The 700 GW interconnection queue figure is being treated as a curiosity or a data point about AI enthusiasm when it is actually a systemic regulatory failure with profound legal and financial consequences that will play out over the next 18 months in ways nobody is pricing. Here is what the coverage is missing:
FIRST: THE FERC ORDER 2023 TRAP. In 2023, FERC finalized Order 2023, the most significant overhaul of the interconnection queue process in two decades, specifically designed to address the explosion of speculative queue positions. It introduced cluster studies, deposit requirements, and readiness milestones intended to flush ghost demand. The critical analytical failure is that most commentary treats the 700 GW queue as evidence of demand when it is actually evidence that Order 2023 has not yet worked — the queue was filed largely BEFORE the new deposit and milestone requirements bite. The first major cluster study results under the new regime are only now beginning to emerge from RTOs like PJM and MISO. When those results land with full financial consequences attached, a significant portion of AI data center queue positions will either pay substantially higher interconnection costs than modeled or withdraw. This is a 2025-2026 event that is not priced into data center REIT valuations.
SECOND: THE PRECEDENT FROM RENEWABLE ENERGY OVERBUILD. Between 2010 and 2018, renewable energy developers flooded interconnection queues with speculative projects at ratios of roughly 10:1 against projects that actually got built. FERC, PJM, MISO, and CAISO all went through iterations of queue reform trying to address this. The pattern is well-documented: speculative queue positions impose real costs on legitimate projects by forcing them to undergo more complex multi-party studies, delaying everyone, and creating uncertainty that raises financing costs across the board. AI data center demand is now replicating this dynamic but with two material differences that make it worse. First, data centers are load, not generation, so they interact with distribution systems and local substations in ways that interconnection queue reform for generators did not anticipate. Second, data center developers are generally less experienced with interconnection process than renewable developers, meaning the withdrawal rate when costs become clear may be even higher and more disorderly.
THIRD: THE DISTRIBUTION SYSTEM BLIND SPOT. All discussion focuses on transmission-level interconnection queues, but the more binding near-term constraint for most data centers is distribution system capacity — the substations, feeders, and local transformers that are the actual last-mile constraint. These are regulated by state utility commissions, not FERC. They have no equivalent of the Order 2023 reform process. State PUCs are only beginning to grapple with large load interconnection and the frameworks are inconsistent, slow, and in many cases legally ambiguous about cost allocation. A data center project can have a favorable FERC interconnection position and still be effectively blocked for years by a state commission dispute over who pays for substation upgrades. This cross-jurisdictional gap — federal transmission reform meeting unreformed state distribution regulation — is the actual bottleneck that will determine which projects get built first, and it is receiving essentially zero coverage.
FOURTH: COST ALLOCATION LITIGATION RISK. When utilities upgrade transmission and distribution infrastructure for large new loads, the question of who pays becomes intensely contested. The legal framework here is genuinely unsettled. Recent cases before FERC and state commissions have produced contradictory outcomes on whether AI data centers qualify for load-serving treatment that socializes upgrade costs versus being treated as large industrial customers who must bear interconnection costs directly. If regulators move toward full cost causation — making data centers pay the full marginal cost of grid upgrades their load requires — project economics for a large fraction of the pipeline break entirely. This is not a hypothetical: several state commissions are actively litigating or investigating this exact question right now, including in Virginia, which hosts the largest concentration of data center capacity in the world. A Virginia SCC ruling that shifts cost allocation methodology could trigger a repricing cascade across the entire sector.
FIFTH: THE DELL ORDER BOOK CONCENTRATION RISK IS MISUNDERSTOOD. The $130 billion AI server backlog is presented as demand validation. It is actually a concentration risk signal. A meaningful portion of that backlog is from a relatively small number of hyperscaler and neo-cloud customers who are ordering ahead of confirmed data center capacity precisely because lead times for AI servers are long. If grid access delays cause even two or three major customers to push out their deployment timelines by 12-18 months, the order book revision could be violent because server vendors extended their supply chains and component commitments based on those schedules. The analog is the 2021-2022 enterprise networking equipment supercycle, where customers double and triple-ordered due to supply chain anxiety, creating an inventory correction in 2023 that hit Cisco, Juniper, and their suppliers hard despite underlying demand remaining real. The AI server market is running the same playbook with larger numbers and a grid constraint trigger that is exogenous to the technology cycle.
SIXTH: NUCLEAR AND GAS POLICY INTERSECTION. The power demand narrative has correctly identified that AI data centers are driving renewed interest in nuclear power, specifically small modular reactors and the restart of facilities like Three Mile Island. What is not being analyzed is the regulatory timeline mismatch. SMR licensing under NRC processes realistically takes 7-10 years from application to operation for first-of-kind units. Gas peakers can be permitted and built in 3-4 years but face increasing state-level opposition tied to climate commitments, plus fuel price risk. The only near-term dispatchable generation option that can actually meet AI data center load growth on a 3-5 year horizon in constrained regions is largely existing combined-cycle gas. But in states with aggressive RPS targets — California, New York, Illinois — adding gas capacity to serve AI data centers creates a direct legal and political conflict with statutory climate commitments that utilities and legislators have not resolved. The result will be regional divergence: states with flexible resource adequacy requirements will attract AI infrastructure while states with strict clean energy mandates will face legal paralysis that effectively caps data center growth regardless of private capital availability.
IN SIX MONTHS: By late 2025 and into early 2026, the first wave of FERC cluster study results under Order 2023 will create genuine queue attrition visible in withdrawal filings. Simultaneously, at least one major state PUC — most likely Virginia or Texas — will issue a significant ruling on large load cost allocation that reshapes data center project economics in that region. Dell and other AI server vendors will begin reporting order book metrics with increased scrutiny from analysts who have internalized the ghost demand argument, creating multiple compression pressure even if near-term revenues hold. The narrative will shift from 'AI infrastructure supercycle' to 'AI infrastructure rationalization,' which is actually a healthier long-term outcome but will create significant mark-to-market pain for investors who bought the supercycle story at peak multiples. The companies that emerge strongest are those with secured power purchase agreements at the project level before the cost allocation questions are resolved — they effectively have regulatory grandfathering positions. The companies most exposed are those with large queue positions, committed server order backlogs, and no signed interconnection agreements, which describes a larger portion of the announced pipeline than the market currently believes.
The market is pricing AI infrastructure as a demand problem; over the next 6-24 months it is more likely to behave as a power-constrained fulfillment problem with embedded cancellation optionality. The correct framework is not server TAM but conversion from announced compute demand to energizable MW. A useful stack is: 1) queue MW, 2) site-controlled MW, 3) utility-committed MW, 4) constructed MW, 5) revenue-producing IT load. Most equity narratives jump from 1 to 5.
Quantitatively, the key translation is server dollars into power demand. At current AI cluster economics, roughly $8-12 million of server and networking capex corresponds to about 1 MW of sustained IT load for accelerated compute deployments, depending on rack density, storage mix, and power usage effectiveness. That means a $130 billion AI server order book maps loosely to 11-16 GW of IT load, and approximately 13-20 GW of facility load at a 1.15-1.25 PUE. Even if only 60-70% of that order book is ultimately deployable within 24 months, the realized load is still about 8-14 GW, which is large enough to tighten several regional power markets but far smaller than queue figures discussed in the public narrative. This gap is the tell: queue data are not a demand forecast; they are a scarce-option land grab on interconnection rights.
That distinction matters for valuation. For AI hardware/OEMs, the right model is not straight-line backlog conversion. Assume base-case annual backlog conversion of 35-45% with cancellations of 5-10%, versus a bull case of 50-60% conversion and under 5% cancellations if power procurement clears. In a stress case where interconnection reform forces project triage and deposit requirements rise, cancellations/deferments could reach 15-25% of the order book tied to unfunded or non-powered campuses. For a vendor with 15-20% incremental operating margin on AI systems, a 10-point change in conversion rate on $130 billion of orders is a $13 billion revenue swing and roughly $2-3 billion EBIT swing across the supply chain. That is material enough to re-rate server OEMs, optical suppliers, power equipment vendors, and GPU-adjacent names even if end-demand for tokens remains healthy.
Utilities are the clearest medium-term beneficiaries, but only selectively. The market is underestimating the earnings torque from accelerated rate-base growth in transmission, substations, transformers, gas peakers, and grid modernization. As a rule of thumb, each incremental 1 GW of large-load data-center demand can require $1-3 billion of generation/interconnection/transmission investment depending on location and existing reserve margins; in constrained regions with new transmission and firming needs, all-in system capex can be higher. If even 10 GW of incremental load becomes utility-committed over 24 months, that implies perhaps $10-30 billion of associated grid and generation capex. For regulated utilities earning 9-10.5% allowed ROE on roughly 50-55% equity capitalization, the earnings contribution from every $10 billion of incremental rate base is on the order of $450-575 million pre-tax equivalent annual return to equity holders once in rate base, though timing lags matter. The market is broadly treating this as politically difficult capex rather than as one of the best visibility growth opportunities in the sector.
The caveat: not all utilities win. Regions with already stretched reserve margins, difficult permitting, or political resistance may face load-moratorium risk, tariff redesign, or requirements for customer-funded dedicated infrastructure. Utilities with clean pathways to recover large-load interconnection costs and add regulated transmission should outperform those forced into contested generation procurement or those exposed to retail-bill backlash. The threshold to watch is whether a utility can convert data-center MOUs into signed service agreements with deposits and cost recovery riders. Without that, the load is narrative, not earnings.
Data-center REITs and developers have more bifurcation risk than the market implies. The right unit is not leased square feet but delivered MW with firm energization dates. In constrained markets, powered shell capacity with signed utility service becomes scarcer and pricing power rises. REITs with land banks but uncertain power should trade at a discount to those with secured substation/interconnection pathways. A 100 MW campus delay of 12 months can defer $100-200+ million of annualized revenue potential depending on pricing, density, and customer mix; at sector EV/EBITDA multiples, that can erase billions in enterprise value if repeated across a development pipeline. Conversely, power-secured operators may capture rent step-ups and pre-lease premiums that consensus has not fully modeled.
Energy markets are likely to experience regional rather than national repricing. The market keeps talking about US power demand in aggregate, but the tradable impact is zonal. If only 8-14 GW of new AI load is realized, that is manageable at national scale yet highly disruptive in specific balancing areas where reserve margins are already thin. Gas demand upside is therefore not a simple national call; it is a capacity-value and basis call. Regions with fast-cycle gas, available pipeline capacity, and supportive permitting should see improved spark-spread optionality and capacity market support. Renewables and storage also benefit, but only where they can be paired with transmission access or firming resources that satisfy utility reliability standards for hyperscale loads. Articles on the topic usually overstate renewables-only feasibility for AI campuses in the next 24 months; from a dispatchability standpoint, many large-load customers will still need utility-backed firmness, gas peaking support, or costly overbuild plus storage.
On options/implied signals: the most likely market implication is not that options are directly pricing interconnection risk, but that cross-sector skew is misallocated. AI hardware names typically embed elevated implied vol around earnings and demand visibility, yet the larger underpriced risk is a medium-dated cancellation/deferment cycle rather than quarter-to-quarter order momentum. A practical read-through is that 6-18 month downside skew in server OEMs and cooling/power-density suppliers should widen if evidence emerges that queue clean-up is reducing executable campuses. Conversely, utility and grid-equipment names often trade with lower implied vol than the magnitude of potential capex upside would justify. In simple terms: the market is overpaying for near-term AI upside convexity and underpaying for medium-term utility/grid upside convexity.
Specific thresholds matter more than headlines:
1) If queue-to-executable conversion falls below roughly 10-15%, then current queue statistics are mostly noise and should not support heroic load-growth assumptions for merchant power, gas, or broad AI infrastructure.
2) If OEM backlog conversion stays above 45-50% with cancellations under 5%, then the power bottleneck is being solved fast enough to sustain AI hardware revenue through 2027.
3) If disclosed utility deposits/interconnection payments per project rise sharply, that is bullish utilities/grid suppliers and bearish speculative developers because it separates real from ghost demand.
4) If delivered data-center capacity in key hubs grows less than about 15-20% annually despite large announced capex, it implies power is constraining realization and the market should de-rate unsecured development pipelines.
5) If regional capacity auctions or bilateral power prices begin repricing for persistent large-load additions, merchant generation and gas transport optionality become investable before national load data show it.
What the coverage is getting wrong, specifically:
First, it treats all MW requests as evidence of end demand. They are better thought of as call options on scarce grid access. In any constrained market, rational developers and hyperscalers will over-file because the cost of missing an interconnection slot exceeds the deposit cost. Therefore queue inflation is endogenous to scarcity; it does not reveal true final load.
Second, it treats server backlog as if power were a back-office issue. Power is now the gating factor for revenue recognition in AI infrastructure. A server order without a utility-backed energization schedule should be haircut substantially in DCF and supply-chain planning. The narrative should move from semiconductor TAM to power-backed deployment probability.
Third, it ignores that the likely regulatory response changes economics nonlinearly. Higher deposits, milestone-based queue retention, and stricter load-served criteria do not merely reduce speculative demand; they transfer value from land/speculative developers toward utilities, incumbents with existing powered campuses, and vendors tied to retrofit densification rather than greenfield sprawl.
Fourth, it misses the retrofit trade. If greenfield power is constrained, enterprises and colocation operators will spend more on densifying existing powered halls: liquid cooling, busway upgrades, UPS replacement, switchgear, transformers, and software to improve PUE and workload placement. Some of the best beneficiaries may be electrical equipment, thermal management, and services providers rather than pure-play server assemblers.
Fifth, it underappreciates balance-sheet risk from mis-sequenced capex. If developers order long-lead electrical and compute equipment before interconnection certainty, working capital and cancellation exposure rise simultaneously. That can pressure not just OEM inventories but also trade-credit, leasing structures, and supplier capex plans.
Base case across sectors for 6-24 months:
- Utilities/grid equipment: positive. Expect estimate revisions higher where signed large-load agreements convert into rate base. Upside most visible in transmission, substation, transformer, switchgear, and standby/peaking generation ecosystems.
- Data-center REITs/operators: mixed to positive for power-secured portfolios, negative for speculative pipelines in constrained regions. Delivered MW commands a premium.
- AI hardware/OEMs: still positive near term, but with materially higher medium-dated downside if power-backed deployment rates disappoint. Backlog quality matters more than backlog size.
- Merchant power/gas: selectively positive in constrained regions; not a blanket national demand call.
- Renewables/storage: positive where integrated with utility procurement and transmission access; less positive for standalone intermittent projects marketed as direct AI solutions without firmness.
Instrument implications: long regulated transmission-heavy utilities and grid equipment versus a basket of unsecured data-center developers is the cleaner expression than simply being long AI. Another pair trade is long retrofit-enabling electrical/thermal infrastructure against names whose valuations assume frictionless greenfield campus energization. For credit, utility issuers with constructive regulatory frameworks may tighten on visible capex growth, while speculative private developers reliant on power queue conversion deserve wider spreads than equity narratives imply.
The narrative ignores the most important data point: the ratio of firm, energized MW to announced MW. That single conversion metric will drive earnings realization across AI servers, utilities, REITs, power markets, and equipment suppliers more than any top-down estimate of AI demand.
The confluence of 700 GW in grid connection requests in some US regions, a figure several times the nation's total installed generation capacity, alongside Dell's $130+ billion AI server order backlog, presents a profound market disconnect. The sheer scale of the power demand signal – particularly the identified 'ghost demand' component – suggests an underlying speculative frenzy that has yet to confront the physical and regulatory realities of energy infrastructure. The market is currently pricing AI infrastructure growth based on an idealized future where compute power is infinitely scalable and immediately deployable. However, the 700 GW figure, even if concentrated in specific regions, functions as a hard, immutable constraint. Building out generation, transmission, and distribution at this scale takes years, not months, and involves astronomical capital expenditure, complex permitting, and significant public pushback. The $130+ billion in AI server orders, while indicative of strong intent, effectively becomes 'vapor demand' if the corresponding data centers cannot secure reliable, sufficient, and affordable power. This creates an environment ripe for capital misallocation, where speculative projects in the data center and AI hardware sectors may absorb significant investment without a clear path to operational viability. Investors are tasked with discerning genuine, power-backed deployments from those driven by the speculative 'land grab' for AI compute capacity, a distinction largely absent from mainstream narratives.