E85:下一次权力转移：亚洲如何构建人工智能时代

Recorded: July 17, 2026 Published: August 6, 2026 Host: Erica Poon Werkun, JANA & Asia South Head of Equity Research, Citi Guest: Kyna Wong, China Technology and Communications Analyst, Citi Guest: Pierre Lau, Asia Utilities Analyst, Citi Transcript: Erica Poon Werkun (0:00) Hi, everyone. Welcome to the Research @ Citi podcast. I'm Erica Poon Werkun, JANA & Asia SouthHead of Research at Citi. Joining me today are Kyna Wong, China Technology and Communications Analyst, and Pierre Lau,Asia Utilities Analyst. Kyna, Pierre, and the broader Asia team have recently collaborated on a Super-Sector report[available to existing clientshere]examining the structural shift to 800VDC powerarchitecture in AI data centers and the investment implications across Asia. Kyna and Pierre, welcome to the podcast, and thank you both for joining us today. Kyna Wong (0:41) Hi Erica, hi everyone. Great to be here. Pierre Lau (0:42) Hi, everyone. Erica Poon Werkun (0:46) Great. I will turn the first few questions to Kyna. Kyna, why is 800VDC becoming the standard fornext-generation AI data centers, and why now? Kyna Wong (0:57) Thanks for that question. It's a little bit complicated and also related to physics. But I wanted to saythe fundamental driver is about the GPU power density. We see that rack power has grown nearly100-fold since 2020—from 10kW, at that kind of energy level, with the NVIDIA Ampere architectureGPU to 120kW with Blackwell now,and projected to even exceed one megawatt with upcomingRubin Ultra and also Kyber platforms. The standards are getting much higher now. At these kind ofpower levels, traditional AC power distribution simply cannot handle it physically and alsoeconomically. So doubling the voltage to 800VDC is massive because it halves the electricity currentfor the same power delivery.That can save on a lot of things.It will reduce power requirements, cableusage, transmission losses. It's all about physics,but it helps a lot. This is also the reason why electricity is transmitted from rural area to rural area through very high-voltage transmission lines in the country and then transferred to a lower voltage to the household. Ithink this is the reason they wanted to save the energy losses during the transmissions. Beyond the headline efficiency improvement and also the copper cable savings, 800VDC also freesup space in the data center by eliminating lots of parts.For example, UPS is Uninterruptible PowerSupply, it's a back-up power system, and also transformers like stepping up or down the powervoltage and the batteriesin the data centers,which already occupy around 30 to 40% of a traditionalfacility's total footprint. Critically, for hyperscalers with carbon-neutrality commitments, they wantto be green so 800VDC is architecturally compatible with DC-power-native renewable energy thatare already using DC power transmissions, and also battery-storage-saving intermediate ACinversion losses that also saves on other parts. And the distributed battery-backup unit, called theBBU, and also the battery energy-storage system, or BESS, is actually replacing some of the otherparts and is better architecture fit for managing these very dynamic loading profiles of GPU AIclusters in data centers than historically in UPS. Erica Poon Werkun (3:26) Thank you, Kyna. With power being such a bottleneck for the AI DC development, this technologicaladvancement surely is very important. So how quickly will 800VDC be adopted globally, and whatdoes the transition timeline look like? Kyna Wong (3:42) That's a good question. The demand for 800VDC is accelerating, driven by NVIDIA's cloud and data-center architecture roadmap. We project 800VDC will account for 16% of new global data-centercapacity additions in 2027, rising to 58% in 2028, 73% in 2029, and even 79% by 2030, based onhigh-end GPU server shipment projections. In absolute terms, that represents incremental 800VDC data-center capacity growing from 4.3gigawatts in 2027 to over 25 gigawatts annually in 2030. The transition is not a single event. It unfolds in phases, beginning withsidecar retrofits of existingfacilities in 2027 to 2028, and progressing toward purpose-built solid-state transformer or SST-nativedata centers from 2029 onward. The distinction in architectures matters because the twomodes create very different equipment-demand profiles. Sidecarsprobably could peak in 2027 to2028; SST may take over from 2029onward. Erica Poon Werkun (4:51) Well, Asia is the center for a lot of the supply chain for AI.So,if we look at the individual sectors,which are the most directly impacted by this 800VDC transition and how? Kyna Wong (5:06) In our report, we summarized the 800VDC theme impact into six distinct sectors across the data-center infrastructure. First, data-center operators are the primary demand aggregators. Their ability to develophigh-density, 800VDC-ready white space will increasingly determinetheirlease pricing power and tenantretentio