The AI infrastructure buildout will cost $7.6 trillion over the next five years, but most investors are focused on the wrong part of the supply chain. While semiconductor companies like NVIDIA capture the headlines and the margins, the actual bottleneck that could slow or derail the entire cycle sits in unglamorous but critical infrastructure: power generation, grid interconnection, cooling systems, and electrical transformers. Understanding which assumptions drive this massive forecast reveals where the real opportunities and risks hide.

What Four Assumptions Drive the $7.6 Trillion AI Infrastructure Forecast?

Goldman Sachs Global Institute built its $7.6 trillion projection on four specific structural assumptions that determine whether the AI infrastructure ramp continues at the projected pace, accelerates, or stalls. The headline number itself is less important than understanding what could change it by hundreds of billions of dollars in either direction.

• Silicon Useful Life: AI accelerators like GPUs and custom chips typically depreciate over four to six years on accounting books, but NVIDIA's annual release cadence with step-function performance improvements pressures operators to replace hardware faster than depreciation schedules suggest. Shortening the assumed useful life from six years to four years directly increases replacement cycles and pushes cumulative capital expenditure significantly higher.
• Data Center Cost Per Megawatt: Traditional cloud-era data centers cost roughly $10 million per megawatt to build, but AI-era facilities now run $15 to $20 million per megawatt because they require advanced liquid cooling, tighter power delivery tolerances, higher rack densities, and tightly integrated system designs. This rising cost directly pushes the aggregate $7.6 trillion number upward.
• Chip Architecture Mix: Most AI compute today runs on NVIDIA GPUs, but custom silicon from hyperscalers like Google's TPUs, Amazon's Trainium, and Microsoft's Maia designs is gaining share. Whether this shift reduces aggregate spending depends entirely on demand elasticity; if cheaper compute unlocks more usage, the total spend remains high even as margins shift away from NVIDIA.
• Physical Bottleneck Elongation: Power interconnection queues, permitting delays, specialized labor shortages, and long lead times for transformers and cooling equipment stretch the timeline between capital commitment and operational capacity. In stress scenarios, supply-side friction can introduce demand-side doubt and defer or downsize investment plans entirely.

Why Is Power Infrastructure the Most Under-Owned Part of the Trade?

The conventional narrative frames NVIDIA as the primary beneficiary of the AI capex cycle, but Goldman's framework points elsewhere: the most under-owned exposure sits in power infrastructure, cooling systems, and grid capacity. The International Energy Agency projects global data center electricity use to roughly double by 2030, with AI workloads tripling their share of that total. U.S. data centers already account for close to half of incremental electricity use, and utilities in PJM, ERCOT, and the Mid-Atlantic are signing decade-long supply contracts at premium prices.

This is not a theoretical constraint. Data centers generate heat at densities that air cooling cannot handle, so liquid and direct-to-chip cooling systems are now standard in every hyperscale deployment. The bottleneck is not silicon supply; NVIDIA is delivering record revenue. The bottleneck is transformers, switchgear, cooling equipment, and grid interconnection capacity. Companies like Vertiv, which supplies power management and liquid cooling systems that show up in nearly every hyperscale request for proposal, sit at the center of this constraint.

How Are Hyperscalers and Nuclear Power Reshaping the Energy Equation?

The convergence of massive AI capex and energy constraints has triggered an unexpected alliance: hyperscalers are now driving a nuclear power renaissance in the United States. Construction began in April 2026 on two long-planned nuclear power projects, one in Wyoming and the other in Tennessee, signaling the start of the "steel-in-the-ground" phase of what industry observers call the second "Nuclear Renaissance".

Meta Platforms has signed agreements totaling 6,600 megawatts with Vistra, TerraPower, Oklo, and Constellation Energy to boost nuclear generation, with commitments to build new next-generation nuclear generators and extend the lives of existing plants in Ohio, Pennsylvania, and Illinois. TerraPower, backed by former Microsoft CEO Bill Gates, is building the Natrium project in Wyoming, a 345-megawatt sodium-cooled fast reactor projected to cost about $4 billion and begin commercial operations in late 2029. Kairos Power, backed by Google, is constructing a 50-megawatt small modular reactor demonstration project in Oak Ridge, Tennessee, estimated to cost about $303 million.

"This is the moment our industry has been working toward for a generation. We're not just breaking new ground on a first-of-a-kind nuclear plant in Wyoming; we're building the next generation of America's energy infrastructure," said Chris Levesque, president and CEO of TerraPower.

Chris Levesque, President and CEO at TerraPower

The Trump administration has strongly encouraged nuclear construction to meet future power needs of data centers powering artificial intelligence applications. The U.S. Department of Energy is authorized to make up to $2 billion in 50-50 cost-sharing for TerraPower's Natrium project and recently announced a $26.5 billion loan package for Southern Company subsidiaries that includes licensing and upgrades for about 6,000 megawatts of nuclear generation.

What Does This Mean for the $500 Billion Annual AI Capex Cycle?

The Mag 7 hyperscalers (Microsoft, Alphabet, Meta, Amazon, Apple, Tesla, and NVIDIA) will spend over $500 billion on AI infrastructure in 2026, with Amazon targeting $200 billion and Alphabet guiding to $180 to $190 billion. Combined hyperscaler capex is tracking above $650 billion for 2026, with roughly $500 billion of that earmarked specifically for AI infrastructure. Investors holding only Mag 7 names are missing the second-order plays that capture the actual flow of capital across the supply chain.

The four cleanest baskets to capture this flow are semiconductors, power, real estate investment trusts (REITs), and cooling and networking equipment. Semiconductors capture the largest share of each capex dollar, led by NVIDIA, Broadcom, AMD, and Arm Holdings. Broadcom sells custom AI application-specific integrated circuits (ASICs) to Google and Meta plus the networking chips inside training clusters. Arm Holdings powers the host CPUs paired with GPU accelerators, including AWS Graviton processors.

The power sector benefits from regulated utilities in Northern Virginia, Phoenix, and central Ohio that see rate-base growth, plus independent power producers with nuclear and gas baseload capacity that capture premium power purchase agreement pricing. Grid equipment makers and turbine original equipment manufacturers sit on multi-year backlogs. The investable angle is broad, but the risk is regulatory; utility commissions decide who pays for grid upgrades, and pressure to shield residential ratepayers can compress utility returns.

How to Position for the AI Infrastructure Buildout

• Diversify Across the Supply Chain: Rather than concentrating exposure in Mag 7 names, spread investment across semiconductors, power infrastructure, data center REITs, and cooling and networking suppliers to capture the full flow of the $500 billion annual capex cycle.
• Monitor Silicon Replacement Cycles: Track whether NVIDIA's annual release cadence continues to pressure operators to replace hardware faster than traditional depreciation schedules suggest, as this is the single most influential variable in the Goldman framework.
• Watch Power Interconnection Queues: Physical bottlenecks in power, permitting, and specialized labor can elongate deployment timelines and introduce demand-side doubt; monitor whether supply-side friction begins to defer or downsize investment plans.
• Follow Hyperscaler Nuclear Commitments: Track announcements from Meta, Google, Microsoft, and Amazon regarding nuclear power agreements and small modular reactor partnerships, as these signal confidence in long-term AI infrastructure deployment.
• Assess Data Center Cost Inflation: Rising costs per megawatt for AI-era facilities directly push the aggregate $7.6 trillion projection upward; monitor whether architectural shifts and cooling requirements continue to increase construction costs.

Industrial Info Resources data show developers and nuclear plant operators have scheduled approximately 111 capital projects for the nuclear sector with an aggregate value of about $211 billion. This represents a dramatic shift from the previous nuclear renaissance attempt, which ended with more than 40 capital projects valued at roughly $250 billion being cancelled or placed on hold. The only announced project completed was the addition of two new units to the Alvin W. Vogtle Nuclear Power Station, delivered years late and at a cost of approximately $35 billion.

The difference this time is structural. Data center hyperscalers and the Trump administration provide two powerful allies that the nuclear industry lacked at the start of the 21st century. Hyperscalers are short-term capex machines and long-term toll collectors; the capex flows to semiconductor suppliers today, but the ongoing AI-as-a-service revenue accrues to whoever owns the deployment layer, and the deployment layer is the hyperscaler stack. This creates a powerful incentive for hyperscalers to secure reliable, long-term power supplies through nuclear partnerships rather than relying on grid interconnection alone.

The $7.6 trillion projection is not a fixed forecast; it is a baseline that depends on four specific assumptions about how infrastructure is built and renewed. The market is trading the headline number, but the opportunity lies in trading the assumptions behind it. The real risk is not overinvestment in AI infrastructure; it is mispricing the power, cooling, and grid capacity constraints that determine whether the $7.6 trillion can actually be deployed at the projected pace.