AI’s growth depends on power and computing capacity that may not arrive as quickly as software does, argues Square One Labs co-founder Amir Fishelov. His warning is that scarcity could raise the cost of running AI, squeezing businesses and providers that assume compute will keep getting cheaper.
Watch Desk analysis
What happened
In an opinion article for CTech, Fishelov says AI’s electricity demand is putting pressure on both compute providers and the power grid. He cites a Berkeley Lab report for the US Department of Energy: data centres consumed about 176 terawatt-hours of electricity in 2023, and scenarios in the report project demand could reach 325 to 580 TWh by 2028. Those are projections, not a guarantee of what will be consumed.
Fishelov also points to Microsoft saying customer demand exceeds available capacity, Amazon saying it will not have enough capacity for projected demand in 2026, and a PJM power auction that fell about 6.8 gigawatts short of reliability requirements. He notes that large power transformers can take up to five years to deliver, while new software can scale much faster.
Why it matters
If capacity stays tight, Fishelov argues, the first consequence is more likely to be higher prices for compute than an immediate blackout. That could affect the cost per token, query or agent-hour, challenging AI products and company valuations built around the expectation that intelligence will become cheaper to run.
His proposed responses are physical as well as technical: build generation closer to computing sites, or design data centres that deliver the same compute with substantially less power. He argues efficiency alone may not reduce total consumption, since cheaper computing can encourage more use.
Our read
The useful point is that an AI business plan has an electricity bill hiding inside it. Fishelov has a commercial perspective as a venture investor, and his article does not establish how quickly prices will rise or which projects will face shortages first. But the examples he cites make power availability a concrete assumption for companies to test, not a distant footnote.
What to watch
- Whether cloud providers report capacity constraints or higher costs.
- How quickly new power generation and data-centre capacity come online.
- Whether efficiency improvements reduce energy demand, or help AI use expand further. Activity teaser: AI’s growth runs on electricity as well as software, argues Square One Labs co-founder Amir Fishelov. He points to rising data-centre demand, reported capacity constraints and power infrastructure that can take years to build. His concern: scarce compute could push up the cost of running AI, putting pressure on businesses that expect it to get cheaper. Are AI companies stress-testing their plans for energy costs, or still treating power as somebody else’s problem?
Discussion spark: Should AI companies have to stress-test their business plans for higher compute and electricity costs, or is that simply part of the risk investors should price in?
Sources and evidence
- What if the AI runs out of electricity? | Ctech (31 October 2026, 06:35 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.