Tomasz Tunguz: AI infrastructure exhibits a long tail effect, with bottlenecks gradually transmitting and driving up costs
Venture capitalist Tomasz Tunguz pointed out that the narrative of AI infrastructure resembles a slow relay race, with bottlenecks sequentially transmitting from GPUs to memory, CPUs, and storage, each link freezing the supply chain of the next for years and locking in higher baseline costs. At the beginning of 2023, the GPU shock caused H100 rental prices to exceed $9 per hour, and server shipments fell by 22%; subsequently, manufacturers shifted capacity to HBM, leading to an 80% quarterly increase in enterprise SSD prices and over a 60% rise in DRAM.By the end of 2025, the workload of intelligent agents will push the CPU to GPU ratio to about 1:1, with the average price of server CPUs rising by 27% year-on-year; in 2026, nearline HDD annual capacity will be sold out. The construction cost of data centers has risen to about $20 billion per gigawatt, with orders for long-cycle equipment such as transformers and turbines scheduled as far out as 2029 to 2031.Tunguz referred to this as the long whip effect in the hardware sector: years of manufacturing delays amplify downstream demand shocks upstream, and when pressure is relieved at a certain bottleneck, it will be delayed in transmitting to the next link, with transformers scheduled for delivery in 2027 to 2028, NAND wafer fabs, and turbine production lines potentially facing the risk of overcapacity.
