Nvidia Corp. CEO Jensen Huang told a Berlin audience Tuesday that prefabricated "AI factories" built inside shipping-container footprints can shrink data-center deployment from years to months, and Quanta Computer Inc. said it is already costing power and cooling for container clusters aimed at European industrial customers.
What Huang showed on stage
The keynote paired a Mercedes-Benz street trial of Nvidia Drive software with a modular data-center mock-up: liquid-cooled racks prewired for Blackwell GPUs, dropped onto concrete pads with on-site substations. Huang argued that physical AI—robots, inspection cameras, and warehouse automation—needs inference close to factories, not only hyperscale clouds in Iowa.
For Taiwan's contract manufacturers, the pitch matters because Quanta, Wistron Corp., and Foxconn Industrial Internet already assemble GPU servers for U.S. clouds. Container packaging could let them ship entire aisles from Taoyuan or Kaohsiung ports without waiting for local construction crews abroad.
Margin and power questions in Taipei
Quanta executives on a media call Wednesday said container deals hinge on who owns the power purchase agreement. If European utilities pass through renewable premiums, Taiwanese builders may capture lower hardware margin but higher service fees for remote monitoring.
Analysts on the Taiwan Stock Exchange floor noted that Huang's Berlin timing landed just as the TAIEX rebounded on AI server demand. MediaTek and ASIC houses were quiet, but rack integrators rose with Hon Hai, reflecting investor belief that physical-AI capex is a second wave after training clusters.
Competition from local players
Wiwynn Corp. and Inventec Corp. are pitching their own modular designs with Taiwanese switch vendors, betting that not every buyer wants Nvidia's reference stack. Still, Huang's container narrative gives sales teams a story for municipalities that lack land zoned for megawatts-scale buildings.
Regulators in Taiwan will watch export licenses for high-end GPUs inside containers, especially when destinations sit near sensitive infrastructure. For now, the Berlin keynote sets the talking points for fourth-quarter ODM earnings calls: ship faster, cool denser, and sell AI as a factory appliance rather than a distant cloud tab.
Energy and export licensing
Taiwan's Ministry of Economic Affairs said exporters must still file strategic-high-tech reviews when shipping the newest GPU classes, even if they leave port inside containers. Power utilities in Taichung warned that demo units pulling 200 kilowatts could strain industrial park feeders unless buyers install dedicated transformers.
European customers on Quanta's call asked about repair loops: whether failed racks would return to Taiwan or be serviced in the Netherlands, a detail that affects warranty revenue recognition next quarter.