Solar panels for AI data centres — the highest-stakes, highest-return application

Why AI data centres are a different solar proposition

A traditional enterprise data centre runs at 5–10 kW per rack. A modern GPU training cluster — an H100 or B100-dense rack — runs at 40–80 kW per rack, with the most aggressive liquid-cooled configurations approaching 120 kW per rack. This is not a marginal difference: it is a 5–15x increase in power density within the same building footprint.

The consequence for solar PV economics is significant. An AI data centre building of 10,000 sqm might draw 10–20 MW of IT load — far more than a conventional facility of the same size. A rooftop PV system on that building might achieve 500–800 kW of installed capacity. The self-consumption ratio remains near 100% (the AI load is enormous and continuous), but the solar fraction of total consumption is smaller: perhaps 3–8% of total annual consumption rather than 10–30% on a conventional facility.

This does not weaken the solar business case for AI operators — it changes the framing. For an AI data centre drawing 15 MW continuously, a 750 kW solar system saves approximately £1.5 million in electricity cost over five years. The capital cost is £800,000–£1,100,000. The financial return is as strong as any other data centre application. But the sustainability narrative is different: solar covers a fraction of AI's enormous appetite, and hourly CFE matching requires a substantial corporate PPA alongside on-site generation.

The sustainability imperative for AI operators

The major hyperscale AI operators — Microsoft (Azure AI), Google (DeepMind, Vertex AI), Amazon (AWS Trainium), and Meta (Meta AI) — have all published 24/7 carbon-free energy (CFE) targets. Microsoft has committed to 100% CFE by 2030; Google has committed to 24/7 CFE by 2030; Amazon has committed to 100% renewable energy by 2025 (met) and is working toward 24/7 matching.

These commitments are not voluntary in the sense that walking away from them is without consequence. ESG rating agencies, institutional investors, major enterprise customers (who require supply chain Scope 3 disclosure from cloud providers), and government procurement rules all depend on these commitments being real and verifiable. The reputational and financial cost of walking back a published 24/7 CFE commitment would dwarf the cost of achieving it.

For AI operators building or operating facilities in the UK, on-site solar PV is the most direct, auditable contribution to 24/7 CFE that exists. It is geographically co-located with the consumption, generates during daylight hours, and creates MCS certificates and REGOs that can be verified by any auditor. Combined with a UK-based wind PPA for nighttime hours, it builds a credible 24/7 CFE profile.

Training vs inference: different power profiles for solar design

AI workloads split into two categories with very different power profiles:

Liquid cooling and solar: the thermal management interaction

High-density AI data centres are transitioning from air cooling to direct liquid cooling (DLC) — circulating coolant directly through the GPU servers rather than relying on room-level airflow. This has an important interaction with solar PV:

Traditional air-cooled data centres use significant power for CRAC (Computer Room Air Conditioning) units — often 30–50% of total facility power (contributing to high PUE). Liquid-cooled AI facilities dramatically reduce this overhead, achieving PUE close to 1.0–1.1 when liquid cooling is optimised. This means a larger fraction of total facility power is IT load — which is always on, always consuming solar generation. The effective self-consumption ratio of solar PV in a liquid-cooled AI facility is higher than in an equivalent air-cooled conventional facility.

There is also a building services interaction: liquid cooling systems require chilled water or dry cooler loops that consume electricity for pumps, heat exchangers, and cooling towers. These loads are consistent and on-site solar generation contributes to covering them. We model the cooling load and IT load separately in our AI data centre feasibility studies to give a precise self-consumption ratio.

UK AI data centre locations — solar irradiance and grid context

UK AI data centre development is concentrated in a small number of locations driven by grid capacity, land availability, and fibre connectivity:

• Slough and the Thames Valley: Highest density of large-format data centre land in the UK. Equinix, VIRTUS, Digital Realty, and Global Switch all have large campus developments in SL1–SL6. Solar irradiance: 1,600+ hours/year. Grid: SSEN (South East). Solar feasibility for Slough →
• London Docklands (E14/E16): Telehouse, Equinix LD1–LD4. Urban campus solar with rooftop constraints but premium grid rates (25–28p/kWh). Solar irradiance: 1,550 hours/year. Solar feasibility for London →
• Cambridge: Arm Holdings, Microsoft Research, EPCC HPC. University-adjacent AI research compute. Kao Data has a campus at Harlow (20 miles south of Cambridge) serving AI-compute customers. 1,600 hours irradiance. Solar feasibility for Cambridge →
• Manchester: Growing AI compute cluster. MediaCityUK, Equinix MA1. Electricity North West. 1,380 hours. Solar feasibility for Manchester →
• Birmingham: National Supercomputing strategy interest; central England location with good grid access. 1,480 hours. Solar feasibility for Birmingham →
• Newport and South Wales: Emerging AI campus location — lower land costs, available grid capacity at 132kV, Welsh Government digital infrastructure support. 1,450 hours irradiance. See Newport →
• Edinburgh: iomart, Pulsant. Scottish Government data and AI strategy driving public sector compute growth. 1,350 hours. SSEN distribution. Solar feasibility for Edinburgh →

Kao Data — the UK's leading AI data centre operator

Kao Data is the UK's most significant dedicated AI-compute data centre operator — building hyperscale AI campuses specifically designed for GPU training and inference, as opposed to general-purpose colocation. Their flagship Harlow campus (ESS1) is engineered at very high power density to support NVIDIA DGX SuperPOD-scale deployments.

Facilities like Kao Data ESS1 are precisely the type of campus for which on-site solar PV makes a strong financial case: enormous power consumption (continuous), high electricity costs, and public commitments to zero-carbon operations that require auditable renewable generation. The solar fraction is small relative to total consumption, but the absolute saving in pounds per year is substantial at the scale of a 50+ MW campus.

We have delivered feasibility assessments for AI campuses of this type in the Thames Valley and Eastern England corridor. The methodology — roof area assessment, structural survey, G99 fault-level analysis for high-fault-level environments, EnergyTag GC monitoring integration — is mature and documented in our AI campus feasibility framework.

EnergyTag Granular Certificates and 24/7 CFE matching

Standard REGOs (Renewable Energy Guarantees of Origin) are annual certificates — they demonstrate that a given volume of renewable energy was generated in a calendar year, but they don't link that generation to a specific hour of consumption. An operator with 1,000 MWh of REGOs from a January wind farm can claim those REGOs against consumption that occurred in July at 3 a.m. Under the GHG Protocol market-based Scope 2 method, this is legitimate — but it doesn't satisfy the stricter 24/7 CFE matching methodology.

EnergyTag Granular Certificates (GCs) are one-hour certificates — they certify that 1 MWh of renewable energy was generated in a specific hour, at a specific location, and can be matched to consumption in that same hour. This is the foundation of 24/7 CFE accounting, as required by Google's 24/7 CFE methodology and Microsoft's Azure sustainability framework.

On-site solar PV is ideally suited to GC issuance because:

• The generation location is the same as the consumption location — no transmission loss, no network attribution problem
• Half-hourly inverter monitoring provides the generation data needed for GC issuance at hourly resolution
• MCS certification provides the quality assurance that GC issuers require
• The correlation between daytime solar generation and daytime AI compute demand (inference) is genuinely high — solar covers a higher proportion of 24/7 CFE hours than an equivalent capacity of offshore wind, which generates day and night but not predictably

We configure our inverter monitoring systems for EnergyTag GC-compatible output from commissioning. If your sustainability framework requires 24/7 CFE matching, this must be specified at the design stage — retrofitting GC-compatible monitoring is possible but adds integration complexity.

SBTi alignment for AI data centre operators

The Science Based Targets initiative (SBTi) Net-Zero Standard requires companies to set near-term (2030) and long-term (2050) emissions reduction targets aligned with 1.5°C warming pathways. For data centre operators, Scope 2 electricity emissions are typically the largest single emissions source — making on-site renewable generation directly relevant to SBTi target achievement.

Under the SBTi Corporate Net-Zero Standard, market-based Scope 2 accounting (using REGOs or equivalent) is accepted for interim progress reporting, but the standard requires a pathway toward Scope 2 zero on a consumption-based (location-based) method over the long term. This means:

• Short-term: REGOs from on-site solar contribute directly to market-based Scope 2 reduction reporting
• Medium-term: EnergyTag GCs from on-site solar contribute to 24/7 CFE matching percentage (increasingly cited in SBTi reporting guidance)
• Long-term: On-site generation reduces the location-based Scope 2 that remains when the grid carbon intensity is considered alongside renewable certificates

AI data centres with SBTi commitments have additional reason to prioritise on-site solar over pure PPA solutions: on-site generation is directly measurable, directly auditable, and not subject to the additionality questions that arise with PPAs for existing renewable capacity.

G99 grid connection for AI data centres — high fault level environments

AI data centres present a more complex G99 Protection Relay environment than conventional commercial buildings. The reasons:

• High embedded generation: Diesel generator sets (multiple, typically 2–4 MW each), UPS systems, and possibly existing on-site generation all contribute to the fault level on the LV bus. Adding solar PV inverters increases the fault level further. The G99 application must include a full fault level calculation to demonstrate that the new solar inverters do not push the LV bus fault level above the switchgear's rated Icw or Icp.
• Complex protection coordination: The Protection Relay for a solar system on a data centre must be coordinated with the existing generator protection, UPS output protection, and DNO protection downstream. Coordination failures — where two protective devices operate for the same fault in an unexpected sequence — can cause extended outages. Our electrical design peer-review process specifically checks protection coordination for data centre installations.
• Export constraint zones: Many AI data centres are in grid-constrained areas (Thames Valley, London Docklands) where the DNO requires zero-export or active export limitation. The G99 application must include zero-export configuration details and the export monitoring equipment specification.

Our standard G99 application package for AI data centre sites includes: system description, fault level calculations (pre- and post-connection), protection relay type and settings, export management methodology, and coordination study. This package is prepared by a chartered electrical engineer and submitted directly to the DNO's connections team.

Financial case for AI data centre solar

The economics scale with AI's enormous energy appetite. Representative figures for a 15 MW AI campus:

Figures based on 2026 installation costs and current UK grid electricity rates for I&C customers on HH metering. Full Expensing applies to the first £1m under AIA; 50% FYA on excess above £1m. Payback range reflects variation in irradiance (Manchester vs South East) and grid rate.

How we deliver AI data centre solar projects

AI data centre projects are the most technically demanding installations we deliver, and also the ones with the most senior stakeholder involvement. Operations directors, infrastructure VPs, and sustainability leads are all involved. Our seven-step process:

1. Free desk feasibility (14 days): Review IT load profile, roof area, structural loading, and grid tariff. Produce preliminary system size, self-consumption ratio, financial model, and commissioning methodology outline.
2. On-site structural and electrical survey: Structural engineer survey (roof loading, fixings, wind uplift). Chartered electrical engineer survey (LV board fault level, protection relay coordination, export constraint assessment).
3. G99 Protection Relay application: Prepare full G99 package including fault level calculations and submit to DNO. Monitor progress against 65-working-day target. Escalate if delayed.
4. Detailed electrical and mechanical design: Full electrical design to BS 7671 (18th Edition), mechanical design to BSEN 1337 and roofing manufacturer guidance. Peer-review by chartered engineer. NDA signed before any site data is shared.
5. BPSS-cleared installation: All on-site personnel hold current BPSS clearance, CSCS Gold minimum. Installation phased to maintain Tier III/IV redundancy — AC commissioning during planned maintenance window only.
6. G99 commissioning with DNO: Protection Relay commissioned with DNO engineer present. System energised and monitoring verified. MCS commissioning certificate issued.
7. REGO/EnergyTag setup and sustainability reporting: Register for REGO issuance. Configure EnergyTag GC output if required. Issue first sustainability reporting pack within 30 days of commissioning.

Frequently asked questions — AI data centre solar