The Hidden Cost of Intelligence · Deep Dive May 2026

Methodology & Scope Note

This report covers global data centre electricity demand and its grid, cost, carbon, and development implications, with particular attention to African markets. The analysis is grounded in primary institutional sources (IEA, PJM Interconnection, IEEFA, national regulatory filings) supported by secondary journalism and market research. Modelled projections are identified as such throughout.

AI load refers to electricity consumed by GPU-accelerated servers for AI training and inference, in dedicated facilities or hyperscale campuses. General cloud computing demand is included in aggregate data centre figures but distinguished where the source permits.
Data centre demand follows IEA methodology: metered electricity consumed at the facility level, including IT load, cooling, and power conditioning. Behind-the-meter generation is counted in facility totals but may not appear in grid statistics.
Hourly matching means carbon-free generation exists on the same grid, same region, at the same hour as facility draw. Annual matching means only that total certificates purchased equal total annual consumption, with no requirement for temporal or geographic alignment.
Behind-the-meter generation is capacity owned and operated by the data centre, outside utility regulation and renewable portfolio standards.
Geography: global aggregate to 2030 and 2035; country case studies cover the US, Ireland, Singapore, Chile, Malaysia, and four African markets. Africa market-size figures are commercial research aggregator estimates and should be treated as indicative order-of-magnitude figures.
Limitations: Corporate energy disclosures vary materially in methodology, making cross-company comparisons imprecise. African grid and data centre data are materially less complete than OECD equivalents. All forward projections assume current policy trajectories and carry wide uncertainty, particularly after 2030.

The Load Grids Were Not Built For

Power engineers use the term "industrial anchor" for a load that is dense, continuous, and geographically fixed. For decades, grid planners modelled steel mills and paper plants running three shifts, built the transmission, and recovered costs over 20-year asset lives. The model worked because anchors moved slowly, negotiated with grid operators, and respected the physics of copper and concrete.

AI data centres operate on a different timescale. Primary The IEA's April 2025 Energy and AI report put global data centre electricity consumption at approximately 415 TWh in 2024, equivalent to roughly 1.5% of world supply. The IEA base case projects that figure reaching 945 TWh by 2030. An April 2026 IEA update revised the 2025 baseline upward to 485 TWh, with the 2030 projection holding near 950 TWh. Japan's total annual electricity consumption is approximately 950 TWh. The projection is the addition of roughly one Japan's worth of demand from a single sector inside one decade.

Source: IEA, Energy and AI (April 2025) and Key Questions on Energy and AI (April 2026). All figures are modelled projections based on near-term investment pipelines, not observed outcomes. Post-2030 projections carry substantially wider uncertainty; the IEA designates them "explorations" rather than forecasts.

Global data centre electricity demand, 2020 to 2035

Terawatt hours (TWh) · IEA base case with high scenario range · Geography: global

The profile of AI load is as significant as its scale. A conventional cloud server operates at roughly 30 to 40% of rated capacity on average, cycling with traffic. A GPU cluster running AI inference or training operates at 80 to 95% continuously, because the hardware economics only function when chips are processing work. Grid operators call this a square wave profile: flat, high, unrelenting, resembling a baseload power plant rather than any commercial load they had previously modelled.

Primary In PJM Interconnection's region, which operates the grid across 13 US states serving 67 million people, projected data centre demand in the Dominion Zone grew from a forecast 5,700 MW of incremental load by 2037 (as modelled in December 2022) to more than 20,000 MW in the revised 2025 forecast (IEEFA, July 2025, citing PJM data). That is not a revision; it is a different forecast for a qualitatively different world.

Secondary Hyperscale operators are reported to be investing between $630 and $700 billion in AI data centre infrastructure in 2026 alone, approximately 62% more than 2025 capital expenditure (Fortune, March 2026). These are analyst-derived estimates from company disclosures with varying accounting periods, not audited totals.

415 TWh

Global data centre electricity demand, 2024

IEA, April 2025 · Observed

945 TWh

IEA base case for 2030, roughly 3% of global supply

IEA, April 2025/2026 · Projected

15%/yr

Projected annual demand growth rate, 2024 to 2030

IEA, April 2025 · Modelled

Bottom Line

If current build trajectories hold, AI load will represent the largest single new industrial demand type added to advanced electricity grids in two decades. The constraint is not scale but speed: grids are engineered and permitted over 10 to 15-year cycles while data centre construction timelines run to 18 to 36 months. If that planning gap is not closed by deliberate regulatory intervention, the default outcome is fossil generation filling the space.

What Is Actually Powering It

Every major hyperscale operator reports that its data centres run on renewable energy. Google reports 100% annual matching. Microsoft committed to the same standard by end of 2025. Amazon reached that target ahead of schedule. These statements are technically accurate in their accounting terms. They are also methodologically incomplete in their physical implications, and the difference is operationally significant.

Annual matching means purchasing enough renewable energy certificates across a full year to equal total consumption. A data centre drawing 300 MW from a coal-heavy grid at 2 a.m. is balanced, in the accounting, by a solar farm that generated power at 1 p.m. in a different region. The electron powering the GPU and the electron from the solar panel never shared a wire or a moment in time. The accounting is clean. The grid is not.

Primary Google has committed to hourly matching by 2030 and publishes its progress. Its 2025 Environmental Report (covering 2024 data) shows a global average of 66% hourly matching, with nine of 20 grid regions reaching at least 80%, but Asia-Pacific averaging 12% and Singapore reaching only 4%. The gap between 100% annual and 66% hourly identifies what physically powers the grid during the remaining hours. For those hours, in many regions, the answer is gas and coal.

The electron powering the GPU and the electron from the solar panel never shared a wire. The accounting is clean. The grid is not.

TCL Editorial synthesis, 2026

What is filling the gap

Primary The IEA's Energy and AI report quantifies the supply mix meeting incremental data centre demand through 2030. Renewables account for approximately 47% of additional generation (roughly 250 TWh); natural gas for 33% (175 TWh); coal for 13% (70 TWh); and nuclear for 7% (35 TWh). Fossil fuels collectively meet over 40% of net new data centre demand in the base case.

These are modelled projections for the IEA base case scenario, not observed outcomes. The renewable share depends heavily on policy environments, interconnection queue clearance rates, and technology cost trajectories that could shift materially in either direction.

How additional data centre demand is being met, 2024 to 2030

Additional TWh of generation by source · IEA base case · Geography: global

Secondary Unable to secure grid connections fast enough, some data centre developers are building their own generation entirely. At least 46 US facilities with a combined reported capacity of approximately 56 GW operate on behind-the-meter gas generation (Cleanview dataset, reported by Grist, February 2026). This capacity sits outside utility regulation and bypasses renewable portfolio standards. Contracts signed in 2025 and 2026 will still be running in 2045. Note: the 56 GW figure covers both operating and planned capacity; definition of "behind-the-meter" varies across data providers.

The lock-in risk

Secondary Interconnection economics favour gas. The American Action Forum (May 2026) estimated interconnection costs at roughly $24 per kilowatt for natural gas, versus $253 for solar and $335 for offshore wind. Gas connects faster, completes more reliably, and does not require storage for firm capacity. Where speed is the binding constraint and renewable interconnection queues run to several years, gas wins the investment decision. These cost differentials reflect the current US regulatory system, which is itself a policy variable, not a technologically fixed limit.

Bottom Line

Technology companies are purchasing renewable energy certificates, and most of those certificates represent real generation. The grid those same facilities draw from during off-peak hours is running on gas and, in high-coal markets, on coal. Both facts are simultaneously true. If hourly matching remains voluntary and behind-the-meter gas is permitted without carbon accounting requirements, the corporate sustainability narrative and the physical grid reality will continue to diverge for at least the next decade.

Who Pays the Bill

Capacity markets operate on a principle that large new loads routinely obscure: when significant new demand enters a grid system, it does not simply purchase power at prevailing rates. It shifts the clearing price in the forward auction that determines what all generators receive for all capacity across the system. A fast-arriving load raises the auction price. Everyone connected pays more, regardless of whether they caused the increase.

Primary PJM Interconnection's capacity market cleared at $28.92 per MW-day in the 2024/2025 delivery year. The 2025/2026 auction cleared at $269.92, a rise of 833% in a single cycle. The 2026/2027 auction hit the FERC-approved price cap at $329.17/MW-day. The December 2025 auction for 2027/2028 cleared at $333.44/MW-day while falling 6,625 MW short of the reliability requirement, the first such shortfall in PJM's 27-year history.

Primary Monitoring Analytics, PJM's independent market monitor, attributed 63% of the 2025/2026 price increase to data centre demand, translating to approximately $9.3 billion in additional costs recovered from customers in higher electricity rates (IEEFA, July 2025). Washington DC Pepco residential customers saw bills rise by an average of $21 per month from June 2025 as a direct result. The 63% attribution is an analytical estimate from Monitoring Analytics' modelling, not a regulatory determination.

833%

PJM capacity market price rise, 2024/25 to 2025/26

PJM / IEEFA · Primary

$9.3bn

Additional PJM customer costs in 2025/26 attributed to data centres

Monitoring Analytics / IEEFA · Estimated

$163bn

NRDC projected cumulative PJM capacity costs 2028 to 2033 without reform

NRDC · Modelled range

Primary Virginia's State Corporation Commission responded in November 2025 with a new rate class for customers drawing 25 MW or more, requiring minimum demand payments of 85% on transmission and distribution and 60% on generation capacity (SCC Virginia, 2025). Even with this structure, the Piedmont Environmental Council calculated that residential ratepayers would still absorb roughly 61% of grid upgrade costs attributable to hyperscale load growth.

Secondary A further mechanism runs in parallel: PJM ratepayers are also funding new transmission lines built to serve Northern Virginia data centres. The Union of Concerned Scientists estimated ratepayer exposure at approximately $4.4 billion from data centre-related transmission projects approved in 2024 alone (Utility Dive, October 2025).

Water belongs in this cost accounting. Primary Google reported 6.4 billion gallons of total fleet water use in 2023, with 31% drawn from medium or high water-stressed watersheds (Google 2024 Environmental Report). Lawrence Berkeley National Laboratory estimated US data centres consumed 17 billion gallons directly through cooling and an additional 211 billion gallons indirectly through their electricity supply chain in 2023 (LBNL, December 2024). Neither figure appears on a renewable energy certificate.

Bottom Line

Data centre expansion is currently being partially subsidised by electricity infrastructure paid for by people who will never operate a GPU. Virginia's rate reform is a structural improvement, not a resolution. If the NRDC projection for cumulative PJM costs through 2033 proves accurate at even half that magnitude, the political pressure for structural reform will intensify well before most current clean energy commitments reach their stated target dates.

What the Footprint Actually Looks Like

Fossil fuel combustion is the largest environmental cost of AI infrastructure. Three additional cost categories are routinely excluded from corporate sustainability reporting: embodied carbon in hardware, thermal effects on surrounding land, and the carbon efficiency implications of hardware obsolescence cycles.

Operational emissions

Primary The IEA projects data centre CO2 emissions peaking at approximately 320 million tonnes around 2030 before a shallow decline to roughly 300 Mt by 2035 as the generation mix improves (IEA, April 2025). Germany's electricity sector produced approximately 180 Mt in 2023. As a share of global emissions, the data centre contribution remains under 1%. As a share of the emissions growth still occurring in a decarbonising world, it is one of the few sectors still moving in the wrong direction.

Embodied carbon in hardware

Secondary Each Nvidia A100-class GPU carries an estimated embodied carbon footprint of approximately 150 kg CO2 equivalent before it is powered on, based on lifecycle analysis by Patterson et al. (arXiv:2104.10350, 2021). Training a model of GPT-3 scale consumed approximately 1,287 MWh and produced an estimated 552 tonnes of CO2 in operational emissions, not accounting for hardware manufacturing (Patterson et al., 2021). Note: these figures derive from 2021 methodology applied to earlier GPU generations and should not be applied without adjustment to current Blackwell-generation hardware.

Thermal effects Preprint

A March 2026 preprint (Marinoni et al., arXiv:2603.20897) mapped 20 years of NASA land-surface temperature data against more than 6,000 hyperscale data centre locations, reporting surface temperature increases averaging 2 degrees Celsius within a radius of up to 10 kilometres from facilities, with estimated effects reaching 340 million people globally. This paper has not completed peer review. Several physicists have challenged the magnitude on energy-balance grounds, arguing waste heat per unit of ground area is insufficient to drive warming at that scale and that site-selection patterns may account for part of the measured signal. The direction of a thermal effect is not contested. Its precise scale and radius will be determined by peer review. This report treats the finding as preliminary.

Hardware obsolescence

No pre-existing generation to serve this load; Entergy is constructing three new gas-fired plants at approximately $3 billion

Policy response

State-level permitting approved without federal environmental review; no additionality or carbon requirements attached to generation contracts (Fortune/Goldman, March 2026)

Strategic implication

Richland Parish is the most visible current example of behind-the-meter gas lock-in at hyperscale. The generation built here will operate on contract through the mid-2040s regardless of how the surrounding grid transitions.

Bottom Line

Every mature data centre market is either managing the consequences of under-regulated expansion or competing for investment on terms designed to prevent those consequences. The policy lesson is consistent and documented across six jurisdictions. Governments that establish conditions before facility construction begins retain meaningful leverage. Those that respond after the jobs are announced do not.

Africa's Arithmetic

Constraint

Policy environment and regulatory certainty for long-term foreign investment commitments

Policy response

15-year Public Data Center Provider licences to attract investment; 12 GW renewable energy target for 2026

Strategic implication

Egypt's grid scale and submarine cable density make it a viable hyperscale anchor for Africa-to-Europe routing. Its regulatory framework is developing faster than those of comparable emerging markets.

The access competition

Primary The continent has approximately 600 million people without reliable electricity. The World Bank and African Development Bank's Mission 300 programme, signed by 48 countries at Dar es Salaam in January 2025, targets connecting 300 million people by 2030. As of October 2025, approximately 32 million had been connected (World Bank, 2025). At the current rate, the 2030 target requires approximately eight times the pace achieved in year one.

Behind-the-meter data centre generation does not directly compete with Mission 300 for the same physical infrastructure. It does compete for the same scarce sovereign fiscal resources, the same engineering capacity, and the same project finance bandwidth. Where those resources are finite, a priority choice is being made even when no one frames it as one.

The opportunity inside the constraint

Olkaria clarified the terms of the problem. It did not close the question. A hyperscale power purchase agreement, structured correctly, provides the credit-rated long-tenor offtake that independent power producers need to raise project finance in markets that have historically struggled to attract it. If data centre demand underwrites new generation that would not otherwise have been built, and if that generation is available to non-facility users, the arithmetic changes. The data centre becomes a catalytic load rather than a competitive one.

Secondary Cassava Technologies launched Africa's first NVIDIA-powered AI factory in Cape Town in 2025 with 3,000 GPUs and committed solar energy across the fleet. The planned Johannesburg expansion targets 20 MW of capacity, with 12,000 to 13,000 GPUs deployed across the continent over three to four years (Cassava Technologies, March 2026). Whether the economics are competitive with Northern Virginia or Singapore at inference scale requires further evidence from actual deployment. The model, behind-the-meter renewables combined with local compute and sovereign data frameworks, is the most operationally plausible form of the catalytic case.

The difference between extractive hosting and anchor-load hosting is not technology. It is the terms on which the connection gets made.

TCL Editorial synthesis

Bottom Line

Africa's grid arithmetic is not a fixed barrier to AI infrastructure investment. It is a set of constraints that determine whether that investment serves African development or runs parallel to it. The Olkaria stall demonstrates that terms matter more than announcements. The moment of commercial negotiation, before the facility design is fixed and before the jobs are announced, is the only moment when governments hold meaningful leverage over what gets built and who benefits from it.

What Actually Works

Mature markets have generated enough evidence to distinguish policy tools that address the structural problem from those that address its political appearance. The following matrix presents five tools ranked by demonstrated impact, alongside feasibility and implementation complexity, for comparison across market types.

Policy Tool Matrix: Evidence-Based Ranking

Ranked by demonstrated impact on structural grid and carbon outcomes · Editorial

Tool	Impact	Feasibility	Complexity	Best example
Hourly matching requirement Carbon-free generation aligned to facility draw by hour and location, not annual certificate average	High	Medium	High	Singapore DC-CFA2 (2025); Google 2030 target
Additionality requirement Hyperscalers must procure renewables that would not have been built in their absence	High	Medium	Medium	Ireland new-build sourcing mandate (CRU, 2025)
Load-indexed rate class Large-load customers bear minimum demand obligations on T&D and generation, reducing cost socialisation onto residential ratepayers	Medium	High	Medium	Virginia SCC 25 MW rate class (November 2025)
Water accounting at permitting Facility environmental assessments must incorporate climate change impacts on local water systems; disclosure required at permitting stage, not in annual reports	Medium	High	Medium	Chile Second Environmental Court ruling (February 2024)
Connection approval as leverage Grid connection approval and data sovereignty mandates used to negotiate additionality, grid access, and community benefit before facility design is fixed	High	High	Low	Nigeria NDPR, Kenya Data Protection Act, South Africa POPIA

These tools are operational in at least one jurisdiction each. None requires new technology or novel legal frameworks. The gap between knowing what works and deploying it is political timing: these conditions are only fully negotiable before the investment is committed and before the facility exists. After the announcement, governments negotiate from a structurally weaker position.

Bottom Line

The effective policy tools are not new inventions. The gap between knowing what works and deploying it is whether governments act before investment momentum makes conditions politically unenforceable, which is almost always before the announcement, not after it.

Three Ways This Resolves by 2035

These are not forecasts. They are analytical descriptions of the logical endpoints of three distinct policy trajectories, each internally consistent. The actual outcome will sit somewhere between them. The base case is the default trajectory in the absence of deliberate policy intervention; it is the path of least political resistance, not the most probable physical outcome.

Best Case

Demand drives the clean build it claims to support

Hourly matching becomes the regulatory standard in major markets by 2028. Additionality requirements fund approximately 200 GW of new renewable capacity that would not otherwise have been built. Gas lock-in is avoided. In Africa, data centre demand underwrites new generation available to non-facility users, converting demand from a competitive load into a catalytic one. Corporate renewable narratives and physical grid reality converge by 2032.

Trigger indicators: Hourly matching adopted by FERC or EU regulators by 2027. A major hyperscaler PPA in Africa with explicit additionality and grid-access clauses. Behind-the-meter gas permitting restricted in two or more US states by 2027.

Low without deliberate intervention. Technically feasible.

Base Case · Current trajectory

Muddled progress with fossil lock-in at the margins

Annual matching remains the dominant corporate standard through 2030. Natural gas absorbs the baseload gap in the US; coal-heavy grids absorb it in Asia. PJM-style capacity price spikes recur in two or three additional regional markets by 2028. African data centre development is uneven. Mission 300 falls materially short of its 2030 connection target.

Trigger indicators: No federal hourly matching standard by 2027. Capacity price events in at least one additional US ISO by 2028. Behind-the-meter gas continues without a federal carbon accounting requirement through 2026.

High. This is the current trajectory.

Worst Case

Speed of connection defeats terms of connection

The gap between AI compute demand and clean generation widens through 2030. Behind-the-meter gas becomes the default for developers facing multi-year renewable queues. Capacity markets fail in multiple regions, requiring emergency government intervention to retain ageing generation assets. African markets approve hyperscale facilities without additionality or grid-access conditions; net-zero commitments are formally redefined to accommodate the gap.

Trigger indicators: A third consecutive PJM auction at the price cap with a reliability shortfall. Behind-the-meter gas capacity in the US exceeds 80 GW by 2028. Two or more African governments approve facilities without additionality or grid-access conditions.

Moderate if current behind-the-meter gas trends continue. Not inevitable.

Bottom Line

The energy future of AI infrastructure is not written in the technology. It is written in the permitting conditions, rate classes, and matching standards that governments set now, while the leverage still exists and the buildings are not yet standing. Each year of delay transfers negotiating power from regulators to developers in a transaction that, once complete, is very difficult to reverse.

Editor's Note: What Changed in This Version

This document is the second revision of the May 2026 deep dive. It incorporates an editorial standards brief and a separate web and platform quality brief. Major changes are as follows.

IEA figures updated to April 2026. The original cited a 2024 baseline of 415 TWh. The April 2026 IEA update revised the 2025 baseline to 485 TWh; the 2030 base case holds at approximately 950 TWh. The 2035 base case is correctly stated as approximately 1,200 TWh (the original cited 1,700 TWh, which is the high scenario only).
Kenya/Olkaria corrected. The original characterised the project as cancelled. Current reporting indicates it is suspended pending renegotiation. Both grid capacity and commercial structure (capacity payment guarantees) contributed to the impasse.
PJM data extended. The 2026/27 auction cleared at the FERC cap of $329.17/MW-day. The December 2025 auction for 2027/28 cleared at $333.44/MW-day with a 6,625 MW reliability shortfall.
Google hourly matching updated. Corrected from 64% to 66%, reflecting the 2025 Environmental Report (covering 2024 performance).
Evidence hierarchy applied throughout. Every major claim carries an inline evidence quality marker.
Comparative framework. All country case studies now use a consistent five-field analytical structure.
Policy matrix added. The What Works section now presents a ranked comparison table for cross-market analysis.
Scenarios upgraded. Each scenario now includes explicit trigger indicators, likelihood language, and signpost conditions.
Platform upgrades. Reading progress bar, sticky accessible sidebar TOC with active section highlighting, skip-to-content link, print stylesheet, mobile hamburger navigation, improved type scale, and accessibility semantics (roles, aria-labels, aria-labelledby).

References

1 International Energy Agency, Energy and AI (April 2025). iea.org/reports/energy-and-ai Primary
2 International Energy Agency, Key Questions on Energy and AI (April 2026). iea.org/reports/key-questions-on-energy-and-ai Primary
3 PJM Interconnection, 2027/2028 Base Residual Auction Results (December 2025). pjm.com Primary
4 IEEFA, Projected data centre growth spurs PJM capacity prices by factor of 10 (July 2025). ieefa.org Primary
5 American Action Forum, AI Data Center Power Surge: Shifting Trends Toward Natural Gas (May 2026). americanactionforum.org Secondary
6 Virginia State Corporation Commission, Order on Dominion Energy Biennial Review 2025 (November 2025). scc.virginia.gov Primary
7 Commission for Regulation of Utilities Ireland, Large Energy User Connection Policy Decision (December 2025). Primary
8 Singapore EDB and IMDA, DC-CFA2 Call for Application (December 2025). imda.gov.sg Primary
9 Climate Case Chart, Municipality of Cerrillos v Evaluation Commission, Second Environmental Court of Chile (February 2024). climatecasechart.com Primary
10 Ember, From AI to emissions: Aligning ASEAN's digital growth with energy transition goals (June 2025). ember-energy.org Secondary
11 Semafor, Energy shortfall problem scuppers Kenya's $1B Microsoft data center (6 May 2026). semafor.com Secondary
12 Data Center Dynamics, Microsoft and G42 data center in Kenya stalled due to lack of power capacity (May 2026). datacenterdynamics.com Secondary
13 Daily Maverick, New data centres set to swallow 34% of Cape Town's current electricity supply (28 April 2026). dailymaverick.co.za Secondary
14 Cassava Technologies, Cassava scales African AI Infrastructure with NVIDIA-Powered AI Factories (March 2026). cassavatechnologies.com Secondary
15 Fortune, Meta's $27 billion AI data centre is causing chaos in small town Louisiana (26 March 2026). fortune.com Secondary
16 Engineering News-Record, $27B Meta Data Center Pushes Louisiana Toward Massive Power Expansion (2 April 2026). enr.com Secondary
17 Patterson D. et al., Carbon Emissions and Large Neural Network Training, arXiv:2104.10350 (2021). Secondary
18 Marinoni A. et al., The data heat island effect, arXiv:2603.20897 (March 2026, preprint, not peer reviewed). Preprint
19 Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (December 2024). Primary
20 Google, 2025 Environmental Report (covering 2024 performance). sustainability.google Primary
21 World Bank, Mission 300 Africa Energy Summit Outcomes (Dar es Salaam, January 2025). worldbank.org Primary
22 Eskom Holdings, Energy Availability Factor update (March 2026). eskom.co.za Primary
23 Citizens Utility Board, Capacity market prices reflect need for data centre reform (December 2025). citizensutilityboard.org Secondary
24 E&E News, AI energy demand by the numbers (December 2025). eenews.net Secondary
25 Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (December 2024). Primary
26 Grist, Data centres are scrambling to power the AI boom with natural gas (February 2026). grist.org Secondary
27 Union of Concerned Scientists data centre transmission cost analysis, cited in Utility Dive (October 2025). ucsusa.org Secondary