The Real Cost of AI: What Google and Amazon's Emissions Spike Signals for Your Token Margins

Key takeaways

• Google's total carbon emissions are up 25% year over year, Amazon's up 16%, and both companies point to surging energy use as AI scales.
• Emissions growth is a proxy for energy demand, and energy is a large, rising share of the cost of serving an LLM token.
• The $/1M-token number on a provider's pricing page is a floor, not a fixed input: treat it as something that can move.
• Builders who model margins only against today's prices are exposed to energy-driven cost shocks and quiet tier changes.
• Stress-test your unit economics against a scenario where inference costs rise 20% to 40%, not just the current rate card.

Why does a sustainability report matter for AI pricing?

Both Google and Amazon released sustainability reports this week, and the headline numbers are blunt: Google's total carbon emissions are up 25% since last year and Amazon's are up 16%. Neither company blames AI outright, but both acknowledge that energy use has climbed sharply as AI workloads have grown.

For a founder shipping an AI product, that is not an environmental footnote. Carbon emissions track energy consumption, and energy is one of the biggest and fastest-growing lines in the cost of running inference at scale. When the two largest cloud providers report that their power draw is outrunning their own efficiency gains, they are telling you something about the direction of your input costs.

What is the real cost of an AI token?

The price you see on a provider's pricing page, say a few dollars per million tokens, bundles compute, memory, networking, and energy into one clean number. What it hides is that the underlying cost of serving that token is not stable. It moves with GPU supply, data-center capacity, and, increasingly, the price and availability of power.

Here is the part the emissions reports make concrete: if the hyperscalers are spending more energy per unit of useful work, and paying more to secure clean power to hit net-zero pledges, that cost has to land somewhere. Historically it lands in one of three places: higher list prices, tighter rate limits, or repackaged tiers that quietly raise the effective price per token.

How should builders respond?

Do not model your margins as if the current rate card is permanent. A simple discipline helps: take your current cost per user, then re-run it assuming inference prices rise 20% to 40%. If your gross margin survives that, you have a durable business. If it collapses, your pricing or your architecture needs work now, while you still have room to move.

Concrete levers, for example: route cheap tasks to smaller models, cache and batch aggressively, cap tokens per request, and price your own product on usage so your revenue scales with the cost that drives it. The goal is to make your margin a function you control, not a bet on provider prices staying flat.

You can model both the current and the stressed scenario side by side in Calcaas, so you see the margin gap before it shows up on an invoice.

The one-line takeaway

Rising emissions at Google and Amazon are an early signal that the energy under every token is getting pricier, so build your margin model around a token price that can move, not one that stays still.