Venture capitalist Tomasz Tunguz said the AI infrastructure narrative is unfolding like a slow relay race, with bottlenecks moving from GPUs to memory, CPUs, and storage. According to ChainCatcher, he said each constraint can freeze the next part of the supply chain for years and lock in a higher cost base.

Tunguz said the GPU shock in early 2023 pushed H100 rental rates above $9 per hour and cut server shipments by 22%. He added that manufacturers then shifted capacity toward HBM, causing enterprise SSD prices to rise 80% in a single quarter and DRAM prices to climb more than 60%.

He said that by the end of 2025, agent workloads will push the CPU-to-GPU ratio to about 1:1, while average server CPU prices will rise 27% year over year. He also said nearline HDD capacity for 2026 is already sold out for the full year.

Tunguz said data center civil construction costs have risen to about $2 billion per gigawatt, and long-lead equipment such as transformers and turbines is already booked through 2029 to 2031. He described the pattern as a hardware bullwhip effect, saying manufacturing delays amplify downstream demand shocks and can later create overcapacity risks for transformer expansion, NAND fabs, and turbine production lines delivered in 2027 to 2028.