Powering Intelligence: The AI Buildout's Next Bottleneck Is Electricity

The Framework: Constraint Migration

Our working model for the AI buildout, the constraint-migration framework, holds that AI demand reveals successive bottlenecks, and each bottleneck re-rates a downstream sector in a wave. The first wave was compute: chips, networking, and the clouds that rent them. As compute supply scaled, the binding constraint moved downstream to electricity, and specifically to the ability to deliver firm power to a specific site on a specific date. This note formalizes the energy leg of that framework into a standalone thesis: Energy + Nuclear + AI.

The thesis in one paragraph: the generation constraint is real and worsening, the near-term supply options are being consumed as inventory, and small modular reactors are, in our assessment, the technology class positioned to add firm, repeatable, carbon-free capacity in the 2028 to 2035 window that near-term supply cannot cover and fusion cannot yet reach. Meanwhile, the equities that express this thesis have de-rated hard from their 2025 highs. That combination, a tightening physical constraint and a loosened valuation, is the setup this note examines.

The Demand Side: The Constraint Is Not Softening

Four observations anchor the demand case, each with a date attached:

Interconnection is the bottleneck. Industry trackers report that more than half of planned 2026 U.S. data centers are delayed by grid connection timelines. Interconnection queues in PJM, ERCOT, and the Southeast stretch 4 to 7 years for large loads. A data center that cannot connect does not train models, whatever its chip allocation says.

Load growth is back. U.S. power demand set records in summer 2026 after roughly two decades of flat load growth, with data centers the marginal demand driver. Hyperscaler capital spending for 2026 is running near $725 billion, up about 77% year over year, and street estimates for 2027 sit above $1.0 trillion. McKinsey-class estimates put cumulative data-center capital spend near $7 trillion by 2030.

The buyers are technology-agnostic and timeline-desperate. Microsoft contracted the Three Mile Island restart (835 MW, targeted online 2027, with Constellation) and holds a fusion power purchase agreement with Helion. Meta signed a binding 1.2 GW PPA with Oklo alongside multi-gigawatt renewables deals. Google holds 500 MW with Kairos Power and a 200 MW fusion PPA with Commonwealth Fusion Systems. Amazon invested $500 million in X-energy with multi-site SMR plans. In our view, that procurement posture, buying every credible electron across every technology, is precisely the demand environment that funds first-of-a-kind deployment.

Behind-the-meter is the strategic prize. Co-located generation bypasses the interconnection queue entirely. This is why the SMR value proposition is not dollars per megawatt-hour alone: it is time-to-power. When the alternative is a 4-year interconnection wait, a reactor delivering in 2029 competes with a grid connection arriving in 2029, even at a premium energy price.

The question that haunted the 2025 nuclear trade, whether anyone would actually pay for this, has been answered by contract. Binding hyperscaler PPAs now exist across fission SMRs, restarts, and fusion. The demand is contracted. What remains is supply execution, and that is where the risk genuinely lives.

The Supply Stack: Why 2028 to 2035 Is the Window

Walk the alternatives inside the middle window. Gas turbines carry multi-year order backlogs at GE Vernova, Siemens Energy, and Mitsubishi that stretch into the 2030s. Renewables plus storage have not yet demonstrated firm, around-the-clock, gigawatt-scale delivery at data-center density. Existing nuclear uprates and restarts are one-time inventory: once Three Mile Island and Palisades are back, there is no second Three Mile Island. Utility-scale new nuclear of the AP1000 class takes a decade or more. Our view, stated as opinion: SMRs are the technology class that matches the load profile AI infrastructure requires, firm, carbon-free, capable of behind-the-meter siting, and deployable in increments that track how a data-center campus actually grows.

Three de-risking thresholds were crossed in close succession, and this is the part of the story we believe the de-rated tape ignores. Regulatory: the DOE Reactor Pilot Program produced four DOE-authorized reactor criticalities by its July 4, 2026 deadline (Antares, Valar, Deployable, Aalo), the NRC approved Oklo's principal design criteria in under half the customary review time, and the ADVANCE Act keeps statutory pressure on the NRC to modernize licensing. Industrial: groundbreakings and long-lead procurement are underway, including Oklo's Aurora-INL construction with Kiewit and TerraPower's Natrium civil works in Wyoming. Commercial: binding offtake exists, and the letter-of-intent-to-PPA conversion pipeline across the sector exceeds 30 GW by our tally of company disclosures.

The SMR Field: Who Matters

Two readings of the July 2026 criticality milestones matter, and honest analysis holds both. The permissive reading: the DOE pathway works, fueled criticality of new designs is achievable in months rather than decades, and the regulatory fast lane that Oklo's commercial-scale Aurora-INL unit also rides has been validated in public. The cautionary reading: zero-power criticality on a test core is a fundamentally easier milestone than commercial first power from a 75 MWe powerhouse, and none of the four pilot-program winners has a commercial order book. We weight the permissive reading more heavily, and we say so as opinion, because the pathway being de-risked is the same one the commercial-scale projects travel.

Among public SMR developers, Oklo is the name we study most closely: by our assessment it pairs the largest disclosed customer pipeline (roughly 14 GW), a dual-track deployment strategy (DOE pathway now, NRC commercial fleet later), a fuel-recycling program we regard as a genuine moat given nascent HALEU supply, and about $2.5 billion of liquidity per company disclosures. After the de-rate, its enterprise value per contracted pipeline gigawatt screens, by our estimates, as the lowest in the public group. That is an observation about the setup, not a recommendation; the risks section below applies to Oklo with full force.

The Fusion Question, Taken Seriously

If a better technology arrives, this thesis has a terminal-value problem, so we handicap fusion explicitly rather than waving at it.

The bull case for fusion is no longer fringe. Helion raised $465 million in June 2026 at a $15.5 billion valuation, per press reports of the round, and is building toward a 2028 delivery for Microsoft in Washington state. Commonwealth Fusion Systems holds a 200 MW PPA with Google against its planned ARC plant, targeting early-2030s grid delivery in Virginia. More than 50 fusion companies globally are funded and building. Capital markets are treating fusion as investable infrastructure, and hyperscalers are signing offtake.

Our assessment, stated as opinion: fusion is a 2030s technology whose success would compress fission SMR terminal values, not strand the 2028 to 2035 window. Three reasons. First, timeline asymmetry: even on sponsor timelines, which slip in this industry as reliably as fission timelines do, fusion delivers first commercial power between roughly 2028 and 2033 at pilot scale, tens to hundreds of megawatts. The AI power gap is now through the early 2030s; fission fleets contracted today serve demand fusion physically cannot reach in that window. Second, the buyers are hedging, not choosing: Microsoft holds both a fission restart and a fusion PPA, Google holds both Kairos and Commonwealth Fusion. That procurement behavior tells you sophisticated buyers assign material failure probability to every individual technology and are buying the portfolio, which funds SMRs regardless of fusion's eventual outcome. Third, the risk classes differ: no fusion developer has yet demonstrated sustained system-level net facility gain on a repeatable basis in public evidence. Zero-power fission criticality is 80-year-old physics executed with new engineering; commercial fusion gain is new physics and new engineering simultaneously.

We commit to changing this assessment on evidence: a demonstrated, repeatable, system-level net-gain result from Helion or Commonwealth Fusion, or a fusion PPA priced at or below firm fission power with delivery guarantees, would materially compress the long-term value we assign to the fission SMR class and shift our attention toward fusion-adjacent exposure. Neither exists today.

The Bear Case, Steelmanned

Credibility in a speculative sector belongs to whoever names the risks before the market does, so here is the strongest version of the case against this thesis.

Nothing here has earned commercial revenue. No U.S. SMR developer has sold a commercial megawatt-hour. Zero-power test criticality is not commercial first power, and the gap between the two has swallowed reactor programs before. The sector trades on milestone probability, not earnings, which means every valuation in it, including any we might construct, rests on assumptions about events that have not happened.

Schedules slip; procurement windows close. Every reactor program in history has slipped, and the sector's 2027 to 2028 first-power promises carry material schedule risk. Meanwhile hyperscaler procurement windows are being filled right now by gas, restarts, and renewables PPAs. An SMR that arrives late does not just lose time, it can lose the marquee offtake that justifies the fleet.

Fusion may simply win. The capital, the talent, and now the offtake flowing into fusion are real. If net gain arrives sooner and cheaper than we handicap, the fission SMR class becomes transitional infrastructure with compressed terminal value.

The fuel chain is immature. HALEU, the enriched fuel most advanced designs require, has nascent commercial supply and uncertain costs into 2028 and beyond. Fuel scarcity can delay reactors that are otherwise ready.

Dilution is structural. Pre-revenue developers issue equity to live. Oklo's share count grew roughly 40% in 2026 alone per company filings, and investors should expect continued issuance across the class. Any analysis that is not fully diluted is fiction.

Policy is a tailwind that can reverse. The current regime, executive orders, the DOE pilot program, an NRC under statutory modernization pressure, is close to maximal support. Elections change policy, and a single safety incident anywhere in the class would re-rate every name in it at once, regardless of whose reactor was involved.

This complex trades violently. The 2025 euphoria and its round trip demonstrated that these equities move at a multiple of the thesis in both directions. A drawdown of the magnitude already observed, roughly 70% peak to trough in the flagship name, can recur on sentiment alone, with no change in fundamentals. Readers should assume that volatility is a permanent feature of the theme, not a phase.

What Would Prove Us Wrong

We publish break conditions on every framework we maintain. For the Energy + Nuclear + AI thesis, the conditions are:

1. Repeatable fusion net gain. A demonstrated, repeatable, system-level net-energy result from Helion, Commonwealth Fusion, or any credible developer, or a fusion PPA priced at or below firm fission power with delivery guarantees, would compress the long-term case for fission SMRs and force a rework of this framework.

2. First-power slippage past the procurement window. If the leading SMR programs slip materially past their 2027 to 2028 first-power targets while hyperscaler procurement windows are filled by gas, restarts, and renewables, the thesis loses its commercial anchor even if the technology eventually works. Watch announced hyperscaler power deals: if new marquee offtake stops flowing to SMR developers for several consecutive quarters, that is the signal.

3. A safety incident anywhere in the class. Any significant safety event at any advanced reactor, in any country, would re-rate the entire complex and likely reverse the policy tailwind. We would reassess everything on that day, not the following quarter.

4. Regulatory reversal. The framework assumes the compressed licensing timeline holds. A change in administration posture, a slowdown in NRC modernization, or a rollback of the DOE pathway would restore the decade-long cycle this thesis assumes is gone.

5. The demand side softening. The constraint case rests on interconnection queues measured in years and hyperscaler capex that keeps rising. Two consecutive quarters of hyperscaler capex guide-downs, or evidence that grid interconnection is debottlenecking materially faster than expected, would remove the scarcity that makes time-to-power valuable, and premium-priced firm power would lose its justification.

Our Opinion

Our view, stated as opinion and not as a recommendation: the generation constraint on AI is real, it is worsening on the evidence through July 2026, and the 2028 to 2035 window belongs to small modular reactors because nothing else can add firm, repeatable capacity inside it. The 2025 to 2026 de-rate repriced the sector's sentiment without repricing its physics, and we think the milestone cadence now observable in public, criticalities achieved, construction mobilized, binding offtake signed, is more consistent with progress than the tape implies. We follow the theme across the whole stack: the SMR developers themselves, the nuclear-heavy independent power producers (Constellation, Vistra) that capture the same demand with earnings today, the fuel bottleneck (Cameco, Centrus), the turbine and EPC complex (GE Vernova), and the private microreactor and fusion cohorts as data sources. We could be wrong in the specific ways enumerated above, and we track each of them. Our full reasoning and estimates are available in our research process description.