Descriptions:
Daniel Beutel, co-founder of Flower Labs, uses his AI Dev SF 2026 talk to introduce Flower SuperGrid, a new platform designed to make building decentralized, federated AI systems accessible to teams without deep distributed systems expertise. The announcement comes alongside a striking milestone: Flower ran on an H100 GPU aboard StarCloud 1—the first H100 ever operated in space—where it powered the first vision transformer training run in orbit.
The motivation behind SuperGrid is the gap between how valuable collaborative AI can be and how painful it is to actually deploy. Beutel notes that setting up a federated AI system previously required working through more than 250 discrete steps—and that was before onboarding external partners. SuperGrid collapses this into a GitHub-like interface: create a federation, invite collaborators, register SuperNodes (the distributed compute endpoints), and run workloads with a single `flower run` command.
Beutel frames the broader opportunity around the 2,000 trillion tokens of data locked in enterprise silos—roughly 133 times the size of all high-quality public English web data—that collaborative AI could unlock without requiring centralized data movement. Flower, already used by Bosch, Samsung, Mozilla, the US Air Force, and top research institutions, is Apache 2.0 licensed and has more than 7,000 community members. SuperGrid represents the company’s move to productize years of federated learning R&D into an accessible developer platform.
📺 Source: DeepLearningAI · Published May 22, 2026
🏷️ Format: Keynote Launch
