The cheapest electricity in Canada was developed for Bitcoin mining — and it is the exact same power advantage that makes local AI inference, on-premise LLM servers, and distributed compute workloads viable around the clock.

D-Central has been operating at the intersection of energy and Bitcoin since 2016, building on the foundational work of ASIC manufacturers like Bitmain and MicroBT, the open-source firmware community, and the Bitcoiners who spent a decade proving that interruptible, location-agnostic compute could monetize power that conventional industries could not reach. That decade of energy discipline is now the entry point for sovereign AI.

Power is the common denominator

Bitcoin mining and local AI inference share one constraint above all others: energy cost. Miners learned this first. Proof-of-work competition collapses margins until only the cheapest kilowatt-hours survive. The operators who built that cheap-power discipline — identifying low-rate jurisdictions, negotiating industrial tariffs, optimizing every watt of infrastructure — are the same operators who can now run local AI servers at a cost structure that cloud providers cannot match at the edge.

Cloud AI is priced to extract margin at scale. Sovereign compute is priced by your electricity bill. If your electricity is cheap, your inference is cheap. The mining community understood this before the AI industry did.

Canada’s energy mix — dominated by hydroelectric in Quebec, Manitoba, and British Columbia, with wind and natural gas contributing elsewhere — creates jurisdictions where commercial electricity rates are structurally competitive for both workloads. Rates vary significantly by province, utility, and commercial agreement; the Energy & Sustainability archive covers the specifics province by province.

ASIC efficiency: the energy discipline mining taught

The ASIC industry has driven mining efficiency from hundreds of joules per terahash to under 20 J/TH on current-generation hardware (manufacturer-rated figures; real-world efficiency depends on firmware tuning, ambient temperature, and power delivery quality). That progression was not accidental — it was forced by economics. Miners who could not keep pace with efficiency improvements were priced out of the network.

The same logic applies to AI inference: energy per token is the operational metric that determines whether a self-hosted model is economically rational. The tools miners already use — power profiling, thermal monitoring, per-watt optimization, efficiency curve analysis — translate directly to managing a GPU inference server.

D-Central’s ASIC power profile database documents efficiency tuning data across 820 operating points on 28+ Antminer model lines, assembled from manufacturer data, open-source benchmark contributions, and field measurements. That data layer is the foundation for understanding how mining loads can coexist with, or give way to, AI compute workloads on the same electrical infrastructure.

For deeper technical context on how ASIC chips achieve their efficiency ratings, the ASIC chip reference covers the semiconductor architecture behind the numbers — a resource built on the published engineering work of Bitmain, MicroBT, Canaan, and the academic community studying proof-of-work hardware.

The second workload: AI inference alongside mining infrastructure

A common misconception: that “pivoting from mining to AI” means repurposing an ASIC to run language models. That is not the bridge. ASICs are SHA-256 engines; they cannot run general matrix multiply workloads. The stranded power and sovereign compute overview is direct on this point.

The actual bridge is the power infrastructure itself. A mining facility’s electrical capacity — the service entry, the distribution panels, the cooling plant — can host a parallel GPU inference server on the same circuit. The ASIC keeps mining; the AI server runs alongside it. The cost advantage flows from the cheap power contract that was negotiated for the mining load.

This parallel-workload model is exactly what distributed compute describes: disaggregating the compute layer so that hash power and inference throughput can coexist, share infrastructure costs, and serve different revenue streams. Bitcoin mining provides a predictable base load that justifies the power contract; AI inference adds optionality without requiring additional infrastructure investment.

GPU power draws vary by model and workload. Modern AI inference GPUs range from under 150 watts (consumer-tier, suitable for small language models) to 700+ watts per card (data-center class accelerators). A single kilowatt of available electrical capacity — a modest allocation by mining standards — is sufficient to run capable open-source models for personal or small-team use. A mining operator running multiple kilowatts has room for meaningfully larger deployments. Sizing depends on your specific GPU selection, model requirements, and cooling capacity; D-Central’s AI sovereignty consulting covers scoping for mining-adjacent deployments.

Province-by-province: where Canada’s energy advantage is strongest

Canada’s energy geography creates distinct profiles for mining-plus-AI deployments. The competitive advantage is real but uneven — specific rates require direct engagement with utilities, and commercial tariff schedules change. The province guides below provide operational context; always verify current schedules independently before sizing a deployment.

Quebec

Quebec’s hydroelectric grid (operated primarily by Hydro-Québec) is the most-cited Canadian jurisdiction for both Bitcoin mining and potential AI compute deployments. Hydroelectric generation provides stable, renewable power. Quebec’s Bitcoin mining energy guide covers the rate environment, grid dynamics, and regulatory context in detail. Quebec’s Loi 25 data-privacy framework also makes on-premise AI deployments a compliance advantage for organizations handling Québécois resident data.

Alberta

Alberta operates a deregulated electricity market, creating both opportunity and volatility. Operators with access to natural gas self-generation or stranded wind resources have historically found competitive cost structures. See the Alberta Bitcoin mining guide.

British Columbia

BC Hydro serves the majority of the province with hydroelectric power. Industrial rate availability varies; some regions have constrained grid capacity. See the British Columbia Bitcoin mining guide.

Ontario

Ontario’s grid uses time-of-use pricing structures that create optimization windows for interruptible loads. The Ontario Bitcoin mining guide covers the rate structure and its implications for both mining and AI compute scheduling.

From energy advantage to distributed sovereignty

The energy angle is not just economic — it is political. Canadian organizations that depend on US cloud providers for AI inference are subject to US export-control law, the US CLOUD Act (which reaches data held by US companies on infrastructure outside US territory), and the unilateral policy decisions of those providers. The Claude Fable export-control event of June 2026 illustrated how quickly that dependency can create operational disruption.

A local AI server powered by cheap Canadian electricity is sovereign in a way that a cloud API subscription cannot be. The same energy advantage that built Canada’s Bitcoin mining industry is now the competitive moat for Canadian organizations that want to own their AI compute stack.

This is the thesis behind D-Central’s Sovereign AI Canada pillar and the local AI versus cloud AI comparison: sovereignty is not an ideological position, it is an operational one. You either own the compute or you rent it on someone else’s terms.

For Canadian organizations evaluating this transition, running a local LLM in Canada covers the practical stack — hardware, model selection, power requirements, and total cost of ownership — without the ideological overhead.

Built on the shoulders of the energy and mining community

D-Central’s energy-for-compute work builds explicitly on the research and engineering that came before it:

• Cambridge Centre for Alternative Finance (CCAF) — the Cambridge Bitcoin Electricity Consumption Index (CBECI) established the methodology for measuring Bitcoin’s energy footprint and identifying mining’s geographic distribution by energy mix. That work is the foundation for understanding why cheap-power jurisdictions attract mining.
• Bitmain and MicroBT engineering teams — the efficiency roadmap from early-generation ASICs to sub-20 J/TH hardware was driven by manufacturer R&D investment. D-Central’s power profile database builds on top of, not in competition with, that published engineering work.
• BraiinsOS and the open-source firmware community — autotuning research and open benchmarking made per-ASIC efficiency optimization accessible to operators at every scale. The firmware comparison covers the landscape D-Central works within.
• The Bitcoin off-grid and stranded-power community — operators who pioneered flared-gas mining, hydro-spillage monetization, and solar-plus-mining proved the off-grid compute model before “sovereign AI” was a phrase. Their field reports are the empirical foundation the current AI-plus-mining thesis rests on.

D-Central’s role in this ecosystem is to translate that energy and hardware knowledge into sovereign compute deployments for Canadian operators — not to claim credit for discoveries made by the community over the past decade. See the mining field manual for the operational knowledge base this work draws from.

Frequently asked questions

Does running a local AI server use more power than a Bitcoin mining ASIC?

It depends entirely on the hardware. A Bitaxe SOLO open-source miner runs at roughly 15 watts. A commercial Antminer S21 series ASIC runs at roughly 3,500 watts under full load (manufacturer-rated; real-world draw varies with firmware tuning and ambient conditions). A GPU inference server card — such as an NVIDIA RTX 4090 — has a manufacturer TDP of 450 watts and typically draws less under light inference workloads. A mining facility built to handle commercial ASIC electrical loads has more than sufficient capacity to host an AI inference server in parallel. The comparison that matters operationally is energy cost per unit of useful work: joules per terahash for mining, or energy cost per inference for AI.

Can I run a local LLM on the same electrical circuit as my ASICs?

Yes, if your circuit has headroom. Total amperage is the constraint — each device’s steady-state draw must fit within your service capacity with appropriate margin for startup transients. Cooling is the second constraint: both ASIC miners and GPU inference servers generate significant heat that must be managed, and a combined load may exceed what a cooling plant sized for mining alone can handle. The operational skills from managing ASIC thermal environments transfer directly to GPU server management. D-Central’s AI sovereignty consulting covers load assessment and cooling design for mining-adjacent AI deployments.

Which Canadian provinces offer the best energy conditions for both mining and AI compute?

Energy rates, availability, and commercial tariff terms vary significantly by province, utility, load profile, and negotiated agreement. Hydroelectric-dominant provinces — Quebec, Manitoba, and British Columbia — have historically offered competitive commercial rates due to low marginal cost of generation, but availability to new industrial customers fluctuates with grid capacity and utility policy. Alberta’s deregulated market creates different dynamics. Specific rates require direct engagement with the utility or an industrial electricity broker. D-Central’s province guides (Quebec, Alberta, BC, Ontario) provide operational context; always verify current tariff schedules before committing to a deployment.

Is Bitcoin mining energy “wasteful” — and does that argument apply to AI compute?

Bitcoin mining consumes energy to produce security for a decentralized monetary network — the energy expenditure is the security mechanism, not a by-product to be minimized. The same structure applies to AI inference: running a local model consumes electricity to produce inference that would otherwise be outsourced to a cloud provider’s data center. Both workloads have real energy costs. The relevant question is not whether energy is consumed, but whether the output justifies the cost and whether the power source is appropriate. Mining’s decade-long push into stranded, curtailed, and renewable power sources has made it one of the most geographically flexible energy consumers on the grid — and that same flexibility extends to sovereign AI deployments built on mining infrastructure.

What is “stranded power” and why does it matter for AI compute?

Stranded power refers to electrical generation capacity that cannot be economically transmitted to conventional demand centers — hydro spillage, flared natural gas, curtailed wind — or capacity that exists in remote locations without industrial demand nearby. Bitcoin mining was the first large-scale use case that could follow stranded power to its source, because mining hardware is modular and location-agnostic. Local AI inference follows the same logic: a GPU server can be deployed wherever cheap power exists, serving the local team or organization rather than requiring a data-center connection. D-Central’s stranded power and sovereign compute overview covers the economics and practical design in detail.

How do I start building sovereign AI compute at my mining facility?

The starting point is your existing power and cooling envelope. D-Central’s sovereign AI consulting service assesses your available capacity, recommends GPU hardware appropriate for your inference requirements, and designs a deployment that coexists with your mining operation rather than competing with it. All deployments are scoped as build-to-order; there are no standard packages because every facility’s power, cooling, and use-case requirements differ. Replacing cloud AI with a local LLM is a practical starting guide for operators evaluating the transition independently.