Two numbers tell you why grid operators are nervous. Data centre electricity demand grew 17% in 2025 and is on track to double again by 2030. Meanwhile, the investment heavy industry needs to decarbonise — cement, steel, chemicals — has to rise three- to five-fold this decade just to stay on a net-zero path. Both are racing for the same grid connections, the same permitting queues and, increasingly, the same pools of capital.
That is not a coincidence of timing. It is a structural collision, and it raises an uncomfortable question for anyone working in climate policy or project finance: is the AI buildout accelerating the energy transition, or is it quietly displacing the harder, less glamorous work of decarbonising industry?
The scale of the data centre demand shock
The numbers from the IEA's 2026 electricity and AI analyses are worth sitting with. Combined capital expenditure from five large technology companies passed USD 400 billion in 2025 and is set to jump a further 75% in 2026 — a single-year increase larger than most countries' entire annual infrastructure budgets. Data centre electricity consumption grew 17% in 2025, with AI-focused facilities growing even faster, and is projected to roughly double to around 945 terawatt-hours by 2030.
To put that alongside the rest of the system: global electricity demand overall is forecast to grow at an average of 3.6% a year through 2030, itself 50% faster than the previous decade. Meeting that demand requires annual grid investment to rise by around 50%, from roughly USD 400 billion today. Data centres are not the only driver of that growth — electrification, EVs and air conditioning matter more in aggregate — but they are the most concentrated and the fastest to build, which makes them the most disruptive at a local grid level.
That concentration is the crux of the problem. Solar and wind projects take one to five years to build; data centres one to three; new transmission and grid infrastructure takes five to fifteen. More than 2,500 gigawatts of renewable, storage and large-load projects are currently stuck in grid connection queues worldwide, and data centres are jumping that queue in practice even where they aren't meant to in theory, because hyperscalers can move faster and pay more than most industrial developers.
Where the crowding-out argument comes from
At IRENA's 2026 assembly, delegates from emerging economies and small island states put the concern plainly: in some regions, data-centre-driven demand has coincided with severe grid congestion and electricity price rises of up to 267%, pulling investment and grid capacity away from other priorities. Research from CEPR frames the underlying risk in similar terms — that under faster-growth AI scenarios, competition for electricity and grid access intensifies enough to delay decarbonisation objectives altogether, forcing a trade-off between digital development and climate goals that most governments never explicitly chose to make.
Even the hyperscalers themselves are conceding the point. Google's 2026 Environmental Report acknowledged that its AI infrastructure buildout is expanding faster than the grid supplying it is decarbonising — a striking admission from a company that has spent a decade building a clean-energy reputation, and a signal of how quickly the arithmetic has shifted.
For anyone tracking industrial decarbonisation specifically, the finance side of this is arguably the sharper edge. UNCTAD's tracking of global investment flows shows data centre and semiconductor capital concentrating overwhelmingly in a handful of advanced economies — France, the United States and South Korea lead the list, with only a few emerging markets among the top recipients. That matters because the OECD has separately calculated that emerging and developing economies, where industrial output is growing fastest, are precisely where financing for low-carbon technology is most limited relative to need. Capital that might have gone into de-risking a green steel plant or an industrial electrification project is instead chasing the much larger, much faster returns of AI infrastructure.
The case that data centres are catalysts, not competitors
It would be too simple to stop there, and the evidence doesn't support a purely zero-sum reading. Large, predictable electricity loads can lower the average cost of generation and grid capacity for everyone else, the same way high passenger volumes let airlines cut ticket prices on popular routes — a dynamic the IEA explicitly notes as a possible upside of data centre growth, provided the policy and infrastructure mix is right.
There's also a growing shift in how the cost burden is allocated. In the United States, regulators have pushed hyperscalers to bear more of the infrastructure cost directly: PJM Interconnection has moved toward requiring tech firms to bid on long-term contracts for new generation capacity, and several major technology companies signed a Ratepayer Protection Pledge committing to fund the grid upgrades their facilities require rather than spreading the cost across ordinary consumers. Some operators are also becoming more flexible: Google has integrated roughly 1 gigawatt of demand-response capacity into its US utility agreements, shifting machine-learning workloads away from periods of peak grid stress — a small fraction of its total load, but a real signal that data centres can behave as grid assets rather than pure liabilities.
In emerging markets, the picture is more mixed than the crowding-out framing suggests. The IFC has identified around 15 priority markets — Brazil, South Africa, India and Malaysia among the leaders — for data centre investment specifically because they see spillover potential for local grids and digital infrastructure, and case studies from Malaysia's Johor corridor and South Africa show hybrid energy strategies and coordinated permitting turning data centre demand into bankable local infrastructure rather than a drain on it.
What this means for industrial decarbonisation finance
Put the two sides together and the honest answer is: it depends on what gets prioritised in the next two to three years, not on some inherent property of data centres themselves.
The risk is concrete and specific. Grid connection queues, permitting capacity and government attention are finite resources in the short term, even where capital in aggregate is not. When a data centre developer can outbid, out-negotiate or simply out-move a steel producer trying to electrify a furnace, the industrial project loses time it often cannot afford — decarbonisation pathways for cement and steel are already tight against 2030 and 2050 targets, and delay compounds. Blended finance structures, which exist precisely to de-risk industrial decarbonisation projects that private capital won't touch alone, depend on a pipeline of bankable projects reaching financial close. If grid access becomes the binding constraint rather than capital availability, blended finance solves the wrong problem.
The opportunity is just as concrete. Where governments have required hyperscalers to co-invest in grid upgrades, fund new generation, or accept flexible operating conditions, the resulting infrastructure — transmission lines, substations, even new generation capacity — is available to industrial users too, not just to the data centre that triggered it. The policy lever that determines which outcome dominates isn't a ban on data centres or an appeal to their goodwill; it's cost allocation and grid connection rules that treat large loads consistently, whatever they're for.
The practical takeaway for policymakers and project developers
A few things follow from this that are more useful than a general worry about AI's footprint:
- Grid connection queues need explicit prioritisation criteria, not first-come-first-served rules that favour whichever developer moves fastest. Several regulators are already experimenting with this, and it is the single most consequential lever available in the next 12–24 months.
- Cost-allocation reform matters more than headline capital totals. The US shift toward requiring large loads to fund their own grid upgrades is a template other jurisdictions can adapt, including for emerging markets where the temptation is to subsidise data centre connection to attract investment.
- Blended finance and concessional capital for industrial decarbonisation need to build grid access into project design from day one, rather than treating it as a downstream implementation detail once financing has closed.
- Flexibility is undervalued. Data centres that can shift load, and industrial processes that can do the same, both reduce the peak-capacity investment the whole system needs — which is good for electricity affordability across the board, not just for whichever sector gets the credit.
The AI buildout isn't going to slow down, and treating it as an adversary to industrial decarbonisation misses where the actual leverage sits. The grid doesn't distinguish between a server hall and a steel furnace when it runs out of transmission capacity. Whether 2026's investment supercycle ends up accelerating or delaying the industrial transition will be decided by connection rules and cost allocation, not by how much anyone disapproves of a hyperscaler's capital expenditure.
Further reading on Greener Wisdom: Blended Finance Explained · Policy Gaps in Industrial Decarbonisation · How Much Electricity Does ChatGPT Actually Use? · Grid Expansion vs Electricity Prices
Sources: IEA, Electricity 2026 and Key Questions on Energy and AI (2026); CEPR/VoxEU, "Powering the digital economy" (2026); IRENA Assembly proceedings, reported by Data Center Dynamics (2026); UNCTAD, global investment trends note (2026); OECD, Financing solutions to foster industrial decarbonisation in emerging and developing economies; Atlantic Council, "Powering data centers in emerging markets" (2026); Google 2026 Environmental Report, reported by Data Center Knowledge.
