Part VI: Risks, Red Flags, and the Bear Case
An exhaustive analysis must address the structural fissures beneath the company's robust top-line growth. While Wall Street is enamored with the AI narrative, several fundamental red flags require strict monitoring.
1. Chronic Margin Compression
Operating a massive digital footprint is becoming exponentially more expensive. A quantitative analysis reveals that over the past five years, Digital Realty's gross margins have compressed by a staggering 46.1 percentage points, while net margins have similarly eroded. This is a severe red flag. It indicates that the rising costs of land, specialized construction materials, high-voltage cooling equipment, and debt financing are increasing at a faster velocity than the rental rates DLR can charge its tenants.
The Offsetting Mechanism: The logical question arises: How is Digital Realty mathematically offsetting this severe margin collapse to maintain bottom-line Core FFO growth? The answer lies in massive absolute top-line scaling. By continually deploying hundreds of new megawatts and driving top-line revenue to absolute record highs (e.g., $1.9 billion in Q3 2026), the company generates enough absolute gross profit dollars to outpace the deteriorating margin efficiency percentage on a per-share basis. However, this brute-force method of masking margin erosion requires continuous, staggering capital expenditure.
2. Shareholder Dilution (The ATM Treadmill)
Because DLR pays out a large dividend (over 70% of AFFO) and requires vast sums of capital for development, it continuously prints new shares. Through its At-The-Market (ATM) equity issuance program—a mechanism allowing the company to continuously sell new shares directly into the secondary market at prevailing daily prices, rather than executing a single, discrete secondary block offering—DLR sold approximately 2.4 million shares in Q3 2026 alone, raising $450 million. Year-to-date, the company has sold a massive 15.9 million shares to raise $2.9 billion, driving shares outstanding from roughly 343.6 million up to 370.0 million [cite: 4]. While this successfully deleverages the balance sheet, it continuously slices the equity pie into smaller pieces, requiring ever-greater aggregate net income just to keep per-share FFO flat.
3. AI Power Constraints and Sovereign Bottlenecks
Data center power demands have skyrocketed. AI workloads require up to 150 kW per cabinet, compared to standard enterprise workloads of 10 to 15 kW. The industry is facing severe local shortages of power, water, and grid infrastructure. If utility companies cannot deliver the gigawatts required, DLR's 9.8 gigawatt future development pipeline may remain unrealized potential.
The Dominion Energy Case Study: A concrete, disruptive example of this vulnerability is the ongoing Dominion Energy crisis in Loudoun County, Virginia—the world's densest "Data Center Alley." Beginning in 2022, Dominion Energy explicitly warned data center operators that it could no longer guarantee the delivery of sufficient electricity via overhead powerlines to Ashburn, threatening to significantly delay power connections for new data center developments until 2026 (DataCenterDynamics) [cite: 10, 11]. Resolving this transmission bottleneck has required the approval of new 185-foot high-voltage transmission lines routed directly through residential neighborhoods, sparking intense regulatory battles, environmental concerns, and massive local backlash (WUSA9) [cite: 12, 13, 14].
While these power delivery delays create a powerful economic moat for DLR's existing energized sites—customers cannot easily leave these hubs, driving DLR’s cash renewal spreads to 11.8%—they also represent a hard ceiling on future campus expansion.
4. The Structural Shift: Training vs. Inference
A subtle but critical shift is occurring in the AI ecosystem. Between 2024 and 2026, the bulk of capital was spent on AI training—which requires massive, centralized data campuses where latency to the end-user is irrelevant. However, as the industry pivots to AI inference (the real-time answering of queries, voice assistants, and robotics), workloads must be distributed to the "edge" to minimize latency. Inference requires proximity to the user and must navigate strict data sovereignty laws. If DLR over-commits capital to massive hyperscale campuses optimized for training, it risks holding misaligned real estate as the computing paradigm shifts toward hyper-localized, edge-inference freight networks.