Owning the AI Optimization Layer
Discovery Loop
The strategic prize here is owning the optimization layer that decides how AI workloads are built and where they run. Today, an AI lab tunes model architecture, a cloud provider tunes cluster usage and instance mix, a chip company tunes silicon for target workloads, and EDA vendors tune chip design flows. A system that searches across all four at once can turn scattered efficiency gains into one control point, and one budget line.
-
In practice this means searching across concrete knobs that usually sit in separate teams, model size and topology, compiler and runtime settings, GPU placement, scheduling policy, and even chip power, performance, and area tradeoffs. Synopsys already sells AI tools for chip design space optimization, which shows this layer already carries real budget and margin.
-
The buyer can justify this spend from several directions at once. A frontier lab cares about faster training and better model quality, a cloud team cares about higher cluster utilization and lower cost per run, and a chip team cares about matching silicon to real workloads. Discovery Loop is trying to package those gains into one product instead of leaving them split across vendors.
-
Comparable companies each capture only one slice. Lila couples AI with physical lab hardware to own experimental execution, while Wafer is extending optimization software toward enterprise, lab, and chipmaker infrastructure. Discovery Loop is aiming at the cross layer search engine above these silos, starting in ML where experiments can run fully in software.
If this model works, the market will shift from selling tools for isolated steps to selling autonomous systems that optimize the whole stack. That would push AI labs, clouds, chip vendors, and EDA incumbents to either expose their layers to a neutral orchestrator or build tighter in house loops to keep that value for themselves.