AI infra scarcity widens
Hyperscalers have locked in huge GPU orders — Nvidia said customers placed more than $1 trillion of orders through 2027 — but firms are now facing shortages that go beyond chips to power, deployment capacity and rising GPU prices. (cnbc.com) (enterpriseai.economictimes.indiatimes.com) (investing.com)
Artificial intelligence companies can still buy Nvidia chips, but many cannot get enough electricity, cooling and rack space to turn those chips on. (cnbc.com) (ft.com) Nvidia Chief Executive Jensen Huang said in March that customers are on track for about $1 trillion in Nvidia revenue through 2027, a sign that the biggest cloud companies are still placing enormous orders. CNBC reported in February that Alphabet, Amazon, Meta and Microsoft could push combined 2026 capital spending toward $700 billion as they build more artificial intelligence infrastructure. (cnbc.com 1) (cnbc.com 2) The shortage now runs past chips. The Economic Times, citing The Wall Street Journal, reported on April 14 that companies are rationing access and raising prices as demand for “agentic” tools strains computing capacity and energy supply. (economictimes.indiatimes.com) A graphics processing unit is the engine that trains and runs modern artificial intelligence models, but the full system also needs networking gear, storage, power feeds and cooling loops inside a data center. Nvidia’s Blackwell systems are built as rack-scale, liquid-cooled machines, and the company says the new unit of deployment is the GB200 NVL72 rack. (blogs.nvidia.com 1) (blogs.nvidia.com 2) That changes the bottleneck from a single chip order to an entire construction project. Microsoft Chief Executive Satya Nadella said the industry’s biggest problem is “power” and getting new builds done fast enough near available electricity, according to the Financial Times. (ft.com) OpenAI said in July 2025 that its Stargate partnership with Oracle added 4.5 gigawatts of United States data center capacity and brought Stargate to more than 5 gigawatts under development, enough to run more than 2 million chips. Those numbers show how artificial intelligence expansion is now measured in utility-scale power, not just server counts. (openai.com) The energy side is tightening at the same time. The International Energy Agency said on April 10 that global electricity demand from data centers is set to more than double to 945 terawatt-hours by 2030, with the United States driving about half of that growth. (spglobal.com) (iea.org) Prices are moving with the scarcity. Investing.com said in early April that high-end graphics processing unit prices were climbing by roughly 40% to 50% in 2026 as artificial intelligence demand outpaced supply growth and buyers accepted longer-term contracts to lock in capacity. (investing.com) (msn.com) Cloud providers are turning those costs into hourly rental rates. CoreWeave’s public pricing page shows access to newer Nvidia systems, while third-party trackers in April listed an eight-chip H100 instance at about $49.24 an hour and an eight-chip B200 instance at about $68.80 an hour. (coreweave.com) (costbench.com) The near-term result is that the winners are the companies that reserved chips, power and construction capacity early. Everyone else is competing for the same transformers, substations, liquid-cooling gear and installation crews needed to make those trillion-dollar chip orders usable. (cnbc.com) (ft.com)
