Epoch AI launches tool tracking critical component consumption in AI chip manufacturing

Epoch AI has released the AI Chip Components Explorer, a model that estimates how much of the three main bottlenecks in advanced chipmaking were consumed by the four largest designers each quarter from Q1 2024 through Q4 2025. The model works in three stages: estimating quarterly production for each designer and chip type based on sales data, adjusted for manufacturing lags, inventory and export-control write-offs; translating chip volumes into component demand using bill-of-materials specifications and yield assumptions; and comparing that demand against global supply estimates.

The three bottlenecks are advanced logic wafers at 3 nm and 5 nm, advanced CoWoS packaging, and high-bandwidth HBM memory. TSMC holds the overwhelming majority of global 3 nm and 5 nm logic-wafer capacity and effectively controls CoWoS capacity as well. HBM is produced by a handful of memory makers — SK Hynix, Samsung and Micron. Because these are the binding physical constraints, understanding AI’s share of each shows how much of the world’s advanced output is being absorbed by AI chips.

The tracker covers four designers: Nvidia, AMD, Google (TPUs) and Amazon (Trainium). Together they account for most CoWoS and HBM consumption and a significant slice of logic wafers. The excluded list is long: Meta, Microsoft, Tesla, Groq, SambaNova, Huawei and Cambricon, with their addition described as a possible future direction. An “Other” row in the tables captures global component supply not attributed to the four tracked designers — capacity consumed by unmonitored designers such as Apple and Qualcomm on logic wafers, Chinese AI-accelerator and HBM stockpiling, and capacity that was available but went unused in a given quarter.

The model counts component consumption regardless of whether the finished chips have shipped to customers, sit in finished-goods inventory, or remain in work-in-process (WIP). Nvidia and AMD are modeled directly from their financial statements, splitting the GPU portion into raw materials, WIP and finished goods and assigning each category to the quarters in which the components were consumed. Google and Amazon do not hold finished-chip inventory (their backend partners do), so the model extrapolates Q1 2026 production volumes from 2025 trends. A concrete example: Blackwell chips in WIP at the end of Q3 2025 had already consumed logic wafers (fab complete) but not yet CoWoS or HBM (packaging incomplete), so consumption appears as Logic > 0 and CoWoS = HBM = 0 for that quarter.