Plan to scale AI on the assumption that compute will not always be cheap, plentiful or guaranteed. A pilot’s running cost can multiply at full volume, and demand for data-centre power is rising fast. Boards should require a unit-cost model, capacity commitments from suppliers and a design that can switch to smaller models when needed.
In a recent interview, Eric Schmidt discussed the power AI will need. We take the topic, not the forecasts. What follows is AIPEX’s own guidance on what compute and energy constraints mean for a business planning to scale AI.
Why does compute belong on the board’s agenda?
For most businesses, AI arrives as a subscription or a usage bill, and the machines behind it are someone else’s problem. That is true until the business scales. At pilot volumes, the cost of running AI is often too small to notice. At full volume, across thousands of users or millions of transactions, it can become a material operating cost that varies with usage, and one that the business does not directly control.
The wider market adds a second reason. The International Energy Agency projects that electricity consumption by data centres worldwide will more than double, from around 415 terawatt-hours in 2024 to around 945 terawatt-hours by 2030 in its base case (International Energy Agency, Energy and AI, April 2025). Rising demand for power and chips can show up for customers as higher prices, usage limits or slower access to the newest models.
How does a pilot’s cost change at scale?
Pilots flatter economics in three ways. They run at low volume, often on the provider’s most capable model because quality matters more than cost in a test. They leave out the supporting costs of production: monitoring, security, retries and the extra processing needed to make outputs reliable. And they rarely test peak demand, when capacity is scarcest.
A scaling plan should therefore model cost per unit of business output, such as per claim handled, per customer query or per document reviewed, at expected and peak volume. If the unit cost is not known, the business case is not yet complete. Our 90-day playbook for moving AI from pilot to production covers the wider production checklist; compute cost is one of the items that most often surprises.
What design choices reduce the exposure?
Much of the risk can be designed out without technical detail reaching the boardroom. The board’s role is to ask whether these choices have been made deliberately.
| Choice | What it protects against | Question for management |
|---|---|---|
| Use the smallest model that meets the standard | Paying premium prices for routine work | Have we tested cheaper models against our quality bar? |
| Keep the option to switch model or provider | Price rises, usage limits or withdrawal of a model | How long would it take us to move to an alternative? |
| Reserve capacity for critical services | Slow or refused service at peak demand | Which AI services must keep running, and is capacity committed? |
| Set usage limits and alerts | Runaway costs from errors or misuse | Who is alerted when AI spend exceeds plan? |
| Include compute in the business case | A return that disappears at volume | What is our cost per unit of output at full scale? |
What should supplier contracts cover?
Compute risk usually reaches a business through its contracts. The terms that matter most are price change notice and caps, committed capacity or service levels for critical uses, notice before a model is withdrawn or materially changed, and the practical ability to leave. Many AI contracts were signed during pilots, on standard terms, before anyone knew how much the business would come to depend on them. Scaling is the moment to renegotiate, while the business still has alternatives. Procurement and legal teams should be briefed on what the board expects before those conversations start, so that commercial terms reflect the scale the business is planning, not the scale of the pilot.
Sustainability commitments add a further angle. Many organisations have public commitments on emissions, and some report under sustainability rules. AI running at scale adds to energy use, even when it runs in a supplier’s data centre. Boards with such commitments should ask suppliers for the energy and emissions information they need for their own reporting, and should expect management to weigh model choice against both cost and footprint. Using a smaller model where it does the job well is often better on both counts.
What should the board ask before approving scale?
- What is the cost per unit of output at full volume and at peak, and how does it compare with the pilot?
- Which AI services are critical, and what capacity or service commitments protect them?
- How much would a 50% rise in AI prices change the business case?
- How quickly could we switch model or provider if we had to?
- Who watches AI spend each month, and what triggers an escalation?
Where to start
Our briefing on how to scale AI across the business, from pilot to production sets out the six conditions that must be true before you scale, four phases from pilot to production and a 100-day roadmap. For boards where pilots are running and nobody has agreed how to scale them, the step on its four-step path that fits is step 3, Plan: an Executive AI Strategy Day (£1,499 + VAT) agrees priorities, named owners, a benefits scorecard, a 90-day plan and a 12-month roadmap. The Scaling AI Board Checklist below is a quicker first pass.
Frequently asked questions
Why does AI cost more at scale than in a pilot?
Pilots run at low volume, often on the most capable model, and leave out production costs such as monitoring, security and retries. At full volume, usage-based costs can become material. Boards should require a cost per unit of business output at expected and peak volume.
How can a business reduce its exposure to AI compute costs?
Use the smallest model that meets the quality standard, keep the option to switch model or provider, reserve capacity for critical services, set usage limits and alerts, and include compute in every business case at full scale.
What should AI supplier contracts say about capacity and price?
Price change notice and caps, committed capacity or service levels for critical uses, notice before a model is withdrawn or changed, and practical exit terms. Scaling is the time to renegotiate. See How to Scale AI Across the Business →
Correct as of 28 September 2026. This content reflects the law, official guidance and deadlines as published on that date. The EU AI Act and its guidance, standards and deadlines are still being amended and clarified. If you are reading this after 28 September 2026, check for later amendments, European Commission guidance and harmonised standards before relying on it. It is general information, not legal advice, and AIPEX accepts no liability for decisions taken on it without specific professional advice.