0.1 · The mapFrom a substation to a screen

Read left to right. Each box is a layer with its own owners and its own economics. The label on each arrow is the contract that moves the asset to the next layer. The two bands beneath are the money and paper threads, which run the length of the chain.

Holders named are examples from public sources, to be verified in the chapters. The highlighted row is our position.

0.2 · The questionsEight questions, asked of every layer

Each chapter answers the same eight questions in the same order, so layers can be compared like for like. The answers to the last question are collected in chapter 9.

0.3 · The chaptersNine chapters and an appendix of cards

0.4 · The playersEight deep, the rest on cards

UK only, and spread across the chain rather than clustered in one layer. The deep eight are the firms whose behaviour sets the terms we will be offered. The cards cover everyone else a tenant, lender or broker might mention. The seed list is the operator and broker target list of 31 October; this study will revise it.

Deep eight

Cards, by layer

0.5 · MethodOnline only, labelled, with the gaps stated

• Sources by layer. Grid: DNO heat maps and long-term development statements, NESO registers, Ofgem decisions, our own power.yeats.com data. Site and hall: RNS announcements, planning applications, agency research, trade press. Compute and capacity: company filings and prospectuses, funding announcements, job adverts (which say where a firm is building before the press does), customer announcements. Money: lender press, bond documents, filings. Paper: law-firm commentary, published lease summaries, government policy papers.
• Labels. Every capacity figure is marked Confirmed, Indicated or Assumed. Every money figure carries its source and date. Dollar and euro figures are converted to sterling and say so.
• What online cannot tell us. The rent actually paid on a UK powered shell; the terms of any backstop; a DNO's offer for a specific site; a tenant's true utilisation. Where we cannot know, we say so and triangulate from the nearest public evidence rather than fill the gap.
• The US as precedent. American structures (miner conversions, lease backstops, gigawatt campuses) suggest where the UK may go. They are presented as precedent, not as UK facts.
• Sequence. One chapter at a time, read and steered before the next starts. Each chapter ends with a "what it means for us" paragraph; chapter 9 collects them.

0.6 · Carried forwardWhat we already know, to be tested

• Ten megawatts is the sovereign and edge inference size in Britain: Civo at about 8 MW a site, Nscale and BT at 14 MW over three sites, a neocloud taking 14 MW at Romford, Nebius at 22 MW at Harlow.
• Confirmed power is the scarce input. Nscale's Loughton site was grid-delayed and lost its anchor customer in October 2026. A site with an agreed supply this decade carries a premium; chapter 2 should try to size it.
• Google has backstopped a neocloud's leases at two US sites in return for warrants. No UK example has been found. Chapter 5 asks whether one is coming.
• Fluidstack moved its headquarters to New York and left a French project; it is a structure to study, not a UK tenant.
• Marketplaces (SF Compute, Lepton) trade GPU-hours, not megawatts. They matter to our tenant's revenue, not to our rent.
• AI Growth Zone benefits attach to designated sites; none is in Kent or London.

0.7 · OpenTo settle before chapter 1

1. Title. "From power to answer" is the working title. Alternatives: "The chain", "Who sells the compute".
2. Voltron. The earlier book names Voltron as the proposed anchor tenant. This study can treat Voltron as one layer 3 player among several, or leave it out and keep the piece purely market. We suggest one card.
3. Length. Twenty-five to thirty pages of story plus the cards, with one infographic per chapter. If that is too long for the readers, the chapter 9 summary and the map can stand alone.

1.1 · What it isThe asset is an agreed supply, not a wire

Every building in Britain is connected to a distribution network at low or high voltage. What matters for compute is not the connection but the agreed supply capacity written into the connection agreement: the MVA the network operator has promised to deliver, firm, at that point of supply. A building's physical substation might be rated at 2 MVA while its agreement says 1 MVA; the agreement governs. Anything above the agreement needs a new application, and that application joins the queue.

The network has four tiers between a building and the national grid. Headroom at any one of them can bind.

A 10 MW demand connection normally lands at a primary at 11 kV or 33 kV. Whether it needs a transmission assessment depends on the headroom at the grid supply point above it, not on a fixed MW threshold.

1.2 · WhoSix groups, fourteen licences, one system operator

Distribution is a regulated monopoly by area. The licences were privatised in 1990 and have changed hands many times since; two changed in the last two years. The regulatory settlement is the same for all six and matters more than the owner, but the owner shows whose capital is behind the reinforcement programme.

Above them all sits NESO, the National Energy System Operator, publicly owned since October 2024. It runs the transmission queue, assesses large demand for transmission impact, and is now being directed by government on which demand projects go first. Ofgem sets the price controls (RIIO-ED2 runs April 2023 to March 2028 with £22.2bn of allowed spend; ED3 from April 2028 is being designed now) and writes the connection rules.

1.3 · What they doHow a connection happens, and what it costs

1. Budget estimate. A desk exercise with no site visit. For 1 MVA and above a DNO must return it within 20 working days under the guaranteed standards.
2. Application and formal offer. For high-voltage work the standard is 65 working days to a quotation. Since 1 June 2026 any application of 5 MVA or more must carry a letter of authority, heads of terms if the applicant does not own the site, a site plan, a single-line diagram where the kit could disturb the network, and a preliminary programme with a credible connection date. "As soon as possible" is no longer accepted.
3. Acceptance and milestones. Offers are open for a fixed period (ENWL's is 180 days; UKPN's should be checked). Acceptance starts the milestone clock. Where the connection date is within five years UKPN works the milestones forward from acceptance.
4. Charges. Since the Access reforms of April 2023 a demand customer pays for the sole-use assets to its site and nothing towards reinforcement of the shared network, unless the project trips the high-cost threshold of £1,720 per kVA or is deemed speculative. Reinforcement is funded by the DNO and recovered from all users through use-of-system charges. In effect, a demand customer in Britain does not pay for the shared network it triggers.
5. Energisation. The physical works. The binding constraint in 2026 is plant: transformer and switchgear lead times have lengthened, in places doubling, and now set the programme.

1.4 · What brokeThe queue, in numbers

Nscale's Loughton campus is the example. Planned at 50 MW rising to 90 MW, it lost its anchor customer in October 2026 when Microsoft withdrew, and is now reported for the 2030s. The constraint was 90 MW of new connection in Essex, not the building.

1.5 · How it is being fixedOfgem, government and the networks

Ofgem: Curate, Plan, Connect

• Curate (consultation 29 July to 16 September 2026, decision expected by the end of 2026): a commitment fee for data centres above 40 MW, proposed at 2.5% to 7.5% of average data centre capital cost, which briefings convert to roughly £240k to £710k per MW of requested capacity, held from offer acceptance to energisation and applied to the existing queue with exemptions near energisation; plus data-centre-specific milestones requiring evidence of customer demand, procurement, finance and technical maturity. Respondents, including the DNOs, asked for the same rules at distribution as at transmission to stop projects arbitraging between the two.
• Plan: Regional Energy Strategic Plans, published January 2026, give the DNOs a demand pathway to invest against in ED3.
• Connect (update 16 June 2026): let demand customers self-build and own, or self-build and transfer, high-voltage assets, with an independent transmission owner licence consulted on this autumn; offer large demand a high-voltage connection rather than forcing a step-down; and review connection terms so data centres can run flexibly, since most can be designed to shed or shift some load.

Government: strategic demand

The DESNZ consultation of March 2026 (closed 15 April; summary updated 28 October 2026) uses the Planning and Infrastructure Act 2025 to: tighten queue entry for data centres across transmission and distribution together; designate a list of strategic demand projects, including AI Growth Zones, to which NESO must reallocate capacity freed when others leave the queue and for which it may reserve capacity ahead of need; and, later, align data centre connections to a national data centre strategy. A Connections Accelerator Service has run since December 2025 for a small number of named projects. Distribution-connected projects are not yet inside the prioritisation rules; the paper says a parallel process will be designed "as soon as possible".

The networks: 5 MVA and above

The ENA's minimum information requirements for demand of 5 MVA and above took effect on 1 June 2026 across all DNOs: letter of authority, heads of terms, site plan, single-line diagram and a dated programme. The rule is aimed at speculators holding capacity on land they do not control. It also means an application in our name needs the site under contract first.

1.6 · Why nowWhy the DNOs will reinforce, and why not fast enough

The DNOs are regulated to a revenue allowance, so they reinforce when Ofgem lets them. ED2 allowed £22.2bn across the six groups for 2023 to 2028. The National Infrastructure Commission put the distribution need at £37bn to £50bn to 2050, at least double current annual allowances, and Ofgem's ED3 methodology decision of May 2026 moves to a build-and-flex model with the regional plans as the demand reference and a higher allowed return on equity. National Grid's chief executive said in 2024 that data centre demand would rise six-fold in ten years. The funding is in place or planned; the constraints are plant and people, and reinforcement at a bulk supply point still takes years.

Meanwhile the DNOs' own 2030 forecasts show headroom falling at 1,820 of the 2,940 primaries that publish one, as heat pumps, vehicles and housing take the slack. Headroom published today may not be there in four years.

1.7 · Where the power isPrimaries with 10 MW or more to spare

Our own schedule reads the published headroom at 4,277 primary substations across the fourteen areas and tests it against each operator's 2030 forecast where one exists. The map and table below count primaries with 10 MW or more spare, the size that matters to us.

power.yeats.com, read 31 Oct 2026, from each DNO's published capacity data and long-term development statement. "Hold to 2030" counts primaries whose own operator's forecast keeps 10 MW or more spare; 209 of the 1,029 publish no forecast. All figures Indicated: a verdict band is never a DNO confirmation.

UKPN's three areas hold 251 of the 1,029 primaries with 10 MW spare, and 178 of the 536 that hold to 2030. For this size of load our DNO is the best placed in the country. Central Scotland has the most spare capacity and the least demand to use it, which is why the AI Growth Zone went to Lanarkshire. The Midlands have the most industrial floorspace but publish no 2030 forecast, so a building there is harder to underwrite than a weaker one in Kent.

1.8 · What it means for usBuy the agreement, not the application

• Agreed supply. A building with 5 MVA in its connection agreement is outside the queue, the fee and the milestones described above. A building with a 2 MVA agreement and a 10 MW application is inside all of them. We should price the two differently.
• Confirmation before exchange. The agreed supply, the point of supply and whether capacity survives a change of occupier are three questions to put to the DNO in writing, and the first gate in our process. Capacity that has been de-energised for years is at risk of review.
• Size. Ten MW lands at a primary, normally below any transmission assessment, below Ofgem's 40 MW fee threshold, and inside the band where 1,029 substations have room today.
• Reinforcement. The DNO pays for shared reinforcement; we pay for the sole-use works. Where a supply needs raising, the cost is modest and the time is the risk.
• DNO. UKPN leads the country on primaries that hold 10 MW to 2030, so Kent and London are the right areas. Its new owner, Engie, is an energy group rather than a financial holder, which may change how it engages on flexible and self-build connections.
• Private wire and self-build. Ofgem's Connect reforms make self-build of high-voltage assets and behind-the-meter supply more legitimate. For a site beside generation, a private wire is a second route to firm power, and the Westwood plan already tests it.
• AI Growth Zones. They will not help us directly, and prioritisation at distribution level is still unwritten. We should watch for the parallel process the government has promised; a cluster of several buildings on one primary might one day qualify as strategic demand.

1.9 · Sources

• DESNZ, Accelerating electricity network connections for strategic demand, consultation Mar 2026, outcome page updated 28 Oct 2026: gov.uk
• Ofgem, Demand connections reform call for input (Feb 2026), Connect update and summary of responses (16 Jun 2026): ofgem.gov.uk; Curate proposals: Ofgem, Addleshaw Goddard, Browne Jacobson, CMS
• ENA demand queue entry requirements, 5 MVA and above, from 1 Jun 2026: Freeths
• NESO queue results and pre-2030 capacity: DCD, Savills, Slaughter and May
• Access SCR, shallow connection boundary and £1,720/kVA threshold: CMS, National Grid ED
• Guaranteed standards timescales: Ofgem 2025, ENWL Apr 2026; UKPN milestones: UKPN
• UKPN sale to Engie: RNS 7 May 2026, Energy Connects; UKPN 30 to 40% utilisation: Utility Week
• Other DNOs: NGED/WPD, ENWL/Iberdrola, Northern Powergrid, SSEN
• Price controls: ED2 £22.2bn, ED3 framework, Oxera on ED3 SSMD
• Market: Burges Salmon via Data Centre Magazine (7-year wait), Sifted and City AM (Loughton), DCD (six-fold, Mar 2024), London figures via Datacentres.com, gas generation: eceee
• AI Growth Zones: gov.uk policy paper, Osborne Clarke
• Capacity map and table: power.yeats.com, DNO published headroom and long-term development statements, read 31 Oct 2026.

2.1 · What it isThree products, one scarce input

A powered site is land or a building with a firm electricity supply already agreed with the network operator. The industry sells it in three forms, and the form decides who funds what and what the rent is called.

Grid MW and IT MW are different numbers. A fitted hall quotes IT load, the power the racks draw. A powered shell quotes the agreed supply at the intake, which must also run cooling, losses and reserve. The gap is the operator's efficiency: a well-run liquid-cooled hall turns roughly 1 MW of supply into 0.8 MW of IT load; an old air-cooled one, less. Our rent is per MW of firm supply. Each comparison below states its basis. UK Power Networks has observed that data centres on its network draw on average only 30 to 40% of the capacity they hold. Operators therefore guard their agreements, and a building with a large, under-used agreement is worth looking at.

2.2 · WhoThe holders of power in Britain

2.3 · How they make moneyWhat the market pays for power

Land

Buildings with power, sold

Our model carries £4.0m per MW all in (£2.5m building, £1.5m works). It sits inside the range paid for converted and edge stock in Britain and America. The Cogent asking prices show what a seller tries for when the site is already fitted and running.

Powered shells and fitted halls, leased

The American figures are contract revenue over critical IT load and may bundle power pass-through and services; they are turnkey, not shell. They are precedent, not UK fact. The CBRE figure is a capacity price whose period is not stated in the release; per month is the industry convention.

Capital per MW and income per MW a year across the first three layers, from the figures above. Our position carries the smallest capital for the smallest rent; the fitted layer is two to four times both; the compute layer is another £26m per MW, funded by the operator. Sterling at October 2026 rates; US and CBRE figures on the bases stated in the tables.

2.4 · The American conversionsMines, switch sites and exchanges, converted

The nearest precedent for buying power as a building is the United States in 2024 and 2025, where bitcoin miners, telecoms carriers and infrastructure funds turned existing powered sites into AI capacity faster than new capacity could be built.

1. The miner conversion. TeraWulf, Cipher, Applied Digital and Core Scientific owned sites with hundreds of megawatts of agreed supply, bought for mining when power was the only thing that mattered. They leased them on 10 to 25 year terms to neoclouds (Fluidstack, CoreWeave) at turnkey rents of roughly $1.5m to $2.3m per MW a year, funding the fit-out themselves at $8m to $10m per critical MW. In the Fluidstack cases Google guaranteed the tenant's lease in return for warrants over the landlord. The land and the power were bought at one price and leased at another.
2. The switch-site conversion. Cogent inherited Sprint's long-haul network and with it switch buildings with power and fibre in nine metros. It fitted six and asked $10m to $17m per MW for them; a buyer walked away wanting more than half the price on vendor finance; I Squared then bought ten for $225m, $4.25m per MW, to seed an inference and edge platform. The asking price and the price achieved were far apart, and the eventual buyer was an infrastructure fund assembling a platform of 5 to 8 MW sites. It is the closest precedent for the Openreach exchanges.
3. The carrier sale-and-leaseback. AT&T sold 74 central offices for $850m and leased them back; Ziply is converting about 200 to colocation; Verizon is fitting some for inference. Telefónica in Spain has turned closed copper exchanges into 1 to 2 MW edge nodes, 100 planned, with a €93m EU grant; FiberCop in Italy plans about 100. The telecom model is many 1 to 2 MW nodes, below our 5 to 15 MW target.

2.5 · Why nowWhy a connection in hand is worth more each year

• Time. A new connection in London takes seven to thirteen years against three to seven elsewhere in Europe (Buro Happold for the GLA, December 2025). The queue reforms in chapter 1 shorten the wait for projects that pay and progress. They do not change the position of a building that already holds its supply, other than to make it rarer.
• Demand. Of the c.66 MW of London capacity completing in 2026, 96% was pre-sold (Savills); vacancy is about 8%. CBRE's London capacity price rose 12% in a year.
• The landlords are leaving the shell. Segro expects most future projects to be fitted halls in joint ventures; Tritax will only build its shell against a pre-let at 107 MW. The listed landlords are moving to larger, fitted schemes on longer timelines. Smaller shell schemes in existing buildings are left to others.
• The exchanges are coming, slowly. About 4,600 buildings with power and fibre leave BT's use from now to the early 2030s, through one landlord that has so far sold to housebuilders. The first exits have slipped a year. It is a channel to cultivate; it is not a supply that will reach the market before us.
• Power is now priced as the asset. Currie & Brown's 2026 note on adaptive reuse states that "power has become the defining constraint". QTS is putting up to 720 MW on the coal yards of Blyth power station. The hyperscalers buy powered land outright. A building with 5 MVA in its agreement now has buyers who were not looking for it before.

2.6 · What it means for usBuy at building prices, lease at power prices

• The premium is real and sizeable. In London the market puts £0.7m to £1.2m per MW on a connection before any building exists. A 10 MW building bought at its ordinary industrial value captures that premium on the day the DNO confirms the capacity. Rule 7 of the Edge method depends on that premium.
• Our rent. It is conservative against every fitted comparable because it buys less: £750k a MW on £4.0m against £1.4m to £1.8m on £9.5m to £18m. A tenant used to fitted pricing pays about 43% of the London rate for the power and funds the hall itself, which it usually prefers to control.
• Say which megawatt. Our rent is per MW of firm supply. A tenant will convert that to IT load at its own efficiency and compare with £ per kW. Put the conversion in the pack so the comparison is made on our basis.
• The gap. Nobody is doing 5 to 15 MW in existing buildings at scale. Gravity Edge is closest, at the fitted layer with BlackRock money. Segro and Tritax start at 50 MVA. TT Group sells to housebuilders. That is the gap shown in the table in section 2.2.
• Vendor finance. Cogent's buyers wanted more than half the price financed by the seller. When a converted site is asked at £7m to £13m per MW and the bidder wants vendor finance, the market does not accept that price for power in a building it did not choose. Our £4.0m sits where deals have closed.
• Next steps. Open a conversation with TT Group about the exchanges with real MVA before the 2028 batch. Put the land premium per MW on the appraisal sheet beside the purchase price, so every site shows what it was bought for and what its power is worth.

2.7 · Sources

• Segro: Data centre update RNS, Mar 2026, Slough pre-let RNS, DCD on the pivot, Park Royal, half-year results July 2026 via LSE
• Tritax: DCD, Construction Enquirer, planning via Sharecast, Baxtel
• Harworth: pipeline update Aug 2026, Skelton Grange RNS, DCD
• Exchanges: DCD analysis, DCD on delays, EG; our BT exchanges brief, 22 Oct 2026
• Land values: Savills via trade press, Apr 2026 and Allwork; Linklaters powered land per MW, 8 Jul 2026 (in our comparables note); Northern Virginia via Data Center Knowledge
• Transactions: Redcentric via W.Media, Oct 2025; Gravity Edge via Bisnow, Sep 2025; I Squared and Cogent via Reuters, May 2026; Loudoun 2026 data centre guidelines; all in our comparables note of 26 Oct 2026
• US leases: TeraWulf release, Aug 2025, TeraWulf JV, Oct 2025, Applied Digital, Core Scientific
• Pricing: CBRE, May 2026; Ofgem £9.5m per MW capex assumption via Linklaters, Sep 2026; Savills Spotlight UK Data Centres H1 2026 (vacancy, pre-sold share, retrofit view); Buro Happold for the GLA, Dec 2025 (connection times)
• Reuse: Currie & Brown, 27 Jul 2026; Southall, Stockport, Truman Brewery
• UKPN utilisation: Utility Week

3.1 · What it isKilowatts, with everything that keeps them on

A fitted hall is a powered site with the critical plant installed: the distribution and transformers beyond the intake, uninterruptible power supplies, standby generators, cooling, fire suppression, monitoring, carrier fibre and the racks themselves. The product is not space but conditioned, backed-up IT load, sold by the kilowatt per month with the electricity recharged at cost.

3.2 · WhoTwo tiers of operator in Britain

The campus operators run the London halls that hyperscalers and neoclouds lease at scale. The regional edge operators run 1 to 5 MW sites in the cities where enterprises and the public sector sit. The campus tier builds its own halls; the regional tier buys existing sites. Ownership is given where it could be confirmed from public sources.

Campus and wholesale operators

Regional edge operators

3.3 · How they make moneyThe hall's economics

A London hall costs £9m to £18m a megawatt to fit and earns up to £1.7m a megawatt a year gross at the CBRE rate, with power passed through. The operator's margin is the gap between that rent and its cost of capital. The margin is wide while vacancy is 7% and every completing megawatt is pre-sold. The risk sits in the plant, which is bespoke to a density and a cooling method that change every two years. A hall fitted for 10 kW racks in 2022 is being refitted for 100 kW racks in 2026.

3.4 · AnatomyOne megawatt of hall, and who pays for each piece

The figure below splits a fitted megawatt into its parts, with the landlord's items in teal and the lessee's in grey. The landlord's figures are from our own scope table; the lessee's are derived from the published shares of a c.£9m London hall (electrical 40 to 45%, cooling about 20%) and the cooling-method ranges quoted in 2026, so they are indicative and will vary by site.

The split at the intake switchgear defines our scope. On the landlord's side, c.£0.5m to £1.3m per MW with £1.5m carried; on the lessee's, £4m to £9m before a single GPU. The lease must set out where the split falls, in terms a lender's engineer can check.

Space is not the constraint. One megawatt at today's average density of 27 kW a rack is about 37 racks; at AI densities of 120 to 230 kW it is five to eight. With coolers, a megawatt needs roughly 650 sq ft. A 10 MW tenant needs about 6,500 sq ft of a building that may have 60,000.

3.5 · DensityRacks are now denser than air-cooled rooms were built for

The cooling decision belongs to the lessee. Whether a hall is air or liquid follows from the tenant's hardware, and it is their £1m to £4m a megawatt, not ours. Liquid cooling at the chip also makes the building less important. The heat is captured in a loop and rejected outside through dry coolers, so the room no longer has to move large volumes of air. An ordinary industrial shell with floor strength, a yard for coolers and a firm supply is a better host for a 2026 rack than a 2015 data hall built for air.

3.6 · Why nowConsolidation, density and a seller's market

• The market is tight. Vacancy is 7% in London, 96% of the 2026 completions are pre-sold, and prices are up 12% with indexation rising. Operators can choose their tenants.
• The regional tier is consolidating with new money: Pulsant's £187m facility, nLighten's £85m and I Squared's $500m, Gravity Edge's BlackRock backing, Latos's 40-site target. Each needs sites. None of them develops shells itself.
• The top tier is changing hands. Ark, with the Crown Hosting channel, is reported in talks with TPG at c.£3.8bn. Global Switch's owners have been trying to sell since 2021. The next owners will pay for growth, and growth needs megawatts that already exist.
• Density is forcing refits. The move from 10 kW to 100 kW racks means every operator is now retrofitting. A retrofit at $2m a MW in a building that already has power is cheaper than a greenfield at $11m that does not.
• Neoclouds pay the premium and move fast. CoreWeave took Digital Realty and Global Switch halls to be live in months; Nebius took 22 MW at Kao on ten years; an unnamed neocloud took 14 MW at Romford. Compute lands in the fitted layer first because that is where it can be live this year.

3.7 · What it means for usThe fitted layer is our tenant, and its numbers are our reference

• Two tenant types. A neocloud takes the shell and fits it for itself (chapter 4). A colocation operator takes the shell, fits it, and sells kW to several customers. The second is the better covenant and the more re-lettable hall; the first pays more and moves faster. The lease heads should work for both.
• The sums work for a regional operator. On our shell at £750k a MW, an operator that fits for £5m to £7m and sells near the London rate of £1.7m gross has hall economics the campus operators would recognise, in a building they would not have considered. That is the pitch to Pulsant, nLighten and Latos, and it should be made in their numbers, not ours.
• The split at the intake switchgear must be written into the lease. Everything in the grey bars of the anatomy is the lessee's to fund, maintain, refresh and remove. Everything in the teal is ours. A lease that leaves the split vague ends with the landlord owning chillers.
• Design for the yard, not the room. Liquid cooling rejects heat outside. A building with a secure yard for dry coolers and a floor that takes a 1,500 kg rack is a 2026 hall; a 2015 air-cooled hall is not. Our screening should weight the yard and the slab, not the raised floor.
• Ask for PUE and density in the first meeting. They decide how much IT load our firm MW carries, and so what our rent is worth to the tenant per kW. A tenant at PUE 1.2 gets 0.83 MW of racks from our megawatt; at 1.5, 0.67. The more efficient operator can afford our rent more easily.
• Watch Ark and Kao. Ark's sale sets a price for UK halls with a public-sector channel; Kao's landlord-of-choice position shows what the neoclouds want in a host. Neither will lease a 10 MW shell in Kent, but both will tell us what their tenants pay.

3.8 · Sources

• Market: Savills, June 2026 and Savills Spotlight UK Data Centres H1 2026; CBRE European Data Centres Figures Q1 2026; CBRE midyear review 2026
• Costs: Turner & Townsend index 2025–26; Ofgem capex assumption via Linklaters, Sep 2026; STL Partners, May 2026; cooling and electrical shares via COR Advisors and Encor; Segro RNS 2025 for Park Royal
• Density: AFCOM 2026 via Encor; Schneider Electric, Jul 2026; Schneider on Vera Rubin, May 2026 (in our demand note)
• Operators: Kao Data shareholders, L&G 2023; Ark and Elliott, DCD, TPG talks, Oct 2026, Crown Hosting; nLighten and I Squared, Proximity; Pulsant profile, Pulsant financing; Latos Cardiff; Global Switch; Vantage UK; Yondr; Equinix DC01UK
• Leases: BCLP on shell-and-core structures; Mayer Brown
• Landlord scope: our scope table of 26 October 2026 (£0.46m to £1.3m per MW)

4.1 · What it isThe unit is a GPU-hour, and the asset wears out

A compute operator buys accelerators, racks them in a hall, and sells their time: by the hour on demand, by the month reserved, or by the year and the cluster under a take-or-pay contract with a hyperscaler or a lab. The unit of sale is a GPU-hour; the unit of capacity is still a megawatt, because the chips are bought to fill a power budget.

Rental figures from the Uvik index of 7 October 2026 and the IntuitionLabs comparison, both vendor-maintained; treat as indicative. Each generation puts fewer, dearer chips in a megawatt and asks the hall for liquid cooling.

4.2 · WhoThe compute operators in Britain

4.3 · How they are fundedThe capital stack of a GPU

The money follows the contract. A neocloud with a signed hyperscaler agreement can borrow; one without cannot, whatever its chips are worth. Maturities are also short. Analysts describe a refinancing wall in 2026 to 2028, as GPU loans fall due before the contracts behind them have paid out. A lease we sign in 2027 puts its tenant inside that window.

4.4 · Life of a GPUWhat a megawatt of chips costs, earns and is worth

Our arithmetic from published prices: c.1,000 H100-class GPUs in a megawatt at c.£26k installed; rental at $1.6 to $4.2 an hour; 70% utilisation; UK commercial power at c.£0.19 to £0.24 per kWh with PUE 1.2. The right-hand panel indexes the on-demand H100 price to its 2023 peak and overlays straight-line book value at three and six years. Vendor indices, not audited figures.

The return on a megawatt of chips depends on rental price, utilisation and depreciation. The rental price fell about 75% from the 2023 peak to 2025 and has since risen: the one-year H100 contract index was up 38% between October 2025 and March 2026, and several first-party list prices rose again in the autumn. Utilisation decides how much of that price is earned. 81% of enterprises report their GPUs busy half the time or less, one vendor claims about 5% in practice, and UK Power Networks sees data centres draw only 30 to 40% of the capacity they hold. Depreciation is disputed. The hyperscalers book servers over five and a half to six years; some neoclouds use three. A well-known short-seller claims the industry has understated depreciation by c.$176bn for 2026 to 2028. The defence is that older chips move from training to inference and keep earning; CoreWeave has said expiring 2022-vintage H100 contracts re-booked at 95% of the original price.

A megawatt of chips must earn back c.£26m in three to six years while paying c.£2m to £2.5m a year for power and £0.75m to £1.7m for the hall. At today's prices and 70% utilisation it does, with margin to spare. At 30% utilisation, or at the 2025 trough price, it does not. The hall rent is the smallest number in the stack. A tenant in trouble therefore stops paying the landlord last and resumes first, because the chips cannot run without the hall's power.

4.5 · Who has stumbledConsolidation, not yet collapse

• No named neocloud bankruptcy was found in 2026. The casualties so far are projects and plans: Microsoft left Loughton for want of power; Fluidstack left France; Microsoft cancelled c.200 MW of US leases in 2025.
• The exits to date are mergers and asset sales: Ori into Brookfield's Radiant; Cogent's converted sites to I Squared at $4.25m a MW after asking $10m to $17m; a small listed company signing a letter of intent for a GPU cloud's assets.
• Forecasts point to heavy consolidation. Gartner expects more than 80% of neoclouds to cease independent operation by 2030; executives quoted in The Information see seven to ten survivors from more than 300 providers, a count up 55% in a year; a prominent founder predicts "more than 200" failures. These are forecasts, not events.
• CoreWeave, the sector's bellwether, is stressed rather than broken. Its bonds traded like distressed debt in October 2026 while its chief executive called the market irrational and Nvidia's chief said no backstop was needed. The contracts are real and the leverage is high.
• Exposure is concentrated. Nscale's largest customer is 52% of revenue, and its two biggest agreements are conditional on delivering capacity it has not yet financed.

4.6 · Why nowWhy the chips still come to Britain

• Inference is served close to its users. Deloitte expects inference to be about two-thirds of AI compute in 2026; McKinsey's inference demand rises from c.21 GW to c.93 GW by 2030, built "in greater metro and surrounding areas". Nebius's Token Factory at Harlow and Civo's 40 planned sites are examples of this in Britain.
• Sovereignty is a buying rule. The AI Research Resource's twenty-fold expansion, the £500m Sovereign AI Unit and the public sector's procurement frameworks create demand that must be served onshore, by operators who can prove where the data sits.
• Nvidia is supporting the neoclouds. Equity stakes, the capacity rent-back from July 2026, the CoreWeave commitment to 2032 and the reported $125bn backstop amount to a vendor underwriting its own channel. For our tenants that is a utilisation floor a lender can see.
• Power decides who delivers. Nscale's contracts are conditional on capacity; Loughton shows what happens when the capacity does not come. A chip without a connection earns nothing.

4.7 · What it means for usWill they still be paying in year eight?

• Underwrite the contract book, not the brand. Ask every bidder for contracted value, customer concentration, the share that is conditional on delivery, and whether the financing for that delivery is signed. Nscale printed all four in its prospectus; a private bidder should be held to the same standard.
• Prefer tenants whose debt is investment grade, or whose customer is. A CoreWeave facility rated on its contracts, a Lambda loan at Baa2, a Brookfield-backed Radiant and a Nebius with Microsoft behind it are different credits from a founder-owned operator with a 1 GW plan and no disclosed funding, however close its thesis is to ours.
• Ask who stands behind the lease. The vendor backstop covers the chips' utilisation; the customer backstop covers the rent. Chapter 5 sets out the Google structure. Ask the question at the first meeting.
• Price the refinancing wall. Our leases start in 2027, inside the 2026 to 2028 window. Where a covenant needs it, a rent deposit or guarantee covering the first two years is cheap insurance against a tenant whose loan falls due before its contract pays.
• Write the re-let right for a chip-owner's exit. If a neocloud fails, the chips belong to its lenders and the hall is stranded. The lease needs the landlord's right to re-let, a short cure period, and a clear line on what the departing tenant removes and what it leaves. The hall fit-out, if left behind, has value to the next operator. The chips will belong to the lenders and should not be counted.
• Keep the £750k in proportion. Our rent is the smallest number in a tenant's stack: about a thirty-fifth of the kit, a third of its power bill. It is affordable in good years and the last cost cut in bad ones. The covenant question is therefore whether the tenant survives, not whether it can pay us.

4.8 · Sources

• Prices: Uvik GPU price index, 7 Oct 2026; IntuitionLabs H100 comparison; Spheron, Oct 2026; SF Compute pricing via chapter 0 sources
• Nscale: S-1 announcement, Computing, Techstrong; Loughton: Windows Central, Baxtel; BT: DCD
• Nebius: Kao Data, 8 Jun 2026, IT Brief, Microsoft deal
• CoreWeave: Reuters via Taipei Times, 18 Oct 2026, Admiral Markets, Capacity, debt at $35bn; UK sites: Tech.eu
• Civo: Civo newsroom, Sep 2026, THG investment, 2021
• Radiant and Ori: Reuters via The Star, DCD, lease finance
• Funding: Lambda term loan, GPU debt as an asset class, Microsoft $33bn, Fluidstack, DataVita
• Depreciation: SiliconAngle, Nov 2025, Barchart, DCD on neocloud economics
• Failures and forecasts: Gartner via ExplainX, The Information via AI Weekly, TechSpot; utilisation: VentureBeat, Sep 2026 and Cast AI (in our demand note); UKPN via Utility Week
• Sovereign: Isambard-AI, one year, Computer Weekly on the compute roadmap, AIRR cloud procurement, Dawn
• Demand: Deloitte TMT Predictions 2026; McKinsey, Dec 2025 (in our demand note of 26 Oct 2026)

5.1 · What it isFour ways to sell a megawatt of chips

5.2 · WhoThe buyers

5.3 · The intermediariesWhat sits between a GPU owner and a buyer

None of the marketplaces buys megawatts. They matter to us only through our tenant's utilisation. A tenant that can sell idle hours into SF Compute or Lepton loses less in a bad quarter.

5.4 · The backstopsThree structures that put a strong name behind a weak one

A. The customer backstops the lease: Google and Fluidstack, 2025

B. The chipmaker guarantees the rent: Nvidia, Nscale and Ionic Digital, October 2025

Nscale's US subsidiary signed a ten-year triple-net lease with Ionic Digital for the Cedarvale facility in Ward County, Texas: 234 MW and about $2bn (c.£1.5bn) of contracted rent, reported at $65 per kW a month. On 14 October 2025 Nscale issued Nvidia warrants over up to 9,476,700 preferred shares, worth $60m (c.£45m) at an exercise price of a cent, and in return Nvidia agreed to guarantee up to $860.34m (c.£650m) of Nscale's obligations under the lease, reported as the first five years' rent, with further "default warrants" that become exercisable if Nscale defaults on a lease Nvidia has guaranteed. All of this is in Nscale's prospectus. It is the clearest precedent for a landlord of a young operator. The chip supplier stood behind the rent, and the landlord could borrow against it.

C. The chipmaker rents back the capacity: Nvidia, July 2026

From 1 July 2026 Nvidia rents back unused capacity from partner neoclouds at a fixed rate, which gives their lenders a utilisation floor; its commitment to buy CoreWeave's unsold capacity runs to April 2032; reports put the programme at up to $125bn within a $500bn financing platform. This protects the operator's revenue rather than the landlord's rent directly, but a tenant with a revenue floor is better placed to pay its rent.

What Europe has instead

No European or UK backstop of either kind was found. CBRE reports a different response. Capacity sold to neoclouds in Europe rose from 133 MW in the first nine months of 2024 to 414 MW in the same period of 2025, and to 420 MW in the first half of 2026 against 89 MW a year earlier, with 57% of the 2025 volume in the Nordics; and operators are "implementing measures to manage the additional risk", chiefly higher rents to offset build costs. Lawyers in the US note the structures are already evolving from full lease guarantees towards residual-value guarantees, where the customer promises the landlord a floor value at the end rather than every payment along the way.

5.5 · The money flowWho pays whom behind one 10 MW lease

Our arithmetic for a 10 MW shell let at £750k a MW to an operator running c.10,000 H100-class GPUs: gross revenue £70m to £200m a year at 70% utilisation on the October 2026 price range; power c.£20m to £25m a year passed through at PUE 1.2; kit c.£260m once; rent £7.5m. The dashed arrows are the three structures in section 5.4.

The operator's largest outgoings are its chips and its electricity; its rent to us is a few percent of its revenue. Its credit therefore depends on the end customer's take-or-pay commitment, and on whether that customer or the chipmaker provides credit support to the landlord or the lender. A tenant with a named off-taker and a backstop is a different credit from the same tenant without them, in the same building, paying the same rent.

5.6 · Why nowWhy the buyers are starting to stand behind the operators

• The buyers need the capacity more than they fear the credit. The miners' sites were the fastest megawatts available in America, and Google and Nvidia guaranteed the leases to secure them.
• The lenders demanded it. GPU-backed debt is lent against contracts; property debt is lent against rent. Both lenders looked through the neocloud to the end customer, and the warrant-for-guarantee trade gives them a named credit.
• Concentration is extreme. Five buyers account for most take-or-pay value, and Microsoft alone for c.$60bn. Any one of them can make or break an operator, as Loughton showed, and any one of them can make an operator bankable by guaranteeing its lease.
• Europe is about a year behind. Neocloud signings in Europe tripled in 2025 and quadrupled again in the first half of 2026. The structures that followed the volume in America are likely to follow it here. A UK landlord asking for one today should expect to be told it is unusual.
• The state is becoming a buyer. The £250m research-compute contract, Crown Hosting's awards, G-Cloud 15's AI listings and the MoD's edge lot add up to a public customer that pays slowly but reliably and needs its capacity onshore.

5.7 · What it means for usLease to the tenant, underwrite the buyer

• Ask about credit support in the first meeting. Who is the end customer, and will they, or the chipmaker, stand behind the lease? Three forms are now in the market: a guarantee of the rent for a period (Nvidia at Cedarvale), an undertaking to assume the lease or pay the fee on default (Google at Lake Mariner), or a residual-value floor. Any of them changes what a lender will advance against our building.
• Decide what we give for it. The US landlords gave warrants; we have no listed equity to give. The alternatives are a right of first offer on the next cluster, a share of exit value above a hurdle, or nothing beyond the rent where the operator wants the power badly enough. The Edge rules exclude revenue share. The decision should be made before the first operator meeting, not in it.
• Do not follow the European approach. CBRE's operators charge neoclouds more rent to cover the risk. Our rent is fixed at £750k a MW and is already a fraction of a fitted rate. We should hold the rent at that level and take a guarantee, rather than raise the rent in place of one.
• A named off-taker is worth more than a well-known tenant. A regional colocation operator with Crown Hosting contracts, or a neocloud whose UK cluster is sold to Microsoft or a lab under take-or-pay, is a credit; an operator with a 1 GW plan and a pipeline is not. Underwrite the end customer's commitment, not the tenant's plan.
• Use the public channel as a tenant's selling point, not ours. We will not win a Crown Hosting award; our tenant might. A building that can meet sovereign requirements, UK-owned, UK-sited, auditable, helps them sell to the state, and that should be in the pack.
• Marketplaces help utilisation and nothing more. A tenant that lists on SF Compute or Lepton fills idle hours. That counts in their favour but says nothing about whether they pay rent in year eight.

5.8 · Sources

• Buyers: Converge Digest, neocloud spending, Trending Topics (Microsoft c.$60bn), Meta and CoreWeave $14.2bn, Meta $21bn and Anthropic, Apr 2026, Capacity (OpenAI $22.4bn), Microsoft and IREN; Nscale contracts from its S-1 (chapter 4)
• Backstops: Nscale Form S-1, 18 Sep 2026 (Nvidia warrant instrument and lease guarantee), Ionic Digital, 14 Oct 2025, Davis Polk; TeraWulf 8-K, The Block, TeraWulf JV; Vinson & Elkins via Bloomberg Law; Measured AI on the guarantee pattern; Nvidia rent-back via Capacity
• Europe: CBRE via Europe RE (414 MW vs 133 MW), CBRE H1 2026 (420 MW vs 89 MW), Verdantix on Loughton
• Marketplaces: SF Compute, DCD, Latent Space wiki; Nvidia Lepton Europe, Tom's Hardware, Spheron on Lepton pricing; Shadeform and Vast.ai
• Inference platforms: Sacra on Fireworks, Baseten, Together AI
• Public sector: AIRR £250m, DCD, Find a Tender, G-Cloud 15, Crown Hosting, MoD DDAD framework, Sovereign AI Unit

6.1 · What it isA request, a budget and a token

At the top of the chain a user sends a request to a model and waits for the answer. What they buy is measured in tokens, roughly three-quarters of a word each. What they notice is latency: the time to the first token and the pace of the rest. Every layer below the user is there to make that wait short and that token cheap. Google alone says it processed 3.2 quadrillion tokens in a month by May 2026, seven times the year before and 330 times the year before that; its APIs run at about 19 billion tokens a minute. Those are company figures, unaudited, but the direction is not in doubt.

Budgets from vendor and practitioner sources (Nvidia, Telnyx, Ultravox, 2026); they differ at the margin but agree on the ranges. Network legs from the same sources and our arithmetic at 5 µs per km of fibre.

6.2 · WhoThe users and the model-makers in Britain

6.3 · How far from LondonThe rings

Our arithmetic. Dots are the 45 largest urban areas of Great Britain, c.33m people, sized by population. Rings are one-way fibre latency from a site at Ramsgate, Kent, at 5 microseconds per kilometre over a route 1.4 times the straight-line distance; double them for a round trip. Within 1 ms: 33% of those people (London). Within 2 ms: 59%, to Birmingham and Southampton. Within 4 ms: 92%, to Manchester, Leeds and Cardiff. Glasgow and Edinburgh sit at about 5 ms.

A Kent site reaches every English city in under 4 ms one way. A Virginia region is about 35 ms away; Amsterdam and Frankfurt are 3 to 5 ms. Latency therefore argues for Britain over abroad, not for one British town over another. For chat, agents and documents, Kent, Slough and Manchester are indistinguishable to the user. The exceptions are metro-scale. Mobile networks, vehicles, robotics and voice at scale want their compute within a millisecond or two, which is a circle of 100 to 150 km around the people, and for 59% of Britain's urban population a Kent site is inside that circle. Residential measurements agree: Edinburgh to London on an ordinary broadband line is about 14 ms round trip, London to Amsterdam about 6, London to Frankfurt about 9.

The map does not show Kent's subsea landings. The cables that join Britain to the Continent land on its coast: EXA's Margate to Ostend route opened in 2025, and a new system is planned for Dumpton Gap by Broadstairs, a landing station used before. A site in east Kent is as close to Brussels and Amsterdam as it is to Birmingham. That matters to a tenant selling into Europe from a British address.

6.4 · SovereigntyThe rules, and what buyers do in practice

As with the credit support behind leases in chapter 5, buyers are moving ahead of the rules. No statute makes a bank or a hospital keep its inference in Britain; most of them choose to anyway.

6.5 · What a token is worthDeflation at the top, inflation underneath

The price of a token has fallen by two orders of magnitude while the number of tokens has risen by more. The hour of compute underneath them got dearer in 2026 (chapter 4). The industry calls this the Jevons effect: cheaper inference leads to more of it. For the layer above us it means thin margins and constant repricing. For the layers below it means demand for GPU-hours keeps rising even as the price of their output falls. The operators who run inference cheapest per token, near their users and with power they already have, keep the work.

6.6 · Why nowWhy the users are pulling the compute towards them

• Use has gone mainstream in two years: business adoption has tripled, half of adults use AI and 29.6m people use it in a month. The workloads that follow are inference, not training, and inference is served near its users.
• The applications that need milliseconds have arrived: robotaxis in London from September 2026, AI-RAN trials on a live mobile network, voice agents in call centres, fraud models in payment flows. Each needs a short network leg, and the slowest 1% of requests is what the product is judged on.
• The buyers are regulated. Finance, health and government are among the heaviest adopters and the most constrained in where their data goes. The critical-third-party regime makes the location of critical services a board question.
• The labs are in London. OpenAI and Anthropic are building London headcount in the hundreds, and their British customers will ask where the compute they buy is run.
• The state has counted the gap. The compute roadmap's evidence annex warns of a 5 GW shortfall by 2030 without intervention, and says ultra-low-latency applications such as vehicles are the ones that cannot wait.

6.7 · What it means for usSell power, sovereignty and timing, not distance from London

• Do not oversell latency. For most inference a Kent site, a Slough hall and a Manchester shed are the same distance from the user. The case to make is that Britain's users need British compute, there is not enough of it, and our buildings have the power to host it now. Chapter 1 made the same case in megawatts.
• Name the tenants for whom distance matters. A mobile operator putting inference in the network, a vehicle company, a voice platform, a trading firm: for them a site within a millisecond or two of London and the south coast is worth something a Scottish site is not. The rings put 59% of Britain's urban population inside 2 ms of Kent. Put that in the pack for those tenants and leave it out for the rest.
• Sovereignty is something a landlord can offer. A UK-owned, UK-incorporated landlord with a UK building and a UK connection answers part of what a bank, a trust or a department will ask a tenant. Put that in the pack.
• Kent's coast is a selling point. The Continental landings at Margate and Broadstairs put Brussels and Amsterdam 3 to 5 ms away. A tenant serving Europe from a British address, or a European operator wanting a British node, should hear it.
• Token deflation is the tenant's problem, not ours. Our rent is per megawatt, not per token. As tokens get cheaper and more numerous, demand for powered megawatts rises and the margin squeeze lands on the layer above us. We therefore underwrite the tenant's contracts (chapter 4) and the credit support behind them (chapter 5), not their price list.

6.8 · Sources

• Adoption: ONS, Artificial intelligence in UK businesses, 2023 to 2026; Ofcom Adults' Media Use 2026 via Advanced Television; DSIT survey via Digital Watch; UKOM and IAB UK figures via ProfileTree; OpenAI global users, eWeek
• Latency: Nvidia on voice latency, Telnyx, Ultravox, Morph; UK and European route latencies: thinkbroadband forum measurements, EXA route, ISPreview; Dumpton Gap landing: SubTel Forum
• Labs and firms: OpenAI London, Tech.eu, Verdict, Knight Frank via IREI, Wayve and Uber, MoJ and OpenAI, EE AI-RAN
• NHS: Digital Health on the £20m X-ray rollout, £30m diagnostics, £900m framework, NHS offshoring guidance, NHS England hosting standard
• Sovereignty and resilience: Pinsent Masons primer, Lewis Silkin on CTP designations, Linklaters, Civo (vendor view), Compute roadmap evidence annex
• Tokens: arXiv 2603.28576, BenchLM token price index, Silicon Data, Google I/O 2026, Implicator on the token figure
• Population and map: 2021 census built-up areas, approximate; GB outline and projection from power.yeats.com

7.1 · The stackWho funds what, and at what price

Debt shares are typical structures by layer, drawn from the sources in this chapter; no single company is depicted. Costs as published in 2026. Leverage falls and the cost of money rises from the regulated grid to the venture-funded application.

7.2 · Layer by layerThe lenders, the equity and the rules they lend by

7.3 · Where the debt is sizedThree sizing rules

1. The regulator sizes the grid's debt at a notional 60% of the asset base and allows a return for which the market has paid a 50% premium. The connection a building holds was financed at this rate, which is why it is cheap to own and expensive to replace.
2. The property lender sizes our debt on the rent: the capacity rent of £750k a megawatt, from a tenant it can underwrite, at a cover ratio it can defend, against a building it could re-let. In our structure, development debt is drawn at purchase and the site is refinanced when the lease is signed. Only rent counts in the sizing. Revenue share, if it ever existed, would be equity upside only; the Edge rules exclude it in any case.
3. The GPU lender sizes the tenant's debt on the tenant's contracts. The spread between a loan at SOFR plus 2.25% and one at SOFR plus 5.50% is the difference between a Meta contract behind the chips and no such contract. The lender lends on the buyer's contract and holds the chips as secondary security.

Rules two and three meet at our lease. The property lender asks who our tenant's customer is, because the tenant's ability to pay rent in year eight depends on its contract book and its own lenders' patience. A lease that gives the landlord sight of the tenant's contracts, a deposit or guarantee where the covenant is thin, a short cure period and a clean re-let right is a lease a property lender will advance more against. The credit support behind the lease, set out in chapter 5, therefore matters to our layer even though we never fund a GPU.

7.4 · The cost of money, in one table

The grid borrows at about six because the regulator guarantees its revenue. A building borrows at six or seven because a lease is a contract and the building has another use. A fitted hall borrows at a little more because the fit-out has one use. The chips borrow at six when a hyperscaler has signed and at ten or sixteen when it has not. At every layer the rate falls when a buyer's commitment stands behind the cash flow.

7.5 · The money that is newThree things that did not exist two years ago

• GPU debt as an asset class. From a 15% loan in 2023 to an A3-rated $8.5bn facility in 2026, with a sector total above $20bn and a Baa2 syndicated term loan to a private neocloud. The rating sits on the contracts and the structure, not the borrower; the same borrower's unsecured paper yields twice as much.
• Data centre securitisation in Europe. Yondr's £532m of AA-rated notes against two let halls in Slough was the first European data centre ABS to reach that rating. It gives operators a public-market exit for stabilised halls and shows a property lender what a let megawatt is worth to the bond market.
• Vendor and customer credit support. Nvidia's rent guarantee at Ward County and its capacity rent-back; Google's lease backstops. This is the credit support behind the lease described in chapter 5. It exists because lenders would not otherwise lend at the rate the operators needed.

The money that is old is also moving. Infrastructure funds that used to buy towers and fibre now buy halls: Antin, I Squared, Macquarie, Infratil, Brookfield, Elliott, and the new L&G and ICG funds. Their cost of capital sits between the property lender's and the venture fund's. Their appetite is for platforms of many sites, and a regional edge operator is one.

7.6 · Why nowWhy the capital is arriving faster than the power

• The buyers' spend has doubled in two years and is heading for a trillion dollars in 2027. Every layer below is financed on the expectation that it continues. The hyperscalers' own filings show it continuing through at least 2027.
• Lenders have learned to look through. The investment-grade GPU loan and the AA data centre bond both exist because a lender agreed to lend on a hyperscaler's commitment several layers away. We expect the same look-through to reach a powered shell's lease.
• The refinancing wall is 2026 to 2028. GPU loans amortise over three to four years and the contracts behind them run longer. Operators who cannot refinance will consolidate, and their halls and leases will change hands. Landlords with a clean re-let right are protected; those without are exposed to the second-hand value of the plant.
• Power is the input the capital cannot buy. Virtus's £2.45bn and Pure DC's $2.7bn are waiting on connections. There is more money in the chain than there are megawatts, so a building with a connection commands a premium (chapter 2).

7.7 · What it means for usSix rules for our layer

• Stay at the building's cost of capital. Our layer borrows at six or seven against rent. If we fund a chiller we borrow at a hall's rate; if we fund a chip we borrow at ten. The boundary of landlord scope is also the boundary of cheap debt.
• Size debt on rent only, and tell every lender so early. Capacity rent, indexed, from a tenant whose contracts the lender can see. No revenue share, no turnover top-up and no uncontracted capacity in the sizing. This is the Edge rule, and it is also how the market's lenders size debt.
• Write the lease with the lender in mind. Sight of the tenant's contract book and concentration, a guarantee or deposit where the covenant needs it, a short cure period, a clean re-let right, and a statement of who stands behind the lease. Each of these moves the pricing of our debt towards the AA notes and away from the 16% bonds.
• Prefer tenants whose own money is cheap. A tenant borrowing at SOFR plus 2.25% has a hyperscaler behind it; one borrowing at SOFR plus 5.50% does not. Ask bidders what their GPU debt costs. It is the most informative single number about a bidder.
• Watch the securitisation market. If a let, powered shell can be financed in the bond market as Yondr's halls were, the exit value of our buildings is set there, not by the agents. The first UK powered-shell ABS will reprice the layer.
• Keep the three gates: the DNO's confirmation of the connection, the operator agreement, and funding. A lender will ask for all three.

7.8 · Sources

• Grid: Ofgem, ED3 Sector Specific Methodology Decision, Finance Annex, 21 May 2026 (Tables 1, 9, 10 and 11); UKPN sale via chapter 1 sources
• Property and halls: Standard Chartered on Virtus, Clifford Chance, Latham on Yondr, Linklaters, Pulsant refinancing, Sidley on nLighten, IPE, Echelon, Mondaq on covenants; Savills yields and Loudoun cap rates from our comparables note, 26 Oct 2026
• Compute: CoreWeave DDTL 4.0, GPU loan pricing tracker, Global Data Center Hub, CoreWeave 10.4% loan, Wing VC, how the GPU became collateral, Lambda; bond yields via chapter 4 sources
• Buyers: Yahoo Finance on 2026 capex, Moody's via DCD, Alphabet's July guidance
• Funds: L&G digital fund, A&M outlook (ICG, BlackRock), Macquarie and Virtus, 3i and Lefdal; venture rounds via chapter 5 sources
• Our structure: the Edge method's financing rules (debt on capacity rent only; three gates), as recorded in the Yeats Edge business plan and the A to Z book of October 2026

8.1 · The paper trailWhat moves the megawatt, and who signs it

8.2 · The connection paperWhose name is on the megawatt

• Named customer. A connection agreement is signed by the party who owns or occupies the site. Under the industry code change DCP181, bilateral connection terms at non-domestic, current-transformer-metered premises now endure on a change of owner or occupier, and the DNO must tell a prospective occupier whether non-standard terms apply. Under DCP114 and DCP115, no reduction in maximum import capacity takes effect without the customer's explicit agreement. Those two rules are the legal basis on which capacity survives a change of occupier. The DNO should confirm them in writing for each site, because a 2026 working group (DCP463) is discussing reducing capacity at sites where the customer cannot be identified.
• Firm supply. The DNO's confirmation should state the agreed supply, the point of supply, that the supply is firm (not curtailable), and that it is not conditional on reinforcement. Ofgem's Connect work is adding flexible and self-build options. Those should be chosen deliberately rather than inherited with the site.
• Application requirements. From 1 June 2026 any application of 5 MVA or more needs a letter of authority, heads of terms if the applicant is not the owner, a site plan, a single-line diagram and a dated programme. Above 40 MW, Ofgem's proposed commitment fee and milestones apply. The 5 MVA threshold is below our size and the 40 MW threshold above it; a 10 MW application sits between the two.
• Holder of the agreement. The landlord holds the connection agreement and the point of supply; the tenant takes a supply contract in its own name through the landlord's intake metering. Segro did this at Slough ("the customer will contract the power directly"). It keeps the capacity with the building when the tenant leaves.

8.3 · The consent paperPlanning, designation and the duties that follow

8.4 · The leaseOne operator lease on one page

Heads for a powered-shell operator lease, drawn from the Edge method's lease checklist, Hogan Lovells' note on UK data centre leases (30 September 2025), BCLP's note on shell-and-core structures (23 October 2025) and Gowling WLG's note on colocation agreements (16 April 2025). Teal markers are the clauses the landlord should lead; grey the ones the tenant will. Solicitors draft; this is the commercial brief.

BCLP reports that leases are "almost always triple net", that deposits of six to twelve months' rent and parent guarantees are typical for smaller tenants, and that the yield-up clause is "one of the most heavily negotiated", with most developers requiring tenants to strip the fit-out back to shell. Our position on yield-up is the reverse. The hall fit-out is what the next operator wants, so the tenant leaves it and removes the compute. Hogan Lovells says the lease must state who maintains the "core infrastructure, such as power and cooling systems". In a fitted hall that is the landlord; in our shell it is the tenant, and the lease must say so without ambiguity because the lender's engineer will read it. Gowling WLG notes that the operator "does not own, provide or manage" the customer's equipment, and that the agreement must give the operator rights if equipment is not removed at the end. The same right belongs in our lease against the tenant's compute.

8.5 · The credit paperWho pays if the tenant does not

8.6 · The power paperPass-through, private wire and the licence

• Pass-through at cost. Electricity is the tenant's largest bill and should not run through the landlord's accounts. The supply contract and MPAN sit in the tenant's name, with landlord metering at the intake for information only. Where a landlord supplies several tenants from one intake, it becomes a distributor and needs an exemption class: the 2001 Order's Class C covers distribution solely to non-domestic consumers, but stakeholders have told government the drafting of "on-site" and "private wire" is unclear, and a review is open.
• PUE and density in the lease. The tenant declares its design PUE and rack density; the lease caps consumption at the agreed supply and gives the landlord the right to require a reduced draw to protect the connection. Colocation operators use the same clause with their customers.
• Private wire and self-build. Ofgem's Connect reforms will allow self-build and ownership of high-voltage assets and are considering class exemptions for large demand; a site beside generation can take a private wire under the existing exemptions, within limits that need checking per scheme. Westwood is the first scheme where this applies to us.

8.7 · The data paperWhat follows the tenant in

• The Data (Use and Access) Act 2025 came into force in the main on 5 February 2026. The test for international transfers is now that protection is "not materially lower" than the UK's; a complaints procedure duty applied from 19 June 2026. These duties fall on the tenant and its customers. The landlord's part is to warrant that it is UK-owned, UK-incorporated and UK-sited.
• The critical third parties regime designated AWS, Google, Microsoft and Oracle from 13 July 2026, and the regulators are to assess AI and cloud providers next. A regulated customer of our tenant will ask where its critical services run and how they fail over. The security schedule to the lease (policies, breach procedure, ISO 27001, controlled access with tenant escort) is what the landlord can offer in response.
• Landlord's data obligations. Hogan Lovells puts UK GDPR and ISO 27001 compliance on the tenant and the security policy on the landlord of a fitted hall. In a shell the landlord's obligations are physical only: the fence, the gate, the cameras and the access list. The landlord should not promise anything further on data.

8.8 · What it means for usSeven rules for the paper

1. Connection agreement. It stays with the building. Before exchange, have the DNO confirm in writing that DCP181 and DCP114 apply to each site.
2. Planning. Before exchange, per building: a lawful development certificate or pre-application view on the compute use, a noise screen for residential neighbours, and the EPC position.
3. Rent. Per MW of firm supply, indexed with a collar and cap, phased only against binding dates. No turnover rent or revenue share; nothing else counts towards the rent the lender sizes on.
4. Scope. Landlord scope ends and tenant scope begins at the intake switchgear. Both are listed in a schedule, in plant terms, with the fit-out guide agreed at the agreement for lease.
5. Yield-up. The hall fit-out stays; the compute goes; the landlord elects what else is removed. Put this in the first heads of terms, because the market default is the reverse.
6. Credit support. Scheduled to the lease, not promised. Parent guarantee always; deposit where needed; the off-taker's or chipmaker's backstop as a schedule, with a direct agreement for the lender.
7. Power. Supply contract in the tenant's name; landlord metering for information only; a reduced-draw right to protect the connection; no multi-tenant distribution until the licence exemption is checked.

8.9 · Sources

• Connection: Ofgem decision on DCP181, DCP114 and DCP115, DCP463 working group minutes, 2026, grid connections legal note; ENA 5 MVA rules and Ofgem Curate via chapter 1; Segro Slough RNS via chapter 2
• Planning: Forsters on the NSIP route, Womble Bond Dickinson via Mondaq, parliamentary answer on use class, OPDC committee report, Premier Park, 2026, FTB Chambers note, RTPI
• Duties: CMS on heat network zoning, parliamentary answer on data centre waste heat, 2026, Knight Frank on the 2026 revaluation, VOA rating manual, computer centres, Browne Jacobson, HKA on lithium-ion hazards, Green Mountain noise assessment, Romford
• Lease: Hogan Lovells, 30 Sep 2025, BCLP, 23 Oct 2025, Gowling WLG, 16 Apr 2025; the Edge method's operator lease checklist
• Credit: chapter 5 sources (Nscale S-1; TeraWulf filings; Vinson & Elkins)
• Power: Electricity (Class Exemptions from the Requirement for a Licence) Order 2001, CMS on the exemptions review, DESNZ and Ofgem licence exemptions guidance, Jan 2026; Ofgem Connect update via chapter 1
• Data: Kennedys on DUAA commencement, Clifford Chance; critical third parties via chapter 6

9.1 · The seatThe chain, and our position on it

9.2 · What each chapter foundEight findings, eight consequences

9.3 · What we never doFive rules the evidence confirms

1. We do not fund the kit. The hall costs £4m to £9m a megawatt and is bespoke to a cooling method that changes every two years; the chips cost £26m and halve in price. Every pound spent inside the intake borrows at a hall's or a chip's rate. Chapters 3, 4 and 7.
2. No megawatt is quoted unless it is confirmed in writing. Nscale lost Microsoft at Loughton over a connection that was never confirmed. The reforms of 2025 and 2026 penalise exactly that. Chapter 1.
3. Debt is sized on rent and nothing else. Revenue share is excluded by our rules and by every lender's. Chapter 7.
4. Power does not pass through our books. The tenant's electricity bill is three times our rent. Chapter 8.
5. We do not depend on one tenant type or one operator. The exits in the compute layer are mergers and asset sales; the lease must let the hall pass to the next operator. Chapters 4 and 8.

9.4 · The tenantThe target list, revised by the study

The list of 31 October stands, re-ordered on credit and with one addition. The first five remain the first five calls; only their order has changed.

CoreWeave, Fluidstack, Lambda, Yondr, Ark and the London campus operators remain references, not targets, for the reasons in chapters 3 and 4.

9.5 · The askTen questions for every bidder, in order

9.6 · What we giveThe decision to take before the first meeting

The American landlords paid for their backstops in warrants. We have no listed equity. Three options remain, and we recommend trying them in this order.

1. Nothing but the power. Where the operator needs a connected building this decade, the building itself is the consideration. Nscale's position after Loughton suggests we try this first.
2. A right of first offer on the next cluster. Costs us nothing today, ties the tenant to our pipeline, and gives the regional operators and the neoclouds what both say they want, which is more sites.
3. A share of exit value above a hurdle, on that site only. The nearest sterling equivalent to a warrant. A lender can accept it because it ranks below the debt. Not a share of rent, and not revenue share.

9.7 · The next ninety days

1. Two DNO confirmations in writing. For each seed site: agreed supply, point of supply, firm, and confirmation that the bilateral terms endure on a change of occupier and that capacity will not be reduced without the customer's consent. Chapters 1 and 8.
2. Planning per site. A pre-application view or a lawful development certificate on the compute use, a noise screen for residential neighbours, and the EPC position. Chapter 8.
3. The one-page pack. The building, the MW with its evidence label, the programme, the lease heads of chapter 8, the conversion from grid MW to IT kW, the sovereignty wording, and for the metro-latency tenants only, the latency rings. Chapters 3, 6, 8.
4. The decision in 9.6, taken and written down.
5. Five calls in one week: Nscale, Civo, Radiant, Pulsant, nLighten, with the ten questions. Chapter 9.
6. One agent retained on the cluster, two sounded. Chapter 0.
7. TT Group: open the exchange conversation with a list of exited buildings screened on power.yeats.com. Chapter 2.
8. Two lenders sounded on the lease heads, before a tenant is chosen, so the heads are drafted to their requirements. Chapter 7.
9. A watch list, reviewed monthly: Ofgem's Curate decision (due by the end of 2026); the DCP463 working group on unidentified-customer capacity; the distribution-level prioritisation process the government has promised; the heat network zoning regulations; the Ark sale; the AIRR cloud-compute award; CoreWeave's cost of capital; and the first UK powered-shell securitisation.

9.8 · What we still do not know

• The rent actually paid on a UK powered shell; Segro's Slough pre-let is undisclosed.
• Whether any UK or European landlord has yet obtained a customer or chipmaker backstop; none was found.
• Civo's financing; Latos's investors; Datum's ownership.
• Whether DCP181 and DCP114 will be read by UKPN as protecting a long-de-energised supply; DCP463 is the risk.
• The government's distribution-level prioritisation process, promised but unwritten.
• The coupon on Yondr's AA notes, which would price our exit better than any agent.
• Whether OpenAI's Stargate UK is paused or dead, and what that does to Nscale's UK contracts.

Each of these needs a conversation rather than further desk research.

9.9 · Sources

Chapters 0 to 8 of this study, each with its own source list; the operator and broker target list of 31 October 2026; the Edge method and its lease checklist.

A.1 · The deep eight

UK Power Networks · layer 0

Segro · layer 1

Kao Data · layer 2

Pulsant · layer 2

Nscale · layer 3

Civo · layer 3

Nebius · layer 3

Microsoft UK · layer 4

A.2 · The cards, by layer

Layer 0 · Grid

Layer 1 · Powered site

Layer 2 · Fitted hall

Layer 3 · Compute

Layer 4 · Capacity sale

Intermediaries and advisers