PowerBank First Lease: 2 MW Ontario Data Center for ~500 NVIDIA GB300 GPUs - News and Statistics - IndexBox

PowerBank's 2MW Ontario lease for 500 GB300s is less a compute announcement than a structural signal about who actually controls AI infrastructure economics. At roughly 4kW per GPU, the facility is liquid-cooled by design, not retrofit. Upstream, CoWoS-L packaging and HBM3E stacking remain TSMC and SK Hynix bottlenecks, but 500 units is trivial for supply chains. The "lease" framing is the real tell: PowerBank is a compute reseller, not an infrastructure operator, with zero chip procurement leverage. On compliance, Ontario's ~90% hydro grid is a quiet competitive moat as CBAM and ESG disclosure tighten across North America. Yet GB300 shipments still sit under US BIS export-control jurisdiction. One policy adjustment could freeze in-transit silicon at the border—a systemic tail risk for every North American AI compute player. Competitively, 500 GPUs is noise for CoreWeave or Lambda. PowerBank's actual arena is regional AI-as-a-Service, where margin lives in the spread between wholesale power and retail inference pricing. The 12-to-24-month wildcard is Rubin. When NVIDIA's next-gen die pushes per-GPU TDP past 2kW, a 2MW shell without electrical headroom becomes a stranded asset overnight. The real long-tail effect isn't silicon—it's electrons. Whoever locks 50MW+ of firm power first owns the next generation of AI compute. The GPU is a commodity; the megawatt is the moat.