AI's Data Center Problem: What It Means for the Planet (and What We're Doing About It)

Every AI answer you've ever received was computed in a warehouse full of servers, drawing power from a real grid and, in many cases, cooled with real water. As AI use explodes, that footprint is becoming one of the technology's defining problems. Here's an honest look at it — including where we fit in, as a small company built on top of these systems.

The scale of the thing

Data centers were already estimated to consume on the order of 1–2% of global electricity before the generative-AI boom. The International Energy Agency has projected that data center consumption could roughly double over just a few years, with AI as a major driver. Training a single frontier model consumes enormous amounts of energy, and inference — the everyday answering of questions like yours — now adds up to even more, because it happens billions of times a day.

There's also water. Many data centers use evaporative cooling, and studies have estimated meaningful water consumption per session of AI use — a cost that lands hardest in the drought-prone regions where data centers often get built.

The honest caveat: precise per-query numbers are hard to pin down and vary wildly by model, hardware, and grid. Anyone quoting an exact "grams of CO2 per prompt" figure with confidence is overselling. But the direction is not in dispute: more AI means more electricity, and the industry is racing to build capacity faster than grids are decarbonizing.

The part nobody talks about: wasted compute

Here's an angle we think deserves more attention. Some meaningful slice of AI's footprint is spent on compute nobody needed:

• Oversized models for tiny questions. Asking a frontier model "how many cups in a quart" is like driving a semi truck to the mailbox. It works, but the energy spent is wildly out of proportion to the task.
• Use-it-or-lose-it psychology. Flat subscriptions that reset monthly quietly encourage burning through capacity before the clock runs out — usage driven by the billing model, not by need.

Efficiency won't solve AI's footprint on its own. But waste is the one part of the problem where product design — not just chip design — can help.

Where Apiary honestly fits

Let's be upfront: we're a small company. We don't run data centers; we build on top of the major model providers, who make their own (increasingly public) commitments on renewable energy and cooling. We're not going to pretend a startup our size moves the grid.

But two things about how Apiary is built genuinely cut waste:

Right-sized compute by design. Apiary's mode system — and the advisor that recommends a mode for your question — exists to match the size of the model run to the size of the task. Quick question, light mode, less compute. That saves you credits, and the same mechanism that saves you money spends less energy. Our incentives point the right way: because we pay providers for every token, waste costs us directly. We will never profit from you burning compute you didn't need.

No use-it-or-lose-it pressure. Rollover credits mean there's never a reason to spend down an allowance before month's end. Nobody should run prompts they don't need because a billing cycle told them to.

What we're committing to

We'd rather under-promise here than greenwash. As Apiary grows, we commit to: being transparent about which providers we route to and what they publicly report about their energy sourcing; continuing to make the efficient path the default path in the product; and, once we reach meaningful scale, putting a defined share of revenue toward a green giveback program — and publishing the details when we do, not before.

That last one is deliberately unfinished. Announcing a program before it exists is exactly the kind of thing this post argues against.

What you can do (without giving up AI)

Match the tool to the task — the biggest single lever an individual user has. Use lighter modes for lighter questions. And favor services whose billing doesn't reward waste. If that sounds like the product we happened to build, that's not a coincidence: Apiary starts at $5/month, routes your questions to the right-sized answer, and never makes you race a reset clock. Try it here.