Nvidia's memory costs soar 485%, latest AI systems now cost $7.8 million to build

As demand from the AI sector surges, semiconductor component prices are rising sharply, significantly increasing the cost of building next-generation AI systems. Morgan Stanley estimates that the upco... Read original →

NVIDIA’s 485% memory cost surge signals a structural shift from compute-centric to memory-bound AI infrastructure. The massive LPDDR5X and HBM4E integration not only inflates BOM but forces TSMC’s 3nm EUV capacity toward SOCAMM2 packaging, intensifying foundry bottlenecks. Geopolitically, U.S. export controls fragment the HBM supply chain—Samsung and SK Hynix can’t efficiently serve both U.S. and Chinese hyperscalers, driving global premiums. In response, AMD and Intel will likely accelerate CXL+UCIe-based modular architectures to bypass NVIDIA’s monolithic pricing power. If memory remains >25% of system cost over the next 18 months, cloud giants will pivot to in-house near-memory computing or RISC-V+NAND co-design to escape NVIDIA’s vertical lock-in.