Bold claim: AI’s demand for power is growing so fast that the real bottleneck isn’t clever code, it’s electricity and capacity. But here’s the twist: a short-lived window emerges where investors can profit from the imbalance between surging AI demand and the limits of centralized infrastructure.
By Jurica Dujmovic
The shortage creates a rare opportunity, but the clock is ticking.
AI workloads could gulp roughly 500 terawatt-hours annually by 2027, about twice the United Kingdom’s total electricity use in 2023.
Rising infrastructure costs and tightening capital are tempering the AI rush. Even giants like the hyperscalers can’t keep pace with the need to expand compute, creating a unique arbitrage moment.
The immediate fix isn’t new data centers. The current chance lies in the temporary gap between exploding AI demand and the physical realities of expanding centralized infrastructure. A small set of companies is already capitalizing on this window, which is likely to last about 24 to 36 months. For investors who dial in the timing, this offers a compelling hedge against the inertia of AI infrastructure.
Physical limits
Forty percent of AI data centers will encounter power constraints by 2027.
The limiting factor for AI isn’t algorithms or data; it’s the brute physics of expanding data-center capacity. Training massive models requires tens of thousands of GPUs, dedicated networking, and vast amounts of power. Gartner predicts that 40% of AI data centers will hit power constraints by 2027.
The math is stark: AI computing workloads could reach around 500 terawatt-hours per year by 2027—roughly double the U.K.’s total electricity consumption in 2023. This surge is already showing up on the grid.
Dominion Energy, the largest utility in Virginia, nearly doubled its data-center power capacity under contract between mid-2024 and end-2024, and the trend has continued.
Even with Microsoft, Alphabet, Amazon, and Meta spending a combined $370 billion on capex in 2025, they can’t build fast enough. A full data-center project typically takes 12 to 36 months to construct and commission, but adding permitting and power-grid upgrades can push the timeline to three to six years.
Time and money
The economics look strong during this shortage window.
This temporary gap is the core of the investment thesis.
When essential resources become expensive and concentrated, parallel markets emerge. History provides examples: electricity co-ops in the early 20th century, independent oil producers during OPEC’s era, and broadband resellers in the early internet era.
With AI, the scarce resource is GPU computing. Several firms are building marketplaces that aggregate idle capacity—from consumer GPUs to academic clusters to enterprise overstock—and resell it at a fraction of centralized data-center costs.
The economics for these companies look compelling during the shortage window:
- Cost-structure advantage: Alternative networks don’t rely on debt-financed data centers. They compensate participants directly for computing capacity, turning spare resources into productive assets. Scaling shifts from capital expenditure to distributed incentives.
- Speed to market: Hyperscalers wait 18 to 36 months for new facilities, while these networks can expand node by node, without large upfront commitments.
- Arbitrage pricing: These platforms capture demand from smaller labs, indie studios, and emerging markets priced out of AWS GPU allocations but still needing compute.
The catch: this growth spurt is finite. Even after constraints ease, these networks remain viable as lower-cost options for cost-sensitive workloads and underserved regions, but the big investment gains compress as growth normalizes and hyperscalers come online.
Read: AI data centers need juice. The next hot stocks give it.
How to play the computing shortage
This isn’t a moonshot. It’s an infrastructure hedge with a defined window. Here are three approaches, ranked by risk:
- Render Network: Aggregates idle GPU capacity from individuals and studios, then resells to the highest bidder for rendering and AI tasks. Think of it as Airbnb for GPUs—idle capacity is monetized, and users pay a fraction of data-center prices. Instead of expensive data centers, Render pays a fraction of that cost to draw capacity from thousands of devices.
- io.net: Specializes in generic GPU computing for AI training and inference. The platform pools capacity from data centers, crypto miners, and consumer hardware, forming a distributed alternative to centralized cloud providers. Its network is newer and more speculative than Render, but it’s attracting AI startups that can’t access or afford hyperscaler GPUs.
- Akash Network: Takes a broader view, offering a marketplace for general cloud computing and storage beyond GPUs. This positions it as infrastructure for the full stack, not just AI workloads. Akash is privately held but has a tradeable token, AKT. This is the riskiest option in this category, but it offers the most diversified exposure if decentralized computing expands beyond AI.
These are crypto-token plays, not traditional stocks
Before proceeding, understand what’s being purchased. All three networks operate via native cryptocurrency tokens, not equity shares. There is no stock ticker, no brokerage account, and no public-equity wrapper.
Direct exposure requires navigating crypto exchanges:
- Render Network (RENDER) trades on Coinbase, Binance, and Kraken.
- Io.net trades on select venues such as Binance and Gate.io, with liquidity varying by location.
- Akash Network (AKT) trades on Coinbase, Kraken, and similar venues.
This means dealing with crypto custody—via exchanges or self-custody wallets—and accepting regulatory uncertainty that comes with token investments. If this setup isn’t comfortable, the thesis won’t work.
For investors who prefer traditional equity exposure, the closest indirect bets are beneficiaries of the same capacity constraint:
- Data-center operators: Equinix (EQIX), Digital Realty Trust (DLR)
- Power infrastructure: Dominion Energy, Duke Energy (DUK), NextEra Energy (NEE)
- GPU supply chain: Nvidia (NVDA), Broadcom (AVGO), Super Micro Computer (SMCI)
But the key distinction is that these public companies benefit from the shortage itself, not from the temporary arbitrage window created by distributed idle capacity. They may do well as the overall AI hardware market grows, but they won’t provide the direct exposure to the current dislocation.
Risk factors
Clarity on potential pitfalls for this arbitrage strategy:
- Performance and reliability: Distributed GPU networks can exhibit performance variability, latency, and quality-control issues. Enterprise buyers expect reliable AI infrastructure; if these networks can’t match centralized performance, the arbitrage loses its appeal.
- Security and compliance: Regulated industries may avoid unknown hardware spread across borders. These networks will be limited to workloads where data sovereignty and compliance aren’t blocking.
- Hyperscaler catch-up: The base scenario assumes constraints ease by 2027–2029 as new data centers and power infrastructure come online. If power limits stretch beyond 2029, the high-growth window could stay open longer.
- Regulatory uncertainty: Some networks operate in gray areas. Government regulation of decentralized computing could raise costs and reduce flexibility.
- Crypto market contagion: Token prices depend on crypto markets. A Bitcoin downturn or regulatory crackdown could impact these assets regardless of fundamentals.
Investment timeline
The window spans from early 2026 through 2027–2028, a core 24–36-month period. The broader infrastructure constraint persists longer, but the outsized arbitrage tightens as hyperscalers ramp up capacity.
- Q1 2026: Start building positions as the 2027 power-constraint view becomes consensus. Use dollar-cost averaging to smooth volatility.
- Q2 2026–Q2 2027: Peak growth as AI demand accelerates while centralized capacity remains tightly constrained. These networks meet long-tail demand priced out of hyperscaler capacity.
- Q3 2027–Q2 2028: Growth continues but begins normalizing as new data centers come online and grid upgrades proceed. Watch hyperscaler capacity announcements closely—each major facility completion reduces the arbitrage.
- Q3 2028–Q4 2029: Maturation phase. These networks settle into specialized roles—emerging markets, cost-sensitive workloads, indie developers. They remain viable businesses, but growth stabilizes.
This isn’t a binary “works or doesn’t” thesis. It’s a maturation curve where networks shift from high-growth arbitrage to steady-state infrastructure alternatives.
The bigger takeaway
If GPU-aggregation networks prove they can deliver reliable computing at competitive prices during the 2026–2028 constraint, they earn legitimacy. Even if hyperscalers eventually reclaim market share, these networks will carve out niches in emerging markets, indie studios, and cost-sensitive workloads.
(MORE TO FOLLOW) Dow Jones Newswires
12-03-25 19:15 ET
Copyright (c) 2025 Dow Jones & Company, Inc.
The articles, information, and content displayed on this webpage may include materials prepared and provided by third parties. Such third-party content is offered for informational purposes only and is not endorsed, reviewed, or verified by Morningstar.
Morningstar makes no representations or warranties regarding the accuracy, completeness, timeliness, or reliability of any third-party content displayed on this site. The views and opinions expressed in third-party content are those of the respective authors and do not necessarily reflect the views of Morningstar, its affiliates, or employees.
Morningstar is not responsible for any errors, omissions, or delays in this content, nor for any actions taken in reliance thereon. Users are advised to exercise their own judgment and seek independent financial advice before making decisions based on such content. Third-party providers are not affiliated with Morningstar, and their inclusion does not imply any partnership, agency, or endorsement.
