AI’s Shortage Moved From Chips to Land and Power
The GPUs already powered on in data centers are fully used, with more work per megawatt as the way forward.

The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch · 20VC: Taking on NVIDIA: Fractile CEO on the Future of Chips | What Everyone Gets Wrong About Capex Spending | The Winners and Losers in Neo Clouds with Walter GoodwinThe tightest limit on artificial intelligence is no longer finding chips. It is finding land to put them on and electricity to run them, says Walter Goodwin, chief executive of chip startup Fractile.
The bottleneck has moved about every quarter for four years. Three years ago it was a shortage of Nvidia graphics processors, the chips that do AI math. Then it became a broader shortage of silicon and supply chains, including HBM, a fast memory stacked close to the processor, and logic chips made by factories such as TSMC.
Now, Goodwin says, it is sites and power. The easy conversions of existing industrial buildings into data centers are largely gone. So are the one-time tricks, such as getting permits in the United States to bring in gas turbines and start a site in about 90 days.
On energy, Goodwin said the queues for grid connection in the United States and Britain are "an order of magnitude more than" the energy in the grid, or roughly ten times as much. He cautioned that not every proposed project would actually be built, but called the demand staggering.
He rejects the comparison to the fiber buildout of 2000 and 2001 because the graphics processors already powered on in data centers are fully used. There are no dark fleets sitting idle, in his account, and there are still many hard problems to use the extra thinking for.
Goodwin says we are not going to get “thousands of times more data centers.” The way forward is getting far more work from existing sites, not building new ones, by combining better models and more efficient chips to deliver much more useful work from each megawatt and each rack.
Questions people ask
How much power does an AI data center actually use?
A large AI data center can use hundreds of megawatts, roughly enough electricity for tens of thousands to hundreds of thousands of homes. That is why a single site needs its own substation and long-term grid planning, not just a normal building hookup.
Why does getting a grid connection take so long?
Often many years. Utilities study grid impact, order transformers and other long-lead equipment, and work through permits. That queue is one reason Goodwin says energy, not chips, may be the most acute limit.
Can data centers just move to where the power is?
In principle, yes, because a data center does not care where it is located. Goodwin points to a huge hydro plant delivering many gigawatts in the middle of nowhere and says to put the computing next to it, adding that renewables can be a huge part of powering AI.
