At a Memphis data center called Colossus, gas turbines run around the clock alongside racks of AI accelerators. A new grid connection at this scale takes four to seven years in the United States. An aeroderivative gas turbine (a jet-derived turbine engine deployable on the customer's own site, bypassing the grid interconnection queue) can go from signed contract to first megawatt in eighteen months. When compute generations turn over annually, that difference in lead time is decisive.
Every well-capitalized AI operator is now trying to secure this supply. SpaceXAI, formed by the April 2026 SpaceX-xAI merger and listed at $135 per share in June 2026, is trying to own enough of it to outpace everyone else. The panel's conclusion, across four analytical frameworks and two rounds of deliberation: the power race will most likely end without a single monopoly winner, and the actual binding constraint in the system doesn't sit at the turbine at all.
Weighted futures
| Scenario | Probability |
|---|---|
| Behind-the-Meter Race: No Monopoly Emerges | 33% |
| Constraint Migration: Upstream and Regulatory Blocks the Strategy | 27% |
| SpaceXAI Power Stack Consolidation | 22% |
| Power Technology Transition Disrupts the Gas Turbine Frame | 18% |
How we got here
Gas turbines had a strange decade. The market for large industrial units collapsed in 2022 and 2023, after years of weak demand. Manufacturers responded rationally: GE Vernova, Siemens Energy, and Mitsubishi all cut manufacturing capacity, reduced workforces, and scaled back expansion plans. Then the AI infrastructure boom arrived, demanding enormous amounts of power with no patience for multi-year permitting timelines.
Behind-the-meter turbines (on-site generation that bypasses the grid queue) became the immediate answer. By early 2025, GE Vernova and Crusoe had confirmed the commercial model with a deal for roughly a gigawatt of aeroderivative units. Crusoe's Abilene, Texas campus reached 1.375 GW by March 2026. Williams and Blackstone closed a $5.34 billion power joint venture in July 2026, targeting over 6 GW. The race was already crowded before Musk's formal bid arrived.
The 2022-23 capacity cuts are what make today's shortage structural rather than cyclical. Gas turbine manufacturing requires skilled metallurgical workforce, long-standing supplier relationships, and decades of institutional knowledge in alloy handling, blade cooling design, and combustion dynamics. None of that rebuilds in a quarter. The GE Vernova backlog extending through 2026-27 is the empirical confirmation: this is not a gap that capital alone can close within the analysis horizon.
The thing nobody is asking about
The public narrative runs like this: Musk sees that electricity is the bottleneck for AI, tries to own the turbine supply, and therefore tries to own the AI advantage. Logical enough. The panel found it incomplete in a specific, load-bearing way.
Gas turbine blades operate at temperatures above the melting point of most metals. The materials that survive (nickel superalloys, precision-engineered metallic compounds) are produced by a small number of specialized foundries: Howmet Aerospace, Precision Castparts, ATI, Haynes International, Special Metals. These suppliers have deep, decades-long relationships with GE Vernova and Siemens Energy. Their forging capacity is structurally tight, not for one year but for six, from 2019 through 2025. And the knowledge embedded in high-temperature alloy forging (what engineers call tacit knowledge, the expertise that lives in practitioners' hands rather than in any procedure manual) takes decades to accumulate and cannot be quickly replicated.
Buying a turbine factory doesn't buy you the metal that turbine factory needs.
Theory of Constraints (Eliyahu M. Goldratt, 1990) has a term for what happens when you improve a step that isn't the system's slowest link: a local optimization error. Output doesn't improve. You've invested real capital in a non-constraining resource while the actual bottleneck sits upstream, unchanged. The dependent-events chain (the sequence of steps that must each complete before the next can begin) governing AI compute output runs: model deployment, then power availability, then turbine delivery, then turbine manufacturing, then nickel superalloy supply. If superalloy is the slowest step, controlling turbine assembly slots generates no additional megawatts.
Four frameworks. Two rounds of deliberation. One convergence: the acquisition, if it happens, targets the wrong node.
The analysis
Theory of Constraints (Eliyahu M. Goldratt, 1990): The five focusing steps of TOC demand that you identify the constraint before deciding how to elevate it. The AI throughput system's binding node in 2024-2026 is power and grid infrastructure. The correct response is to elevate that node as efficiently as possible, by whatever pathway works fastest. TOC doesn't privilege any particular actor or technology for doing this; it measures success in megawatts reaching AI compute, not in ownership structures. Multiple operators deploying behind-the-meter turbines simultaneously is, in TOC terms, a better outcome than monopoly capture, because it collectively elevates the constraint faster.
TOC favors the no-monopoly scenario (33%): distributed elevation is more efficient. It favors constraint migration (27%) as the predicted outcome when the wrong node is targeted. It argues against consolidation (22%): until superalloy is also addressed, the drum (the constraint that sets system pace in the Drum-Buffer-Rope scheduling method) hasn't been correctly identified, and GW-scale deployment claims outrun what the chain can actually deliver.
Dynamic Capabilities (David J. Teece, 2009): Gas turbine generation capacity is a cospecialized asset at the AI value chain's bottleneck, an asset that generates more value inside a specific ecosystem than as a standalone business. SpaceXAI's post-merger organizational architecture, with concentrated decision authority, pooled capital across entities, and tolerance of irreversible commitment, gives it the structure that DC theory identifies as necessary for Chandlerian bottleneck-asset seizing at speed. Public company boards, constrained by quarterly earnings visibility, cannot match this.
The DC framework's sharp counterargument: without reconfiguration into upstream superalloy supply, the acquisition delivers Ricardian scarcity rents (the returns available to any well-capitalized buyer in a tight market) rather than Schumpeterian rents (returns that come from genuine innovation others cannot replicate). The advantage is real but bounded, and it erodes by 2027-28 as GEV's Greenville expansion (+25%) and Mitsubishi's capacity doubling start delivering. DC favors consolidation (22%) only under the condition that upstream reconfiguration happens. It favors constraint migration (27%) as its own upstream warning vindicated.
Complexity Economics (W. Brian Arthur, 2014): The AI power market is in a non-equilibrium window under positive feedback. Increasing returns mean early power access compounds: more compute, faster model training, earlier revenue, more capital for further power priority. Under this dynamic, accepting below-market returns on a turbine acquisition to establish early position is rational strategy, because the compounding value accrues above the turbine, not in turbine margins.
CE is honest about what it can't predict: who wins. Path-dependent processes (processes where the sequence of past decisions limits present options) are explicitly non-predictable as to outcome from initial conditions. Small events determine who locks in. CE argues against the no-monopoly scenario (33%) because increasing returns should produce structural concentration, not stable diffuse competition. But it concedes that GEV's existing customers, Crusoe's early gigawatt, and the Williams/Blackstone capital may already be in the self-reinforcing phase. CE favors the technology-transition scenario (18%) as a legitimate competing pathway: if nuclear SMRs and long-duration storage enter the AI power niche with their own increasing-returns trajectory, gas turbines could lose the power race to a different technology before the turbine race itself resolves.
Wardley Mapping (Simon Wardley, 2022): Each component in the AI power value chain sits at a position on the evolution axis, running from genesis (novel, uncertain, high value) through custom-built, product, and commodity. The value-capture sweet spot is the product stage: scarce enough to command rents, proven enough to deploy at scale. Aeroderivative gas turbines for AI data centers sit at product stage today. The climatic pattern Wardley identifies is that components evolve along this axis whether actors want them to or not, and that competitive pressure from multiple large buyers accelerates the journey toward commodity.
Wardley favors consolidation (22%) as the textbook product-stage capture play: seize the scarce product before it standardizes. It argues against the no-monopoly scenario (33%) because competitive pressure from SpaceXAI, Crusoe, hyperscalers, and Williams/Blackstone simultaneously gives GEV and Mitsubishi strong incentive to expand capacity, commoditizing turbines faster than a monopoly scenario would. It favors the technology-transition scenario (18%) through a different mechanism: Meta's nuclear portfolio represents a genesis-stage component entering the value chain, potentially bypassing the product-stage turbine market entirely rather than competing on its terms.
What the scorecard says: all four frameworks are reporting first-call positions on this topic. No prior scenarios have resolved. Hit rates, calibration scores, and predictive edge are unavailable across the board. The opening position from each lens should carry the epistemic weight appropriate to first analysis, not validated track record.
The evidence gap that matters most: as of August 2026, no turbine OEM acquisition by SpaceXAI has been confirmed. No upstream superalloy investment has been disclosed. Gas turbines appear in SpaceX SEC filings for Colossus Memphis as a behind-the-meter deployment, not an ownership structure. The Fermi Inc. power REIT structure may be an intermediary procurement vehicle, but this is unconfirmed. The gambit as currently observable is a BTM deployment strategy, and the lock-in ceiling for a deployment strategy is meaningfully lower than for OEM control.
The four futures
33%, Behind-the-Meter Race: No Monopoly Emerges. SpaceXAI deploys BTM turbines at meaningful scale, but so does Crusoe, Williams/Blackstone, and the major hyperscalers. GE Vernova and Siemens prioritize their existing contractual customers through 2026-27; new entrants join a queue. The power advantage is real and meaningful for early movers across the field, roughly a 2-4 year lead over grid-dependent competitors. No single actor achieves monopoly allocation. Post-2027 OEM capacity expansions begin eroding the scarcity premium as supply responds to sustained price signals.
The key dissent: Complexity Economics argues this outcome should drift toward concentration rather than stable diffuse competition. The CE counter is that multiple actors may have already initiated the self-reinforcing sequence simultaneously, making "no monopoly" a description of several early movers each locking in partial positions rather than a genuinely competitive market.
27%, Constraint Migration: Upstream and Regulatory Blocks the Strategy. The nickel superalloy supply chain becomes the visible system bottleneck as turbine deployment attempts pile up against material availability. Turbine OEM acquisition, if it occurs, delivers no additional megawatts because OEMs themselves cannot accelerate output against structural material scarcity. Separately or in combination, FERC or EPA enforcement actions analogous to the Amazon-Talen nuclear rejection challenge BTM permitting, creating a second constraint on the pathway.
The key dissent: Wardley Mapping argues that a full value-chain analysis would have flagged the upstream constraint before any committed capital. A sophisticated actor conducting proper map-before-act doctrine would not target the downstream node while the upstream is tighter.
22%, SpaceXAI Power Stack Consolidation. SpaceXAI executes the full gambit: long-term turbine supply contracts or OEM acquisition, partial management of the superalloy constraint via upstream offtake agreements, and a compounding 2-4 year power advantage. The IPO capital funds the irreversible commitments that public company boards cannot make.
The key dissent: Theory of Constraints argues directly and forcefully against this branch. Until superalloy is also addressed, the acquisition cannot move the drum. Moderate-to-high confidence in this dissent.
18%, Power Technology Transition Disrupts the Gas Turbine Frame. Meta's 6.6 GW nuclear portfolio, accelerating SMR licensing, and maturing long-duration storage reach commercial viability at AI data center scale by 2028. The AI power constraint is resolved by technology transition rather than by ownership of any particular generation technology.
The key dissent: Dynamic Capabilities warns that committed capital becomes a stranded asset if this transition arrives on schedule.
Milestones and forward timeline
The September-October 2026 cluster is the most diagnostic window. An executed supply contract above 500 MW, or a disclosed superalloy investment, shifts meaningful weight toward the consolidation scenario. Absence of both through Q4 2026 shifts weight toward the competitive race and narrows the path to structural advantage for any single actor.
Confidence and caveats: as of August 2026, no turbine OEM acquisition has been confirmed. No upstream superalloy investment has been disclosed. All four frameworks carry first-call positions with no resolved scenarios and no track record to draw on. The September-October milestone window is the first real test of these weights.
The bottom line
The panel's most likely single scenario gives you a race with multiple winners. No monopoly. Real but diffuse power advantage. A closing window as OEM capacity expansions arrive post-2027.
If you are building AI infrastructure: the BTM turbine play is real and meaningful. Executing a large-scale behind-the-meter deployment in 2026-27 gives a real lead over grid-dependent competitors. The structural tightness is baked in through at least 2027-28, and the queue ahead of you includes Crusoe's gigawatt, the Williams/Blackstone pipeline, and existing GEV customer priority.
If you are investing in the energy infrastructure underneath AI: the more interesting asymmetry is in the superalloy foundry layer. Howmet Aerospace, Precision Castparts, ATI, and Haynes International sit at the actual binding node and are almost absent from the public conversation about who wins the AI power race.
The most important question in the AI power race isn't who owns the turbine factory. It's who owns what the turbine factory needs.