Videos flag datacenter limits
Two recent YouTube videos emphasise that AI deployment is being constrained by datacentre power, cooling and capacity rather than model capability alone. One piece explores the economics and timelines of AI datacentre buildouts, while another consumer GPU roundup is presented as an early signal of silicon supply and retail pricing. The coverage highlights physical infrastructure and procurement as near‑term constraints on AI rollouts. (youtube.com) (youtube.com)
Artificial intelligence is running into a physical bottleneck: power hookups, cooling gear and building capacity are now limiting how fast new systems can go live. (deloitte.com) A June 24, 2025 Deloitte report estimated United States power demand from artificial intelligence data centers could rise to 123 gigawatts by 2035, up from 4 gigawatts in 2024. Deloitte said some grid interconnection requests already face waits of as long as seven years. (deloitte.com) Data centers are warehouses full of servers, and artificial intelligence versions pack in more graphics chips, which are specialized processors that draw more electricity and dump more heat. Deloitte said a five-acre site can jump from 5 megawatts to 50 megawatts when central processing units are augmented with graphics processing units. (deloitte.com) Google said on April 29, 2025 that it was pushing new rack designs from 100 kilowatts toward 1 megawatt and contributing its fifth-generation liquid-cooling system to the Open Compute Project. Google also said machine-learning racks may need more than 500 kilowatts each before 2030. (cloud.google.com) That shift changes the economics of rollout. Bloom Energy said on January 21, 2025 that data center operators are increasingly buying onsite power because utility supply is too slow, and its survey of about 100 decision-makers found 30% of sites are expected to use onsite power as a primary energy source by 2030. (bloomenergy.com) Bloom said the United States has about 25 gigawatts of existing data center information-technology capacity and could add 55 gigawatts more in the next five years, with about 20 gigawatts already announced. The company said operators are now prioritizing “time-to-power” alongside cost and reliability. (bloomenergy.com) The consumer graphics card market offers a rough read on the same supply chain. Nvidia said its GeForce RTX 5090 and RTX 5080 launched on January 30, 2025, with Founders Edition cards available only in limited quantities through Best Buy. (nvidia.com) Retail prices then moved above list prices for months. TechPowerUp, citing Tom’s Hardware pricing data published March 26, 2025, said the GeForce RTX 5070 was selling for about $700 against a $550 suggested price, while the GeForce RTX 5080 was more than 50% above its suggested price. (techpowerup.com) By September 22, 2025, VideoCardz reported that Walmart had listed the GeForce RTX 5090 at its $1,999 suggested price and the GeForce RTX 5080 below its launch price, a sign that retail supply had started to loosen. The bigger data center buildout, though, still depends on substations, transformers, chillers and permits that do not ship like graphics cards. (videocardz.com)
