Enterprises and cloud providers deploying AI inference face three competing pressures on where to place capacity: low carbon, low cost and proximity to users.

• Low carbon is important because emissions from AI are under growing scrutiny from investors, customers and regulators, and because many operators have made public net-zero and renewable energy commitments that rapid AI growth now puts at risk (see Cutting AI carbon requires more than energy efficiency).
• Low cost matters because AI spending is spiraling while returns remain elusive: many organizations have invested heavily in AI infrastructure and services but have yet to demonstrate a clear return. The economics of tokens add to the pressure. Even as the price per token falls, reasoning models and agentic workflows consume far more tokens per task, so total bills keep rising. In addition, providers selling inference by the token need to produce each one for less than they charge (see A culture of token abundance collides with financial reality).
• Proximity to users is relevant because interactive applications such as chatbots, agents and real-time assistants need low latency to feel responsive. Data residency and sovereignty rules increasingly require that data, and the processing of it, remains within national or regional borders (see What cloud sovereignty really means).

Latency and sovereignty requirements can constrain where inference is deployed, limiting operators' ability to choose locations based on cost or carbon. For workloads that are less sensitive to latency, there is considerably more flexibility.

Cost versus carbon

Figure 1 compares the cost and carbon of generating tokens across a range of locations. These figures isolate the effect of electricity prices and carbon, assuming identical infrastructure, utilization and operating conditions across locations. The figure is based on a depreciated data center with no new labor required (assumptions can be found in Where to deploy AI inference: a guide to the economics).

Figure 1 Carbon by cost per token across locations; quadrants divided by median values

Across all 40 locations, the cost of producing a million tokens ranges from $0.73 in Quebec, Canada to $0.84 in the UK, a spread of about 15%. Carbon per token, by contrast, ranges from 0.0004 mgCO₂e in Quebec to 0.27 mgCO₂e in South Africa, a difference of several hundredfold.

Cost varies little because hardware capital dominates it. At around $0.72 per million tokens, hardware accounts for 86-99% of the cost, while power makes up between 1% of the total in Quebec and 14% in the UK. Moving workload to a cheaper power source therefore delivers only modest savings, and utilization and hardware efficiency remain the most significant cost levers. Moving workload to a cleaner grid, however, can cut local emissions by one or two orders of magnitude.

Trade-offs

The quadrants in Figure 1 show how locations trade off one goal against the other. Customers can achieve low cost and low carbon in some regions, but those locations might not be the ones proximate to users.

• Cheaper, lower carbon. These grids are dominated by hydro and nuclear power. These are long-lived, often long-paid-off assets with low running costs, so their power is both clean and cheap. Grids with high shares of wind and solar, such as Spain's, sit just outside this group but may move into it as renewable energy grows.
• More expensive, lower carbon. These grids are clean, but gas plants often set the wholesale price. Network charges, levies and taxes add to the cost.
• Cheaper, higher carbon. These are typically fossil-heavy grids where abundant domestic coal, oil or gas keeps power inexpensive. Subsidized tariffs also help reduce cost.
• More expensive, higher carbon. Costly imported gas or EU carbon pricing on coal drives up costs.

Unfortunately, many of the lowest-carbon, lowest-cost locations, such as Quebec, Norway, Sichuan and the US Pacific Northwest, are some distance from the largest concentrations of users.

Several major demand hubs sit on more carbon-intensive grids, including Singapore, Japan, Taiwan, Germany and much of Australia's east coast.

However, Virginia's carbon generation-based figure is below the median, helped by nuclear power, although its data centers also draw on power imports from a more fossil-heavy regional grid. Dublin and London are relatively low carbon, but among the most expensive locations.

Less carbon and cost

Latency-sensitive inference may have to stay close to users and endure a higher carbon cost. Workloads that can tolerate delay could move to the clean, cheaper corner of Figure 1 and cut emissions dramatically for almost no cost penalty. Examples include batch inference, background agent tasks and offline evaluation.

If a workload can endure slightly more latency, siting it elsewhere within a country can reduce carbon and/or cost. For operators bound by data residency rules, this is an opportunity: they can often meet sovereignty requirements and still cut emissions sharply by choosing the right region within a country.

For example, in Canada, carbon per token ranges from almost zero in Quebec to 0.09 mgCO₂e in Alberta. US states vary by about fourfold, from 0.033 mgCO₂e in Washington to 0.131 mgCO₂e in Ohio. In China, Sichuan's hydro-heavy grid is only a fifth as carbon-intensive as Inner Mongolia's coal-heavy grid. In Australia, the state of Victoria is about four times as carbon intensive as the island state of Tasmania.

Cleaner locations can have a surprisingly small cost penalty under the model. If carbon reduction is paramount, reductions can be made for a relatively small increase in token cost.

Assumptions and sources

These results come with some assumptions:

• Price bases differ. US and Canadian prices are large industrial rates, excluding tax. Other countries use a smaller-business benchmark that includes taxes, which probably overstates their costs relative to North America. China and Australia use one national price for all their regions.
• Carbon figures are based on official grid emission factors. They reflect direct emissions on a location-based (Scope 2) basis.
• Carbon intensity sources. European Environment Agency, 2023 (EU countries); UK DESNZ greenhouse gas conversion factors, 2026; national government grid factors for other countries, 2023-2024; US Environmental Protection Agency, eGRID2023; Canadian government sources, 2020-2024; China Ministry of Ecology and Environment, 2022; Australian National Greenhouse Accounts Factors, 2025.
• Electricity price sources. GlobalPetrolPrices.com, business rates averaged over 2023-2025, including taxes; US Energy Information Administration, 2024 industrial rates; Hydro-Québec, 2025 large-customer rates, excluding taxes; Eurostat, second half of 2025, non-household rates.