AI Data Center Electricity Costs: Why Infra Bills Matter for Startups

A real but narrower story than the headlines

Electricity is a genuine constraint on AI growth. The IEA’s Energy and AI report estimates that data centres consumed around 415 TWh in 2024, roughly 1.5% of global electricity use, and projects that figure could reach about 945 TWh by 2030. AI-optimised servers are among the fastest-growing parts of that demand.

Those numbers matter for grid planning and sustainability. For a startup buying tokens, the question is narrower: does electricity show up in your price, and what can you do about it?

Power is one cost among several

Running AI at scale needs chips, buildings, cooling, networking, land, financing and people as well as power. Capital costs dominate. Microsoft, for instance, spent $115.9 billion on property and equipment in its fiscal year to June, and in the September quarter capital expenditure rose about 30% to $45.2 billion, roughly three-quarters of it on short-lived assets such as chips and networking gear. Chips that wear out in a few years raise the financing burden far more than the electricity to run them.

How it might reach your invoice

Three channels:

1. API pricing. Providers set prices competitively. Efficiency gains have been lowering the cost per token. If energy or capital costs rise faster, providers may stop cutting or may raise prices, particularly where current prices are subsidised by investors. See the price war analysis.
2. Availability. Capacity limits show up as rate limits, waitlists or higher prices for the newest models.
3. Cloud bills. If you run your own GPUs on rented cloud capacity, power and chip costs are in the hourly price.

What you control

You do not control grid prices. You do control how many tokens a task consumes.

• Cache aggressively. Repeated context, such as system prompts and documents, can often be cached for a discount.
• Right-size models. Route classification, extraction and simple replies to smaller, cheaper models and reserve large models for hard tasks.
• Shorten prompts and outputs. Remove boilerplate, limit retrieved context and set output limits.
• Batch non-urgent work. Many providers discount asynchronous batch processing.
• Measure cost per completed task. A cheap call that fails twice is more expensive than a better model that succeeds once.
• Cap agent loops. Unbounded retries and tool calls can multiply cost with no benefit.

A worked example

A support assistant handles 100,000 tickets a month. Each ticket sends 6,000 tokens of context to a large model. The team caches the shared 4,000-token policy text, routes simple tickets to a smaller model, and trims retrieved snippets. Tokens per ticket fall by more than half, and cost per resolved ticket drops accordingly, without waiting for any change in electricity prices.

When self-hosting changes the picture

If you run your own models, you are exposed to power and hardware costs directly, and you can optimise them: efficient quantised models, consolidating workloads and choosing regions carefully. The trade-offs are in open-weight models.

Sustainability reporting

Enterprise customers increasingly ask about the emissions of AI features. Keep a rough estimate of token volumes by workload and ask your providers for emissions data. Efficient design helps both cost and carbon.

Link to the financing story

The scale of capacity being built ties electricity to the commitments covered in AI circular financing: capacity ordered ahead of revenue must be paid for. If demand slows, power and chips are stranded costs for someone, and prices could shift. Whether that makes this a bubble is examined in the bear case.

Questions to ask your provider

You will not get every answer, but asking shapes your risk picture:

• How much notice do you give before retiring a model or changing prices?
• Do you offer committed-capacity options that guarantee throughput at peak times?
• What discounts apply for cached and batch workloads, and how are they metered?
• In which regions is my data processed, and can I choose?
• Can you share emissions or energy-use estimates per workload for my sustainability reporting?
• What happens to my limits if capacity is constrained?

Write the answers into your vendor file and compare them again at renewal.

Your next step

Pull last month’s model bill, divide it by completed tasks and look at the three most expensive workflows. Try caching or a smaller model on the most expensive one this week and compare cost and quality. The savings will be visible before the next energy price headline.

Bottom line

Electricity is a real constraint but only one slice of what AI costs, and its effect on your prices is uncertain. Watch the capacity story, but spend your effort on the controllable part: fewer tokens per task, the right model for each job, and cost measured per outcome.