Meta Platforms is designing what it says will be the world’s largest artificial-intelligence training installation, a tent-based project dubbed “Prometheus.” According to internal planning details circulated on 1 August, the system is expected to be equipped with about 500,000 Nvidia Blackwell GB200/300 graphics processors, draw roughly 1,020 megawatts of IT power and deliver more than 3.17 trillion floating-point operations per second. The company aims to finish the build-out by the end of 2026, overtaking the compute scale of OpenAI and other rivals. The disclosure underscores how fast the capital-intensive AI infrastructure race is accelerating. Power requirements for large-scale model training already rival those of medium-sized cities, and industry analysts say total AI electricity demand could double New York City’s consumption before the decade ends. Google, meanwhile, has quietly brought online a 100,000-chip cluster of its sixth-generation Tensor Processing Units to run Gemini 2.0 and newer models. On the same day Meta’s plans surfaced, the company began rolling out Gemini 2.5 “Deep Think,” a multimodal reasoning upgrade that it says outperforms competing systems on standard benchmarks. Access is restricted to subscribers of Google’s AI Ultra tier, which is priced at $250 a month and allocates a fixed number of daily prompts. The parallel moves by Meta and Google signal growing willingness among Big Tech firms to commit multi-billion-dollar budgets to bespoke computing complexes as they vie for dominance in generative AI. Analysts expect the resulting demand to shape semiconductor supply chains, power-grid planning and regulatory scrutiny over the next several years.