Every conversation about AI capacity starts with chips and ends with chips. That was correct for about eighteen months. It is now the wrong place to look, because the supply of accelerators is no longer the slowest link in the chain — and the slowest link always sets the pace.
What actually takes years
Ask anyone building capacity at scale what their critical path is and you get the same list, in roughly this order:
| Input | Typical lead time | Can money fix it? |
|---|---|---|
| Accelerators (H-class or equivalent) | weeks–quarters | Mostly yes, at a premium |
| High-voltage transformer | quarters (often 1–2 years) | Barely — foundry capacity is booked |
| Utility interconnection study + agreement | 2–5 years in congested regions | No |
| New transmission line | 5–10 years | Absolutely not |
| Local permits and community consent | unpredictable | Sometimes, with concessions |
Notice the pattern: the items money can't accelerate are all electrical. A hyperscaler with unlimited capital can outbid everyone for chips; it cannot outbid physics for a transformer that doesn't exist or a study queue that hasn't reached its number.
Why this changes the shape of new capacity
When you can't build the plant you planned, you go where power already exists and is underused. That's not a rhetorical flourish — it's the actual playbook of every serious operator right now:
- Behind the meter. Generating on-site (gas, increasingly fuel cells, occasionally geothermal) sidesteps the queue entirely. Slower to permit in some jurisdictions, but measured in quarters, not years.
- Stranded and curtailed power. Regions with surplus hydro, wind or gas flaring, and grids that curtail renewables at peak generation, have energy they cannot sell. Compute is the highest-value way to consume it locally.
- Second-life industrial sites. Retired smelters, paper mills and — yes — former mining facilities already have the three things that take years to get: a high-capacity connection, a substation, and neighbors who've seen heavy industry before.
- Smaller, unglamorous builds. 1–20MW incrementally added at many sites instead of 500MW at one site. Less elegant, far faster.
Every one of those options produces distributed, heterogeneous, modest-sized capacity. Which is a description of exactly the supply side decentralized GPU networks have been aggregating — and exactly the kind of supply a single large buyer can't easily control or monopolize.
The economics get interesting next
There's a second-order effect that matters even more than availability. Power is now the dominant share of the cost of an AI megawatt-hour, and players with cheap or stranded power sit on a structural margin that hardware depreciation no longer erases. In markets where these operators can sell to whoever bids highest — rather than to a single contracted counterparty — you get genuine price competition on compute for the first time. This is the mechanism we kept describing in our rate survey, supercharged by the fact that power, not silicon, is now the scarce input.
It also explains a pattern that puzzles observers: why so many AI-adjacent infrastructure deals in 2026 are really energy deals wearing a data-center costume. Follow the electrons and the strategy becomes legible.
What could break this thesis
Honest version, with the counterarguments stated at full strength:
- Grid upgrades eventually land. Interconnection queues are a policy artifact as much as a physics one; regulators facing pressure to speed up AI infrastructure will plausibly succeed in places. When they do, some of the distributed advantage compresses.
- Efficiency moves the goalposts. If per-token energy keeps falling fast, effective capacity can grow without any new megawatt. Demand has historically outrun efficiency — but that's a trend, not a law.
- Co-location incumbents adapt. The big operators aren't stupid; they're signing behind-the-meter deals and buying distressed industrial sites faster than most people realize.
None of these erase the structural point: for the next few years, the fastest new capacity comes from power that someone else already built. Whoever can aggregate that power — and turn it into verifiable, usable compute — sets the price of the marginal hour.
Why we're writing this on a compute forum
Because the decentralized compute argument used to rest on ideology and arbitrage: cheap GPUs nobody's using, coordinated without a middleman. The power wall upgrades that argument into something structural. It's no longer just "there is idle hardware." It's "the cheapest way to add AI capacity for the next several years is distributed power sites with local compute — and those don't fit inside a single hyperscaler's procurement process."
You can buy the chips. You cannot buy the electrons. Whichever network can reliably turn stranded power into verified compute hours will look, in hindsight, like it was standing in exactly the right place.
Comments