The Zhitong Finance App learned that Goldman Sachs analysts attended the DCD Connect Data Center Conference held in Loudon County, Virginia from November 3 to 4. Goldman Sachs summed up the general consensus at the conference on four points: “monetization time” has become the primary standard for data center site selection. The main implementation risks of development are power, licensing, community support, and labor; modular solutions can shorten the construction and replacement cycle; the impact of severe fluctuations in AI training loads (and lesser inference loads) on equipment and power grids must be taken into account during the project design phase; 800V DC power supply technology is likely to be launched in stages.
The three most debated issues are: whether the power supply behind the meter can be used as a long-term main power supply, whether AI inference will spawn edge computing requirements, and what role traditional telecom operators play in AI connectivity.
Grab time: lock the equipment first and then find a tenant. The construction period and cost are under pressure
Electricity remains the number one constraint, and licensing and community support are closely watched. Negotiations on the tenant's overall asset portfolio, the expected electricity consumption curve of the project, and communication with the community are now beginning earlier and earlier in the development process. Labor is also listed as a bottleneck — around 30% of veteran electricians are expected to retire by 2030.
In terms of construction schedule, an insurance company mentioned at the meeting that 85% to 90% of the approximately 60 data center projects under construction (construction period of 18 to 40 months) it tracks are longer than expected, but some delays are only two or three months, and are related to the supply chain. To cope with the lengthening procurement cycle, developers began to target long-term equipment early in the process, sometimes even before anchoring tenants — which meant investing equity capital earlier and on a larger scale. Interestingly, there are also vendors who have given the opposite anecdote: they have delivered devices to hypercloud vendors with a delivery time of as little as two weeks. Short delivery times are one of the topics of discussion: in the face of unrealistic construction schedule requirements, developers and manufacturers often charge a premium before trying their best to catch up, driving up the cost per megawatt and often still being delayed.
Modularity and flexibility are seen as strategic priorities. Participants pointed out that the industry is shifting from “building” data centers to “assembling” data centers, and the certainty of the modular solution shortens replacement time and improves overall site efficiency; the next possible shift is to replace IT hardware inside the computer room every two to three years instead of directly replacing external module units.
“Flickering” of AI load: 60% fluctuation within 5 ms, and the device broke at the early AI site
The AI training load may fluctuate around 60% in 5 ms, and this impact must be absorbed by the design. Some participants mentioned that early AI sites already experienced breakage of drive shafts and gearboxes — devices originally designed for smooth loads are now being loaded on a large scale at high frequencies. The redundant design ensures that no one engine is running at full load, thereby extending life, reducing shock, and partially hedging the risk of AI load swings. The training load is easy to recover after being cut off, while the inference load requires higher reliability and less swing but is still obvious. Retail hosting facilities use an average of only 35% to 50% of available capacity due to peak conservatism and strict service agreements; participants discussed freeing up surplus capacity from existing sites using use cases that are less sensitive to delays.
In terms of power supply technology, participants generally agreed that 800V DC will not be the majority solution for building new AI data centers before 2028, but will be implemented in stages. DC power has not passed zero, making arc protection and charging difficult; industry standards are scarce, and few operators and maintenance personnel have received operation training in 800V data lobbies.
Three major debates: power, edge computing, and telecom operators after the table
The back of the meter (BTM) power supply is the first controversy. It is estimated at about twice the cost of interconnecting power grids, is less resilient, and is more likely to cause licensing disputes (such as air emission permits). There were mixed responses from participants, but the general consensus was that power models will increasingly be hybrid.
Edge computing is the second game. Potential catalysts include sovereign AI use cases and the advent of AI models that are lightweight enough to fully operate at the edge; however, there are also cautions that if a technology can bring the latency of long-distance transmission close to zero, edge computing will become less necessary for AI inference.
Telecom operators are in the third game. Participants believed that traditional telecom companies could benefit from AI construction, but the main path was the last mile and existing capabilities. They expect that the latest Nvidia GPUs will bring about a significant increase in north-south IP traffic, and telecom companies can create value for private traffic and network routing between large cloud vendors and consumers; the value of interconnected data rooms (meet-me-rooms) is also rising — low latency requires neutral interconnection systems, and it is unlikely that hypercloud vendors will build their own. However, participants also warned that telecom companies have higher capital costs than giant cloud vendors, which may limit the investment they need to benefit from AI. Additionally, fiber has been discussed as a potential next bottleneck: the wait time to excavate pipelines in some regions has reached 36 months; however, others think that compared to obtaining electricity infrastructure or labor, it is relatively easy to excavate optical cables in rural areas.
Geographic restructuring: 77% of pipelines in frontier markets, “demarketized” data center markets
Participants discussed the significant constraints on green space development in traditional first- and second-tier cities: 77% of current construction pipelines are in frontier markets. Limited availability in Tier 1 and 2 cities, compounded by overall tight supply and favorable price dynamics, has spawned what some participants called the “de-marketized” (de-marketized) data center market — capacity growth in the frontier market will continue to outperform Tier 1 and Tier 2 cities, and power supplies will further accelerate this trend.