The story everyone is reading wrong is that AI's real bottleneck is electricity, not chips. On June 18, 2026, the US Federal Energy Regulatory Commission (FERC) ordered six major grid operators to give data centers a fast lane onto the grid, per TechCrunch. What it did not do is add a single megawatt of supply.
I think that gap matters more than the order itself, and it changes how I'd plan any AI project from here, including one run from a laptop in Colombo.
⚡ What FERC actually ordered
The regulator told the six operators to prove that data centers can connect "in a timely and orderly manner," and to consider alternative transmission technologies like solid-state transformers and superconducting lines. There are deadlines attached:
| Requirement | Deadline |
|---|---|
| Report spare generating capacity | 30 days |
| Defend or revise electricity rates | 60 days |
| Speed up large-load interconnections | Ongoing mandate |
Key takeaway: This is a paperwork-and-priority order. It reshuffles who gets to the front of the queue. It does not build new power plants, and faster access to a constrained resource is not the same as more of it.
📊 A queue problem, not a supply solution
Here is the number that reframes the whole thing. As of late 2023, grid-connection requests for new power plants already exceeded the total capacity of the existing US fleet. The waiting list to add supply was longer than everything currently running.
Now stack demand on top:
- Data center electricity demand is projected to rise nearly 300% through 2035.
- Wholesale electricity rates in some regions have already climbed 267% over five years.
- The administration spent $2.6 billion to cancel offshore wind projects, including a $765 million payout to scrap leases.
So we are accelerating demand, accelerating queue priority for the biggest consumers, and simultaneously removing planned supply. You do not need a grid model to see where rates go.
🌐 Why a Sri Lankan builder should care
It is tempting to file this under "American problem." It is not. The cost of the AI you call from here is downstream of that grid.
When I rent a model through an API, I am renting a slice of a data center that pays for power at those wholesale rates. If electricity is the binding constraint, three things follow:
- Inference prices have a floor that is rising. Cheap tokens were partly a land-grab. Energy economics pull the other way.
- Free tiers get stingier first. The cheapest customers are the easiest to throttle when power is the limiter.
- Efficiency becomes a feature, not a nicety. Smaller models, caching, and batching stop being optimizations and become how you stay solvent.
If you are budgeting around free quotas, it is worth tracking what each provider actually gives you. I keep a running comparison on the AI free-tier comparison tool so I can see who is squeezing limits before my bill does.
🛠️ How I'm adjusting my own stack
None of this means stop building. It means build like power is expensive, because for the people upstream of you, it is.
| Habit | Why it pays off now |
|---|---|
| Cache aggressive, recompute rarely | Every cached response is an API call you never pay for |
| Default to the smallest model that passes your eval | Fewer tokens, lower energy footprint per request |
| Batch overnight jobs | Off-peak compute is cheaper and less contended |
| Keep a local fallback (Ollama, small open weights) | Insulates you from price shocks on hosted APIs |
There is a local mirror to this story too. Our own grid runs on time-of-use pricing, and if you are running anything power-hungry overnight, the Sri Lanka TOU tariff calculator shows exactly how off-peak versus peak rates stack up. The principle is identical at both scales: schedule heavy compute when power is cheapest.
Bottom line: The same discipline that lowers your CEB bill at home is the discipline that will keep hosted AI affordable. Treat compute like electricity, because increasingly it is electricity.
💡 What this means for you
FERC moved data centers to the front of the line without lengthening the line. For builders far from that grid, the signal is clear: the cost curve of AI is being set by power, and power is the part nobody fixed today.
So plan for it:
- Assume hosted AI gets more expensive, not less. Architect so you can downgrade models or self-host without a rewrite.
- Measure tokens and calls like you measure rupees. Waste here is now energy waste, and energy is the scarce thing.
- Lean on open weights and free tiers while they last, but never let a project depend on a free quota staying generous.
I am not betting against AI. I am betting that the winners over the next few years are the people who learned to do more with less compute, because the grid is going to force that lesson on everyone eventually. Better to learn it now, on your own terms, from a small team that never had cheap power to begin with.
