Goldman Sachs has released a landmark report titled Powering the AI Era, highlighting that electricity supply has emerged as the single biggest bottleneck constraining the rapid development of artificial intelligence. The report reveals that building a typical 250-megawatt AI data center can cost up to $12 billion, and global data center electricity demand is projected to surge by 160% by 2030—driven by the immense power requirements of AI-driven GPU clusters. The report traces the evolution of infrastructure investment across history, drawing parallels between today’s AI boom and past transformative eras. In the 19th century, railroads were the backbone of economic growth in the U.S., with over 215,000 miles of track built and capital mobilized from European investors through early investment banks. By the 1920s, the U.S. expanded its grid capacity by 50 gigawatts—equivalent to $295 billion in today’s dollars—fueling industrialization and enabling innovations like Ford’s assembly line. The rise of public utility holding companies allowed for more efficient capital allocation and large-scale expansion. The internet era followed, with over $800 billion invested in fiber optics, broadband, and server infrastructure between the late 1990s and early 2000s. By 2001, 39 million miles of fiber had been deployed, creating the digital highway that underpins modern economies. Capital markets evolved accordingly, with venture capital, M&A activity, and credit instruments supporting this wave of innovation. Now, AI is the next major infrastructure shift. According to Goldman Sachs, the cost of constructing a 250 MW AI data center—complete with computing hardware—averages $12 billion. The energy demands of AI workloads are growing exponentially, with a single AI server rack expected to consume 50 times more power by 2027 than a typical cloud computing rack from five years prior. This leap in density has made traditional air cooling obsolete, pushing the industry toward advanced liquid cooling systems. The report outlines four generations of data centers: Cloud 1.0 (5–15 kW per rack), early AI upgrades (40 kW), transitional AI centers (130–200 kW), and the new “AI factories” that now exceed 500 kW per rack. These facilities are no longer just data storage hubs—they are high-power industrial plants. Global spending on AI infrastructure is projected to reach $1 trillion by 2027, driven by tech giants and emerging “new cloud” players. Importantly, AI demand is not cannibalizing traditional cloud spending—it is creating a massive new market. OpenRouter data shows that weekly token consumption—measuring AI computation volume—rose over 42-fold between June 2024 and May 2025. AI’s two core phases—training and inference—present different infrastructure needs. Training, which requires massive compute and is less sensitive to location, is increasingly concentrated in low-cost, energy-rich regions like Iowa, Nebraska, and parts of Northern Europe and Southeast Asia. The Stargate project, a $50 billion joint venture by OpenAI, Oracle, and SoftBank, exemplifies this trend, with its first data center located in Abilene, Texas. In contrast, inference—where users interact with models like ChatGPT—demands low latency, requiring data centers to be close to end users. This creates a geographic mismatch: training centers are built far from population centers, while inference needs to be local. As AI applications grow faster than expected, the value of remote training facilities may be at risk. Meanwhile, global data center vacancy rates have fallen to a historic low of 3%, with some markets near zero. New power supply projects are not expected to come online until 2028, leading to over 50 million square feet of new data center space under construction—double the level from five years ago. Goldman Sachs stresses that the key constraint is no longer capital—it is electricity. The world’s power grids, with an average age of 40 years, are unprepared for the surge. In the U.S., regulatory hurdles, lengthy permitting, and supply chain issues slow down new power projects. Decades of cheap natural gas, falling renewable costs, and low interest rates kept electricity prices low, but also weakened the financial viability of baseload power plants like coal and nuclear, accelerating their retirement. The result is a growing gap between supply and demand. The “duck curve”—a spike in electricity demand during evening hours when solar generation drops—has strained grid stability. To meet AI’s 24/7 power needs, the U.S. will require over $700 billion in grid investment by 2030. Utilities face a dilemma: they must modernize aging infrastructure while investing in new capacity for AI. Many are hesitant due to past overinvestment risks. However, AI demand is expected to keep growing, driven by enterprise adoption and agentic AI. To manage risk, utilities are exploring new pricing models—such as “take-or-pay” contracts, upfront capital contributions, and long-term capacity agreements. Partnerships between tech firms and utilities are increasing. Entergy and Meta are collaborating on joint power projects. Meanwhile, the acquisition of Calpine by Constellation Energy for $29.1 billion in January 2025 reflects renewed interest in natural gas as a bridge to meet immediate demand. Despite the challenges, solutions are emerging. Goldman Sachs projects that by 2030, 60% of new data center power will come from new generation facilities: 30% natural gas combined cycle, 30% gas peaking plants, 27.5% solar, and 12.5% wind. While renewables are the fastest-growing source, their intermittency requires complementary storage and backup. Nuclear energy is regaining attention. Companies like Alphabet and Microsoft are securing long-term power deals for advanced nuclear projects. Microsoft’s 20-year agreement to restart the Three Mile Island Unit 1 is a key example. Though historically plagued by cost overruns and delays—like the Vogtle 3 reactor, which doubled in cost and was seven years late—new technologies like small modular reactors (SMRs) offer safer, more scalable options. An innovative strategy gaining traction is “behind-the-meter” power generation—where data centers build their own microgrids or locate directly next to power plants to bypass grid congestion. This reduces time-to-market and avoids long interconnection queues. In Texas, during the 2021 blackout, PowerSecure’s microgrid provided over 2.26 gigawatt-hours of reliable power. However, such models raise concerns about environmental impact and equity. xAI’s Project Colossus in Memphis faced local opposition over emissions, and some nuclear co-location plans have been blocked by regulators over potential rate hikes. Beyond technology, this infrastructure race has become a geopolitical priority. The ability to attract AI data centers is now a key indicator of national competitiveness. The U.S. faces a growing supply gap, with projections of over 10 gigawatts of unmet demand by 2028. This may drive offshore training workloads to regions with abundant clean energy, such as Brazil—where over 90% of electricity comes from renewables. Financial innovation is keeping pace. In 2024, U.S. tech giants spent an average of $8 billion per day on capital. Utilities invested $200 billion annually. New financing models—such as joint ventures between real estate trusts, pension funds, and sovereign wealth funds—are emerging. The $15 billion Equinix partnership with Canada’s CPPIB and Singapore’s GIC is one such example. With over $4 trillion in available capital, institutional investors are eager to deploy funds in AI infrastructure. Goldman Sachs’ new Capital Solutions Group aims to provide integrated financing, advisory, and investment services tailored to the sector’s complexity. The report concludes that while the future remains uncertain, the trajectory is clear: AI is reshaping the global economy through a convergence of vision, infrastructure, and capital. The challenges—grid limitations, regulatory delays, technological risk, and geopolitical tensions—are immense. But history shows that every great infrastructure wave begins with doubt and ends with transformation. The winners will be those who can turn complexity into opportunity.