Semiconductor billings surge

The number looks abstract at first. It is not. In 2025, chipmakers and foundries spent $135.1 billion on semiconductor manufacturing equipment, up 15% from $117.1 billion a year earlier, according to SEMI’s latest worldwide billings data. That is money for the machines that etch, deposit, inspect, package, and test chips. It is the clearest measure of how hard the industry is still pushing to build more capacity for the AI boom (finance.yahoo.com, seaj.or.jp). The surge was not evenly spread. China remained the biggest equipment market at $49.3 billion, basically flat from 2024. Taiwan jumped 90% to $31.5 billion and moved into second place, overtaking South Korea, which rose 26% to $25.8 billion. North America fell 20% to $10.9 billion. Europe dropped 41% to $2.9 billion. That pattern says a lot. The center of gravity is where advanced logic and memory are actually being built, not where AI software gets the headlines (seaj.or.jp). That spending wave did not come out of nowhere. SEMI had already forecast in mid-2024 that equipment sales would set records in 2025 because wafer fab spending was shifting back into growth, with especially strong investment in DRAM and HBM for AI systems. Back-end tools were also expected to rebound fast because advanced AI chips are not just hard to fabricate. They are hard to package and test, too, and those steps now matter much more than they did in the era of simpler processors (semi.org). That is the part casual coverage often misses. The bottleneck in AI hardware is no longer just the GPU die. It is the stack around it. Nvidia’s Blackwell-based DGX B200 systems are built around GPUs with 192 GB of HBM3e memory, which is far more memory than earlier generations needed. Every jump in memory per accelerator multiplies pressure on DRAM fabrication, advanced packaging, and the substrate supply chain at the same time (resources.nvidia.com, blogs.nvidia.com). That is why the equipment number matters. A modern AI accelerator is a small systems-engineering project disguised as a chip. High-bandwidth memory has to be manufactured, stacked, bonded, and integrated next to the compute die with advanced packaging technologies such as CoWoS. When demand for those parts spikes, the answer is not one new factory. It is a whole chain of new tools across front-end fabs, memory lines, packaging houses, and test floors (semi.org, techinsights.com). Even after that spending, memory is still tight. Micron said in December 2025 that its HBM volume for calendar 2026 was already sold out, with customer negotiations for both volume and pricing completed. SK hynix said in its January 2026 results that record performance was being driven by AI demand for HBM and DDR5. Those are not warning flares from analysts. They are supply signals from the companies making the memory (fool.com, skhynix.com). So the strange combination in this story is real. Equipment billings are surging because the industry is spending furiously to remove AI bottlenecks. Memory constraints are still real because the bottlenecks moved faster than the factories. Taiwan’s 90% jump in equipment billings is what that looks like on the ground: not a pause, not a correction, but a scramble to build enough advanced logic and packaging capacity before the next rack of GPUs arrives (seaj.or.jp).