The people building your AI just told you what it costs

The companies you build on top of just put a number on the table, and it is enormous. OpenAI says it will spend $750 billion. Google is spending up to $190 billion this year alone. Anthropic keeps signing gigawatt-scale chip deals. All that money has to come back out of the product somewhere, and that somewhere is your pricing and your margins. Let me catch you up on what changed and what to do about it.

The bill your vendors are pre-paying

Start with the raw size. OpenAI now plans $750 billion on infrastructure through 2030, about 25% more than it said earlier this year. Its first move is a $20 billion campus in Georgia pulling at least 3.2 gigawatts from the local utility. AMD is putting up to $5 billion into Anthropic, which will deploy up to 2 gigawatts of AMD chips starting in 2027.

These are not vanity numbers. They are the cost base under the API you call every time a user hits your AI feature. When a vendor commits hundreds of billions to compute, that spend eventually shows up in token prices, rate limits, or the tier you get pushed onto. Plan your unit economics like the cheap inference you enjoy today is a promotion, not the permanent price.

Why demand keeps getting worse

The power side is moving faster than the forecasts. BloombergNEF now expects data centers to use one-fifth of U.S. electricity by 2035, four times today's draw. That estimate is 83% higher than the same firm's guess from just December. Nearly half of the new capacity goes to training and running models.

That demand hits grids that are already tight. In the PJM region, electricity prices jumped 76% in a year, and data centers made up 38% of charges in the last capacity auction. Constraint like that does not stay upstream. It reaches you as higher API costs and slower access to the newest, hungriest models. If your roadmap assumes you can always get more compute on demand, build a backup that does not.

The receipt Google just showed

Here is the part that should shape how you defend your feature. Google's cloud revenue jumped 82% to $24.8 billion, driven by companies buying enterprise AI. Its backlog of signed-but-unbilled cloud work hit $514 billion. That is the proof that the buildout is paying off for the seller: businesses are writing long-term checks.

But notice who is paying. Sundar Pichai points to "long-term deals" from enterprises as the reason to keep spending. The money coming back is coming from customers like you locking in multi-year commitments. When you sign one, you are helping fund the next data center. Read those contracts for what they are, and make sure the price you lock in still leaves you a margin if your own pricing has to hold flat.

The pledge that is a pinky promise

There is real backlash to all this, and the response so far is thin. Nearly 200 utilities and developers signed Trump's "rate payer protection pledge", covering about 80% of U.S. power. It is meant to keep AI's energy costs off regular consumers' bills. The catch: it is voluntary, carries no penalty, and rates are set by state regulators, not the White House. PJM is still expected to add $6.3 billion in costs for consumers across 13 states.

The environmental toll is local and real, too. Researchers estimate AI could reach 1,300 premature deaths a year from air pollution by 2030 and use hundreds of billions of liters of water. Some projects are already being downsized or blocked. That means supply risk for you, and it means your "AI-powered" branding may start drawing questions you should be ready to answer.

The deep cut

Watch the timelines. Georgia Power's electricity for OpenAI's campus does not start flowing until 2028. AMD's first gigawatt for Anthropic lands in the first half of 2027. The huge spending is buying capacity that is years out, which means the tight, pricey period is now and next year, not later. If your growth plan needs a lot more inference in the next 18 months, you are competing for compute that has not been built yet.

So do two things before your next review. Model your feature's cost at two or three times today's token price and see if it still stands. And find the parts of your product that do not need the biggest, newest model, because cheaper and smaller is where your defensible margin will live while everyone else fights over scarce top-tier compute.

Three questions for your team

1. If our vendor's API price doubled next year, which features stay profitable and which quietly become charity? Get the number, not a guess.
2. Where are we paying for a frontier model when a smaller or cheaper one would do the job? Move those calls now, before the price forces you to.
3. What is our answer when a customer or regulator asks about the energy and water behind our AI features? If we don't have one, who owns writing it this quarter?