Microsoft Left a Site the Grid Can Power Only in the 2030s. AI Infrastructure Is Duration Risk.

简而言之： Britain’s grid now decides when an artificial intelligence data centre opens, and Ofgem’s own model shows that date costs more than its proposed fee, so investors should price compute as long-duration utility risk.

Microsoft Left a Site the Grid Can Power Only in the 2030s

The clearest price signal in AI infrastructure this autumn came from a tenant leaving. Microsoft has withdrawn from a deal to anchor Nscale’s planned AI data centre at Loughton in Essex, after grid officials told Nscale they expected to supply enough power to the site only in the 2030s, Reuters reported on 18 October 2026¹. The site was due to begin operating early in 2027 with 50 megawatts (MW) of AI computing capacity. A Microsoft spokesperson told Reuters that “this decision relates to a single site”, and Nscale said it remains committed to the project and is “exploring alternative power sources for the site”¹.

The power problem was public three weeks earlier. UK Power Networks, the local distribution network operator, had told Nscale that the grid could supply enough electricity to the site only in the early to mid-2030s², for a facility planned to grow from 50 MW to 90 MW. Nscale was already looking at generating power on site².

Read it as an investor would. The site was planned, the anchor customer had a deal and the hardware story was intact. One date moved: the day the grid could energise the load. When that date slid from 2027 into the 2030s, the customer’s need for capacity and the site’s ability to deliver it stopped overlapping, and the anchor deal went with them. The technology held, and the project lost its anchor on a utility timetable that sat outside the control of everyone at the table.

That is the argument of this piece. AI data centre investment is still discussed and underwritten in the language of technology: the speed of chips, the growth of demand and the fund’s exit date. The assets themselves now behave like regulated utility assets, whose value is set by grid dates, connection terms and regulators, and which run for decades. The mismatch between those two readings is a duration risk, and Britain’s regulator has just written down what it costs.

Ofgem Has Priced a Place in Britain’s Grid Queue

On 29 July 2026, Ofgem, the Office of Gas and Electricity Markets and Great Britain’s energy regulator, published its Curate consultation on demand connections. It proposes that data centre projects above a 40 MW threshold secure a returnable financial commitment from the point they accept a connection offer until energisation³, set at 2.5% to 7.5% of average data centre capital expenditure (capex), “approximately £237,500 to £712,500 per MW of requested capacity”. For a 100 MW campus that is £23.75 million to £71.25 million, posted on the day the offer is signed. The fee is returned if the project proceeds as planned and forfeited if it terminates or fails the relevant requirements; a project that reduces its requested capacity would be liable for the fee on the megawatts it gives up³.

The motive is queue hygiene. Between November 2024 and June 2025, contracted offers in Britain’s demand queue rose from 41 gigawatts (GW) to 125 GW. Ofgem’s analysis puts around 73 GW of that queue in roughly 315 data centre projects, against peak electricity demand of 45 GW, and values the pipeline at £693 billion of implied capex, around 23% of United Kingdom (UK) gross domestic product (GDP)³. A queue place had become a cheap option, and the fee is designed to make holding one expensive.

The proposal is still a proposal. The consultation closed on 17 September 2026 and Ofgem’s page marks it “Closed (awaiting decision)”⁴; the regulator states that it intends to take decisions later in 2026 and deliver them through changes to industry codes, licences and contracts³. The level is open too: Ofgem calls the percentage of capex one of the most critical points it is seeking feedback on³. The direction is set, and the consultation’s own numbers already show what it does to a model. Mezyan has set out what the same proposal means for boards, and why a grid place sits outside any green power contract, in its reading of the commitment fee⁵. The investor’s question is different: what the wait itself is worth.

Ofgem’s Own Model Shows the Connection Date Costs More Than the Fee

Ofgem stress-tested its range on a hypothetical 100 MW data centre, built to earn an unlevered pre-tax internal rate of return (IRR) of about 15% before any fee, with a weighted average cost of capital (WACC) of 10.95% and capex of £9.5 million per MW³. It assumed the fee is deposited as cash on 31 December 2026 and measured net present value (NPV) from that date. It then ran three connection years.

Read the table down a column before reading it across a row. With no fee at all, moving the connection from 2029 to 2035 cuts the project’s NPV from £210 million to £127 million, a loss of £83 million, or about 40%. Adding the full 7.5% fee to a 2029 connection costs £18 million. In Ofgem’s own numbers the grid date is worth more than four times the fee. The fee matters most where the date is late: at 2035 it takes NPV from £127 million to £84 million and IRR from 14.7% to 13.0%, the 170 basis points that Ofgem attributes to “the negative cost of carry until the fee is returned at project energisation”³.

The IRR column hides the same point. The no-fee IRR is 14.7% in every row, so the delay shows up only in NPV, measured from the end of 2026. Ofgem concludes that impacts from 20 to 170 basis points “would suggest that our fee range would not materially impact viable projects”, and adds that viable projects may gain as speculative ones leave the queue and waits shorten³. Both statements can hold for the regulator’s question, which is whether a good project still clears its hurdle. They answer a different question from the investor’s, which is what the asset is worth and when its cash arrives.

Two modelling choices sharpen that gap. The case study values the data centre on a ten-year cash-flow forecast and books everything after year ten as a single terminal value, calculated with the Gordon Growth method³; for a building that runs for decades, that one figure carries the years an infrastructure buyer cares about most. And a deposit posted at the end of 2026 for a 2035 energisation sits idle for eight years or more before a single graphics processing unit (GPU) draws power. For a closed-end fund with a ten-year term, eight or more years of posted security is most of the fund’s life.

Four Signals Show Compute Now Behaves Like a Regulated Utility

Curtailment is on the regulator’s list as a backstop

The fee is one half of the shift. In its Connect Update of 16 June 2026, Ofgem said it is exploring whether an additional backstop operational control measure, such as mandatory curtailment, may be required⁶ to keep confidence that the system can handle stress events, and that it expected to consult on these measures during autumn 2026, confirming then “what, if any, mechanism is required”. The same document records operator and trade-association feedback on the challenges mandatory measures present to “overall investability”, and notes that non-firm connection agreements “may be a route to an earlier connection date”⁶. A load that the system operator can interrupt in exchange for faster access earns utility-style economics: revenue adjusted for availability, regulated terms and a negotiated place in the order of who gets switched off first.

Power sets the opening date, and developers are building around the grid

Loughton has company. Ofgem reports that data from the gas networks shows around 117 active cases, principally from data centres “aiming to achieve earlier operation, in advance of connection to the electricity network”⁶. Those developers are seeking a gas connection to get ahead of the electricity timetable, the same kind of on-site answer Nscale is studying at Loughton². On-site gas changes the asset again: it adds a fuel contract, a generation plant and an emissions profile to what was underwritten as a building with servers in it.

Debt finances what is powered; announcements describe what is planned

The gap between a headline and an energised hall shows in one company. In May 2025, MGX, Bpifrance, Mistral AI and Nvidia announced a joint venture for an AI campus in the Paris region expected to ultimately reach a capacity of 1.4 GW⁷, with construction expected to start in the second half of 2026 and operations by 2028. In March 2026, Mistral raised $830 million of debt from a consortium of seven banks to buy 13,800 Nvidia chips⁸ for a data centre at Bruyères-le-Châtel, and said it would seek 200 MW of capacity across Europe by the end of 2027. DatacenterDynamics, quoting Mistral’s statement, put the site at 44 MW of powered capacity⁹. The financed slice is about 3% of the headline gigawatt. Lenders underwrote the powered megawatts; the gigawatt remains a plan with a grid timetable in front of it.

Offtake runs in years while buildings run in decades

Ofgem is minded to require projects on its lease-or-sale pathway to show a binding customer commitment for at least 20% of compute capacity for at least one year³, evidenced at the milestone that tests readiness before final investment decision and construction. Read it from the lender’s side. A contract covering a fifth of the hall for twelve months is enough to prove seriousness to the network operator, and it is the revenue certainty set against a shell that lasts three decades. Loughton shows how quickly even an anchor contract can fall away when the grid date moves. Every year beyond the first contract is re-leasing risk, priced at whatever the chip cycle says compute is worth at the time.

The Duration Mismatch Is the Real Risk, and the Grid Makes It Visible

Four clocks govern every AI infrastructure deal, and they run at different speeds. The silicon clock is the fastest. Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five from 1 January 2025, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning”, which increased its 2025 depreciation and amortisation expense by $1.4 billion¹⁰. The building clock is the slowest: Amazon’s chief executive described data centres “which have 30-plus year useful lives”¹¹, hosting at least five to six generations of servers. The capital clock sits between them: in the Goodwin Terms Database, 56% of private equity funds have a ten-year term¹², and half of infrastructure funds are designed to run for twelve years or more. The grid clock now belongs to the regulator and the network operators, with connection dates that Ofgem’s own scenarios run to 2035³ and that Loughton’s network operator has put in the early to mid-2030s².

Put the clocks side by side and the risk is arithmetic. Take an illustrative fund that holds its final close at the end of 2026 with a ten-year term, and commits to a British site whose security is posted that December and whose energisation date is 2035. The fund reaches the end of its term in 2036, about a year after the first GPU draws grid power. It will have financed a building that runs past 2060, posted a security for most of its own life, and sold the asset before the first server generation has depreciated once. The exit buyer pays for whichever clock the market watches that year. When it watches silicon, the asset trades like technology. When it watches the grid, it trades like a utility. The volatility sits in the multiple, and the multiple is set by duration.

Allegory Capital’s analysis of Big Tech’s AI capex outrunning the cash flow it once minted¹³ reaches the same point from the hyperscalers’ income statements: the spending is front-loaded and the payback is long. Neumarz found the operating-company version of the gap in private equity’s AI bet, with plenty of pilots and limited proof in the profit and loss account¹⁴. Infrastructure is the part of the AI trade where proof arrives on a date, and the date is now set outside the deal.

The Four-Clock Test Gives Investment Committees Five Questions

Neumarz’s Four-Clock Test is a one-page discipline for any AI infrastructure memo, whether the position is equity in a developer, debt on a powered shell, a stake in a platform or a fund commitment. It asks the committee to name each clock, date it and price the gaps between them.

1. Which clock sets the exit? Name the clock the buyer will price at exit, and model the multiple that clock implies, with a utility-style multiple as the downside case.
2. Who carries the security until energisation? Map every commitment fee, bond and letter of credit, its holder, its release date and its forfeiture triggers, and charge the cost of carry to the deal itself.
3. Is curtailment priced into revenue? Model a curtailed year and a non-firm connection alongside the firm case, and test covenants against the curtailed case.
4. Does the tenant’s start date survive a later grid date? Read the lease for its long-stop date, its walk-away rights and its rent start, and test them against an energisation two and five years later than planned.
5. Is the price paid on powered or announced megawatts? Value energised capacity at infrastructure yields and treat queued capacity as an option, worth less each year the grid date moves.

A memo that answers all five with dates and numbers is an infrastructure memo, and it can be priced as one. A memo that answers them with growth rates and chip roadmaps is a technology memo for an asset that now behaves like a utility.

Re-Date the Pipeline to Energisation Before Ofgem Decides

Four steps put the test to work while the rules are still being written. First, re-date every AI infrastructure asset and pipeline project to its expected energisation date, and rerun valuations on that date instead of the construction schedule. Second, inventory exposure to Britain’s queue: any data centre project of 40 MW or more that connects to the transmission system, or that connects to a distribution network and goes through a transmission assessment, falls inside Ofgem’s proposed scope³, and Ofgem’s decision is due later in 2026. Third, ask each manager in writing which of the four clocks sets their exit, and how their fund term and extension rights cover the gap to energisation. Fourth, review structures: long-hold vehicles, continuation funds and project-level debt matched to building life suit assets whose value compounds over decades.

British data centres remain investable on these terms. Grid access is becoming scarce, priced and regulated, and scarce regulated access is the classic foundation of durable infrastructure returns. Ofgem’s own case study still shows a viable project clearing its cost of capital at a 2035 connection. The investors who earn those returns will be the ones who underwrite compute on the grid and building clocks, hold it for as long as those clocks run, and stop paying technology multiples for an asset whose opening date belongs to the regulator.

Neumarz, a Kainjoo SA venture, advises investors, acquirers and family offices on structuring and financing long-duration assets, including the duration profile of AI infrastructure positions. Investment committees reviewing an AI data centre exposure can request a confidential Four-Clock review of the position.

This article has been reviewed by Haider Alleg, Managing Partner at Kainjoo SA (Terraxis).

This article is general commentary prepared for professional and qualified contacts. It is general information within the meaning of the Swiss Financial Services Act (FinSA), separate from any offer, solicitation, investment advice or personal recommendation, and advice on individual circumstances remains a matter for a dedicated mandate. Regulatory proposals are described as consulted on and may change before any decision. Company figures are taken from public filings and reports and express no view on any security. The fund example and the Four-Clock Test are illustrative.

参考资料

1. Reuters (Paul Sandle), via The Star. “Microsoft pulls out of UK data center deal with Nscale, Telegraph reports,” 18 October 2026. https://www.thestar.com.my/tech/tech-news/2026/10/18/microsoft-pulls-out-of-uk-data-center-deal-with-nscale-telegraph-reports
2. Euronews (Roselyne Min). “UK’s biggest AI supercomputer may have to wait until 2030s for enough power,” 29 September 2026. https://www.euronews.com/2026/09/29/uks-biggest-ai-supercomputer-may-have-to-wait-until-2030s-for-enough-power
3. Ofgem. Proposed data centre connection reforms: Curate – Demand Connections Reform, consultation document, 29 July 2026. https://www.ofgem.gov.uk/sites/default/files/2026-07/Proposed-data-centre-connection-reforms-curate-consultation-document.pdf
4. Ofgem. Proposed data centre connection reforms, consultation page (status: Closed, awaiting decision). https://www.ofgem.gov.uk/consultation/proposed-data-centre-connection-reforms
5. Mezyan. “Britain Wants Up to £712,500 a Megawatt to Hold a Grid Place. A Green PPA Doesn’t Buy One,” 27 October 2026. https://mezyan.ch/ofgem-data-centre-commitment-fee-grid-connection-ppa/
6. Ofgem. Connect Update: Demand Connections Reform, 16 June 2026. https://www.ofgem.gov.uk/sites/default/files/2026-06/demand-connections-reform-update-June.pdf
7. Abu Dhabi Media Office. “MGX partners with Bpifrance, Mistral AI and NVIDIA to build Europe’s largest AI campus in France,” 19 May 2025. https://www.mediaoffice.abudhabi/en/technology/mgx-partners-with-bpifrance-mistral-ai-and-nvidia-to-build-europes-largest-ai-campus-in-france/
8. Reuters (Supantha Mukherjee and Leo Marchandon), via Yahoo Finance. “France’s Mistral raises $830 million in debt for AI data centre build-up,” 30 March 2026. https://finance.yahoo.com/sectors/technology/articles/frances-mistral-raises-830-million-060341046.html
9. DatacenterDynamics. “Mistral AI raises $830m in debt financing for data center in Paris, France,” 30 March 2026. https://datacenterdynamics.com/en/news/mistral-ai-raises-830m-in-debt-financing-for-data-center-in-paris-france
10. Amazon.com, Inc. Annual Report on Form 10-K for the fiscal year ended 31 December 2025. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm
11. Diginomica (Stuart Lauchlan). “Growth accelerates at AWS, so does capex spend on AI,” 31 July 2026, reporting Amazon’s second-quarter 2026 results. https://diginomica.com/growth-accelerates-aws-so-does-capex-spend-ai-lots-more-come-warns-ceo-andy-jassy
12. Goodwin. “How long do closed-ended funds last?,” 13 November 2024 (Goodwin Terms Database). https://goodwinlaw.com/en/insights/publications/2024/11/insights-practices-pif-how-long-do-closed-ended-funds-last
13. Allegory Capital. “Big Tech’s AI Capex Now Outruns the Cash Flow It Once Minted,” 25 August 2026. https://allegory.capital/big-tech-ai-capex-outruns-cash-flow/
14. Neumarz. “Private Equity’s AI Bet: Plenty of Pilots, Limited Proof in the P&L,” 22 June 2026. https://neumarz.com/private-equity-ai-value-creation-gap/