Sign in Subscribe

Research Parks

Power you can use today is the AI question regions can answer

Chamath Palihapitiya argues that power is the binding constraint in AI, and not power promised for 2028 but power today. What that means for research parks, economic developers, tech transfer offices, and science founders deciding what a site can count on.

Tung Nguyen 5 min read

Four identical gray buildings on a prairie horizon; a power line reaches only the first, glowing teal, while a dotted line runs on to the others
A site with power today next to sites on a promise, as a risograph print · November 2026

On November 1, Chamath Palihapitiya posted a short argument on X about who will shape artificial intelligence. His answer: whoever shapes "power, data centers, chips and models into an economically feasible endpoint." The rest of the post explains why, and every step of the explanation comes back to one word. Power.

audio-thumbnail
Listen to this article
0:00
/439.728

He posted it on top of a big number. The post quotes Elon Musk saying there will be "around 1.2M Blackwells at Colossus 2 by end of year." That is not the scale a research park works at, and most regions will never host a campus like that. But the core of the argument travels well beyond the largest AI companies. It lands on anyone whose job is to say what a site, a building, or a region can deliver, and when.

The argument, in his words

Palihapitiya, whose X profile lists Social Capital and 8090, makes the case as a chain. Each link depends on the one before it.

  • Power comes first. He calls it "THE binding constraint in AI/SI" and adds: "Few really have it."
  • Data centers need power. "Data centers don’t work without power."
  • Chips need a working data center. Chips, he writes, cannot operate profitably unless they are installed in one. "If you can install chips, it’s because you’ve solved the power problem."
  • Models need cheap, reliable infrastructure. That requires a functional data center "with low input costs and reliable management."
  • The return depends on all of it. AI returns do not pencil out, in his telling, unless models produce profitable outputs, which need functional data centers. "Again, power."

His conclusion: whoever can put data centers, chips, and models together into "a tightly coordinated stack" has solved power, and "Solving power allows you to disproportionately shape AI/SI right now."

The line that matters most for this audience is about time. "This is not about promises of power tomorrow. Or forecasted power in 2028. Power TODAY."

To be fair to the post, it is an argument, not a dataset. It does not cite figures beyond the one it quotes, and it is written about the companies racing to build the largest AI systems. I am taking one idea from it, the gap between power promised and power available, and applying it somewhere he was not writing about.

What the energy data says

The International Energy Agency's Energy and AI report makes a similar point in plainer language: "there is no AI without energy." It adds that countries "that can deliver the energy needed at speed and scale will be best placed to benefit."

The report also puts numbers on the gap Palihapitiya describes. According to the IEA, a typical AI-focused data center uses as much electricity as 100,000 households, and the largest ones under construction will use 20 times as much. It estimates that around 20% of planned data center projects could be at risk of delays unless grid risks are addressed. It describes grid connection queues as long and complex, says new transmission lines can take four to eight years in advanced economies, and reports that wait times for transformers and cables have doubled in the past three years.

That is the distance between a forecast and a working connection, measured in years.

The same report adds two cautions worth keeping. First, the pressure is local more than global. Data centers used about 1.5% of the world's electricity in 2024, but nearly half of U.S. data center capacity sits in five regional clusters. Second, the outlook is uncertain. Across its scenarios, the IEA's range for data center electricity demand in 2035 runs from 700 to 1,700 terawatt-hours. Power is a hard constraint today. How hard it stays is an open question.

What this means for research parks and economic developers

My understanding is that most research parks will never be in the market for a frontier AI campus. But the distinction between power today and power in 2028 applies to every deep-tech tenant: a lab with heavy equipment, a compute room, a pilot line. Specifically, three things follow.

  • Know your power in two units: capacity and months. A site with capacity on a utility forecast is, in Palihapitiya's terms, a 2028 site. A site that can energize a tenant this year is a today site. Both can be honest offers, but they are different offers, and a prospect will find out which one they are getting.
  • Say it in the listing. In an earlier piece we suggested that a founder reading a park listing is asking, among other things, whether there is enough power and network for the compute the product needs. That question is getting sharper, and the answer belongs on the first page, not in a follow-up call.
  • Weigh what a large load leaves for everyone else. The IEA warns that meeting data center load growth could entail "trade-offs with other goals such as electrification, manufacturing growth or affordability." A region weighing a large data center project should ask what capacity is left for the manufacturers and lab tenants it is also trying to recruit.

The IEA also lists, as one way to reduce delays, "locating new data centres in areas of high power and grid availability." For a region with available power, that is a real advantage. It only counts if the region can document it.

What this means for tech transfer offices and science founders

For a science founder whose product depends on AI, Palihapitiya's chain is a cost structure. If models are only viable on "cost effective infrastructure," then the price of the compute a startup buys is set, in part, by power decisions made far upstream. My read is that it belongs in a founder's plan and in an investor's diligence: where the compute comes from, what it costs, and how exposed the company is if that price moves.

For tech transfer offices, the constraint points the other way. A shortage creates buyers. University inventions in grid hardware, storage, power electronics, cooling, and energy efficiency are aimed at the exact bottleneck Palihapitiya describes. That does not make any single invention licensable. It does mean the market question for those disclosures, who needs this and how soon, may have a clearer answer than usual.

The IEA makes one more point that fits this audience closely. Discussing AI as a tool for energy innovation, it says policy will be needed to support AI-led invention and to "accelerate commercialisation, which is often a bigger impediment to new products than the discovery phase." Faster science does not move the market by itself. The step from lab to market still has to be built.

What's next

Palihapitiya wrote about who will shape AI. Most regions will not be in that contest. They can still answer a narrower question with the same discipline: how much power can a tenant use on day one, and can we show it? His two words are a fair test for any site: "Power TODAY."

Sources