Why hyperscale data centres are the fragile choice

This is article 3 of 8 in The Third Number series. Here is article 2.

There is a kind of operations executive I have spent thirty-five years sitting across from, and for most of that time I thought they were right.

They can tell you the unit cost of every input on their line to four decimal places. They have spent a career hunting down slack and taking it out, because slack is what an unexamined budget hides in, and every gram of redundancy they removed showed up on the quarterly report as margin they had earned. Removed are the second supplier, second port, second line, second set of tooling: because each one is a cost with no revenue attached, defended by someone who cannot prove it will ever be needed. So, it goes. That is not laziness. That is thirty years of being rewarded, correctly, for exactly that judgment about costs based on only two numbers: the cost to acquire/build and the cost to operate.

Then came 2021, and a year of not being able to build anything.

One semiconductor fab went down for a few months and reshaped automotive production schedules on three continents. One canal blocked for less than a week and the ripples took months to clear. And the executive who had spent a career proving that redundancy was waste discovered, in the space of about nine months, that they had not been removing waste at all. They had been removing the only thing standing between an efficient system and a stopped one.

None of that was exotic risk. It was ordinary risk, sitting where it always sits, in systems that had optimised every gram of slack out of themselves because slack shows up on a quarterly report as cost and never as the thing that saved you.

The belief was not stupid. The belief was that efficiency and resilience are the same property, measured from different angles. In fact, they are close to opposites, and the bill for confusing them almost always arrives all at once.

A forest a season after a bad fire has no control room in the ashes, because there never was one. No headquarters burned down. No single point of failure took the whole system with it. The mycelium underneath is already rerouting around the dead patches, and seeds that needed the heat to crack open are already germinating in the scorched soil, because some species build the fire into the plan instead of pretending it will never come. Within a season there is green where there was black. Within a decade you need a trained eye to find the scar.

Industry has spent a century building the opposite of that, on purpose, because bigger looked cheaper. And we are now building the largest version of that bet any of us has ever seen.

Nature does not scale the way we scale

Nature never solves a hard problem at our kind of scale by building one enormous thing and asking it to hold its position and performance level. It solves hard problems by replicating small things that already work, and letting them talk to each other.

A coral reef is not one organism scaled up to reef size. It is billions of small polyps, each running its own show, networked into something that reads from a distance as a single structure and behaves up close as a federation.

A termite mound is not a building with a thermostat. It is millions of insects, each following a handful of local rules. Between them they throw up a lattice of thin outer flutes and deep central chimneys that turns the daily swing between hot afternoon and cold night into a kind of slow breathing. Air driven up the flutes by day and back down them after dark, flushing the colony’s carbon dioxide out through the walls.

Closer to a lung than to an air conditioner, running on nothing but the fact that the sun goes down, across a structure that dwarfs its builders by a factor of a thousand, with no blueprint and no foreman.

Nobody is in charge. Everybody is doing the same simple job in their own corner, and the ventilation shafts, the humidity gradient, the fungus gardens deep in the core, all of it emerges from replication rather than from enlargement.

Scale, in the natural world, is achieved by making more of the small working unit. It is almost never achieved by making the unit bigger.

And there is physics under it

That observation has a physics behind it, and the physics matters because it turns an analogy into a prediction.

In 1996 Adrian Bejan, at Duke, proposed what he called the constructal law: for a finite-size flow system to persist in time, its configuration must evolve so that it provides easier access to the currents flowing through it. It came out of a heat-conduction problem, the question of how best to connect one point to an infinity of points across an area or a volume.

Solve that access problem and what the law predicts is a hierarchy rather than a monolith. A few large channels and many small ones. One of Bejan’s own examples is the vascular system, a few large arteries feeding many small capillaries. He points out that river basins, tree branches and highway networks all resolve into the same shape for the same reason.

The branching is not decoration. It is what minimising resistance looks like when a flow needs to reach everywhere.

One qualification, because the argument is stronger for it. The constructal law is widely applied and widely cited, and it is not universally accepted as a law on the footing of thermodynamics. Some read it as a powerful organising principle instead, and there is a live argument between constructal accounts of biological scaling and the metabolic-theory accounts that compete with them.

I am using it for what it does regardless of that dispute, which is to explain why the shape keeps recurring over and over, and why the Constructal Law Conference is in its 16th year.

And two things follow that are important for you to digest, please.

The first is that the pattern stops being something nature just happens to do and becomes something flow systems have to do. A forest is redundant and networked because a system that has to move water and nutrients and information across a large area, and persist while doing it, has no better tested configuration available.

The second is more useful. The constructal law makes no distinction between the living, the inanimate and the engineered. It is stated about flow. And a data centre is a flow system: power in, heat out, water through, information across.

Which means the argument for distributed architecture over enlargement stops being an analogy to biology and becomes the same physics, applied to any kind of processing site.

The same belief, built over and over

Go back to the executive from the opening, because the belief that cost them a year is not confined to a supply chain. It is how we expand everything.

A semiconductor fab that runs beautifully at one scale gets pushed to the next by adding capacity to the same lines, the same utility loops, the same single points of control, because the capital case for expansion always looks better than the case for a second, smaller, redundant facility. It works, until a single contamination event or power interruption or bad batch of ultrapure water takes the entire facility down at once, because everything was coupled to everything else in the name of efficiency.

Systems optimised past the point of redundancy, coupled tightly enough that a failure anywhere becomes a failure everywhere, do not degrade. They collapse. A distributed system loses a node and keeps running. A centralised system loses its core and stops.

And here is the part that should unsettle anyone who lived through 2021 and concluded they had learned the lesson. What most firms did afterward was add a second supplier and call the problem solved, which treats the failure as a sourcing error rather than as a design principle. The executive who now holds two suppliers, one port and one enormous facility has bought insurance on the smallest of their exposures.

Which is why supply chain is now a misleading name for what the better operators are actually building. A chain has one path and fails at its weakest link, and that is precisely the property that made 2021 so expensive. What comes after is closer to what a forest runs: local feedback loops that rebalance flows in real time, sensor-rich and edge-aware, with more than one route between any two points.

A chain optimises for cost per unit under stable conditions. A network optimises for the continued existence of the flow, which is the constructal criterion, and it is what you want when conditions stop being stable.

Which brings us back to the resented neighbour

Nowhere is this fork showing up faster, or getting resented more openly, than in the buildings we have started stamping across the American landscape.

A hyperscale data centre arrives with a jobs number and a tax revenue projection, and leaves, a few years later, with a fight over water rights, a fight over grid capacity, and a fight over whose electric bill went up to subsidise somebody else’s cloud storage.

Both sides of that fight are telling the truth. The industry got there by building the way industry always has when nobody made it price the third number. Concentrate everything, extract locally, push the strain onto whoever lives downstream or downgrid, and call the difference between the site’s cost and its market value a win.

Here is a number with its provenance attached, because the provenance is the checkable part. One assessment, a Microsoft and Biomimicry 3.8 study of industrial development in the Netherlands, found a built-out site delivering only about 13 per cent of the ecosystem-service benefit the same acreage would deliver intact. The water filtration, the flood buffering, the carbon sequestration, the habitat value. Everything the land was quietly doing before concrete and racks arrived.

That is a proprietary regional figure about industrial development in general, not a measured constant for data centres, and it is offered here as indicative rather than settled. Even read cautiously it points one way: the large majority of what that ground could be contributing is gone. Not banked anywhere, not offset anywhere. Lost, and it stays lost for the operating life of the facility unless somebody designs it back in on purpose.

That number is easy to read as an indictment. The correct reading is that an 87 per cent gap between what a site could do for its watershed and what it currently does is an opportunity that has sat unpriced for as long as the industry has built this way.

What closing the gap actually looks like

One large operator has been running a regenerative infrastructure programme across its global footprint, and it supplies checkable numbers rather than mood indicators.

Their biome-based site playbooks, meaning designs matched to the actual ecosystem a site sits inside, the local BiosVerse, rather than a single template stamped everywhere from Iowa to Ireland, now cover the large majority of a footprint of roughly 600 sites worldwide. That is past the pilot stage and close to becoming the organisation’s default way of building.

And the clearest design of the set did not go onto a data centre at all.

It went onto the new United States Coast Guard Headquarters, a building designed by HOK with Biomimicry 3.8. The roof and stormwater system was inspired directly by beaver dams, an animal that has spent millions of years solving exactly the problem a large campus has. How do you slow water down, hold it, and let it infiltrate and recharge, instead of sheeting off a hard surface straight into a storm drain and out of the watershed for good.

The beaver builds a structure that captures water in a series of small impoundments instead of one big reservoir, so the system degrades gracefully in a flood rather than failing catastrophically the way a single retention pond does when it is overwhelmed.

Applied to a terraced roof and to the site’s grading and drainage, that logic cut stormwater runoff by 47 per cent. It saved something on the order of 520,000 gallons of potable irrigation water every year, dropped surface temperatures 12 degrees against a conventional rubber roof, and put eight times more native trees and seven times more woody plants on the ground than the competing designs did.

A federal office building is not a data centre. But keep in mind the roof does not care. Grading is grading, a storm drain is a storm drain, and a hard surface shedding water into a watershed it does not recharge is the same failure whether the thing under the roof is a filing cabinet or electronics rack.

That transferability is the entire reason for organising this knowledge by function and pricing the Third Number rather than by building type.

The credit does not stop at the fence

Pollinator habitat planted around another site produced a measurable benefit out to 15 kilometres from the property line.

That tells you something important about how this works. The BiosVerse compatibility index credit radiates outward into the county the way the liability used to. A data centre that was a pure extraction point, water in, heat out, power drawn, nothing given back, becomes instead a node that is measurably improving the ecology of a 15-kilometre radius around it.

That is not charity. It is the same physics that made the reef and the mound work. A small unit, replicated and networked, doing real work in its own corner, adding up to something bigger than any one site could produce by trying to be bigger.

And it is not a credit that sits still. A pollinator population recovering this year is seed corn for a watershed recovering over a decade, which is seed corn for a county that stops fighting the next site and starts asking for one.

What should matter most to whoever holds the capital budget? Every one of these moves posts to the ledger twice.

Waste heat once dumped into the air becomes district heat piped to homes, or warmth sold into a greenhouse. A liability becomes a credit and revenue on the operator’s books at once. Water reclaimed and returned cleaner than it arrived is a credit to the watershed and water the operator does not buy twice from a municipality already fighting them over allocation. Power generated on site is a credit against the grid strain that turned the county against the project, and power the operator is not exposed to buying on a spot market when demand spikes.

Run that all the way through and you stop having three separate ledgers. The operator’s cost curve, the host community’s grid and water position, and the watershed’s ecological account are not three different books. Held to the same standard over generational time, they are one ledger, and every design choice that raises compatibility lowers generational cost on all three sides at once.

How it arrived, and what that tells us

None of this came from one operator building one perfect campus and everyone else copying it.

It arrived the way the forest arrives. A playbook developed for one biome, tested against real weather and a real water table, then adapted, not copied, for the next site, until roughly 600 of them run variations on the same logic without any two being identical.

That is replication of a working unit, not the enlargement of one.

And the standard itself will have to follow that pattern. A single central authority cannot hold a compatibility standard for every biome any more than a single retention pond can hold every flood. The standard has to be evidence-gated and community-gated, held by a body with nothing to sell, so it can be replicated site by site without becoming the single point of failure this whole article was written to warn against.

Two more numbers belong next to each other. Every person on earth now brushes up against a data centre many times a day, in every text, card swipe, video call and search, and public trackers put the number of these facilities in the many thousands worldwide, upward of ten thousand by some counts, with more breaking ground every quarter.

That is not a specialty industry. It is closer to civilisational infrastructure, sitting under nearly every digital act of a normal day.

A system that touches every human being so often and now so pervasively cannot afford to be built the way industry defaults to building today. Concentrated, single-sourced, optimised for one steady state, blind to the Third-number account until the county shows up angry about the water table.

Four billion years of evidence says that system will not degrade gracefully. It fails all at once, and at that scale the failure does not stay inside the fence line.

The forest we started in is still there because it was never one thing to begin with. It was millions of small things, redundant, networked, individually replaceable, collectively unstoppable, each one posting its own small credit to an account that adds up, generation after generation, to something no single enormous thing could ever hold.

That is not a lesson about trees. It is a design brief.

What we are asking for

The site-level numbers that are the basis for this piece came from operators willing to publish them, and there are not nearly enough of them. The BiosVerse-Compatibility Index (BCI) scores whether a design fits. The Generational Cost of Ownership (GCOO) model prices what the fit or the misfit costs across generations. Both exist and both run today. What is not settled is how they are governed: who holds the standard, what counts as evidence, and who has the standing to fail a project. That is the work of Biomimetics26, on November 9 and 10 at the Minnesota Landscape Arboretum, and it is open. Bring what your own site actually measured. Runoff, recharge, surface temperature, habitat, distance to the last credit that left your fence line.

Next: a chemist stood in front of the American Chemical Society, inhaled a lungful of his new refrigerant, and blew out a candle with it. The room was floored. He had just shipped one of the largest liabilities of the twentieth century.

illuminem Voices is a democratic space presenting the opinions of leading Sustainability Thought Leaders, their views do not necessarily represent those of illuminem.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.