Solar panels for AI data centres — the highest-stakes, highest-return application
AI data centres consume more power per square metre than any previous infrastructure. They also carry the most demanding sustainability commitments. Solar PV is not optional for AI operators — it's foundational to their CFE obligations.
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:
Training clusters
Large language model and foundation model training runs at maximum GPU utilisation for weeks or months — 90–100% GPU load, continuous. Power draw is flat, predictable, and enormous. A 1,000 H100 cluster (approximately 3.5 MW GPU power) runs at essentially constant draw 24 hours a day during training. This is ideal for solar self-consumption: any solar generation is consumed immediately, self-consumption ratio ≈ 100%, and the flat load profile means no battery storage is required to capture generation.
Inference clusters
Inference workloads — serving model predictions in response to user queries — have more variable load profiles. Peak inference demand follows user activity patterns: high daytime, lower at night. This correlates usefully with solar generation: peak solar production occurs mid-afternoon, which aligns reasonably well with high-traffic inference demand periods. Self-consumption ratios are high (inference load is still substantial even at off-peak), and the correlation between solar peak and inference peak is a genuine benefit for inference-focused facilities.
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 and solar irradiance
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. 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 →
Financial case for AI data centre solar
The economics scale with AI's enormous energy appetite. Representative figures for a 15 MW AI campus:
| Metric | Value |
|---|---|
| Typical rooftop capacity | 800 kW – 1.5 MW |
| Annual generation (South East) | 720,000 – 1,350,000 kWh |
| Annual savings (at 24p/kWh) | £173,000 – £324,000 |
| Capital cost | £700,000 – £1,350,000 |
| Full Expensing tax relief (25% CT) | £175,000 – £338,000 in year 1 |
| Simple payback (post-tax) | 3.2 – 5.0 years |
| CO₂ avoided (year 1) | 101 – 189 tonnes CO₂e |
| CFE contribution (annual hours) | ~1,500 hours of 8,760/year |
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.
Our approach to AI data centre solar
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 approach:
- Feasibility study includes IT load profile analysis (if available) and cooling system power breakdown — we size the solar system against actual consumption data, not generic assumptions
- Electrical design accounts for the high fault level environments of large AI data centres — inverter short-circuit current contribution is modelled against switchgear ratings
- BPSS-cleared teams as standard; NDA signed before any site data is shared
- Sustainability reporting pack includes EnergyTag GC-compatible monitoring for operators targeting 24/7 CFE
- We can model the on-site solar contribution alongside proposed corporate PPAs to produce a combined CFE matching profile — essential for operators modelling against a 2030 24/7 CFE target