When power is scarce, computation becomes an allocation problem.
AI data centres and Bitcoin miners look like different industries. Increasingly, they are bidding for the same physical stack: electricity, grid connections, substations, transformers, cooling, land and dispatchable capacity.
The question is no longer whether either industry uses a lot of power. It is which load a grid should serve first when not everyone can have what they want.
What Matters
The IEA says global data-centre electricity demand reached roughly 485 TWh in 2025 and could approach 950 TWh by 2030. AI-focused data-centre demand grew about 50% in 2025. Cambridge estimates Bitcoin mining at roughly 138 TWh in its 2025 industry study, while its live index has recently shown annualised consumption above 150 TWh.
Those numbers are globally manageable. The constraint is local concentration. A single AI campus or mining cluster can hit a grid node much harder than its global share suggests.
The real bottleneck is not energy. It is connection.
The IEA estimates more than 2,500 GW of generation, storage and large-load projects are stalled in grid connection queues worldwide. Transmission investment has not kept pace with generation and load growth, while transformer and cable lead times have stretched sharply.
That means the scarce commodity is increasingly not a kilowatt-hour in the abstract. It is a deliverable kilowatt-hour at the right substation, with the right voltage, at the right time, backed by sufficient transmission and generation.
AI and Bitcoin want different kinds of electricity
| Attribute | AI data centre | Bitcoin mining |
|---|---|---|
| Load profile | Dense, high-utilisation, increasingly spiky | Steady but highly interruptible |
| Revenue sensitivity | High value per compute job; downtime can be expensive | Directly tied to BTC price, difficulty and power price |
| Mobility | Moderate to low once built | Higher; containerised/mobile designs possible |
| Grid flexibility | Improving via storage and workload shifting, but limited for frontier training | High; miners can curtail nearly instantly |
| Connection urgency | Very high because GPU capex depreciates quickly | High, but machines can often be moved |
| Best power source | Firm, reliable, high-quality power | Cheap power, including stranded or volatile supply |
Bitcoin’s surprising grid advantage: it can disappear
Cambridge found surveyed miners curtailed 888 GWh of load during 2023. In Texas, ERCOT analysis released in 2026 showed crypto-mining loads becoming sharply more price-responsive as wholesale prices rose, with reductions that appeared almost instantaneous during high-price events.
That flexibility matters. A miner can turn machines off and stop earning. An AI training cluster may have expensive jobs in flight, tightly coupled accelerators and service commitments that make interruption materially more costly.
AI’s advantage: it can pay more
The economics on the AI side are different. The IEA says the largest hyperscalers are set to spend more than $700 billion on capital expenditure in 2026, mostly for new data centres. Those facilities are extraordinarily capital intensive, and idle GPUs can destroy project economics quickly.
That can justify paying more for firm power, long-term generation contracts, dedicated substations, batteries, gas turbines, nuclear power agreements or private transmission arrangements.
In a straightforward auction for scarce firm capacity, AI may simply outbid miners.
The migration is already happening
The industries are not waiting for policy to decide the contest. Bitcoin miners themselves are becoming AI infrastructure companies.
TeraWulf, for example, expanded into AI/HPC hosting and in November 2026 secured Google credit support for an Anthropic data-centre project in Kentucky. The planned campus is expected to provide roughly 401 MW of critical IT capacity. The strategic logic is simple: if the same powered site earns more from AI than mining, the electrons get reassigned.
This creates a new category of asset: the compute-switchable power site. The highest-value use of the interconnection can change over time.
DN Power Scarcity Allocation Index
DN proposes a framework for thinking about which large compute load is most likely to retain access when electricity becomes constrained.
Revenue per firm MWh — 25%
How much economic value can the workload support per reliable unit of electricity?
Interruptibility — 20%
Can the load curtail rapidly without destroying disproportionate value?
Location flexibility — 15%
Can the compute move toward stranded generation or underused grid capacity?
Grid-support value — 15%
Can batteries, generation or flexible load provide useful system services?
Connection sunk cost — 15%
How much capital is stranded if the site cannot obtain firm capacity?
Policy priority — 10%
How strongly do local policymakers prioritise the project for strategic, fiscal or employment reasons?
DN Power Scarcity Calculator
Score a project from 0 to 100 on each dimension. Higher scores imply greater ability to retain or secure scarce grid capacity.
The likely outcome is not AI or Bitcoin. It is segmentation.
Prime firm power
AI and other high-value, low-interruptibility industries are likely to dominate where reliability commands the highest premium.
Stranded and volatile power
Bitcoin can remain competitive where electricity is cheap because it is difficult to transport, intermittent or periodically surplus.
Flexible-load markets
Mining can become valuable when grids explicitly reward rapid curtailment and demand response.
Hybrid compute campuses
Sites may allocate the same power infrastructure dynamically between mining, AI/HPC and grid services depending on economics.
The policy conflict gets harder from here
Governments increasingly see AI infrastructure as strategically important, while communities worry about electricity prices, water use and who pays for new grid infrastructure. Bitcoin mining faces similar scrutiny but can argue that flexible load is easier to integrate than always-on demand.
The IEA’s October 2026 work on powering AI makes the central problem explicit: policy must balance affordability, energy security, sustainability and digital competitiveness at the same time.
That means the political question may become less about whether compute is “good” or “bad” and more about whether a project pays for the grid capacity it needs, provides flexibility when stressed and creates enough local economic value to justify scarce infrastructure.
DN Alpha Thesis
Electricity is becoming a strategic input for computation in the same way chips became strategic inputs for AI. But power is harder to scale quickly because transmission, substations and generation are physical systems with multi-year lead times.
Bitcoin mining may be pushed away from the best firm-power locations by AI. That does not necessarily kill mining. It may force miners toward the places where their flexibility is most valuable: stranded renewables, gas that would otherwise be wasted, curtailed generation and grids that need interruptible demand.
AI gets the premium electrons. Bitcoin gets the electrons nobody else can use efficiently all the time.
Sources
- IEA — Key Questions on Energy and AI (2026)
- IEA — Powering Artificial Intelligence (2026)
- IEA — Electricity 2026
- U.S. DOE — 2026 National Transmission Needs Study
- Cambridge Digital Mining Industry Report
- Cambridge Bitcoin Electricity Consumption Index
- ERCOT — Demand Response and Crypto Mining Load, 2026
- Reuters — TeraWulf / Google / Anthropic AI infrastructure project, Nov. 2026
FAQ
Does AI use more electricity than Bitcoin?
Data centres as a whole already consume substantially more electricity globally than Bitcoin mining, and AI-focused data-centre demand is growing much faster. Exact AI-only totals remain uncertain because operators do not disclose energy use consistently.
Can Bitcoin miners help electricity grids?
Yes. Mining can act as highly flexible demand because machines can reduce consumption rapidly when electricity prices spike or grids are stressed.
Will AI replace Bitcoin mining at existing power sites?
In some locations, yes. Several mining companies are already converting or expanding sites for AI and high-performance computing where those workloads offer better economics per unit of scarce power infrastructure.
