This does not imply that more gigawatts automatically produce AGI. The resource gate makes the opposite point: without efficiency and accessible infrastructure, even a strong architecture cannot scale into everyday use.
About this eventLeading accelerator and HBM production remains concentrated and cannot instantly follow demand.
Modern AI performance is set by the whole system: accelerators, high-bandwidth memory, networking and software. If one layer cannot scale, expensive compute units sit idle, so nominal FLOPS are not the same as usable capacity.
Source: NVIDIA · Rubin platformAbout this eventConnecting a new data centre to the grid can take longer than building its computing halls.
The IEA identifies electricity as a local infrastructure constraint for data centres: capacity is concentrated in a few clusters while grids and generation expand more slowly than compute halls. The next scale-up is therefore timed not only by chip delivery but by the physical connection of megawatts.
Source 1: IEA · Energy and AISource 2: IEA · key questions 2026About this eventSparse computation, quantisation and specialised accelerators reduce the cost of one useful result.
About this eventPower, grids and cooling scale with systems without externalising the cost into energy and water scarcity.
Primary material behind this dossier: papers, lab publications and official reports.