🧠 Food for thought

Implications, context, and why it matters.

Foxconn’s 1,000 racks a week could mean $48B to $52B in quarterly AI server revenue

• If every system were a GB300 NVL72 at $3.7 to $4.0 million, revenue would reach $48 to $52 billion 1, a big share of the $184 billion Wall Street pegs for Nvidia’s fiscal 2026 data center sales 2. GB200 NVL72 systems carry a 132 kilowatt thermal design point 3 and each GB300 NVL72 rack needs about $50,000 of liquid cooling gear 4. Delivering 13,000 in a quarter would require roughly 1.7 gigawatts and about $650 million for cooling parts. The shift from dual-rack NVL36x2 to single-rack NVL72 or NVL36, with NVL36x2 pairing two 36-GPU racks while NVL72 and NVL36 are single-rack options, ties to supply chain snags 3.

Taiwan’s 27 MW GB300 plan creates openings for liquid cooling and power vendors

• Visionbay.ai plans a 27 MW Taiwan facility using GB300 chips with a first half 2026 launch target, and Zettabyte with Chief Telecom proved legacy sites can reach 100 kW per rack in 45 days 5. Each GB300 NVL72 rack needs about $49,860 of liquid cooling parts 6, so a roughly 27 MW build with about 200 racks would mean near $10 million of cooling hardware. Teco and Foxconn aim to start AI data center work as early as 2026 7. Cold plates are metal heat spreaders that move chip heat into liquid coolant at about $300 per CPU or GPU unit 6.