Productscience

Product Science is a Los Angeles-based company building end-to-end orchestration for decentralized foundation model training. Its hardware-agnostic approach is intended to let enterprises, research labs and public institutions train specialized models across fragmented, geo-distributed compute resources, spanning general-purpose GPUs through to specialized ASICs, with configurable data sovereignty. The company describes itself as a group of researchers and engineers working on distributed training and decentralized AI infrastructure, and states it previously incubated Gonka, a decentralized network for AI training and inference.

The company frames the problem it addresses as the concentration of frontier AI capability in centralized data centers: large infrastructure requirements and the scarcity and cost of NVIDIA chips restrict frontier-scale training to a small number of large technology firms; running training across heterogeneous, geo-distributed hardware creates communication overhead, fault-tolerance and coordination challenges; and existing decentralized training protocols address isolated problems rather than providing full orchestration.

Earlier in its history the company operated as a performance management software business for mobile and cloud applications. That product line analyzed application code alongside runtime data to identify performance issues such as crashes, freezes and errors, and was delivered through tools and IDE plugins that paired video recordings of app sessions with performance traces. Documented products from that period include CodeTuner for Cloud, CodeTuner for Mobile, and Regression Analysis for Mobile.

Founding story

The company was founded by four siblings, David, Daniil, Anna and Maria Liberman; David and Daniil served as co-CEOs during the performance management phase of the business.

Business model

The company positions itself as an infrastructure and protocol provider, offering orchestration for decentralized model training rather than operating its own centralized data centers. In its earlier performance-management phase it sold software tooling to enterprise customers, including an enterprise offering.

Traction

The company states it started and incubated Gonka, a decentralized AI infrastructure network for model training and inference, and cites 10,000 NVIDIA H100 GPUs in connection with that network.

Latest developments

The company is architecting protocols intended to move decentralized AI training beyond traditional GPU clusters toward trustless, permissionless training environments, following its incubation of the Gonka network.