Musk Buys More Generators - Structural Changes in the Global Gas Turbine Industry, Expansion Constraints, and AI Data Center Era Recently, there have been reports that Musk's xAI has signed a gas turbine procurement agreement with a South Korean company for the power supply of its new generation AI data center. 1、 How big is the bottleneck of gas turbines? Over the past decade, gas turbines have been almost forgotten assets by the market. But in 2024-2025, it suddenly returns to the center stage because three forces are simultaneously superimposed. Firstly, the exponential growth of AI data center load. Big model training and inference directly transform "computing power demand" into "sustained, high-density, and predictable electricity demand". Secondly, looking back at "electricity", what is truly scarce is "dispatchable and rapidly deployable base load power sources". The wind and solar power are unstable, the expansion of the power grid is slow, and the nuclear power cycle is too long. In the short to medium term, the only practical solution is natural gas+gas turbines. Thirdly, under the narrative of energy transition, the gas turbine industry has been underestimated globally for a long time. For over a decade, this industry has been almost in a "compressed state"—— Insufficient investment Conservative production capacity Synchronized contraction of materials, talent, and supply chain The result is that when demand suddenly increases, it exposes the supply rigidity of the global supply chain. 2、 Where are the real global hard constraints? ——Materials and Elements one ️⃣ A first-order, globally unified hard ceiling: rare elements and high-temperature materials From a global perspective, the only unavoidable hard constraints lie in materials, minerals, and rare alloy elements. The core of gas turbines - high-temperature blades - relies on nickel based single crystal high-temperature alloys, which cannot do without a set of extremely scarce elements. Taking rhenium (Re) as an example: Key elements of nickel based single crystal alloys Not forming independent mineral deposits, just a byproduct of copper/molybdenum mines The global annual production is only in the tens of tons range The main suppliers are concentrated in a few companies such as Molymet and KGHM In addition, it also includes: Tantalum (Ta), tungsten (W), cobalt (Co), nickel (Ni), etc. are used to enhance high temperature resistance, creep resistance, and corrosion resistance. The expansion speed is very slow New mine development cycle: 10-15 years Most of them are by-products and cannot be expanded solely for the sake of gas turbines The recycling rate is approaching the engineering limit Short term (1-3 years): Almost non accelerating Mid term (3-7 years): up to 20-40% relief Long term (7-15 years): It is only possible to change the supply curve through technological routes However, the global total production capacity of these elements and materials is actually very large, far exceeding the supply required for gas turbine expansion by ten times, but most of them are used by aviation generators. Therefore, the most likely thing to happen next is that the gas turbine supply chain will squeeze out the aviation engine supply chain. two ️⃣ Second order constraints: process, production line, certification, labor force This layer is the friction coefficient of each country, including: Single crystal casting and thermal barrier coating (TBC) process Skilled Engineer Density Military/power grid certification freeze Burden of historical supply chain structure They determine: Who expands faster, who delivers first instead of: Can the system be expanded This is also why there are significant differences in the expansion speed of industrial systems and OEMs in different countries. 3、 The expansion capacity of major global gas turbine OEMs one ️⃣ US, Japan, Europe: Global Main OEMs 美国 GE Vernova The world's largest gas turbine supplier After expansion: Annual production of approximately 21-32 GW Constraints: Occupation of military, power grid, and nuclear energy orders; Aging Labor Force and Compliance System 日本 Mitsubishi Heavy Industries / Mitsubishi Power Clearly declare that production capacity will double within approximately 2 years After doubling: Annual production of approximately 25-30 GW Advantages: Complete industrial system, stable supply chain Bottleneck: also limited by rare materials 欧盟 Siemens Energy Expansion range:+30-40% After expansion: Annual production of approximately 14-20 GW Constraints: high labor costs, complex certification and processes Total of three major OEMs (after expansion): About 60-80+GW/year Driven by AI data centers, potential demand may surge to over 100 GW/year. This is the fundamental reason why the delivery cycle has been extended to 5-8 years. two ️⃣ 韩国 South Korea: Doosan's' Asymmetric Value '(Why Musk Bought Them) Doosan Enerbility Current ~6-8 units/year ~2-3 GW/year The upper limit supported by the order: ~20 units/year ~7-8 GW/year The key advantage of South Korea lies not in its scale, but in its structure: Not in the main supply chain of the US military/power grid Adequate supply of engineers Customer structure is cleaner and more suitable for AI data center customization Conclusion: South Korea may not be the biggest player, but in the era of AI, it is an important marginal unlocking force. 4、 Conclusion: AI has once again pushed the problem from the "economic cycle" to the "physical limit" The global bottleneck of gas turbines is not in factories, but in materials and elements Rare elements are the only globally unified and unavoidable hard ceiling Craftsmanship, production lines, certification, and labor are "frictions of each country" that determine speed rather than whether or not The United States, Japan, and Europe are still the main suppliers, but after expansion, they are still not enough to meet the needs of the AI era South Korea (Doosan) is an important marginal complementary force, especially suitable for AI data centers AI, energy, and materials are pushing the world from a "business cycle problem" to a "physical limit problem" once again. (Last time it was storage)