PDF 原檔：報告_GS_AI基建_20260519_original.pdf
原始內容
Explore >
Data Centers: Asia Communacopia + Technology - Key Takeaways: AI Infrastructure: Powering the Next Intelligence Cycle
Presenters : Mark Fong, CEO (Empyrion Digital), Yaniv Ghitis (Chief Commercial and Investment O ffi cer, Digital Edge), Stephen Watts (CEO, PaleBlueDot AI)
Bottom Line : 1) Power as the key bottleneck a ff ecting construction timelines and delivery capability, with public-private partnership on grid infrastructure builds needed to bridge the gap. 2) AI compresses timelines , with single-tenant GPU customers signing contracts in 6 weeks vs. 4-6 months for traditional hyperscalers. 3) Data center operators preserving fungibility to cater toward both traditional and AIfi rst customers , e.g. hybrid air/liquid cooling, variable rack densities, and delaying data hall fi t-out to the last possible moment. 4) Supply would sell out within 3-6 months at current demand levels if it existed , but expansion is constrained by power, permitting, water, and increasingly community acceptance. 5) Geopolitics raises compliance overhead but has not dampened underlying demand. Japan, India, the Philippines fl agged as key upside growth markets.
Key Takeaways
Power as the key bottleneck
- While GPU shortages should normalize by 2028-2029 ; power is the key n constraint a ff ecting construction timelines and delivery capability, because resolving grid infrastructure issues requires 5 to 10 years.
- Public-private partnership is the key to answer who bears responsibility for n infrastructure builds : Governments should take the lead given power's nature as a national utility, with private operators and hyperscalers increasingly investing directly in power generation and renewables to bridge the gap while contributing through utility tari ff s. Malaysia was highlighted as a market where government-utility collaboration is working e ff ectively.
- Power e ffi ciency (PUE) vs water tradeo ff : Di ff erent markets demand tailored n approaches based on local resource scarcity and pricing. Digital Edge highlighted
Goldman Sachs does and seeks to do business with companies covered in its research reports. As a result, investors should be aware that the fi rm may have a con fl ict of interest that could a ff ect the objectivity of this report. Investors should consider this report as only a single factor in making their investment decision. For Reg AC certi fi cation and other important disclosures, see the Disclosure Appendix, or go to ed as research
Ronald Keung, CFA +852-2978-0856 |
ronald.keung@gs.com Goldman Sachs (Asia) L.L.C.
Timothy Zhao
+852-2978-2673 | timothy.zhao@gs.com Goldman Sachs (Asia) L.L.C.
Kelsey Santoso
+65-6889-2473 | kelsey.santoso@gs.com Goldman Sachs (Singapore) Pte
Eunice Liu
+852-2978-7472 | eunice.liu@gs.com Goldman Sachs (Asia) L.L.C.
e92c7a75ab8b4efbba794e6b187208c8
- its Mumbai facility as a case study where it is sourcing treated wastewater from a nearby puri fi cation plant rather than potable water. The site features the largest water-based chiller plant in its class and delivers PUE below 1.25.
- Empyrion foresees that the four pillars of infrastructure builds (power, water, n connectivity, permits) are being joined by a fi fth: community acceptance . Operators who successfully engage local communities will gain a competitive edge.
How shifting customer mix is a ff ecting data center operators
- Single-tenant AI customers vs. traditional multi-tenant - di ff erent lead times : n Empyrion noted that single-tenant GPU customers show commitment on orders 1524 months before construction starts, locking in bespoke designs early. Traditional multi-tenant customers, by contrast, expect only 3-6 months of lead time and more standardized infrastructure. Empyrion aims to delay data hall fi t-out to the last possible moment to cater to both groups.
- Hyperscalers vs. neoclouds - radically di ff erent commercial speed : New AIfi rst n customers and neoclouds are pushing for modular and prefab designs to compress delivery timelines. They also move faster commercially, with Empyrion noting new contracts can close in 6 weeks vs. 4-6 months with hyperscalers. As a GPUaaS provider who transforms energy into tokens, PaleBlueDot cited a Japan project that delivered 6,000 GPUs in 89 days. Speed to ful fi ll the demand has become a key di ff erentiator from a customer perspective.
- Operator response - modularity, fungibility : The panel agreed that hybrid n air/liquid cooling is now the default design choice to support both GPU-intensive AI and traditional cloud workloads in the same facility. Digital Edge engineers its facilities to support rack densities from 20kW to over 1MW - a wide envelope that avoids over-optimizing for AI at the expense of other workload types.
Growth outlook
- Biggest upside markets : India is expected to grow rapidly, supported by favorable n demographics, deep engineering talent, and proximity to the Middle East. Digital Edge is already executing on this thesis via its 300MW Navi Mumbai campus. Japan was also fl agged as a standout with government-backed initiatives driving aggressive expansion; alongside the Philippines with reduced red tape and scalable power.
- Training vs inference : PaleBlueDot noted that c.70% of its AI compute demand is n now inference. Company expects the demand to continue to scale as application-layer consumption has yet to be fully realized.
- Boldest predictions and forward-looking themes : Empyrion framed the AI-driven n demand surge as non-zero-sum where traditional cloud workloads continue to grow alongside new GPU and AIfi rst demand, expanding total market size rather than redistributing it. On the supply side, consolidation among DC operators is expected to improve e ffi ciency and reduce power/chip constraints.
Geopolitics
- Regulatory shifts have caused operational adjustments (e.g. enhanced due n
e92c7a75ab8b4efbba794e6b187208c8
diligence now extending beyond direct customers to ultimate end-users and supply chain participants) but have not deterred underlying demand momentum, with demand resilience across regions . Empyrion commented that supply would sell out within 3-6 months at current demand levels if it existed.
e92c7a75ab8b4efbba794e6b187208c8