IBM Research’s own reported results. Each setup was tested in a single run, the speedups are estimates from a design tool rather than measurements on a finished chip, and the costs cover only the AI model’s running costs.
IBM Research says a team of AI agents, running on open AI models, made a hardware design for quantum computing 52 times faster for about $26 in AI costs [1].
A setup using a proprietary “frontier” AI model did better, reaching a 204 times speedup, but cost $299. That is 11.5 times more than the open-model setup, under the same search budget [1].
IBM does not say which AI models it used, open or proprietary.
The key terms, explained
- Agent factory. A group of AI agents organised to work on one problem. In IBM’s setup, a “coordinator” agent plans the search and hands ideas to four “worker” agents working in parallel. The workers change the design, a checker confirms the changes still work correctly, and the results guide the next round [1].
- llm-d. Open-source software for running AI models across many servers at once. It sends requests to the servers that already hold the relevant context, so work is not repeated, and it splits the processing of the prompt from the writing of the answer [1].
- Quantum error-correction decoding. Quantum computers make errors. Measurements give clues about those errors, and a “decoder”, which runs on ordinary (classical) hardware, has to read those clues and work out the corrections quickly. IBM calls this “an essential part of building fault-tolerant quantum systems” [1]. The decoder IBM used is called Relay-BP, for a type of error-correcting code known as the [[144,12,12]] bivariate bicycle code.
- High-level synthesis (HLS). A way of designing chips by writing in C or C++ instead of a low-level hardware language. A tool then turns that code into a hardware design and estimates how fast it will run and how much of the chip it will use [1].
What IBM tested
IBM compared three setups, each given 100 ideas (it calls them “hypotheses”) to try [1]:
- a proprietary model as both coordinator and workers;
- an open model in both roles, run on IBM’s own servers through llm-d;
- a mix: a proprietary coordinator with open-model workers.
All three found verified improvements. IBM says the open and mixed setups found compact designs, while the proprietary one was better at turning extra hardware into speed [1]. The all-open setup produced “the smallest area design”, meaning the design that takes up the least chip space [1].
Where the savings came from
The open setup actually used more AI tokens (the units models process text in): about 241 million, against 216 million for the proprietary one [1]. It was cheaper because those tokens cost less to serve.
IBM says llm-d reused shared context about 98% of the time and brought the blended cost to about $0.11 per million tokens [1]. Running an open model on its own servers also let IBM keep its design files in-house.
Doubling the open setup’s budget to 200 ideas raised the speedup from 52 times to 60 times. IBM says the gains “continued, but they diminished” [1].
“An agent factory turns hardware optimization into a coordinated cycle of exploration: A coordinator agent directs parallel worker agents to develop and synthesize alternative designs within a defined search budget,” said Ruchir Puri, IBM fellow and IBM Research chief scientist [1].
The caveats IBM gives
- One run each. “Each configuration was evaluated in one search run, so these results characterize the observed searches rather than establish a universal model ranking” [1].
- Estimates, not chips. The speedups are high-level synthesis estimates [1].
- Inference costs only. “Costs cover model inference, rather than the full hardware development workflow” [1]. IBM says future work must also count the cost of synthesis and verification.
What this means
- AI agents can explore chip designs cheaply. IBM’s result suggests useful hardware improvements are possible on a small AI budget.
- The proprietary model still went further. It reached 204 times against 52 times, at about 11.5 times the cost.
- It is early work. These are IBM’s own figures from single runs, and no one has independently repeated them.
Sources
- IBM Research (Peter Hess), “Making AI-driven chip design affordable with open agent factories and distributed inference at scale”, 13 October 2026. https://research.ibm.com/blog/open-agent-factories-and-distributed-inference-for-chip-design
