IBM reports lower inference costs for AI-led chip optimization

IBM Research reports that agents using open models optimized a quantum error correction hardware design to deliver a 52-times speedup in high-level synthesis estimates for approximately $26 in inference cost. Under the same 100-hypothesis search budget, a proprietary-model configuration achieved a 204x speedup for $299. A coordinator directed four worker agents to modify the design, using correctness checks and synthesis results to guide further experiments. The open models ran on premises through llm-d, which reuses shared context and allocates prompt processing and token generation separately. The open-model search produced the smallest area design. Each configuration received only one search run, and the cost comparison covers inference; evaluation of the complete workflow must also account for synthesis and verification costs.