- Essays··6 min read
You can buy GPUs but not memory Memory 🙈
The constraint on AI infrastructure in 2026 is not compute but memory bandwidth adjacent to the silicon, and the CoWoS packaging required to deliver it. TSMC's advanced packaging capacity is fully booked through 2027. Adding GPUs does not reduce LLM inference latency when the bottleneck is HBM bandwidth.
ai-infrastructuregpu-memoryhbmhardware-supply-chainRead - Essays··5 min read
The Interconnect Decision That Moved Upstream
CoWoS packaging capacity — not wafer fabrication — has become the binding constraint on AI hardware, and the shortfall has reorganised how clusters are designed. Network architects who once specified fabrics after compute teams locked the BOM now sit in allocation meetings, because the interconnect bandwidth budget and the packaging timeline are the same constraint. Memory, topology, and fabric, previously sequential decisions, now collapse into one.
ai-infrastructureadvanced-packaginghbminterconnectRead - Essays··6 min read
The Interconnect Nobody Can Buy
UALink 2.0 was ratified in April 2026; no production-ready silicon exists. For procurement decisions targeting Q1 2027 delivery, that leaves NVIDIA NVLink over InfiniBand or RoCE on Ethernet — each with known performance trade-offs, HBM supply constraints, and operational costs that never appear on the CapEx comparison slide. The interconnect you want and the silicon you can order are not the same thing.
ai-infrastructureinterconnectnvidiaualinkRead