Better experiences. Better-informed decisions.Work with Smarter BI

Software · Research

Conceptual illustration connecting electrical infrastructure, server halls and customer networks; not a real site or engineering plan.
Smarter BI · Conceptual illustration, not a real site.

Can the UK compete for the next wave of AI infrastructure?

Can the UK compete in AI infrastructure? Compare workload needs, location value, electricity costs and delivery evidence across competing markets.

Abstract

This article examines how the UK can compete for AI infrastructure when electricity cost is only one part of the location decision. It combines workload comparisons, an electricity cost sensitivity and a review of four projects to distinguish potential location value from evidence of delivery. The aim is to help readers assess what a site can provide, for whom, at what cost and by when. The analysis offers a framework for scrutiny, not a country ranking or a calculation of investment returns.

Structure and purpose

  1. Workload and location value. Introduces the four investment tests and explains why megawatts do not all serve the same customer.
  2. Location comparison. Examines how training, inference and colocation change the questions to ask across seven regional lenses.
  3. Electricity cost sensitivity. Shows how an assumed price difference affects annual energy costs, without treating it as a profit difference.
  4. Delivery evidence. Separates announcements, development, construction and operation so projects face a consistent test.
  5. Four project reviews. Applies that test to London, Stockport, Bosquel and Lappeenranta, identifying disclosures and remaining gaps.
  6. Implications for the UK. Draws together the evidence and identifies the commercial questions that must be answered before comparing returns.

Scope: project sources establish what their publishers disclose, not independently verified physical delivery, private contracts or profitability.

The UK’s AI infrastructure case is not simply a contest for the cheapest electricity. A location advantage matters only if a project can deliver it: the right service, at the right time, at a cost customers will support.

01. Not every megawatt competes for the same customer.

The cheapest electricity does not necessarily produce the most competitive data centre. Some workloads can move towards available power; others gain value from proximity to customers, networks and existing systems. Location sensitivity and power sensitivity are overlapping constraints, not two separate markets.

The UK’s 2025 Compute Roadmap explicitly recognises that training is generally less location-dependent, while inference may benefit from proximity to data sources or end users. It therefore anticipates different architectural needs across AI Growth Zones. This supports a workload-specific approach; it does not establish a London requirement or prove that a zone can deliver its planned capacity.[8]

Training requires tightly connected computing equipment, but need not sit near the model’s eventual users. NVIDIA’s NCCL technical discussion makes the importance of fast GPU communication clear for both training and inference.[9] That is a different question from the distance between a facility and its customers. Batch inference and interactive services should not inherit the same location assumption.

Two distances. Two sets of constraints.

Where does the waiting happen?

Follow the work inside a cluster, then a customer’s request to a facility and back. Highlighted steps show where communication can hold up progress.

Inside the computing cluster

Equipment ↔ equipment

  1. ComputeGPUs work on parallel tasksEach GPU performs its part of the computation.
  2. Cluster networkExchange data and coordinateA task may wait for required data or for other GPUs to reach a synchronisation point.
  3. ContinueDependent work can proceedLate exchanges can leave computing resources waiting rather than doing useful work.

What constrains progress?

Bandwidth (data per second), latency (communication delay) and coordination between GPUs.

Location lesson: being closer to customers does not fix a slow cluster network. Fast GPU communication matters to both training and inference.

Between facility and customers

Service ↔ customer

  1. Outward network journeyRequest travels to the facilityNetwork travel time and congestion can add delay before processing starts.
  2. At the facilityRequest queues and is processedAvailable capacity and the model’s work affect how long a response takes.
  3. Return network journeyResponse travels to the customerThe return path adds network delay too. Faster GPUs do not remove this journey.

What constrains response time?

Network round-trip time, congestion, queuing and processing. The importance of each depends on the application.

Location lesson: proximity may reduce network delay, but cannot eliminate queuing or processing time.

Two different problems: moving closer to customers does not fix slow GPU links. Faster GPU links do not remove the customer’s network journey.

Simplified paths, not measured timings: box sizes do not represent duration. Computing and communication can overlap; real services may involve multiple exchanges or streamed responses. Training and inference can involve both constraints. Policy [8] · GPU communication [9]

London’s investment case rests on whether its location advantages justify its higher costs. For some customers, proximity to users, data and established networks may be worth paying for. For others, access to scalable power may matter more. A regional British site therefore needs to compete with relevant alternatives in France, Finland or elsewhere on the cost, timing and reliability of the service it can deliver. These are site-level comparisons, not fixed national roles: France, too, offers both customer-proximity and power-led opportunities.

CBRE’s research illustrates this trade-off. Its midyear 2026 UK report describes neoclouds willing to pay premium rents near London for core locations and speed to market, alongside trading and quantitative investment firms seeking low-latency computing. But it also identifies developments accepting greater latency in exchange for secure, scalable power near London.[10] Its global report describes pricing premiums in parts of West London near cloud availability zones, where wholesale space is limited.[11]

This supports a differentiated London investment case, not a London-wide premium. These are broker assessments of particular market segments. They do not establish how much additional revenue a specific project can earn, whether that outweighs its costs, or what return investors should expect.

To move from a location argument to an investment case, ask four questions.

The investment lens

Beyond the electricity price: four investment tests

  1. Workload fit

    What service does the customer need?

    Define the computing task, performance requirements and service commitments.
  2. Location value

    Why would the customer choose this site?

    Identify the value of proximity, network connections and jurisdiction.
  3. Full service economics

    Can the revenue justify the total cost?

    Include capital and operating costs, with explicit assumptions about utilisation and pricing.
  4. Deliverability

    What can be delivered, and by when?

    Check power availability, permissions, construction and readiness to serve customers.
A credible investment case must address all four. Market growth and low vacancy provide context, but neither proves a location premium or a project’s return. These are connected diligence questions, not country scores or a formula for choosing a winner.

02. Change the workload. Change the investment question.

Choose a workload to see what each location must prove. These are screening questions, not measured performance or tariff comparisons. Every decision ultimately needs a named site, customer and delivery date.

Location explorer / Seven comparison lenses

Where does the location earn its keep?

All three workload cases are shown. Every location requires project-level verification.

United Kingdom / Demand hub

London & Thames Valley

Training · Challenge the premium. For a workload that can move, establish why this location offsets the cost of metropolitan land and power. Do not assume proximity adds value to training.

Inference · Potential proximity advantage. Test measured end-to-end latency and access to the customer’s existing systems against regional alternatives. UK data residency does not require London.

Colocation · Potential ecosystem advantage. Customer networks, cloud connections and migration costs may justify the location. Confirm the connections and contracted demand at the actual facility.

Key test: will customers pay for this facility’s connections? Connectivity examples [2]

United Kingdom / Site-specific opportunity

Regional Britain

Training · Conditional opportunity. A site can compete if it secures suitable power, dense-compute infrastructure and a customer. A policy designation is not an electricity connection.

Inference · Test the regional alternative. Some UK-facing workloads may meet their service targets outside London. Measure network and model-serving delays rather than using distance as a proxy.

Colocation · Establish demand first. Available land does not create a customer ecosystem. Test anchor demand, carrier access and the cost of moving existing workloads.

Key test: can power and anchor demand arrive together? UK project evidence [3]

France / Demand hub

Paris & Île-de-France

Training · Separate city from campus. Compare a metropolitan proposal with power-oriented French alternatives. A national energy advantage does not establish a site’s delivered price or connection date.

Inference · Potential customer proximity. Evaluate access to intended French and continental users and their cloud estates. It is not automatically a substitute for UK workloads with different requirements.

Colocation · Test ecosystem value. Existing enterprise and network relationships may matter more than the lowest energy price. Obtain facility-level service and customer evidence.

Key test: does the customer need the metropolitan location? Market context [1]

France / Power-led comparator

Northern & industrial France

Training · Potential power-led advantage. Investigate sites near existing energy infrastructure. Existing grid assets can be helpful, but do not by themselves prove available import capacity.

Inference · Workload-dependent. Latency-tolerant and batch services can consider a wider geography. Interactive services still need measured performance to their intended customers.

Colocation · Customer case required. Industrial land and electricity do not replicate a metropolitan carrier ecosystem. Test anchor-led development separately from speculative retail colocation.

Key test: is nearby grid infrastructure actually available to the site? French project evidence [4]

Germany / Demand hub

Frankfurt & Rhine-Main

Training · Challenge location cost. For movable compute, test whether connectivity or a specific customer offsets metropolitan costs. Include applicable cooling and heat-reuse obligations in the site design.

Inference · Potential ecosystem advantage. Existing enterprise and cloud connections can support the case. Geography alone cannot establish an application’s latency or availability.

Colocation · Test interconnection value. Carrier choice and customer architecture may justify a premium. Verify actual connections and migration alternatives rather than ranking countries.

Key test: does interconnection value offset site costs and applicable environmental obligations? Connectivity examples [2]

Finland / Power-led comparator

Finnish candidate sites

Training · Potential power-and-cooling case. Test long-term energy procurement, climate, construction costs and cluster connectivity together. The project announcements reviewed here are not a verified inventory of operating capacity.

Inference · Separate interactive from batch. Nordic customers and latency-tolerant workloads may have different location needs from UK-facing interactive services. Benchmark each case.

Colocation · Avoid extrapolating from hyperscale. An owner-occupied campus does not prove third-party colocation demand. Identify the customer ecosystem relevant to the proposed service.

Key test: do procurement and cooling benefits survive the full service-cost comparison? Finnish project evidence [5]

United States / Ecosystem benchmark

Northern Virginia

Training · Not the cheapest-US benchmark. Compare any ecosystem benefit with other locations available to the same customer. Include utility commitments and delivery risk, not merely an average energy price.

Inference · A different customer geography. Use it to examine US East Coast demand and network value, not as a like-for-like UK inference location.

Colocation · Test willingness to commit. An established ecosystem can support location premiums, but neither historic growth nor utility rules establish the returns of a new project.

Key test: will customers commit on terms that cover utility obligations? Electricity-data limits [6]

Method: compare workload fit separately from evidence completeness. No quantitative attractiveness score or verified delivered-cost ranking has been calculated.

03. What is a power-price difference worth?

A recurring electricity premium can be material. But national wholesale prices are not a project tariff: procurement, taxes, network charges, losses, fixed commitments and currency can change the comparison. The tool below isolates the energy-linked price difference. It does not estimate project returns or assign prices to countries.

Sensitivity model / User assumptions, not market data

The annual electricity cost of a location premium

Electricity cost only, not profit. Compare two otherwise identical sites using illustrative inputs, not site quotes.

Price difference = candidate minus alternative. Negative means cheaper.

Annual electricity-cost difference+£47.3m/ year
Facility electricity / year946,080 MWh
For each £10/MWh difference£9.46m / year

At these assumptions, the candidate location costs approximately £47.3m more per year in electricity.

How the price difference changes annual cost

Vertical scale: ±£1,000m/year for 100 MW. Fixed while PUE, electrical load or price changes.

Annual electricity-cost difference (£m)

+1,0000−1,000 −3000+300 Additional costSavings +£50/MWh→ +£47.3m/year

Candidate minus alternative (£/MWh)

The circle marks the selected assumption. Its dashed guide leads down to the price-difference axis.

Selected: +£50/MWh → +£47.3m/year (additional cost).

Electricity cost only, not profit or investment return. The line holds capacity, electrical load and PUE constant while varying the electricity-price difference. The price axis stays at −£300 to +£300/MWh. The vertical range covers 100% electrical load and PUE 2.50 at the selected IT capacity. Only changing capacity can rescale it; a notice identifies the change. Compare slopes only on the same scale. Without JavaScript, this chart shows the default example.
Input guidance and calculation

IT electrical capacity: enter 1 to 2,000 MW, in whole MW.

Average IT electrical load: 0 to 100% of IT capacity, in whole percentage points. Not billable or GPU utilisation.

Annual power usage effectiveness (PUE): facility energy divided by IT energy. A PUE of 1.20 means 20% extra energy for supporting systems. Enter 1.00 to 2.50, in steps of 0.01.

Electricity price difference: candidate price minus alternative price. Use a negative value if the candidate is cheaper. £100/MWh versus £80/MWh → enter 20. Range: −£300 to +£300, in whole £/MWh.

Illustrative starting assumptions: 100 MW, 90% average electrical load, annual PUE of 1.20 and a +£50/MWh price difference. These are not site quotes.

Annual electricity (MWh) = IT capacity (MW) × average IT electrical load (%) ÷ 100 × annual PUE × 8,760 hours.

Annual electricity-cost difference (£) = annual electricity (MWh) × electricity price difference (£/MWh).

100 MW × 90 ÷ 100 × 1.20 × 8,760 hours = 946,080 MWh/year. 946,080 MWh × +£50/MWh = +£47,304,000/year.

Electrical load is not GPU utilisation or billable utilisation. Idle equipment still uses electricity. Annual PUE compares total facility energy with IT energy. This simplified scenario uses a non-leap year and assumes the two locations have identical capacity, electrical load and PUE.

Use a comparable price basis: the same currency, period and treatment of energy-linked charges and taxes for both locations. Subtract alternative from candidate. Only charges proportional to electricity consumption are represented.

This is an electricity-cost comparison, not a complete location business case. Fixed network commitments, capex, financing, taxes not included in the entered price, revenue, hardware replacement, delay, hedging and discounting are excluded. Headline values are rounded, not estimates of forecast precision. The static example remains available without JavaScript; printing records the current scenario.

The hypothetical £47.30 million energy premium is not a profit difference. Earlier service, committed customers or lower migration costs could offset it; financing and delivery delays could widen it. The next test is whether those benefits are evidenced.

04. Apply the same evidence test everywhere.

Do not scrutinise a British connection application while treating a French or Finnish announcement as delivered capacity. Equally, do not confuse an owner-operator’s internal investment plan with a third-party lease. Both can be commercially meaningful, but the evidence and risks differ.

Delivery explorer / Open each evidence state

What does the claimed MW actually represent?

01. Announced or requested

A proposed campus, policy zone or connection request establishes intent. Record the capacity boundary, source date and dependencies. Do not count it as available capacity or convert a queue total into a demand forecast.

02. Development-stage

Land, planning or connection work may be progressing, but verify each separately. Obtain consent conditions, the connection point, reinforcement scope and who pays. A power-purchase agreement alone does not prove that electricity can reach the site.

03. Contracted and under construction

Inspect construction progress, utility obligations, funding conditions and customer commitments. Distinguish first-MW and full-campus dates. Customer credit, termination rights, energy exposure and residual-value risks remain alongside commissioning risk.

04. Energised and operating

Verify commissioned IT capacity and the ability to deliver the specified service. Energised does not mean fully equipped, occupied, profitable or available to a new customer. Keep operating evidence separate from commercial performance.

These are evidence categories, not a guaranteed sequence or valuation formula. A project can have different phases in different states.

Goldman Sachs frames these constraints within the investment cycle; J.P. Morgan connects time-to-power with financing and monetisation risk. These are useful analytical perspectives, not proof of any project’s viability.[7]

05. Four projects. Four different evidence states.

Nine primary documents reveal four distinct evidence problems: undisclosed electrical capacity in London, unreconciled capacity and dates in Stockport, programme-versus-site timing in Bosquel, and unspecified first-phase capacity in Lappeenranta. These are disclosure findings, not a ranking of the projects.

Source review / 29 September 2026

Verification boundary: the documents below were retrieved and read. Their statements are attributed to the publishers; physical delivery, legal approvals and private contracts were not independently certified. “Not established” means not established by this review, not that an agreement or capacity does not exist.

London / Existing service offering

Cloud House · LHR19

Digital Realty, London Docklands. The operator markets colocation and connectivity; its linked brochure describes incoming utility redundancy and campus cloud connections.

Disclosed capacity
No electrical MW figure found in the reviewed page or brochure. Both also discuss Cloud House West or wider campus services: those boundaries cannot be silently combined.
Delivery evidence
An existing service offering is described, but no commissioning date or independently confirmed operating MW was established in these documents.
Still needed
Firm import and commissioned IT MW; capacity available to new customers; actual occupancy, tariff and site-specific AI rack/cooling capability.
Read the London sources

Both reviewed 29 September 2026. Operator specifications are not utility confirmations or customer contracts.

Regional Britain / Disclosures to reconcile

KMAN-01 · Stockport

Kao Data, Kenwood Point. The operator reported planning approval in March 2024 and work starting in October 2024; the latter specifically describes demolition and site preparation.

Disclosed capacity
33 MW on the current location page, versus 40 MW in the 2024 releases. IT, gross facility and connection boundaries are not explicitly reconciled.
Delivery evidence
The current undated page, observed on 29 September 2026, says “under development” and targets Q3 2027. The 2024 releases expected operation/customer workloads in 2026.
Still needed
Capacity definitions, first-service scope and an explanation of the date difference; the council decision and conditions; current construction and utility milestones; binding customer commitments.

Do not infer a 7 MW downsizing or a cause of delay. The figures may have different boundaries; the sources reviewed do not explain the discrepancy.

Read the Stockport sources

All reviewed 29 September 2026. Historic site preparation is not proof of the current construction phase or energisation.

Northern France / Selected developer, proposed campus

Bosquel · Hauts-de-France

SoftBank Group / Sesterce. A majority-SoftBank joint venture announced its selection to develop and operate a campus in Bosquel. Developer selection is not a building permit or an electricity connection.

Disclosed capacity
1 GW proposed campus capacity; its precise electrical boundary is not explicit. It sits inside, not in addition to, the overlapping 3.1 GW regional and 5 GW French programmes.
Delivery evidence
The announcement carries a 30 May 2026 dateline. “By 2031” is the regional programme target across three sites, not an established Bosquel first-MW date.
Still needed
Project permits, construction evidence, allocated grid capacity, reinforcement responsibilities, phased dates, site tariff and binding customer/offtake commitments.
Read the Bosquel sources

Both reviewed 29 September 2026. Programme-level or Dunkirk-specific grid statements are not Bosquel connection evidence.

Finland / Operator-announced build

Lappeenranta · South Karelia

Nebius. The AI-cloud operator announced construction of a new campus in March 2026. A recruitment page continues to promote the site; recruitment is not commissioning evidence.

Disclosed capacity
Up to 310 MW. The recruitment page calls this AI-optimized compute capacity, but a precise IT/gross/utility boundary was not established.
Delivery evidence
The operator expects first capacity in 2027. The amount of that first tranche and the full-campus delivery date are not specified.
Still needed
Permits, site-specific power agreement and independently checked progress; first-phase MW, customer allocations, applicable tariff and phased commissioning dates.
Read the Lappeenranta sources

Both reviewed 29 September 2026. Do not transfer Mäntsälä’s operating results, portfolio power totals or unrelated energy agreements to this site.

No like-for-like site tariffs or binding end-customer commitments were established across these four cases. Their disclosed MW must not be summed or ranked as comparable operating capacity. An unknown is not zero.

Selection limits: Cloud House is a colocation offering; Stockport is a metropolitan regional development; Bosquel is a large-campus proposal; Lappeenranta is an AI-cloud build. They illustrate evidence problems, not representative national returns.

06. What would make a UK project competitive?

For London, demonstrate the premium. A connected facility may offer customer value, but Cloud House’s public materials do not establish how much customers pay for it or how many new MW are available. The ecosystem argument still needs commercial evidence.

For regional Britain, reconcile the project before modelling it. Stockport’s published capacity and date differences need explaining. Apply the same discipline overseas: Bosquel’s programme target is not a site commissioning date, and Lappeenranta’s first capacity is not the whole campus.

From delivery evidence to economics

Three questions before comparing returns

What does it really cost?

Align energy, fixed network charges, taxes, security and indexation terms. Include construction, cooling and any planning, water, noise or heat-reuse obligations, not just the electricity price.

Who commits to pay?

Identify the service, customer credit, contract length, break clauses and ramp-up. Internal demand is not third-party revenue; there is no universal MW-to-revenue conversion.

Who carries the downside?

Test delayed energisation, weaker demand and hardware replacement. Identify who absorbs redesign costs, idle capacity and financing obligations.

The UK has a conditional case, not a demonstrated national advantage. London must turn connectivity into paying demand; regional sites must turn power and land into timely service. The evidence reviewed here is not sufficient to identify a lowest-cost location or compare investment returns.

Applying these questions to an AI hosting decision? Our AI hosting requirements review starts with your workload, budget and location constraints. It is a separate service, not a condition of access to this research.

Sources, method & open evidence

The four-project review links nine primary documents read on 29 September 2026. It establishes what their publishers disclose, not physical delivery or private contracts. The location comparisons are editorial hypotheses; calculator inputs are assumptions and outputs are arithmetic.

The supplied research manuscript and PDF bibliography informed the wider questions, but their unresolved citation tokens are not evidence. References [1], [2] and [6] were checked on 1 October 2026 for market context, connectivity examples and electricity-data definitions. References [3] to [5] point to the dated project documents reviewed in section 05, not national delivery assessments. Unverified follow-up links have been removed. The bank material in [7] was reviewed for analytical context; its relevant passages were rechecked on 1 October 2026. Sources [8] and [9] were retrieved and their relevant passages read on 30 September 2026 for the opening’s policy and technical distinctions; they do not establish project economics. CBRE sources [10] and [11] were reviewed from a user-supplied PDF and pasted report text on 30 September 2026; live-page access was blocked, and the global report’s charts were not supplied.

  1. Market framing: JLL, 2026 Global Data Center Market Outlook, 5 January 2026; sections on on-site power and construction costs. JLL discusses power availability, latency and proximity as location considerations. This is broker market context, not a Paris-specific assessment, site tariff or delivery guarantee. No forecast totals, construction-cost benchmarks or location scores are adopted here.
  2. Connectivity examples: LINX LON1, undated; introduction and “Our Services”, describing London peering, private connections and cloud access. DE-CIX, Introducing: DE-CIX VirtualPNI, 1 July 2019, describes private connections within and between metro regions, including a Frankfurt example. These are provider descriptions, with the latter a historical example, not evidence of current availability at a particular building, measured application latency or willingness to pay.
  3. UK project example: Kao Data, Manchester, undated, observed 29 September 2026; opening and Key Facts. Read alongside the 2024 releases in the four-project review. Operator-reported development and target dates are not utility-confirmed delivery or binding anchor demand. No UK connection-fee or queue-total claim is adopted here.
  4. French project example: SoftBank Group and Sesterce to Develop 1 GW AI Data Center in Bosquel, France, page dated 31 May 2026, announcement dateline 30 May. Reviewed 29 September 2026; see the Bosquel review for programme and site boundaries. Developer selection does not establish allocated grid capacity, a tariff or a commissioning date.
  5. Finnish project example: Nebius to construct 310 MW AI factory in Finland, 31 March 2026, reviewed 29 September 2026; opening capacity and first-availability paragraphs. The Lappeenranta review separates first capacity from the full campus and company-wide commitments from site evidence. No Finnish tax-change figures, comparative tariff or quantified cooling advantage are adopted here.
  6. Electricity-data limits: EIA, Electric Power Monthly, Table 5.6.A, average electricity prices by state and end-use sector. The version reviewed on 1 October 2026 covers July 2026 and July 2025, released 24 September 2026; the linked table updates over time. Its notes describe preliminary sample estimates and possible customer reclassifications. State industrial averages are not a site tariff. No values from this table enter the calculator, and no Virginia utility-order, minimum-payment or contract-duration claim is adopted here.
  7. Investment context: Goldman Sachs, GS SUSTAIN: AI/Data Centers, supplied redacted report dated 6 August 2026, cover and key takeaways; J.P. Morgan, Powering progress: Financing the infrastructure behind US data center growth, 10 August 2026, sections on time to power and over-capacity risk. Both banks participate in this industry. Their analysis informs questions, not an endorsement of their market outlook.
  8. Workload and location: UK Compute Roadmap, section 4, Action 6. The publication record gives 17 July 2025, last updated 23 April 2026; the latter update concerns the evidence annex. The page identifies the document as published under the government of 2024 to 2026. This is a dated policy statement, not evidence that planned infrastructure has been delivered.
  9. Communication within compute systems: NVIDIA, Enabling Fast Inference and Resilient Training with NCCL 2.27, 14 July 2025; introduction and “Unlocking new performance levels”. Vendor technical material about GPU communication for training and inference, not evidence that a workload needs London or a measurement of customer-network latency. No vendor performance multipliers are adopted here.
  10. London customer demand: CBRE, UK Real Estate Market Outlook Midyear Review 2026, supplied PDF, printed pp. 17 to 18. Page 17 distinguishes premium rents and low-latency customer demand from developments accepting greater latency for available power. Page 18’s take-up and vacancy figures are forecasts, not observed outcomes; they are not used here.
  11. Submarket pricing: CBRE, Global Data Center Trends 2026, 17 June 2026, user-supplied text, the Europe subsection of “Rental Rates”. The report describes West London cloud-availability-zone proximity alongside limited wholesale supply. Its pricing comparison concerns 250 to 500 kW requirements, not hyperscale campus energy tariffs. CBRE participates in the market it analyses; these observations do not isolate a causal location premium or establish project returns.

Limits: no country scores, project tariffs, verified operating-capacity totals or investment returns are established here. Tax, utility and regulatory details require original documents and site-specific applicability checks before entering a project model.