Microsoft and Nscale: The Price of Waiting for Grid Power

A data centre can have land, financing discussions and an intended customer, yet still lack the one input without which it cannot operate: electricity. That makes a grid connection date an investment variable, not a logistical footnote.

In September 2025, Nscale announced an AI campus in Loughton, Essex, with 50 MW of AI capacity, the option to expand to 90 MW, and an initial 23,040 Nvidia GB300 GPUs expected in the first quarter of 2027. On 18 October 2026, reporting emerged that Microsoft had withdrawn from the site's customer arrangement. Reuters subsequently reported that the connection could slip into the 2030s and quoted Microsoft as saying the decision concerned one site. The full commercial reasons and the termination mechanics remain undisclosed. [1–3]

The question is not whether gas-generated electricity costs more than grid electricity. Often it will. The question is how much more a developer can pay to begin producing something saleable before its commercial opportunity expires.

The value of an earlier megawatt

The economics of an AI data centre have an unusual clock. Land and buildings can survive delays for years; processors may lose relative economic value much faster. A delayed grid connection can therefore impose several separate losses: site-holding expenses, postponed cash flow, weaker ability to offer current-generation compute and, potentially, the disappearance of an anchor customer.

That last loss changes the calculation. If a customer will accept service whenever power eventually arrives, delay mainly discounts future profits. If delivery dates matter to the contract or the customer's expansion plans, the chance of a profitable sale may fall with each year. Paying a premium for earlier power then preserves a commercial option, not just a few months of revenue.

A developer has at least three choices: keep the site and wait; install on-site supply that bridges the period until the grid becomes available; or move computing capacity to a different site that can connect earlier. Each uses different capital and exposes the investor to different risks. A realistic choice also includes declining to spend on hardware when no viable buyer remains.

An illustrative 50 MW investment

We model a representative project, not Nscale's confidential accounts. The 50 MW public figure is treated by assumption as IT nameplate capacity. With average draw at 80% of that nameplate and power-usage effectiveness (PUE) of 1.25, annual whole-site electricity consumption would be 438,000 MWh. Peak site load would be 62.5 MW. Nscale has not disclosed those particular operating parameters, and its announced AI capacity is not automatically its total grid import requirement. [1, 4]

The model uses a hypothetical six-year computing-services contract, £6 million annual revenue per MW of IT capacity, £95/MWh grid electricity and a 12% real investment hurdle. Equipment costs £475 million initially, other facility works £85 million, with a subsequent equipment refresh. All of these commercial figures are analyst scenarios, not price quotations or Nscale figures.

The commercial hinge is a simple assumed retention rule: 85% probability of retaining an anchor contract if power arrives by 2028, decaying after that deadline. A lost contract may be replaced after another year, at a lower price, but the investor can refuse uneconomic deployment. The assumption is an ex ante modelling device. Since Microsoft has reportedly already withdrawn from Loughton, it must not be read as a probability that Microsoft will return.

Under the central case, on-site generation is ready in 2028, another energised site is available in 2029, and the original grid connection arrives in 2033:

These are probability-weighted, unlevered project values in constant 2026 pounds, not measured company losses. Crucially, no strategy is assumed physically available just because its spreadsheet NPV is positive.

How expensive can self-generation become?

A representative gas-generation calculation starts with an assumed £42/MWh thermal gas price and 42% electrical efficiency. Allowing for variable maintenance, a £60-per-tonne illustrative carbon charge and other costs yields about £144/MWh of variable electricity cost. This excludes approximately £77 million of assumed generating-plant capital expenditure, fixed servicing, and a smaller up-front readiness expense. Carbon pricing and environmental-permit applicability would need to be determined for the actual installation. [5–7]

The relevant ceiling is where the probability-weighted value of on-site generation just equals the better of waiting and relocating. With the central inputs, that is about £156/MWh of on-site variable electricity cost, or £61/MWh above the modelled grid price. Generator capital and fixed costs have already been included separately in that comparison. The currently assumed variable premium is approximately £49/MWh, leaving only about £12/MWh of further headroom before relocation becomes more valuable.

Allocated over the discounted, probability-weighted electricity produced before the grid arrives, generating-plant capital, fixed costs and readiness add roughly £52/MWh on top of the central variable premium. This makes the illustrative current effective all-in premium about £101/MWh. That allocation is a modelling comparison, not a quoted electricity price or a market-wide shadow price.

When does relocation beat waiting?

Changing only the date of the original grid connection reveals another threshold. In this model, waiting is worth approximately £110 million if power arrives in 2029, still above the £64 million relocation alternative. By 2030, waiting has declined to about £62 million, slightly below relocation. The crossover is not an engineering rule or a forecast of where data centres will move. It follows from the assumptions about customer timing, alternative-site readiness and relocation cost.

The comparison changes when on-site supply is feasible. At the central £144/MWh variable cost, generation remains ahead of relocation for a 2033 connection. But if generation takes longer to permit, gas is not reliably available, its power quality is unsuitable or equipment cannot be financed, that option may disappear entirely. The Environment Agency's generator guidance makes clear that routine power generation cannot simply be treated as exempt emergency backup. [7]

Ofgem's July 2026 consultation reported roughly 73 GW of data-centre connection requests in Britain's queue and proposed new project-commitment requirements. The queue measures applications for access to the network, not 73 GW of future electricity actually consumed. For developers, the economic point is that queue position and power readiness can influence the value of both sites and customer commitments. [8]

The commercial distinction

The incentive to pay for early power is strongest when the developer has an anchor buyer, can procure the required hardware just in time, and has a credible bridging supply. The benefit shrinks if revenue per unit of compute falls, equipment replacement becomes more expensive, the deadline does not endanger customer demand, or an early-powered alternative site is readily available.

An already-paid site cost should not be charged again when choosing the next step. Conversely, a building already under construction does not justify buying hundreds of millions of pounds of GPUs if no profitable service contract survives. The financial decision concerns avoidable future cash flows, while historical spending matters separately for understanding who bears the loss.

This research establishes a conditional method for pricing an earlier connection and locating the point where relocation makes more sense. It cannot establish why Microsoft made every part of its decision, that on-site gas generation is viable at Loughton, or the financial consequences for Nscale. Those claims require contracts, engineering evidence and private project accounts that are not publicly available.

Research note. Figures are rounded from a formula-based real-terms cash-flow model. The 2027–2035 grid-date sweep, fuel-price construction, hardware renewal assumptions, replacement-customer branches, sensitivity cases and computational checks are documented in the supporting technical report and editable workbook. No measured loss, project-level forecast or investment recommendation is claimed.

Sources

[1] Nscale, Nscale announces UK AI infrastructure commitment in partnership with Microsoft, NVIDIA and OpenAI, 16 September 2025. https://www.nscale.com/press-releases/nscale-uk-ai-infrastructure-announcement

[2] Financial Times, Microsoft pulls plug on Nscale data centre deal, 18 October 2026. https://www.ft.com/content/35cc5cc6-f9b6-49cc-8fc2-bd61e09c8c40

[3] Reuters, Microsoft pulls out of UK Nscale data centre deal over power delay, 19 October 2026. https://www.investing.com/news/stock-market-news/microsoft-pulls-out-of-uk-nscale-data-centre-deal-over-power-delay-4957808

[4] Uptime Institute, Global Data Center Survey 2025, PUE discussion. https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2025.Annual.Survey.Report.pdf?version=0

[5] DESNZ, Industrial energy price statistics, September 2026 update. https://www.gov.uk/government/collections/industrial-energy-prices

[6] UK Government, Taking part in UK ETS markets, 2026. https://www.gov.uk/government/publications/taking-part-in-the-uk-emissions-trading-scheme-markets/taking-part-in-the-uk-emissions-trading-scheme-markets

[7] Environment Agency, Specified generator: when you need a permit. https://www.gov.uk/guidance/specified-generator-when-you-need-a-permit

[8] Ofgem, Ofgem acts to free up grid capacity by tackling speculative data centre projects, 29 July 2026. https://www.ofgem.gov.uk/press-release/ofgem-acts-free-grid-capacity-tackling-speculative-data-centre-projects