AI Factories and Gigawatt Datacenters

This panel, moderated by Mohamed Wahib of RIKEN, brings together Geetika Gupta (NVIDIA), Jason Haga (AIST), a EuroHPC Joint Undertaking infrastructure lead, and an Argonne Leadership Computing Facility deputy director to debate the bottlenecks and implications of AI factories and gigawatt datacenters. Opening positions diverge: Gupta argues power and hardware supply are already being solved while policy and talent will be the enduring constraints; the EuroHPC representative points to memory prices, supply chains, and European sovereignty; Haga highlights orchestrating AI workflows across heterogeneous accelerators, describing AIST's NEDO-funded NAPA project, a roughly 180 million dollar five-year effort to build an AI testbed with Japanese industry; the Argonne panelist stresses multi-year lead times in power and cooling infrastructure. The discussion then turns to the role of public HPC infrastructure in an era where private AI industry leads at scale, with consensus that public centers should complement rather than compete, offering workforce training, long-horizon science, and open access. Panelists note that AI-optimized systems diverge from traditional HPC in precision requirements, denser GPU-level networking, multi-tenancy, containers, and richer software stacks. Further exchanges cover how scientific impact should be measured beyond FLOPS or tokens, sustainability of AI demand growth, data retention and archival challenges, trust and verification of AI-generated code, and the influx of non-computational scientists onto HPC resources. EuroHPC's 19 AI factories and the forthcoming AI Gigafactories program, defined around an equivalent of 75,000 H100 GPUs with mixed public-private funding, are presented as concrete European responses.