Xilinx held generic chips and let customers finish them, cutting inventory risk

Xilinx sells field-programmable gate arrays whose logic is defined by software after the chip is built. That is product postponement: customers configure final functionality, so thousands of possible designs collapse into a few physical products.

Xilinx added process postponement by moving the push-pull boundary to a die bank: fabrication partners hold generic dies, and back-end partners customize them to orders. Because 50–150 finished goods often share one die and their demands are nearly independent (measured correlation 0.018), aggregating into die demand pools risk.

The analysis in the Interfaces paper showed that for the same expected back-order service, die-bank postponement needs much less inventory than no postponement. Finished goods cost about 40% more than dies, and nearly all of the industry's ~5%-a-year obsolescence cost hits finished goods, so holding generic inventory is cheaper on both counts. Xilinx adopted a hybrid of die-bank and partial postponement.

Postpone differentiation until the customer configures

Software-configurable chips erased forecast-driven product variety.

Die-bank demand pooling shrank uncertainty sharply at low correlation.

Generic dies are cheaper and rarely obsolete, unlike finished goods.

Managers oversee ~100 dies instead of ~10,000 finished goods.

Design product and process so differentiation happens last: if customers can do the customization, aggregate demand pools risk and the generic form is cheaper to hold — same service, less inventory.

The paper reported that postponement became a key part of Xilinx's overall supply-chain strategy, supported by inventory modeling, supply-mix prediction and back-end cycle-time reduction. Xilinx remained one of the two largest suppliers of programmable logic, and the case became a standard reference for postponement.