The number circulates with the confidence of a compiled binary: 38 gigawatts. Morgan Stanley's projection of the AI data center power gap by 2028 has become the industry's favorite citation, a tidy figure that justifies everything from nuclear energy stock purchases to grid upgrade budgets. But as with most tidy figures in this sector, the methodology behind it remains opaque. Code compiles, but context reveals the exploit. And the context here is a structural mismatch between exponential compute demand and the physical reality of power generation that no amount of narrative engineering can patch.
Let me be precise about what we actually know. The 38 GW figure represents the projected shortfall between AI data center electricity demand and available supply, primarily in the United States. The underlying assumption is that AI compute demand continues its current growth trajectory - roughly 50% annual growth in GPU shipments, with each generation of accelerators consuming more power per unit. The H100 runs at 700W. The B200 exceeds 1000W. The trend line is unambiguous, and it points toward a constraint that no software optimization can fully resolve.
I have spent the better part of a decade auditing infrastructure claims in this industry, from ICO whitepapers to DeFi yield models. The pattern repeats: a compelling narrative, a lack of verifiable methodology, and a market that prices in certainty where none exists. The 38 GW figure deserves the same scrutiny I applied to EtherGem's voting mechanism in 2017 - the same scrutiny that revealed arithmetic overflow vulnerabilities the team ignored until the rug pull made them irrelevant.
The structural problem is not compute. It is physics.
Consider the arithmetic. A single H100 GPU at full load draws 700W. A data center housing 100,000 such GPUs requires 70 megawatts for the IT equipment alone. Apply a PUE of 1.3 - the industry average - and the facility demands 91 megawatts from the grid. Now multiply across the global AI buildout: hyperscale facilities under construction by Microsoft, Amazon, Google, and Meta, plus the emerging sovereign AI projects in the Middle East and Southeast Asia. The cumulative demand curve outstrips grid expansion timelines by a factor that no transformer order backlog can close.
The transformer supply chain tells the story more honestly than any analyst projection. Global lead times for power transformers have stretched from 40 weeks in 2020 to over 120 weeks in 2024. This is not a demand signal; it is a bottleneck signal. The equipment required to deliver power to AI data centers cannot be manufactured fast enough, regardless of how much capital flows into the sector. The constraint is physical, not financial.
The energy arms race is already reshaping competitive dynamics.
Microsoft signed a nuclear power agreement with Constellation Energy. Amazon became the largest corporate purchaser of renewable energy globally. Google committed to 24/7 carbon-free energy by 2030. Meta invested $800 million in an Ohio data center with dedicated natural gas generation. These are not ESG gestures; they are survival strategies. The cloud providers understand that power supply is now the binding constraint on their AI ambitions, and they are securing it through long-term contracts, direct investment, and vertical integration.
The implications for the rest of the market are severe. AI startups without energy partnerships face rising compute costs and supply uncertainty. The cost of electricity represents 20-40% of data center operating expenses. A 30% increase in power prices translates to a 5-8% increase in inference costs for GPT-4-class models. For companies operating on thin margins, this is not a rounding error; it is a competitive disadvantage that compounds over time.
The regional distribution problem is the hidden variable.
The 38 GW figure obscures a critical detail: the gap is not evenly distributed. The United States, Europe, and Southeast Asia face acute shortages. The Middle East has surplus power but lacks AI infrastructure. Africa has neither. This geographic mismatch will drive a reconfiguration of data center placement, with power availability replacing network latency as the primary siting criterion. The 'East-to-West data computing' model in China - moving compute to renewable-rich western provinces - is a template that will be replicated globally.
I built a SQL dashboard in 2020 to track Aave's yield sustainability against treasury reserves. The data showed the yields were debt traps, not organic growth. The same analytical discipline applies here: the 38 GW figure is a projection, not a measurement. It assumes current trends continue, but it does not account for inference efficiency gains from quantization, model distillation, or speculative sampling. It does not fully price the PUE improvements from liquid cooling, which can reduce facility overhead from 1.4 to below 1.1. It does not model the potential for chip architecture innovation - photonic computing, neuromorphic designs - that could fundamentally alter the power curve.
The contrarian angle: the bulls may be right about the direction, wrong about the magnitude.
The energy sector is genuinely benefiting. Power equipment manufacturers - Schneider Electric, Eaton, Vertiv - have order backlogs that validate the demand thesis. Nuclear energy providers like Constellation and Vistra have seen their valuations re-rate on AI power demand expectations. These are real, verifiable signals. The question is whether the 38 GW figure represents a floor or a ceiling. If efficiency improvements and renewable deployment outpace the projection, the gap narrows. If AI adoption accelerates beyond current models - particularly in agentic AI and multimodal inference - the gap widens.
My 2021 forensic analysis of Bored Ape Yacht Club floor prices traced 15% of weekly volume to wash trading clusters. The apparent market cap was inflated by at least $40 million in artificial volume. The subsequent correction wiped out 90% of speculative value. The parallel here is uncomfortable: how much of the current AI infrastructure investment is priced on verified demand, and how much is narrative momentum? The energy companies have real contracts. The AI companies have real compute needs. But the gap between projection and reality is where capital gets destroyed.
The regulatory dimension cannot be ignored.
The EU Energy Efficiency Directive requires data centers to disclose energy consumption and carbon emissions starting in 2025. China's East-to-West Computing project mandates a minimum 30% renewable energy share for data centers. US states like Virginia are reviewing data center power consumption and environmental impact. These regulatory pressures will constrain the pace of AI infrastructure buildout, potentially narrowing the gap - or shifting it to jurisdictions with weaker oversight.
Microsoft's carbon emissions increased approximately 30% since 2020, driven primarily by data center construction. This is the uncomfortable truth: AI expansion and carbon neutrality targets are in direct conflict. The 38 GW gap, if filled by natural gas, would add significant emissions. If filled by renewables, it requires storage solutions that are not yet commercially viable at scale. If filled by nuclear, it requires a decade of regulatory approval and construction. There is no easy answer, and the market is pricing in the absence of one.
The takeaway is not about the number. It is about the accountability.
Morgan Stanley's 38 GW projection is a useful starting point, not a conclusion. The methodology is undisclosed. The regional distribution is unspecified. The time profile is unclear. What is clear is that power has become the binding constraint on AI infrastructure, and the companies that secure reliable, cost-effective electricity will dominate the next phase of the industry. The companies that treat power as an afterthought will find themselves priced out of the market.
I have seen this pattern before. In 2022, I audited Frax Finance's partial collateralization model against Terra's algorithmic failure. The comparative analysis showed that reliance on market confidence rather than hard assets was a systemic risk. The same logic applies to AI infrastructure: reliance on grid power without dedicated supply agreements is a systemic risk. The 38 GW figure is a warning, not a prediction. The question is whether the industry treats it as such, or waits for the blackout to confirm the analysis.
Disillusionment is the price of entry. The data is available. The analysis is straightforward. The only variable is whether the market chooses to verify before it trusts.
