chips.sanjay920.comA nine-layer deep-dive on the AI hardware supply chain, from EUV to tokens.https://chips.sanjay920.com/L01 · EUV — How a tin droplet becomes a chip pattern.https://chips.sanjay920.com/l/01-euv/https://chips.sanjay920.com/l/01-euv/A laser strikes tin droplets to make 13.5 nm light. That light prints the smallest features on advanced logic chips.Wed, 13 May 2026 00:00:00 GMTEUV Lithographyλ 13.5 nm · wattsL02 · Wafer — How many useful accelerator dies fit on one wafer?https://chips.sanjay920.com/l/02-wafer/https://chips.sanjay920.com/l/02-wafer/A 300 mm wafer offers many candidate reticle sites, but edge exclusion and defects reduce the number of good large dies.Wed, 13 May 2026 00:00:00 GMTWafer Fabrication858 mm² · % yieldL03 · Packaging — Why packaging matters as much as the chip.https://chips.sanjay920.com/l/03-packaging/https://chips.sanjay920.com/l/03-packaging/The package connects compute dies to HBM through substrates, bridges, and interposers. Those layers determine bandwidth and yield.Wed, 13 May 2026 00:00:00 GMTAdvanced Packagingμm stack height · mm shorelineL04 · HBM — How HBM gets to 2.5 TB/s.https://chips.sanjay920.com/l/04-hbm/https://chips.sanjay920.com/l/04-hbm/Stacked DRAM, a base die, and thousands of IO lines add up to the bandwidth modern accelerators depend on.Wed, 13 May 2026 00:00:00 GMTHBM MemoryGB/s · TSVs · 12-highL05 · Accelerator — Why modern accelerators are becoming multi-die systems.https://chips.sanjay920.com/l/05-accelerator/https://chips.sanjay920.com/l/05-accelerator/The accelerator is no longer just one rectangle of silicon. It is a tightly integrated package of logic, memory interfaces, and die-to-die links.Wed, 13 May 2026 00:00:00 GMTAccelerator Dietransistors · tFLOPSL06 · Scale-Up — How 72 GPUs behave like one machine.https://chips.sanjay920.com/l/06-scale-up/https://chips.sanjay920.com/l/06-scale-up/NVL72 combines compute trays, switches, copper links, power, and cooling into a rack-scale training unit.Wed, 13 May 2026 00:00:00 GMTScale-up RackGPUs / pod · NVLinksL07 · Data Center — What it takes to power a dense AI data center.https://chips.sanjay920.com/l/07-data-center/https://chips.sanjay920.com/l/07-data-center/As racks get denser, data-center design shifts from floor space to substations, conversion losses, and cooling.Wed, 13 May 2026 00:00:00 GMTData CenterMW · PUEL08 · Cloud — How hardware becomes cloud revenue.https://chips.sanjay920.com/l/08-cloud/https://chips.sanjay920.com/l/08-cloud/This layer connects spending on compute infrastructure to rental pricing, utilization, and payback time.Wed, 13 May 2026 00:00:00 GMTCloud Layer$ / hr-Hopper · depreciationL09 · AI Lab — How tokens inherit the cost of everything underneath them.https://chips.sanjay920.com/l/09-ai-lab/https://chips.sanjay920.com/l/09-ai-lab/The lab is where the hardware supply chain turns into a user-facing service, with compute cost flowing into every token served.Wed, 13 May 2026 00:00:00 GMTAI Lab → Modeltokens / s · gross margin