Nvidia와 Digital Realty의 공동 AI 공장 건설이 전력망 수준의 자금 조달 병목에 막혀 프로젝트 규모 또는 일정이 위협받고 있다.

The push by Nvidia and Digital Realty to build flexible, large-scale AI computing infrastructure is encountering obstacles that extend beyond mere capital availability, according to recent coverage. The core challenge lies in determining whether their shared “AI factory” vision can operate within the physical constraints of the electrical grid while still delivering the computing performance customers expect.

A $150 million financing round for Emerald AI, announced on August 25, injected fresh capital into the initiative. The oversubscribed Series A valued the startup at $1.05 billion, pushing total funding past $220 million. Co-led by Energize Capital and DCVC, the round included Nvidia alongside strategic backers Samsung Ventures, Siemens, GE Vernova, RWE, Aramco Ventures, and Salesforce Ventures; twelve Fortune Global 500 companies now hold stakes in the firm. Emerald’s orchestration technology coordinates computing workloads with onsite energy resources, and the company stated the funding will support commercial expansion. It also highlighted a nearly 100-megawatt research facility in Manassas, Virginia, being developed jointly with Nvidia and Digital Realty, with an expected launch later this year.

The Virginia site is not a conventional data center. Nvidia announced it in October 2025 at its GTC Washington event as an AI factory research center at Digital Realty. It represents the first deployment of infrastructure built around the Vera Rubin platform—the generation succeeding Blackwell—and serves as the proving ground for a design blueprint intended for gigawatt-scale buildouts. Known as the Vera Rubin AI Research Factory, Emerald’s facility is being tested in collaboration with the Electric Power Research Institute, utility Dominion Energy, and PJM Interconnection, the grid operator whose capacity market the site must participate in.

The timing underscores the significance of these developments. An earlier announcement in October 2025 had targeted the first half of 2026 for the Virginia project. The revised schedule indicates a delay, though the companies have not publicly disclosed the reasons. Reports from CryptoBriefing and Yahoo Finance characterize the situation as a financing wall that additional capital alone cannot breach, highlighting deeper structural friction rather than a simple funding shortfall.

AI data center projects require massive upfront investment in land, power infrastructure, and specialized hardware. Nvidia supplies the graphics processing units that drive AI workloads, while Digital Realty operates the physical facilities that house them. Together, the two firms represent complementary halves of the AI compute supply chain. When financing hurdles extend beyond mere dollar figures, they typically signal underlying issues such as the pace of securing electrical power, permitting timelines, real estate availability, or the difficulty of aligning long-term capital commitments with uncertain AI demand curves.

Those constraints carry a tangible cost, which is now evident. PJM Interconnection operates the grid across 13 states and the District of Columbia, encompassing the Manassas site, and conducts an annual auction to procure the generating capacity required to maintain grid reliability. Three consecutive auctions have cleared at the maximum price permitted by federal regulators:

- 2026/2027: $329.17/MW-day, at cap; met by 139 MW

- 2027/2028: $333.44/MW-day, at cap; short by 6,623 MW

- 2028/2029: $325/MW-day, at cap; short by 6,831 MW

The 2027/2028 auction marked the first time in the market’s history that the entire footprint fell short of the reliability standard, and the most recent results widened the gap further. The operator has indicated it will seek approval for a special procurement this month to address the near-term shortfall. Data centers are driving this pressure, accounting for nearly 5,100 of the roughly 5,250 megawatts by which the load forecast for the 2027/2028 delivery year was increased. “These auction results show that demand for electricity continues to grow faster than electricity supply,” David Mills, the operator’s president and chief executive, said in announcing the latest results. The International Energy Agency projects that data centers will account for nearly half of U.S. electricity demand growth through 2030.

The grid operator’s response marks the intersection of Emerald’s thesis and market reality. PJM is developing “Connect and Manage” frameworks that would allow large new loads, such as data centers, to connect to the system while operating flexibly during periods of grid stress. In June, the operator began facilitating bilateral supply contracts for large customers capable of committing to terms of ten years or longer. It has also launched an expedited interconnection track for shovel-ready generation projects sponsored by state governments. This trajectory points toward a regulatory environment where a data center’s connection speed hinges on its willingness to curtail demand—a capability Emerald markets.

Meanwhile, interconnection remains the slower bottleneck. Berkeley Lab reports approximately 2,061 gigawatts of generation and storage projects actively seeking grid connections in the United States, a 10% decline from a year prior and below the peak of nearly 2,600 gigawatts recorded at the end of 2023. Only 13% of projects that filed interconnection requests between 2000 and 2020 had reached commercial operation by the end of 2025, while 75% had been withdrawn. The median project now requires well over three years merely to secure an interconnection agreement and more than five years to achieve commercial operation, with over 40% of projects still dropping out after an agreement is signed. This represents the structural friction referenced in recent coverage: a queue that additional capital alone cannot bypass.

The proposed architecture enables facilities to modulate their electricity consumption during periods of grid strain. Nvidia’s software and infrastructure controls integrate with Emerald’s orchestration technology to form the system. For the chipmaker, successful deployments could alleviate a key constraint for customers deploying and operating its hardware, thereby sustaining demand through tangible infrastructure availability rather than relying solely on venture capital injections into AI startups.

Emerald has already reported commercial deployment across an entire California data center, validating the concept even if it does not yet dictate the Virginia project’s final economics. The company states it has completed five demonstrations at commercial sites in Arizona, Illinois, Virginia, Oregon, and London, collaborating with Nvidia, Oracle, Nebius, and National Grid. It has also partnered with Silicon Valley Power on what it calls the nation’s first flexible-load interconnection program, which grants data centers expanded grid access in exchange for verified, dispatchable flexibility. A system that reduces power consumption while disrupting critical workloads would present a weaker value proposition. Broader adoption will require a viable balance between grid responsiveness and computing performance.

Digital Realty pursues a distinct path to value. Assisting customers in securing power and maintaining reliable operations could enhance the attractiveness of its facilities. More efficient utilization of infrastructure carries particular weight for a data center REIT, whose invested capital must ultimately generate rental yields. Nevertheless, the project announcement outlines no quantified earnings impact for Digital Realty, nor does Emerald’s financing transaction establish a new order backlog for Nvidia. Both companies are engaged in the same technical initiative without deriving identical financial returns.

Investor positioning reflects sustained interest in both equities. According to Insider Monkey’s database, 285 hedge funds held Nvidia in the second quarter of 2026, up from 275 in the first quarter. Holdings in Digital Realty increased to 73 from 46 over the same period. Ken Fisher’s firm maintained a position in Nvidia, while Andreas Halvorsen’s Viking Global held Digital Realty. These allocations predate the August financing round. As of August 14, short interest accounted for 1.23% of Nvidia’s float and 2.07% of Digital Realty’s.

The strategic backer lineup indicates where demand for this capability is anticipated to originate. Among them is the venture arm of JERA, Japan’s largest power generation company, which produces approximately one-third of the nation’s electricity; the fund maintains a stated strategic investment capacity of $300 million. In announcing the investment, JERA noted that grid interconnection constraints are emerging as a primary hurdle for data center development, and emphasized that software optimizing the interaction between computing workloads and the power grid can play a critical role in resolving them.

The AI infrastructure sector has drawn massive capital in recent years. Hyperscalers and specialized data center operators have committed billions of dollars toward expanding computing capacity. Nvidia's chips remain central to that expansion, given their dominant position in AI accelerator hardware. A financing wall at this stage would be notable precisely because so much capital has already flowed into the sector. It suggests that even well-capitalized, high-profile