Power Grids Crack Under AI Server Demand, Forcing Companies to Choose Between Growth or Blackouts

The numbers are not abstract. A standard enterprise server rack draws between 5 and 10 kilowatts. An AI-optimized rack running GPU clusters pulls between 40 and 100 kilowatts or more, according to analysis published by Entrepreneur on June 19, 2026 . That is not a marginal increase. It is a 4x to 20x step change in power density per unit of deployed compute, applied across every hyperscaler, every regional cloud provider, and every enterprise that has rushed to deploy large language models since 2023. The global grid was never designed to absorb this load curve.

Gregory Brew, senior analyst for Iran and energy at the Eurasia Group, framed the geopolitical overlay in terms that investors cannot ignore: "The Iranians have considerable leverage here. I don't see them backing down and, honestly, time is probably on their side." Brew was speaking to Fortune on July 17, 2026, in the context of a conflict now entering its fifth month, a collapsed 60-day ceasefire, and oil prices that surged back above $88 per barrel after briefly touching $68 at the start of July .

For any business that relies on cloud-hosted AI services, these are not separate stories. They are one compounding risk: electricity demand from AI is growing faster than grid capacity can respond, and the energy commodity that powers the gap between current generation and future build-out is subject to a geopolitical choke point that the U.S. Strategic Petroleum Reserve, now at a 43-year low, cannot adequately buffer .

Dan Pickering, founder of Pickering Energy Partners, put the macro framing directly: "All of the signs point to higher prices and a longer duration. We're in the fifth month of this. We have fewer strategic reserves. We have less flexibility, less optionality. It's a more precarious starting point for round two."

AI Compute Density Is Rewriting the Economics of Power Procurement

The power consumption gap between legacy and AI infrastructure is the foundational fact every CFO needs on their desk. The Entrepreneur analysis is precise: standard racks at 5 to 10 kilowatts versus AI-optimized GPU racks at 40 to 100 kilowatts or more . Scale that differential across a 50-megawatt data center campus and the implied power procurement requirement shifts from a budget line item to a capital structure decision.

Our view: this is not a technology cost. It is a real estate and utilities cost dressed in a technology budget. The implication for institutional capital is that data center REITs, power purchase agreement originators, and grid infrastructure operators are now structurally embedded in the AI trade, whether or not they appear in any AI-themed index.

The Entrepreneur piece makes a point that most enterprise technology buyers have not internalized: AI is not software. Behind every AI-powered workflow is a physical machine running at sustained high-intensity load, often 24 hours per day . The capex and opex consequences of that physical reality flow downstream to every business that buys AI services via the cloud. Pricing pressure and reliability risk are not hypothetical. They are already present in the supply chain.

The Strait of Hormuz Variable Is the Risk No AI Model Has Priced

The global oil benchmark reached $124 per barrel in early May 2026, fell to $68 at the start of July following a temporary truce, and has since recovered above $88 per barrel as of July 17 . That is a 29% rally in under three weeks. The Strait of Hormuz, through which nearly 20% of the world's energy flows pass, remains the controlling variable .

The relevance to the AI energy crisis is direct. Data center construction pipelines depend on natural gas peaker plants to bridge renewable intermittency. Industrial electricity rates in high-compute markets correlate with natural gas spot prices. When oil and gas prices spike due to Persian Gulf disruption, the cost per kilowatt-hour at the grid edge rises with them. Hyperscalers with long-term power purchase agreements are partially insulated. Enterprises buying compute at spot prices through cloud contracts are not.

Brew told Fortune that the Persian Gulf is unlikely to return to a free flow of energy and trade regardless of how the current conflict resolves . That is not a tail risk. That is the base case from a named Eurasia Group analyst with a stated position. Institutional capital allocating to AI infrastructure on the assumption of normalized energy costs is allocating against that base case without knowing it.

Supply Chain Bottlenecks Are Not Confined to Semiconductors

The Entrepreneur analysis identifies a second-order transmission mechanism that receives less attention than chip shortages: energy costs and supply chain constraints flow downstream to any business relying on cloud-hosted AI services, creating direct pricing pressure and reliability risk .

This is the operational risk that does not appear in most enterprise technology due diligence frameworks. A company that has outsourced its AI infrastructure to a hyperscaler has also outsourced its exposure to power procurement costs, cooling water constraints, and grid reliability in the markets where those hyperscalers concentrate their compute. The physical concentration risk is real and growing.

The implication for PE portfolio companies is specific: portfolio businesses that have built AI-dependent workflows without securing contractual protections against cloud pricing escalation are carrying unmanaged operating cost risk. As power costs rise, hyperscalers have both the contractual right and the commercial incentive to pass those costs through. The businesses that understand the full picture, digital and physical together, will make sharper investment decisions and carry less unmanaged risk .

Data Broker Infrastructure: A Parallel Demand Signal Worth Watching

A separate data point from the current news cycle reinforces the scale of institutional compute demand. ICE is renewing its contract with a Thomson Reuters subsidiary at up to $25 million per year for up to five years, a contract that includes continuous monitoring of up to one million individuals and entities with event-driven monitoring, real-time alerts, and model-based risk scoring . The previous equivalent contract was worth $24 million total over a five-year period .

The math: a previous five-year contract at $24 million total versus a new structure at up to $25 million per year represents a potential 5x increase in annualized contract value. Terms and final scope were not disclosed in the contract register document cited by Wired , but the directional signal is clear: government demand for real-time, AI-assisted data processing at scale is accelerating, not plateauing.

Our view: government AI workloads are not typically included in hyperscaler demand forecasts that analysts use to model data center capex cycles. They should be. Federal agencies running continuous monitoring platforms against populations of one million or more individuals represent a persistent, non-cyclical baseload demand signal for compute and therefore for power.

The Plocamium View

The market is pricing the AI energy crisis as a data center construction problem with a five-to-seven year resolution timeline. Plocamium reads it as a two-layer risk with a much shorter fuse.

Layer one is the structural mismatch between AI compute density and existing grid capacity. The 4x to 20x power differential between legacy and AI racks is not a trend. It is an installed fact that compounds with every new GPU cluster that goes live . Grid operators in the highest-density AI markets, Northern Virginia, Singapore, the Netherlands, are already rationing new interconnection. That rationing creates a supply constraint on AI capacity expansion that has no software fix.

Layer two is the geopolitical overlay. A Strait of Hormuz that Gregory Brew characterizes as unlikely to return to free flow is a structural shift in the cost baseline for natural gas-dependent power generation. The two layers interact: AI demand pushes power consumption up; Persian Gulf disruption pushes energy commodity costs up. The product of those two vectors is a cost escalation that hits data center operators, cloud pricing teams, and ultimately enterprise AI budgets simultaneously.

The second-order play that the source material does not make: this creates an investable wedge between operators with long-dated, fixed-price power purchase agreements and those buying power at market rates. The former are structurally protected against the cost vector described above. The latter are not. In a PE context, the quality of a data center asset's power contracts should now rank alongside location and connectivity as a primary underwriting criterion, not a due diligence footnote.

The precedent is the 2021 to 2022 natural gas price shock in Europe, which exposed industrial operators with unhedged energy exposure to margin compression that no operational improvement could offset. AI infrastructure operators without locked power costs are running the same unhedged book against a more concentrated demand shock.

Businesses that understand the full picture, digital and physical together, will make sharper investment decisions, carry less unmanaged risk, and build infrastructure that scales without breaking . The ones that do not will discover their AI budgets are actually energy budgets when it is too late to restructure the contracts.

The Bottom Line

The AI energy crisis is not a future problem. It is a present operating cost embedded in every enterprise AI deployment. The Strait of Hormuz conflict, now in its fifth month with emergency petroleum reserves at a 43-year low , has introduced a supply shock into the energy cost stack that AI's physical infrastructure cannot avoid. Businesses and investors who separate AI strategy from energy strategy are making a category error. The power bill is the AI bill. The forward claim from Plocamium: the next 18 months will produce a visible bifurcation in AI infrastructure economics between operators with contracted power certainty and those without. That bifurcation will be the defining valuation spread in the data center asset class by 2027.