Advanced nodes live in the compute tray — the rest of the rack runs on mature silicon

SIA and Deloitte classify semiconductor dies by process node: advanced (below 10 nanometers), current-generation (10 to 28 nanometers) and foundational or mature (28 nanometers and above). In a representative AI server rack, advanced-node dies are concentrated almost entirely in the compute tray, where they handle the AI workload. The other trays — power, cooling, platform management and interconnect — rely mostly on mature and current-generation nodes.

One tray carries the value

The compute tray accounts for about 95% of the rack's semiconductor content value, nearly 90% of the total number of dies and more than 80% of the number of individual chips. Other trays carry more than 15% of the chip count but a small fraction of the value, because their dies are fewer and made on more mature technology.

Why this matters

The split explains two features of the industry at once. Leading-edge manufacturing capacity determines the performance of the whole system, which is why advanced fabrication attracts so much attention and capital. But the rack also depends on a long tail of mature-node chips, so disruptions in older process technologies can still hold back finished AI systems.

Limitations

The node distribution reflects one representative rack configuration in the report's analysis. The report defines advanced nodes as sub-10 nm generations typically fabricated with EUV lithography. Shares will vary with accelerator choice and system design.