Fixed infrastructure, variable input
What a waste plants tell you about data center power infrastructure design.
A split image of a materials recycling facility and a data hall
Fifteen plus years ago I worked in the field of waste treatment infrastructure development. Here a facility was designed based upon a “waste composition study” - an analysis of what local bins contained at a given point in time.
A design may have assumed the incoming material was one percent “sand”. That figure came from the composition table and every piece was designed against that and a margin of tolerance. Every piece of equipment in the plant was sized against it. Trommels, which are rotating cylindrical screens that separate waste by particle size. Fixed screens. The conveyors and separation stages that pull recyclable material out of the stream. The digestion capacity that treats what remains.
If during operation, the sand arrived at three percent, on a facility handling 100,000 tonnes per year, that is an additional 2,000 tonnes of abrasive, sedimenting material that much be handled and disposed of every year, even through equipment sized for a third of it.
Nobody had underwritten the disposal cost. Nobody had allowed for the accelerated wear. The plant was not badly designed. The design basis was simply wrong, in the way design bases usually are, and it was wrong from the first day of operation.
That is the ordinary condition of fixed infrastructure with a long life. You commit capital against an assumption about an input you do not control, and then you live with the difference for twenty years.
The data center industry is doing the same thing now, at considerably greater scale, and it is treating the problem as a forecasting exercise. It is not a forecasting exercise. It is a specification problem, and the waste sector has already run the experiment.
What was specified in West Sussex in 2010
In July 2010, Biffa signed a 25-year contract with West Sussex County Council to treat the county's residual municipal waste. The facility was built next to the Brookhurst Wood landfill site near Horsham. It received planning consent in April 2010 and became operational in 2013. It was designed to handle up to 327,000 tonnes a year of residual municipal and commercial waste using mechanical biological treatment, a process that mechanically separates a mixed waste stream and then treats the biodegradable fraction through anaerobic digestion. The contract was reported at over £1 billion across the full 25-year term (Waste Management World 2010).
That is the part that was widely quoted. The more interesting part of the same announcement is what was left open.
The refuse derived fuel produced by the sorting plant had no committed buyer. The release states it would be sent to a third-party outlet or used in a thermal treatment facility, subject to a “joint exercise by Biffa and the council to find a long-term sustainable market”. The thermal treatment facility itself was conditional, described as something that would be built at Brookhurst Wood "if required" and subject to planning and operating consents. The digestate offtake was a discussion with an unnamed housebuilder. The biogas was to be used for electricity generation and "may in the future be harnessed to produce fuel for vehicles."
Four downstream pathways. None of these were fully settled at financial close on a billion-pound, twenty-five year commitment.
I do not read that as loose contracting. I read it as an accurate statement of what was knowable in 2010, written down rather than assumed away. The parties committed to the parts they could fix and left structured room on the parts they could not.
What happened in year thirteen
In April 2026 the facility was reconfigured. The anaerobic digestion capacity was expanded to process source-segregated household food waste, with capacity for the roughly 35,000 tonnes a year expected to be collected across the county. The council describes the driver plainly: the upgrade helps West Sussex meet new national recycling requirements West Sussex CC, 2026)
The requirements are the reason this matters. Under England’s separate food waste collection rules, households are now required to have food waste collected separately from residual waste (Gov.uk 2026).
Consider what this does to a mechanical biological treatment plant. The entire purpose of the mechanical stage is to separate the biodegradable fraction out of the mixed residual waste. Legislation now separates a significant part of that fraction at the kitchen bin instead. The function the facility was built for is now, in part, moved upstream by law, thirteen years into a twenty-five year contract.
The facility was converted rather than replaced. Biffa’s managing director for treatment services is candid about what that took “converting the existing plant has been a huge undertaking”. ENDs Waste and Bioenergy pointed to a GBP2.4m for remedial works for the MBT site to help the council meet the food waste obligations (Perella, M. 2026)
Two kinds of wrong
An incorrect specification of sand and a change in government legislation are not the same category of problem.
Sand at three percent instead of one percent is a first-order variance. This parameter moved within a range that could be anticipated. It is expensive but absorbable. Competent engineering already prices it through margin, oversizing and contingency. Most design practices treat this as the entirety of the input risk question.
Legislation changing the feedstock is a second-order change. This parameter was not exceeded, like the sand, it caused replacement via legislation. The rules governing what arrives at the gas were changed by an authority the asset owner doesn’t control. No amount of design margin on the original design basis protects against that change, only architecture does. Is there physical space and can the process stages be separated? Can the electrical and control systems accept a configuration that no one specified.
First-order variances are a design margin problem. Second-order changes are an architecture problem. Assets are rarely closed due to the first problem, but they can be readily lost to the second.
What this looks like in a data center.
The similarity is closer than it seems at first glance. The equivalent of a waste composition table is the operating load profile of the data center. It is produced during commercial negotiations, by people optimizing a leasing or financing agreement. It then results in a fixed asset with a twenty-year operating life. The developer does not know the ultimate load profile, nor does the power system supplier. In many cases the tenant also doesn’t know, because the hardware generation that will occupy the data hall has not yet been shipped.
This is a first-order problem and it is real. Facilities with high-density AI racks specified in 2021 had a draw of 10-20kW per rack. New facilities are being asked to accommodate racks drawing 100-140kW per rack (AFCOM 2021-26). Load behavior has changed as well as load magnitude. Large training clusters produce synchronized step changes in the region of 40% in seconds (EPRI 2025) across thousands of accelerators, which presents the power system with ramp rates that traditional data center load profiles never generated.
Density drift is the sand, the first order variable. It is the version the industry is already arguing about and margin handles a good deal of it.
The second-order version is different and it is moving faster. State cost allocation proceedings are redefining who pays for the network reinforcement a large load triggers (e.g., PUCO Case No. 24-0508-EL-ATA establishing AEP Ohio’s 85% minimum demand charge for data centers). Large load tariffs are being written that did not exist when facilities currently under construction were specified. Furthermore, the federal and state treatment of co-located generation - and whether an on-site asset can export, participate in grid services, or run outside emergency conditions - is actively litigated in landmark regulatory actions such as FERC Docket Nos. ER24-2172 and EL25-49, alongside FERC's AD24-11 technical proceeding on co-located loads.
None of those proceedings change how much power the facility needs, they change the terms on which it is allowed to receive it, and what the onsite generation is allowed to do. The food waste rule, arriving mid-life, at an asset that was financed on the assumption the rules were finalized.
The specification test.
If the design basis will be wrong, the useful question is not what the load will be. It is how wrong the design basis can be before the asset requires a capital event to keep serving it. That is answerable when the facility is being designed and is answerable as a contractual obligation on whoever supplies the power system rather than a claim about the equipment.
The questions are:
What load step can be accepted and rejected as a percentage of rated output, within what time, and inside what frequency and voltage envelope?
What is the sustained minimum operating point and what does running there do to efficiency and to maintenance intervals?
How many starts per year are permitted before warranty or overhaul terms change?
What headroom exists in the electrical architecture for sources that have not been selected?
What physical and civil provision exists for capacity that has not been specified?
What would have to change for the fuel pathway to change, stated as a condition rather than as an aspiration?
What can the control system accept that was not a priority at commissioning?
The answers belong in the contract, not in the proposal .
There is a more challenging version of the test and it comes directly out of the 2010 press release. Ask a developer to identify the conditional clauses in their own design basis. The parts they have written down as unresolved, with a stated mechanism for resolving them later. If there are non, the design assumes every downstream variable is settled. None of them are.
The objection
Flexibility costs money and the market rarely pays for it. That it true and should be clearly stated. A developer selling at commercial operation is rewarded for speed and capital efficiency. They have no economic interest in year-thirteen adaptability. The discipline described above adds costs and lengthens specification. Anyone arguing otherwise has not sat in the meeting.
Biffa did not build optionality into Horsham out of engineering virtue. They held a twenty-five year obligation to a public authority. The contract term created the incentive and the architecture made the incentive actionable, and both were required. Neither would have been sufficient alone.
That is what the data center transaction structure currently lacks. The party bearing the consequence of inflexibility is not the party specifying the asset. But the cost does not disappear at sale. It reprices, at diligence, at refinancing, and at residual value, against a holder whose obligation runs fifteen to twenty-five years while the contracted load underneath it runs for a fraction of that. A developer who cannot answer the specification questions is not avoiding the cost. They are moving it into the exit multiple, where it will be found.
One further case is worth noting, with a caveat attached. The ArrowBio demonstration facility at Hiriya in Israel used a water-based separation stage ahead of digestion, which coupled the process more tightly to the characteristics of the incoming material than a sole mechanical separation train does. It is no longer operating.
What to take away
Two facilities do not establish a rule. They illustrate a mechanism, and the mechanism is this: the more tightly a process is coupled to an assumed input, the less room there is to absorb a different one.
This is the argument behind the Structured Transition Model. Deploy for the constraint you have, preserve the ability to serve the constraint you will have, and treat the second as a specification rather than an aspiration. It is also why Infrastructure Half-Life matters as a planning concept. Horsham was reconfigured at almost exactly the midpoint of its contracted life. That is not unusual. It is what the midpoint of a twenty-five year asset life looks like when the world outside the fence keeps moving.
The plants still running twenty years on are not the ones that got the composition table right. None of them did.
Closing Comment
A waste plant is obliged to accept whatever arrives at the gate. A data center operator has some influence over what hardware goes into the hall. Control is precisely why the load profile gets fixed too early.
Five Nines and Fast Power
Making Better Decisions in AI Data Center Investments
References
AFCOM (2021-2026) State of the Data Center Report (Annual report series; e.g., 5th Edition 2021, 10th Edition 2026), Informa Markets (Data Center Knowledge)
DEFRA, Gov.uk (2026) Simpler household recycling rules come into force across England
Electric Power Research Institute (2025–2026) Key Characteristics of Data Center Power Demand (Speed to Power Initiative & EPRI/Epoch AI Report Scaling Intelligence: The Exponential Growth of AI's Power Needs), EPRI
Federal Energy Regulatory Commission (2024–2026) PJM Interconnection, L.L.C. & Talen Energy Susquehanna Co-Location Proceedings (Docket Nos. ER24-2172-000 & EL25-49-000 regarding behind-the-meter generation, non-firm transmission options, and ISA terms), FERC
Federal Energy Regulatory Commission (2024–2025) Commission Staff Technical Conference on Co-Located Large Loads at Generating Facilities (Docket No. AD24-11-000 regarding wholesale transmission cost recovery and on-site generation export rights), FERC
North American Electric Reliability Corporation (2025) White Paper #1: Characteristics and Risks of Emerging Large Loads (Reliability and Security Technical Committee, Large Loads Task Force), NERC
Perella, M. (2026) West Sussex approves £2.4m remedial work for MBT site, ENDS Waste & Bioenergy
Public Utilities Commission of Ohio (2024–2025) In the Matter of the Application of Ohio Power Company for Approval of a New Tariff Schedule DCT (Case No. 24-0508-EL-ATA establishing AEP Ohio 85% minimum demand charge for data centers), PUCO
Public Utility Commission of Texas (2025–2026) Rulemaking Implementation of SB 6 Regarding Large Flexible and Co-Located Loads (Project No. 58479), PUCT
Uptime Institute (2020–2025) Global Data Center Survey (Annual report series; e.g., 10th Annual 2020, 15th Annual 2025), Uptime Institute Intelligence
West Sussex Council (2026) Horsham recycling plant turns thousands of tonnes of food waste into green energy
Waste Management World (2010) Biffa Signs 1 Billion Waste Treatment Deal With West Sussex Council
Questions and answers
What is the difference between first-order and second-order design risk?
First-order risk is when a design parameter moves within a range that could have been anticipated. A waste facility designed for one percent sand receiving three percent is first-order. It is expensive, but margin, oversizing, and contingency absorb it. Second-order risk is when the parameter is replaced rather than exceeded, usually because an authority outside the asset owner's control changes the rules governing what the facility receives or what it is permitted to do. No amount of design margin protects against this. Only architecture does. First-order risk is a margin problem. Second-order risk is an architecture problem, and it is what strands assets.
Why did the Horsham waste facility need to be converted after thirteen years?
The facility was built to mechanically separate the biodegradable fraction out of mixed residual waste, then treat it through anaerobic digestion. England's separate household food waste collection requirements now remove a significant part of that fraction at the kitchen bin instead. The function the plant was designed to perform was partly moved upstream by legislation, thirteen years into a twenty-five year contract. The anaerobic digestion capacity was reconfigured in April 2026 to accept source-segregated food waste rather than the plant being replaced.
What does a waste treatment plant have to do with data center power?
Both are capital-intensive assets with lives of twenty years or more, sized against an input specification set by parties the owner does not control, and fixed at the moment of least information. In waste, that specification is a composition study. In a data center, it is the operating load profile, which is typically produced during a leasing or financing negotiation and then hardens into a power system design. In both cases the specification is wrong from day one, and in both cases architecture rather than margin determines whether the asset can respond when the rules change mid-life.
What is the Structured Transition Model?
The Structured Transition Model is a framework for infrastructure decisions under uncertainty. It holds that assets should be deployed against the constraint that exists today while preserving the ability to serve the constraint that will exist later, and that this second capability should be written as a specification rather than stated as an aspiration. It applies where capital is committed for decades against operating conditions, regulations, and technologies that will change within the asset life.
What is Infrastructure Half-Life?
Infrastructure Half-Life is a planning concept describing the point in an asset's contracted life at which the external conditions it was designed for have changed enough to require reconfiguration. The Horsham facility was reconfigured at almost exactly the midpoint of a twenty-five year contract. That is not an unusual outcome. It is what the midpoint of a long asset life looks like when regulation, technology, and markets outside the site boundary continue to move.
What questions should a data center developer ask a power system supplier?
The useful questions establish how wrong the design basis can be before the asset requires a capital event. What step load can be accepted and rejected, as a percentage of rated output, within what time, and inside what frequency and voltage envelope. What is the sustained minimum operating point, and what does running there do to efficiency and maintenance intervals. How many starts per year are permitted before warranty terms change. What headroom exists in the electrical architecture for sources not yet selected. What physical and civil provision exists for capacity not yet specified. The answers belong in the contract, not the proposal.
Why do developers not design data center power systems for flexibility?
Because flexibility costs money and the current transaction structure does not reward it. A developer selling at commercial operation is rewarded for speed and capital efficiency, and has no economic interest in the asset's adaptability in year thirteen. The party bearing the consequence of inflexibility is not the party specifying the asset. The cost does not disappear at sale. It reappears at diligence, at refinancing, and in residual value, against a holder whose obligation runs fifteen to twenty-five years while the contracted load underneath it runs for a fraction of that.