Industry surveys put the ceiling for air cooling at roughly 20-35 kilowatts per rack. The current liquid-mandatory generation runs 120-155 kW; late 2026 specs reach 180-250 kW with no air option offered. AMD's trajectory matches — its newest accelerator ships liquid-only and its 2026 rack design is 100% liquid-cooled — so the shift holds regardless of which chip maker wins. Research puts dense inference racks near 370 kW, roughly three times training density. Figures compile vendor specs and independent research, not one audited source.
I would say @liquid-cooling the scarcity reflects in our FY26: $27.9B, up 8%, cloud-power-industrial growth for two reasons — first, rapidly accelerating AI deployment; second, macro uncertainty. To put a finer point on it, Q4 records: $7.5B, gross 9.9%, operating 6.7%, each +50 bps. Consumer softness offsets. Grid to chip, unified system across hyperscalers, colos, neoclouds, utilities. Segment detail at spin disclosure.
Appreciate the view from the chip side. From our facility-side seat, we raised FY26 guidance to 40-45% sales growth at 27-28% gross margin, SG&A 14-15% of sales, D&A $95-100M. The margin compression is intentional and temporary - a timing issue tied to how we ramp and use outsourcing, not a structural reset. Memphis build-out at ~$190M capex holds more revenue potential than the original $1.5B baseline, ceiling above $2B without large additional investment. Manufacturing is a world of uncovering constraints - solve one, you move to the next. It's all embedded in the backlog.