysis put the sector at roughly 0.25% of Pennsylvania employment, with as few as 27–111 permanent staff per facility—numbers that make the “economic engine” framing hard to sustain.¹⁰ A working greenhouse is the opposite kind of asset: operationally labor-intensive, year-round, and rooted in place. It does not fix the data center’s job math by itself, but it adds the visible, local, recurring employment that a server hall structurally cannot—closing the fourth and final gap between what a data center offers and what a community is asking for.

The honest limits.

A document that argues only one side is marketing, not analysis—and a skeptical public can tell the difference instantly. So here is where waste-heat agriculture is genuinely hard, stated plainly, with our response to each constraint rather than a wave of the hand.

The risk we take most seriously: greenwashing

The sharpest objection to a project like this is not technical at all. It is that a greenhouse becomes a green fig leaf—a photogenic distraction that buys approval for a facility whose real impacts continue unabated. We think that critique is correct about the danger, and that it is precisely why the work has to be done a particular way. If the greenhouse is allowed to become theater, it deserves to fail, and it will.

The policy is moving our way.

A first mover usually has to fight the rulebook. Here, the rulebook is being rewritten toward exactly this model—by mandate abroad, by incentive at home—which turns a headwind into a tailwind for whoever executes first.

Europe has stopped asking politely. Under Germany’s Energy Efficiency Act, new data centers must reuse a rising share of their energy—10% from July 2026, climbing to 15% in 2027 and 20% in 2028.¹⁴ That is the clearest signal anywhere that waste-heat reuse is migrating from optional to expected, and it explains why the densest cluster of working examples sits in Europe rather than the United States.

The American approach is the carrot rather than the stick—but the carrots worth counting on are the durable ones. The clean-energy investment tax credit that once recognized “waste energy recovery property” was sharply curtailed by the July 2025 federal reconciliation law, which put the technology-neutral credit on a hard phase-out for projects not under construction by the end of 2025.¹⁸ We do not build the case on a credit that is sunsetting. The incentive that actually underwrites a project like ours is older, broader, and far more bipartisan—federal and state farm and rural-development capital, the subject of the next section—because the benefit it rewards is exactly the one we deliver: capture heat, site on legacy industrial land, employ local people, and grow food. On heat reuse itself, meanwhile, the direction of travel at the state level is unmistakable.

None of these is law yet—we flag that plainly, because a document about verifiable benefit cannot cite a bill as if it were a statute. Together they mark only the direction of travel: reuse is becoming a planning default, a credited resource, and a permitting expectation. The capital we actually build on is already enacted and funded, and it is the subject of the next section.

The money is already there.

Here is the part most waste-heat pitches skip: who pays, and what happens in a bad year. We will not skip it—because the answer is the strongest thing about the model.

Start with the reframing that changes everything. The greenhouse is not the asset. The permit is. The greenhouse is the price of acquiring it—and we price it like insurance, not like a farm that has to turn a profit. That single move is what makes the model durable. A farm that must out-earn its own costs is fragile; a deliberately funded cost, sized to be trivial against the asset it unlocks, is not. The greenhouse is allowed to have a hard year. The data center it protects has already been built.

We fund it the way America has always funded farms: with a blend of public and private capital. The private half is an operator facing a project that local opposition has stalled to a halt, buying a few percent of permission insurance against an asset worth orders of magnitude more—the easiest math in the room. The public half is not a hope or a pending bill. It is already enacted, already appropriated, and built precisely for rural value-added agriculture. The first reference build leans on public capital to generate the proof; operator economics carry it at scale once that proof exists.

So the capital question has a concrete answer. These are not proposals; each program below is enacted and administered today.

Two cautions, in the open: REAP saw application pauses and a temporary hold on controlled-environment-agriculture applications during 2025–26,²¹ and grant cycles open and close—so we treat these as a funded foundation to build on, not a check to cash tomorrow. The base case still pencils on operator capital alone; public capital simply makes the first build faster and the rest easier.

The downside, drawn to scale

An honest model shows its worst case—and names the number before anyone else can. So here is the greenhouse’s entire annual cost, in good years and bad, set against the one thing that actually matters: the operating profit of the facility a single approval unlocks. We mark our own figures as design targets; the facility baseline is illustrative. The shape of the picture does not change with the inputs.

Read the table down the right-hand column and the point lands: there is no branch where the operator regrets it. Worst case, a sliver of a loss against a sea of profit—and they built the thing that was blocked, employed the town, and stopped being the bully in the room. They walk into the hearing as an extraction story and walk out as a farmer for their own community. That is not a vegetable stand bolted onto a server hall. It is the cheapest insurance policy in infrastructure.

Sources

References.

Reference

Glossary.

The working vocabulary of waste-heat agriculture, in plain terms.

Reference

Index of terms.

Where each idea appears, by section.

Links jump to the relevant section. “Summary” is the abstract; numerals are sections 01–11.