Compute is the bottleneck
Rented hours for Nvidia’s new Blackwell GPUs have jumped sharply, raising the cost of running large models and experiments almost overnight. That spike sits beside reports that TSMC’s advanced nodes are largely sold out through 2027, suggesting both cloud renting costs and silicon lead times are tightening capacity for AI projects ((intellectia.ai), Seeking Alpha, Yahoo Finance).
Renting Nvidia’s newest Blackwell chips now costs $4.08 an hour, up 48% from $2.75 just two months ago. (techmeme.com) That price jump came from the Ornn Compute Price Index, which the Wall Street Journal cited on April 13, 2026, as demand for “agentic” artificial intelligence systems pushed up rentals for Blackwell graphics processing units. (techmeme.com) A graphics processing unit is the rented engine behind most large-model training and inference, and Nvidia’s Blackwell line is the newest version. Nvidia says its GB200 NVL72 system links 72 Blackwell graphics processing units in one liquid-cooled rack and can deliver 4 times faster large-language-model training than the Hopper H100 generation. (nvidia.com) The supply side is tight too. EE Times reported in January that Taiwan Semiconductor Manufacturing Co. planned $52 billion to $56 billion in 2026 capital spending, yet analysts still expected demand for 5-nanometer-and-below wafers to run 25% to 30% above capacity in 2026 and stay short in 2027. (eetimes.com) Those wafers matter because Taiwan Semiconductor Manufacturing Co. makes the advanced chips used in many artificial-intelligence accelerators, including Nvidia’s. The company’s chief executive, C.C. Wei, said on the earnings call cited by EE Times that revenue from artificial-intelligence accelerators was on track for a mid- to high-50% compound annual growth rate from 2024 to 2029. (eetimes.com) Even new United States capacity arrives slowly. Taiwan Semiconductor Manufacturing Co. says high-volume production on its first Arizona fab’s N4 process started in the fourth quarter of 2024, while its second Arizona fab is targeting N3 volume production in the second half of 2027 and its third fab is targeting N2 and A16 by the end of the decade. (tsmc.com) That timing leaves artificial-intelligence developers squeezed from both ends in 2026: cloud rentals are getting pricier now, while fresh leading-edge chip supply will take quarters or years to show up. (techmeme.com, tsmc.com, eetimes.com) The result is a more ordinary bottleneck than the marketing around artificial intelligence suggests. The limiting factor is not new model ideas alone, but how many advanced chips can be rented today and how many more can actually be manufactured by 2027. (techmeme.com, eetimes.com, tsmc.com)
