WNTD outlines shift to gigawatt-scale AI infrastructure and liquid cooling

The transition toward gigawatt-scale AI factories indicates that the underlying hardware for video processing and recommendation engines is becoming more capital-intensive and complex. As streaming platforms integrate more generative AI for content creation and personalization, their reliance on specialized GPU infrastructure and liquid-cooled data centers will grow. This shift forces a move away from generic cloud instances toward purpose-built AI factories that can handle massive computational loads. The industry must now navigate a talent gap for professionals capable of managing these hyperscale environments while contending with two-year lead times for critical hardware. Watch for whether major streaming providers pivot toward sovereign AI infrastructure to secure long-term compute capacity.

NVIDIA has positioned its GPU architecture as the foundational compute layer for gigawatt-scale AI factories, a framing that directly shapes how streaming platforms plan their inference and training workloads. In March 2025, NVIDIA CEO Jensen Huang announced the company's Vera Rubin platform targeting AI factories at the GTC conference, projecting that next-generation data centers would require liquid cooling from the rack level up. That announcement reinforced the industry consensus that air-cooled facilities cannot support the power densities demanded by modern AI training clusters, a constraint WNTD's Blaine Daws highlighted in his discussion of gigawatt-scale planning.

CoreWeave has emerged as one of the most aggressive builders of GPU-dense infrastructure purpose-built for AI workloads. In March 2025, CoreWeave filed for an IPO valuing the company at approximately $23 billion, with its S-1 disclosure showing that the company had secured long-term contracts with major AI labs requiring dedicated liquid-cooled capacity. The filing revealed that CoreWeave's data center portfolio was concentrated in facilities designed for NVIDIA's H100 and Blackwell GPUs, underscoring how AI infrastructure providers are locking in multi-year commitments that mirror the strategic-national-infrastructure framing WNTD described. Volta's 133MW Norway facility fits into this same pattern of European sovereign AI capacity buildout, where the company secured funding to expand its Nordic data center footprint for AI workloads with a focus on renewable energy and low-latency connectivity to continental European markets.

The technical requirements for these AI factories are pushing cooling and power delivery into new territory. In February 2025, NVIDIA published specifications showing its GB200 NVL72 rack requires 120kW of liquid-cooled power per rack, a density roughly ten times higher than traditional air-cooled server deployments. That specification has forced data center operators to redesign their entire power distribution and thermal management stacks. For streaming companies that rely on GPU-accelerated video encoding, recommendation models, and generative AI pipelines, the implication is clear: the infrastructure they depend on is being built at a scale and complexity that demands specialized engineering talent and multi-year capital planning, exactly the challenges Daws outlined for the WNTD team. Recent market shifts toward further highlight the competitive pressure to optimize these massive compute investments, a trend further complicated by constraints. As these massive facilities expand, has become a critical bottleneck for developers, a challenge compounded by the fact that now dictates site selection.