There’s a 40,000-square-foot building in central Washington state that still has its cooling towers, its raised floors, and its backup generators. It has been empty since 2021. The fiber is dark, the racks are gone, and the building’s owner is now leasing it to a mushroom farm that uses the remaining chillers to keep humidity stable. The mushrooms are thriving. The GPUs are somewhere in a landfill in Ghana.
That building is not an anomaly. It’s the leading edge of a wave that the AI industry has barely begun to reckon with: the afterlife of infrastructure that was built for a compute generation that is already obsolete.
We talk constantly about how fast AI is moving. We talk almost never about what gets left behind. The industry is building at a pace that no physical infrastructure system has been built at since the interstate highway program, and the depreciation schedules haven’t caught up with the pace of obsolescence. The question is not whether data centers will be abandoned. It’s whether we’ll have the imagination to repurpose them before they rot.
This is a piece about the archaeology of a computational era. And it starts with a sacred cow that the industry holds dear: the belief that data centers are long-lived assets.
The Sacred Cow: “Data Centers Are 20-Year Assets”
Here’s the belief I want to dismantle, and I want to be specific about what I’m claiming.
The claim, repeated in investor decks, real estate prospectuses, and procurement documents, is that a data center is a 20- to 30-year asset. The building shell, the power infrastructure, the cooling plant, the fiber runs, all amortized over a generation. That framing underpins the entire financial logic of the hyperscale buildout. It’s also wrong for a growing fraction of the fleet.
The reason is simple: the useful life of the building is now decoupled from the useful life of the compute it houses.
A building designed for 5 to 10 kW per rack cannot economically support 80 to 150 kW racks. The power distribution, the cooling, and the floor loading all need to change. That’s not a retrofit. It’s a rebuild.The GPU generation cycle is 2 to 3 years. The building cycle is 20 to 30 years. That ratio is unprecedented. It means that by the time a data center finishes its first depreciation cycle, it has hosted 8 to 10 generations of hardware, and at least 2 to 3 of those generations required infrastructure changes that the original design didn’t anticipate.The location decisions made in 2018 (cheap land, available power, tax incentives) are now being reversed by interconnection queues, water rights, and grid constraints. A building in a place that made sense for 2018 workloads may be in the wrong place for 2026 workloads.
The upshot: a large fraction of data center capacity built between 2015 and 2023 is now functionally obsolete for frontier AI workloads. It’s not empty. It’s just serving a different tier of compute: inference, storage, enterprise SaaS, workloads that don’t need 100 kW racks.
The buildings that can’t be repurposed for those workloads will be abandoned. Some already are.
What Actually Happens to an Abandoned Data Center
There’s a small but growing literature on this, mostly from real estate and environmental remediation, and it’s worth knowing because the pattern is consistent.
Stage 1: Decommissioning. The hardware is removed, either to resale, refurbishment, or e-waste. The raised floor is often left in place. The cooling plant is sometimes drained and mothballed, sometimes sold for scrap, sometimes just left running at low power because the cost of shutdown exceeds the cost of idling.
Stage 2: The shell sits. The building has power, fiber, and cooling. It’s a valuable shell. Some get re-leased to smaller cloud providers, crypto miners, or enterprise customers. Others get bought by real estate developers who strip the infrastructure and convert to warehouses, offices, or in a few notable cases, indoor farms.
Stage 3: Decay. For buildings that don’t find a second tenant, the infrastructure starts to fail. Chillers seize. Generators fail their monthly tests. Fiber gets cut and not repaired. The building enters a liminal state, still standing, still powered, but functionally dead.
Stage 4: Repurposing or demolition. The successful cases convert to something else. The failures get demolished, and the demolition is where the environmental cost shows up, because the cooling systems contain refrigerants, the batteries contain heavy metals, and the generators contain decades of oil and diesel residue.
The mushroom farm in Washington is a Stage 4 success story. The building wasn’t designed for mushrooms, but it was designed for climate control, and mushrooms need climate control. The fit was accidental, but it worked.
The question is how many accidental fits there are. The answer, based on the real estate data, is: fewer than the industry needs.
The Cooling System Is the Interesting Part
Here’s the part that doesn’t get enough attention, and it’s where the engineering details get genuinely interesting.
Data center cooling systems are among the most sophisticated HVAC plants ever built. They have chillers, cooling towers, economizers, hot and cold aisle containment, and increasingly direct-to-chip liquid loops and immersion tanks. When the data center is abandoned, this plant is often the most valuable remaining asset, and it’s also the hardest to repurpose.
Some examples of what’s actually being done:
District heating. Stockholm’s data centers already reject waste heat into the city’s district heating network. The cooling plant becomes a heat source. This is the highest-leverage reuse, but it requires proximity to a heat customer, which most abandoned data centers don’t have.Greenhouse heating. A handful of projects in the Netherlands and Iceland use data center waste heat to warm commercial greenhouses. The fit is good because greenhouses need steady, low-grade heat, which is exactly what data centers produce.Aquaculture. Fish farms need temperature-controlled water. Data center cooling loops can provide it. Pilot projects in Norway and Singapore have shown promise.Industrial process cooling. Some manufacturing processes need chilled water. A repurposed data center can supply it, provided the building is near the industrial customer.Desalination pre-heating. This is speculative but real: data center waste heat can pre-warm seawater before desalination, reducing the energy required. Pilot projects in the Middle East are underway.
The pattern: cooling systems are surprisingly adaptable, but only if someone is thinking about the second life before the first one ends. In practice, almost nobody is. The default outcome is scrap or abandonment.
If you’re an engineer working on data center design, the implication is direct: design the cooling plant for disassembly and reuse, not just for the initial workload. Use standard fittings. Document the refrigerant charge. Leave service manuals in the building. The second life of the plant is a design variable, and right now it’s not being treated as one.
The Archaeology of Compute
There’s a discipline emerging here that doesn’t have a name yet, but I think of it as the archaeology of compute. It’s the study of what a society leaves behind when its computational infrastructure moves on, and what those remains tell us about how we thought, what we valued, and what we couldn’t see.
Consider what a future archaeologist would find in an abandoned 2024 data center:
Racks that were designed for a specific power density, now useless for current hardware.Cooling loops sized for 30 kW racks, now undersized for 100 kW.Fiber runs that were overprovisioned for 2018 workloads, now barely adequate for 2026 inference at the edge.Backup generators that were sized for grid reliability assumptions that no longer hold.Security systems designed for physical threats, now obsolete against the threats that matter.
The building is a fossil record of a moment when the industry believed certain things about scale, location, and longevity. Those beliefs were reasonable in 2018. They’re wrong now. The building doesn’t know that yet.
The interesting question is not whether the archaeology will happen. It’s what we’ll learn from it. The data centers we’re building today will be the abandoned sites of 2040. What do we want the archaeologists to find?
What To Actually Do
Three moves, concrete enough to start on Monday.
- Treat the second life as a design requirement.
If you’re involved in data center design or procurement, add a “second life” section to every project brief. Where will the cooling plant go? Who will buy the chillers? What’s the refrigerant recovery plan? What’s the realistic reuse scenario for this building in 10 years? Very few projects have answers. The ones that do will be cheaper to decommission and more valuable to sell.
- Track the infrastructure lifecycle in your cost model.
Most teams model data center cost as a single capex number amortized over 20 years. That’s wrong. Break it into building shell (20 to 30 years), power infrastructure (10 to 15 years), cooling plant (10 to 15 years), and IT hardware (2 to 4 years). Once you separate the layers, the economic case for modular, replaceable infrastructure becomes obvious, and the case for monolithic builds becomes much harder to defend.
A rough starting point in code:def data_center_tco(
shell_capex,
power_capex,
cooling_capex,
it_capex,
shell_life_years,
power_life_years,
cooling_life_years,
it_life_years,
annual_opex,
years=20,
):
annualized = (
shell_capex / shell_life_years
- power_capex / power_life_years
- cooling_capex / cooling_life_years
- it_capex / it_life_years
- annual_opex ) replacements = { "power": max(0, years // power_life_years - 1), "cooling": max(0, years // cooling_life_years - 1), "it": max(0, years // it_life_years - 1), } total_replacement_capex = ( replacements["power"] * power_capex
- replacements["cooling"] * cooling_capex
- replacements["it"] * it_capex ) return { "annualized_cost": annualized, "replacement_capex": total_replacement_capex, "total_20yr": annualized * years + total_replacement_capex, }
If your model doesn’t include replacement capex for the layers that turn over multiple times, you’re underestimating cost and overestimating the building’s useful life.
- Build the reuse network before you need it.
If you’re an operator, start conversations with district heating authorities, greenhouse operators, and industrial cooling customers now, before the building is obsolete. The reuse deals that work take 3 to 5 years to negotiate. The ones that fail are the ones that start after the building is dark.
The Mental Model
Here’s the lens I want you to keep.
Data centers are not long-lived assets. They are stacked short-lived assets on a long-lived shell, and the mismatch between the shell’s life and the compute’s life is the defining economic and environmental problem of the AI infrastructure era.
The industry has spent two decades optimizing the shell for the compute of the moment. That worked when compute turned over slowly. It doesn’t work now. The buildings we’re constructing today will host 8 to 10 generations of hardware before their shells are ready to retire, and most of them weren’t designed for that.
The interesting question is not whether data centers will be abandoned. They will. It’s whether the industry will build for the second life before the first one ends. The teams that do will have cheaper decommissioning, more valuable real estate, and a defensible story for regulators and communities. The teams that don’t will leave behind buildings that nobody wants, in places nobody can use them, for reasons nobody will remember.
The archaeology of compute is already being written. The question is what we want it to say.
What’s your read? If you’ve worked in data center operations or real estate, what’s the realistic reuse scenario you’ve seen? And if you haven’t, what would it take to make the second life a design requirement in your organization? Drop it in the comments.
Top comments (0)