Energy, Environment and Physical Limits

The scale, from a primary source

The International Energy Agency's Energy and AI analysis provides the most useful published figures for a product manager who needs to reason about this rather than repeat headlines.

Two comparisons from the same analysis put it in proportion. Data centres account for around one-tenth of global electricity demand growth to 2030 — less than industrial motors, air conditioning or electric vehicles — but more than 20% of demand growth in advanced economies, where the concentration bites. And by 2030 the United States alone is projected to consume more electricity for data centres than for producing aluminium, steel, cement, chemicals and all other energy-intensive goods combined.

AI is identified as the most important driver of this growth, though the analysis does not isolate a single figure for AI's share of data centre power as distinct from other digital services — a distinction worth preserving when you quote it, because secondary sources frequently blur it.

Why this becomes a planning constraint

Not primarily a corporate responsibility matter. Three practical routes by which it reaches a product roadmap:

Capacity is gated by power. A data centre needs a grid connection, and in constrained regions those take years. This is upstream of the capacity discussion in module 5 — the reason regional capacity is finite is often that the power is.

Location follows energy. Providers build where power is available and affordable, which shapes which regions exist — and interacts directly with the residency requirements in module 6. The compliant region and the cheap-power region are not always the same place.

Procurement asks. Enterprise and public-sector buyers increasingly require environmental disclosure in tenders, and "we do not know" is a weak answer in a competitive process.

What a product team can actually influence

Most of the footprint is determined by choices above you — the model, the hardware, the data centre's power mix. But the levers you do control are the same ones that reduce cost, which makes this unusually tractable:

The alignment is worth naming explicitly: the module 4 cost levers and the energy levers are the same levers. A product that has done the work to control unit economics has already reduced its footprint, and can say so with a straight face.

The last row generalises a point the managers course makes: an always-on pattern that reprocesses continuously, where periodic processing would serve equally well, consumes far more for little benefit. Choosing the scheduled design is both cheaper and lower-impact, and it is a design decision rather than an efficiency initiative.

Answering the procurement question honestly

You will be asked. What a credible answer contains:

• What you can measure — your request volume, your model choices, your provider's published regional carbon intensity where available.
• What you cannot — a precise per-request figure, unless your provider publishes one. Say so rather than estimating something unfounded.
• What you have done — routing, caching, context reduction, scheduled rather than continuous processing. These are concrete and verifiable.
• What your provider reports — their published commitments and regional energy mix, cited as theirs rather than as yours.

Do not invent a number. Fabricated per-request carbon figures are increasingly checked, and being caught overstating environmental credentials is a worse commercial outcome than admitting the limits of what you can measure. The honest version — here is what we control, here is what we have done, here is what our provider reports — is both more defensible and, in most tenders, sufficient.