AI Power Shortage Dilemma: Solutions and Investment Targets

I. The Bottleneck in AI Infrastructure Is Shifting from GPUs to Power

Over the past two years, market discussions on AI infrastructure have primarily focused on three things:

First, who has the GPUs.
Second, who has the models.
Third, who has the customers.

But entering 2026, a more fundamental issue is beginning to surface:

Even if you've bought the GPUs and built the data centers, where does the power come from?

This is not an abstract narrative; it's a real constraint now entering corporate orders, capital expenditures, and project delays.

A DOE/LBNL report shows that U.S. data center electricity consumption grew from 58 TWh in 2014 to 176 TWh in 2023, and is projected to potentially reach 325-580 TWh by 2028. IEA's "Energy and AI" report further projects that global data center electricity consumption will grow from about 415 TWh in 2024 to about 945 TWh by 2030.

In other words, AI data centers are pulling up the power demand curve that had seen moderate growth for many years.

This also explains why companies like Oracle, Google, AWS, Meta, Microsoft, and OpenAI are no longer just buying servers; they are starting to lock down power generation, land, substations, gas turbines, long-term nuclear power purchase agreements, energy storage, and on-site power sources.

AI infrastructure is transitioning from a "computing power procurement cycle" to a "power-first infrastructure cycle."

II. The Real Shortage Isn't Total Energy, But Power-On Speed

The U.S. isn't lacking in energy sources, nor will it lack generation capacity in the long term.

The real shortage is:

During the 2026-2030 window, who can turn MW into data-center-usable power the fastest, most reliably, in a financeable and permit-able way.

This is the time-to-power.

For an AI data center, waiting four years for grid connection is almost the same as having no power at all. Because GPUs, server rooms, cloud contracts, and customer demand won't wait.

This is also why we can't use traditional single metrics from the power industry to understand this cycle. The lowest cost per kilowatt-hour is certainly important, but in AI infrastructure, time itself is a cost.

If a 1GW-scale AI data center is delayed by a year, the loss isn't a few tens of millions in electricity cost differences; it could be billions in lost cloud revenue, computing power market share, and model iteration speed.

NVIDIA articulated this logic directly in an earnings call: AI factories are essentially revenue machines; computing power converts to tokens, and tokens convert to revenue. Each data center is power-constrained, so performance per watt determines how much revenue a customer can generate from a fixed MW. Oracle's stance is similar: AI infrastructure demand continues to outstrip supply, and it needs to lock down over 10GW of power and data center capacity for the next three years in advance. Google treats computing power allocation as a high-frequency management task, because Search, Gemini, AI Overviews, AI Mode, and Cloud backlog are all competing for the same pool of computing power.

Therefore, for cloud providers, a power shortage isn't a matter of "slightly higher costs," but of "delayed revenue, a generation-late model lineup, and losing a step with customers." AI competition isn't about adding capacity after demand arrives; it's about who can first organize power, facilities, networks, cooling, and chips into deliverable computing power.

Thus, the main investment theme of this cycle shouldn't be simplified to "buy renewables" or "buy natural gas," but should focus on five types of scarce capabilities:

Scarce Capability	Representative Path	Representative Investment Targets
Fast On-Site Power	Fuel cells, gas engines, aeroderivative gas turbines	`BE`, `GEV`, `CAT`
GW-Scale Stable Generation	Combined Cycle Gas Turbines (CCGT), Gas Turbines, Nuclear, Independent Power Producers (IPP)	`GEV`, `CEG`, `VST`, `NRG`, `TLN`
Grid Connection & Distribution	Transformers, switchgear, High-Voltage Direct Current (HVDC), Uninterruptible Power Supplies (UPS)	`ETN`, `VRT`, `PWR`, `GEV`, Siemens Energy
Power Dispatch & Buffering	Energy Storage, Microgrids, Energy Management Systems (EMS), Control Systems	`FLNC`, `TSLA`, `GEV`
Pre-Locked Power & Sites	Mining facility conversions, existing high-voltage connections, campus development	`HUT`, `WULF`, `CORZ`, `IREN`, `CIFR`

These five lines are not mutually exclusive. A large AI data center campus might use gas turbines for baseload, substations for grid connection, energy storage for load buffering, fuel cells for phased rapid deployment, and might also directly acquire or lease high-voltage connection sites originally serving Bitcoin mining.

III. How the Giants Are Voting with Their Feet

Analyzing the latest news and announcements, cloud providers' stance on the power issue is already clear:

Oracle: The Most Aggressive "Bring Your Own Power" Route

Oracle is one of the clearest case studies for the on-site power route.

Oracle expanded its strategic partnership with Bloom Energy, planning to deploy up to 2.8GW of fuel cell systems, with an initial 1.2GW already contracted and under deployment. Oracle's logic isn't buying the cheapest power, but buying certainty: rapid deployment, low pollution, low water consumption, and reduced reliance on grid queueing.

Oracle's earnings call also indicates it has locked down over 10GW of power and data center capacity for the next three years. For Oracle, AI infrastructure isn't about building the facility first and waiting for power; it's about planning data centers and generation capacity together.

Google: Buying Energy Development Capability Directly

Alphabet's acquisition of Intersect essentially brings energy development capability in-house. Google is no longer just signing long-term Power Purchase Agreements (PPAs); it's directly controlling multi-GW-scale energy and data center project development capability.

This indicates that energy has shifted from a procurement item to a strategic asset. Just as Google developed its own TPUs to avoid being locked into a single supplier for AI chips, Google also doesn't want to be held back by grid queues, energy developers, or project permitting for power.

Google's moves can be broken down into several layers:

Layer 1: Continue increasing capital expenditures. Alphabet's Q1 2026 earnings call raised the 2026 CapEx guidance to $180-190 billion and explicitly stated 2027 would be significantly higher than 2026.
Layer 2: Buy energy development capability directly. The Intersect transaction was $4.75 billion in cash, with the core value being multiple GW-scale energy and data center projects, and a development team to push data centers and generation capacity online faster.
Layer 3: Co-locate data centers and power sources. The Haskell County, Texas project embodies "data center + dedicated power" co-development, rather than the traditional approach of waiting for utility power.
Layer 4: Maintain multiple technology pathways. Google is simultaneously focused on gas, renewables, energy storage, geothermal, long-duration storage, gas with carbon capture, and advanced nuclear projects like Kairos.
Layer 5: Continue using chips and software to reduce power consumption per unit of computing power. TPU in-house development, model efficiency optimization, and AI-assisted grid integration all aim to reduce the power consumed per token.

This set of actions shows that Google's strategy isn't "buying one type of power," but integrating energy development, chip efficiency, data center siting, and grid coordination into a single infrastructure system.

AWS: Paying in Advance for Land, Power, and Construction

AWS management explicitly mentioned that in generative AI-related capital expenditures, land, power, construction, chips, servers, and networking equipment all require investment 6-24 months in advance.

AWS's partnership with Siemens Energy also indicates it is exploring GW-scale generation, microgrids, backup power, and turnkey substation solutions.

Microsoft and Meta: Nuclear Power PPAs as the Long-Term Answer

Microsoft supports the restart of Three Mile Island Unit 1, and Meta has signed long-term agreements with multiple nuclear projects and generation assets. Their commonality: for 24/7 clean firm power, they are willing to sign multi-year, even multi-decade agreements.

However, nuclear power has longer cycles and is more of a mid-to-long-term solution for post-2028 and even the 2030s.

OpenAI / Stargate: Power, Land, Permitting, and Finance Are an Integrated Problem

OpenAI's Stargate project illustrates that AI infrastructure is no longer a single data center problem, but a systems engineering challenge involving power, land, permitting, transmission, workforce, community support, and partner readiness.

OpenAI and SoftBank each invested $500 million in SB Energy, OpenAI signed a 1.2GW data center lease agreement, and SB Energy will build and operate the Milam County data center and related energy infrastructure.

This sends a dual signal for the entire power supply chain:

On one hand, the AI power shortage is real.
On the other hand, major customers don't necessarily just buy one type of equipment; they prefer the total package of "Power + Data Center + Financing + Delivery." Bloom Energy's fuel cells are one approach, but GEV, Siemens, Independent Power Producers (IPPs), utilities, mining sites, and energy developers are all competing for this value chain.

IV. Seven Main Paths to Address the Power Shortage

1. Wait for the Grid: Most Traditional, but Slowest

The most direct way is, of course, to wait for utility expansion and grid connection.

The problem is, the grid isn't software; you can't just hire more engineers to instantly scale it up. Transmission lines, transformers, switchgear, permits, and interconnection queues all take time, easily stretching a single project's timeline to several years.

LBNL's interconnection queue data shows that the US has a massive backlog of generation and storage projects, but whether they connect on schedule is another matter.

The primary beneficiaries of this path are not single generation technologies, but grid equipment and the engineering chain:

ETN, PWR, GEV, Siemens Energy, HUBB, POWL, VRT.

2. Gas Turbines: The Most Realistic GW-scale Mid-term Answer

If the goal is to provide 24/7 power for large data centers from 2026-2030, gas turbines remain the most realistic mainstay solution.

GE Vernova's Q1 2026 data is very telling:

Total orders: $18.3 billion, up 71% year-over-year;
Total backlog: $163 billion;
Gas turbine backlog and slot reservation agreements reached 100GW;
Management expects most slots to be sold before 2030;
Data center-related electrification equipment orders for the quarter: $2.4 billion, exceeding the full year 2025.

Siemens Energy Q2 FY2026 also gives a similar signal:

Orders: €17.7 billion, a record high;
Backlog: €154 billion;
Gas Services orders: €8.87 billion, with a book-to-bill ratio of 2.55;
US demand is primarily driven by data centers, with very favorable pricing conditions;
H1 data center-related revenue exceeded €1.8 billion, up over 45% year-over-year.

These numbers show that gas turbines are not obsolete assets, but are regaining pricing power within the AI data center cycle.

However, their issues are also clear: tight capacity slots, lengthening lead times, air permits, natural gas pipelines, carbon emission pressures, and community acceptance.

So gas turbines are a mid-term, large-scale solution, but not the fastest for all scenarios.

3. On-site Fuel Cells: One of the Fastest Power-on Tools

Bloom Energy's core value isn't "cheapest power," but "fastest power."

Bloom's SOFC fuel cells can be modularly deployed on-site at data centers, featuring low NOx, low noise, low water usage, small footprint, and phased deployment. These characteristics directly address the pain points faced by AI data centers in certain regions.

Oracle's up to 2.8GW partnership and Nebius's 328MW project show that Bloom's solution is not just a paper story.

But we must be clear-eyed:

Bloom's advantage is a window-of-opportunity advantage, not necessarily a final, monopolistic advantage.

When GE Vernova, Caterpillar, Siemens Energy, Crusoe, SB Energy, IPPs, utilities, and infrastructure funds all start offering the "Power + Data Center + Financing + Delivery" package, Bloom will face two types of pressure:

First, price pressure.
Second, diminished platform premium.

Bloom can continue to grow, but it should be understood as part of the "rapid on-site power supply toolkit," not as the sole answer to the AI power problem.

4. Nuclear Power: Highest Quality Long-term, but Long Cycle

Nuclear power is one of the power sources best aligned with the long-term preferences of hyperscalers: 24/7, low-carbon, baseload, and suitable for long-term Power Purchase Agreements (PPAs).

Companies like Constellation, Vistra, and Talen, which own nuclear or high-quality generation assets, are thus being revalued.

But nuclear's drawbacks are equally clear: approval, construction, and grid connection cycles are all long.

In the US, new nuclear projects typically must pass through multiple gates: federal nuclear regulatory approval, environmental assessment, state-level cost recovery arrangements, local permits, transmission access, fuel supply, EPC contracts, etc. Even if a project goes smoothly, the timeline from initiation to commercial operation is often 7-10 years or longer, with cost overruns and delays being recurring risks in nuclear investment.

Therefore, the more realistic nuclear path in the near term is not "immediately building a bunch of new plants," but three types: First, restarting existing but idled nuclear units, like Microsoft supporting the restart of Three Mile Island Unit 1; Second, signing long-term PPAs with existing nuclear plants to lock in 24/7 clean baseload power; Third, betting on Small Modular Reactors (SMRs) or advanced nuclear in the 2030s.

Restarting old nuclear and securing PPAs with existing nuclear plants are more realistic than building new ones. SMRs are an important direction, but more like an option for the 2030s.

Thus, nuclear power is suitable for expressing the scarcity of long-term 24/7 clean stable power, not for solving urgent power needs in 2026.

5. Energy Storage: Not Baseload, but Will Become a System Stabilizer

Energy storage is not generation. It cannot create power out of thin air.

But in AI data centers, the importance of storage will rise significantly because AI training and inference loads create rapid fluctuations, requiring millisecond-level response and power buffering.

Storage's role is not to replace gas turbines, nuclear, or fuel cells, but to make the entire system more stable:

Smooth load fluctuations;
Support microgrids;
Reduce backup power pressure;
Integrate with renewables;
Provide short-term transition power.

The corresponding tickers are FLNC, TSLA, ENR, etc. But the issue with the storage chain is low margins, strong competition, and volatile revenue recognition, making it more suitable as a satellite position than a core holding.

6. Energy Efficiency & Load Scheduling: The Most Underestimated "Invisible Power Source"

NVIDIA repeatedly emphasizes performance per watt in its calls. For cloud providers, how much token revenue each watt generates is already a core economic metric.

Google's TPU, AWS's Trainium, and NVIDIA's Blackwell are essentially doing one thing:

Convert each MW of power into more compute revenue.

Improving energy efficiency won't eliminate power demand. The Jevons effect might even increase total demand. But in situations of local power constraints, efficiency itself becomes a form of "invisible power."

This line corresponds not to traditional power stocks, but to the compute and data center infrastructure chain: NVDA, AVGO, ANET, VRT, ETN, etc.

7. Miner Collaboration: Converting Existing Power & Sites into AI Data Centers

Another easily overlooked path: partnering with Bitcoin miners, or acquiring/leasing their existing high-voltage connections, power contracts, land, and facility infrastructure.

The core assets of miners in the past were not the mining rigs, but:

Already secured or in-progress power capacity;
Land near power sources or transmission nodes;
High-voltage connection, substation, and cooling infrastructure;
Operational teams familiar with power markets and load scheduling;
Facilities capable of switching from crypto mining to high-performance computing.

The logic of this path is: Instead of queuing up again for power, buy or lease sites that already have power.

Several cases are already clear:

CoreWeave attempted to acquire Core Scientific, with the strategic rationale being to gain ~1.2GW of existing total power capacity and over 1GW+ expansion potential. The deal was later terminated due to Core Scientific shareholder disapproval, but it shows AI cloud providers are willing to pay a strategic premium for miner power sites.
TeraWulf positioned high-performance computing colocation as a primary growth engine in 2025; Lake Mariner disclosed a 200MW+ long-term AI colocation agreement, and Abernathy disclosed a 168MW HPC joint venture project, with Fluidstack as a core customer and Google providing credit support.
Hut 8 partnered with Anthropic and Fluidstack, planning to deliver at least 245MW, with a long-term potential of up to 2,295MW, of AI data center infrastructure at sites like River Bend; in 2026, it announced a 15-year lease for the 352MW IT capacity Phase 1 at Beacon Point.

This path is not a generation technology path, but a "power site arbitrage" path. It addresses power-on speed and site scarcity.

Corresponding tickers include HUT, WULF, CORZ, IREN, CIFR, etc. But these companies also carry higher risks: high volatility in original business, strong financing dependence, complex project conversion, high customer concentration, and valuations prone to large swings with AI sentiment.

V. Stock Selection: Don't Bet on Just One Technology

If we stratify this round of AI power investment, it can be divided into five layers.

Layer 1: The Highest Certainty "Pick-and-Shovel" Plays

This layer does not bet on whether data centers will ultimately use gas, nuclear power, fuel cells, or energy storage. As long as data centers are being built, they will need to buy their products.

GE Vernova (GEV)

GEV is one of the most core stocks in this theme.

It has exposure to gas turbines, electrification equipment, transformers, high-voltage direct current (HVDC) transmission, and energy management systems (EMS). The 100GW backlog and slot reservations for gas turbines in Q1 2026, along with $2.4 billion in data center electrification orders, indicate it benefits from both the generation side and the grid side.

The risk is that its valuation is no longer low. As of 2026-06-01, GEV's stock price is approximately $968, with a market cap of about $263.4 billion and a P/E ratio of around 28x. This valuation is not cheap, but compared to the visibility of its order book, it remains more solid than many pure concept stocks.

Eaton (ETN)

Eaton is a stable representative of electrification and power distribution equipment. It doesn't need to judge which generation technology wins, because all routes require switchgear, power distribution, circuit breakers, UPS, and electrical systems.

As of 2026-06-01, ETN's stock price is approximately $401, with a market cap of about $155.9 billion and a P/E ratio of around 39x. Its elasticity is lower than high-beta stocks, but its certainty is stronger.

Vertiv (VRT)

Vertiv directly benefits from the power management and thermal management demands driven by the high power density of AI data centers.

The problem is its valuation is expensive. As of 2026-06-01, VRT's stock price is approximately $316, with a market cap of about $123.8 billion and a P/E ratio of around 79x. It's a good company, but not a cheap stock.

Quanta Services (PWR)

PWR represents the grid engineering and construction chain. Transmission, substation, and energy engineering are not short-cycle; they will accompany AI data center expansion for many years.

However, as of 2026-06-01, PWR's P/E ratio is close to 98x, and its valuation already reflects a lot of optimistic expectations. It's suitable as a long-term watchlist candidate, requiring a more stringent entry point.

Layer 2: Revaluation of Generation Assets

This layer bets on 24/7 dispatchable power assets themselves becoming scarcer.

Constellation (CEG)

CEG is one of the core stocks for nuclear asset revaluation. Microsoft's support for the Three Mile Island restart logic indicates that high-quality nuclear assets are being repriced by cloud providers.

Vistra (VST)

VST owns nuclear and gas assets, making it more like a comprehensive expression of "dispatchable power asset revaluation." Compared to pure nuclear plays, it benefits more from capacity prices and tight power markets.

Talen (TLN)

TLN has high elasticity but also high regulatory risk. Direct connections between data centers and power plants raise issues with FERC, state regulation, and cost allocation. Suitable as a satellite position for investors with high risk tolerance.

NRG (NRG)

NRG, along with GE Vernova and Kiewit, is advancing over 5GW of new CCGT projects, indicating it is transitioning towards an AI power supply platform. However, project commercialization is more oriented towards 2029 and beyond; in the short term, it relies more on the market pricing of future projects.

Layer 3: High-Elasticity On-Site Power

Bloom Energy (BE)

BE is one of the purest and most aggressive high-elasticity stocks in the rapid on-site power supply route.

Its advantages are very clear:

Up to 2.8GW partnership with Oracle;
Initial 1.2GW already signed for deployment;
328MW with Nebius validates demand from new cloud providers (neocloud);
Scarcity of rapid on-site power-up capability;
Low NOx, low water consumption, low noise suitable for constrained areas;
Q1 2026 revenue significantly up year-over-year, full-year revenue guidance raised.

But its risks are equally clear:

Current stock price already prices in a bull case scenario;
5GW of existing plant expansion footprint does not equal stable delivery of 5GW;
Product gross margin, installation gross margin, and service gross margin all need continuous validation;
High customer concentration with Oracle;
GE/CAT/Siemens/IPP/SB Energy may suppress its platform valuation;
Energy-as-a-Service (EaaS) model may mean revenue growth does not equal synchronized free cash flow growth.

As of 2026-06-01, BE's stock price is approximately $285, with a market cap of about $91.1 billion and a negative P/E ratio. This price is not pricing "real AI power shortage," but rather pricing "Bloom can convert the power shortage into 3.5-4.5GW shipments by 2028 and maintain relatively high gross margins."

Therefore, the strategy for BE is not to negate it, but to avoid chasing highs.

Fluence Energy (FLNC)

FLNC is a high-elasticity stock in energy storage and power control systems. It is not a baseload power source, but a stabilizer within microgrids and AI load fluctuations.

As of 2026-06-01, FLNC's stock price is approximately $18.88, with a market cap of about $2.5 billion and a negative P/E ratio. Its advantage is high elasticity; its problem is that profitability quality and gross margins still need validation.

Layer 4: Miner and Power Site Transformation

This layer bets not on a specific generation technology, but on "who already controls the power sites."

Bitcoin miners, in the past, secured a large amount of low-cost power, high-voltage connections, and remote land in advance for mining. With the rise of AI data centers, these sites suddenly have a new use: replacing low-value mining loads with high-value AI training and inference loads.

Representative stocks in this layer include:

Stock	Logic	Risk
`HUT`	Hut 8 is transitioning from a miner to a power, digital infrastructure, and AI compute platform; Beacon Point and River Bend are core examples	Project financing, delivery, customer concentration
`WULF`	TeraWulf is transitioning from mining to high-performance computing leasing; Lake Mariner and Abernathy have long-term lease agreements	Construction progress, financing structure, customer concentration
`CORZ`	Core Scientific owns data center and power capacity that can pivot to AI and HPC; was once a CoreWeave acquisition target	Independent execution post-failed M&A, valuation volatility
`IREN` / `CIFR`	Own power and site resources, can pivot to AI and HPC	Transformation certainty and customer quality need validation

These types of stocks are suitable for offensive positions or a watchlist, not as stable core positions. Their upside comes from "power site revaluation," and their downside comes from financing, conversion, and customer realization risks.

Layer 5: Long-Term Options

OKLO, SMR, BWXT, the hydrogen chain, long-duration energy storage, etc., all belong to long-term options.

They may be very important, but most cannot solve the urgent power shortage problem from 2026-2028. For investment, they can be watched but should not replace stocks like GEV, ETN, CEG, VST, or BE that are already receiving industrial orders.

VI. The Biggest Variable Lies in How the Power Supply System is Combined

When discussing AI power investment, it's easy to fall into debating which is better.

This question itself is not very accurate, because they occupy different positions in the power supply system.

BE sells "rapid on-site power-up."
GEV sells "GW-scale power generation and electrical infrastructure."
Miners and energy developers sell "already secured power sites."
Nuclear and generation assets sell "long-term stable power."
ETN, VRT, PWR sell "electrical systems and engineering capabilities that no route can bypass."

A 1GW data center could very well first use BE's modular fuel cells to quickly bring a portion of capacity online, then use GEV or Siemens' gas turbines and grid equipment as the long-term main power source, while also using energy storage to smooth loads and deploy some AI loads to former mining power sites.

The real difference lies in investment attributes:

Type	Representative	Suitable Role	Biggest Risk
Large-scale generation & electrification platform	`GEV`, Siemens Energy	Core position, covering gas turbines, grid, and system integration	Valuation, delivery cycles, overly optimistic cyclical expectations
Electrical equipment & engineering chain	`ETN`, `VRT`, `PWR`	Common bottleneck across technology routes	High valuation, slowdown in AI capex
Generation assets	`CEG`, `VST`, `TLN`, `NRG`	Long-term stable power and capacity price revaluation	Regulation, electricity prices, project cycles
Rapid on-site power	`BE`	High-elasticity satellite position, betting on speed to power	Gross margins, delivery, service, valuation overextension
Miner power sites	`HUT`, `WULF`, `CORZ`, `IREN`, `CIFR`	High-elasticity power site revaluation	Financing, conversion, customer concentration, stock price volatility

If one can only choose stable core positions, types like GEV, ETN, CEG are more suitable.
If betting on the highest elasticity, BE and miner transformation stocks are purer, but position sizing should be more restrained.
For a portfolio, the core should be placed in cross-route electrical equipment, generation platforms, and quality generation assets; BE, FLNC, and miner transformation stocks are more suitable as satellites.

VII. A More Practical Portfolio Framework

Conservative Type

Suitable for those who don't want to bet on a single technology route:

GEV + ETN + CEG + PWR

This combination covers gas turbines, grid equipment, nuclear power assets, and engineering & construction. The drawbacks are less flexibility compared to BE, and most are not cheaply valued.

Balanced Type

Suitable for those seeking both certainty and some flexibility:

GEV + ETN + VRT + VST + BE small position + Miner transition small position

Here, GEV/ETN/VRT represent the infrastructure core, VST provides power generation asset flexibility, BE offers high flexibility for rapid power deployment, and miner transition targets provide revaluation flexibility for power sites.

Offensive Type

Suitable for those who can withstand high volatility:

GEV + BE + VRT + FLNC + TLN + HUT/WULF/CORZ

This portfolio offers the highest flexibility but also carries the greatest risk. BE, FLNC, TLN, and miner transition targets are not low-volatility assets; if orders, regulation, financing, or gross margins fall short of expectations, drawdowns can be swift.

VIII. Eleven Key Metrics to Watch in the Future

Whether GEV's gas turbine backlog and slot reservations continue to grow.
Whether Siemens Energy's gas services business order-to-revenue ratio remains above 2.
Whether data center electrification orders continue their high growth.
Whether lead times for transformers, switchgear, and HVDC transmission ease.
Whether BE secures a second GW-scale customer besides Oracle.
Whether BE can maintain product gross margins above 34% and repair installation margins.
Whether BE's planned 2GW annualized capacity by the end of 2026 is delivered on schedule.
Whether AWS, Google, Oracle, Meta, and Microsoft continue to revise their capital expenditure guidance upward.
Whether nuclear power PPAs and direct data center-to-power plant connections face regulatory headwinds.
Whether lease agreements for miner transitions to AI data centers convert to revenue as scheduled.
Whether AI model efficiency improvements genuinely reduce total power demand or trigger even more usage.

The most critical factors are not individual company metrics, but three sets of signals:

First, whether orders for electrical equipment and gas turbines continue to prove that demand has not receded. Second, whether on-site power supply and miner site conversion routes can translate "access to power" into real revenue. Third, whether regulation allows data centers to bypass traditional queues via self-generation, direct plant connections, and dedicated power sites.

IX. Conclusion: The Thesis Holds, But Price Cannot Be Ignored

AI's power shortage is not a short-term theme; it is one of the most critical infrastructure constraints from 2026 to 2030.

However, from an investment perspective, the most dangerous approach is to directly translate this conclusion into "all power stocks are buys."

A more accurate conclusion is:

The AI power shortage is real, but the solution will be a multi-pronged approach; there won't be just one technology or a single winner.

GEV and Siemens Energy have demonstrated the scarcity of gas turbines and electrification equipment. BE has demonstrated the strategic value of rapid on-site power deployment. CEG, VST, and TLN have demonstrated the revaluation of 24/7 power generation assets. ETN, VRT, and PWR have demonstrated that electrical equipment and engineering & construction are common bottlenecks across all pathways. FLNC has demonstrated that energy storage and control systems will be essential complements in the microgrid era.

To put it more conservatively:

The outcome of AI infrastructure depends not only on who has the most GPUs, but also on who can organize land, power, equipment, financing, and permits into deliverable capacity. For investors, power is not a single track, but an infrastructure chain spanning generation, grid integration, distribution, storage, sites, and operations.

References

2024 LBNL Data Center Energy Usage Report — Lawrence Berkeley National Laboratory report, source for the data: "Data center electricity usage: 58 TWh in 2014 → 176 TWh in 2023 → Projected 325-580 TWh in 2028"
IEA: Energy and AI Executive Summary — International Energy Agency report, source for the prediction: Global data center electricity usage ~415 TWh in 2024 → ~945 TWh in 2030
DOE: Data Center Electricity Demand Assessment Report — U.S. Department of Energy press release, confirming AI data center load pressure from a policy perspective
Bloom Energy × Oracle Expand Strategic Partnership (Up to 2.8GW) — Original announcement for Oracle's on-site fuel cell power route, the core basis for the BE case study in the text
GE Vernova 2026 Q1 Earnings Release Materials — Official source for key GEV figures like quarterly orders, gas turbine backlog, and slot reservations