Xcelsa Labs
Reimagining the future of silicon design.
We build AI systems for the most challenging problems in silicon design, so that the chips that will power the next decade of computing get designed in a fraction of the time they take today.
Demand for custom silicon is growing faster than the world's ability to design it. The tools have not kept pace, and the expertise is scarce.
Most of a chip's power, performance, and area is determined long before anything reaches a foundry --- right when the description is still written, restructured, and tuned largely by hand, by engineers whose judgment takes years to acquire. The search space they are exploring is astronomically large. They see a vanishing fraction of it.
Modern language models are extraordinary generators of candidate designs and unreliable judges of their own work. So we let them generate, and we never let them judge. Our systems propose, and then prove.
Propose
We search over rewrites of a design's source: restructuring datapaths, sharing and retiming operators, rebalancing pipelines, exposing parallelism a human would not have time to try.
Prove
Every candidate is checked against the original by a formal equivalence engine before it is allowed to survive. There is no heuristic reviewer, no sampled test bench standing in for a proof, and no reliance on a model's confidence in itself.
Measure
What survives is scored on power, performance, and area using the industrial synthesis and implementation tooling the field already trusts. The optimizer is guided by the numbers that actually decide a tapeout, not by a proxy for them — and every accepted result arrives with a proof of correctness attached.
Xcelsa is venture-backed, with support from leading VC funds and from industry veterans who have spent their careers building the tools and the companies that shaped modern semiconductors.
Three NYU professors who have helped establish AI for chip design as a research field; founded and run multiple companies; and work at the frontier of reasoning models, evaluation, and agentic systems. We have spent years on opposite sides of this problem — the learning side and the silicon side — before concluding it was a single problem. The proposals are only as good as the models generating them, the results are only as good as the proofs behind them, and none of it counts until it survives a real design flow.
We're hiring for multiple positions across AI and silicon design. If you love small teams, hard problems, and being early enough that what you build defines the field, send a note to careers@xcelsa.ai.
Leave your email below and we'll be in touch as we have news to share.