Shipboard generators meet
data-center power plants
As grid connections and gas turbines stretch out, some on-site power demand is shifting to four-stroke engines · In Korea, part of that demand runs into ship-generator orders on the same medium-speed manufacturing and test base
U.S. data centers can wait years for a grid connection, while new heavy-duty gas-turbine slots are already pushed out to 2031 deliveries. In the gap, some on-site generation demand is moving to four-stroke gas engines. Korean engine makers entered this cycle with thick marine order books from the shipbuilding boom. On parts of their medium-speed four-stroke lines—and in the test and component supply chains around them—ship-generator orders now meet land-based power orders. This report follows where those two demand pools actually collide, and who gets to book the revenue when capacity is scarce.
Two-stroke marine propulsion engines and four-stroke power-generation engines are different products, so we do not add their manufacturing capacity together.
Hanwha Engine’s estimated 2027 headroom for data centers: about 360 MW (SK Securities) · a single 400 MW project delivered in one year would be large enough to force a rethink of marine allocation
Published 2026.10.03 · Information checked through 2026.10.02 20:00 KST · Overseas-company figures use each company’s latest reported period: 2Q26, except Siemens Energy FY26 Q3 and Mitsubishi Heavy Industries FY26 Q1 · Korean broker estimates and manufacturer claims are labeled as such · Figures we could not verify in public materials remain flagged in the source ledger.
A U.S. data-center project can run into three separate waits: the wait to connect to the grid, the wait to receive its own generating equipment, and the wait until every supporting system is ready for commercial operation.
The first queue is already long. ERCOT, the Texas grid operator, has more than 438,000 MW of large-load interconnection requests in its pipeline, and 89% of that volume is data centers.1 SK Securities estimates that a U.S. AI data center can take about five years from applying for grid interconnection to commercial operation.2 The LBNL figure often cited alongside it—“more than five years,” with a 2025 median of 61 months—measures interconnection time for generation and storage projects. It is not the same statistic as connecting a load such as a data center.3
That is why developers are building power plants on site. In a forecast DS Investment & Securities cites from BCG and Goldman Sachs, on-site generation rises from 10% of U.S. data-center power in 2026 to 30% by 2030.4 Then the second queue begins. Book a new GE Vernova heavy-duty gas-turbine slot today and the delivery year is 2031.5 In Korea, Doosan Enerbility has booked 12 cumulative orders for 380 MW-class gas turbines from U.S. customers. The latest seven are scheduled to ship at a pace of one per month starting in May 2029.6
Engines come with a shorter queue. SK Securities puts four-stroke lead times at one to two years for high-speed engines and two to three years for medium-speed engines.2 Wärtsilä’s 412 MW Ohio engine plant, ordered in April, is targeting commercial operation in early 2028—roughly two years from order to start-up.9 Not every way around the grid requires a new engine. On Oct. 1 in Japan, JERA, Dell, and RHAELM signed an MoU to place up to 400 MW of AI infrastructure at JERA’s existing Chiba thermal plant and use plant power behind the meter. They say the structure can come online years ahead of a conventional grid-connected development, with operation targeted around 2028.10 Putting a data center next to an existing power asset is another answer to the same grid bottleneck.
Faster equipment does not automatically mean an earlier power-on date. The gas pipeline, emissions permits, and electrical systems all have to be ready before commissioning can begin. At Oracle’s Project Jupiter in New Mexico, pipeline completion slipped to February 2027 and an air permit for the fuel-cell system had not been issued, prompting Oracle to notify the developer of force majeure.11, 12 Oracle says the project remains on schedule.12
The companies announced a U.S. data-center “Generator-to-Chip” collaboration: Wärtsilä provides modular engine power plants; Schneider provides electrical equipment, automation, and power management; Stanley provides design, permitting, construction management, and commissioning oversight. The pitch is to design the pieces together from the start and compress schedules versus the traditional sequential model. No firm capacity or quantified schedule saving was disclosed.13
The collaboration goes straight at the problem that shortening generator lead time alone may not move commercial operation forward. The critical constraint can migrate from engine procurement to the integration of electrical design, permitting, and commissioning.
PAGE 01 On-site generation is an attempt to bypass the first queue. Engines shorten the second. The third queue remains.
NEXT02 When does an engine win?Fuel cells, diesel generators, and rental power are also options for on-site supply. This report focuses on three large gas-based alternatives: banks of four-stroke medium-speed gas engines, banks of aeroderivative gas turbines, and—if the schedule allows it—a large combined-cycle plant. Break the choice into four questions and each option wins under different conditions.
Separate equipment lead time from installation time
Equipment lead time: 2–3 years · multiple units can be installed in sequence, allowing partial operation to start sooner.2
Equipment lead time: up to 3 years · GE says packaged units can be installed in under 30 days.2, 14
Heavy-duty turbine lead time: 5–6 years · then add construction of the heat-recovery steam generator (HRSG) and steam turbine.2
Distinguish a single machine from the whole plant, and simple cycle from combined cycle
Separate the behavior of one machine from a plant made up of many machines
Turn individual units on and off while keeping plant efficiency around 45–48% across a 20–100% plant load · add capacity in increments from a few MW to roughly 20 MW.2
A 50 MW-class unit is about 25% efficient at 30% load · GE cites a five-minute fast start.2, 15
Capacity comes in large blocks · cold starts take hours.2
Compare at the same required capacity and reliability target
Smaller unit size keeps the cost of one spare unit lower · less high-temperature derating and lower water use.2
Tens of MW per unit · the spare-capacity percentage depends on unit count.
Large unit size makes redundancy more expensive · bigger footprint and typically cooling-water needs.2
The efficiency figures are manufacturer-derived data from Wärtsilä cited by SK Securities · actual performance varies with plant configuration and operating conditions.
AI compute loads swing on a millisecond timescale. The first device to absorb that shock is the UPS.16 An engine or aeroderivative turbine reaching full output within minutes is a statement about minute-scale load following; power-quality problems live on a much faster clock. That is why on-site plants pair generators with power-conditioning equipment such as UPS systems.
Baseload supply and load following · output adjusted by starting and stopping units.
Absorbs fast swings and momentary outages · the Bergen–Crusoe contract also includes separate power-conditioning equipment.17
Aeroderivative LM2500XPRESS gas turbines for data centers: 29 units cumulatively, nearly 1 GW · GE emphasizes five-minute fast-start capability.15
Gas engines: about 750 MW · roughly 438 MW is under firm contract (27 × 12.5 MW + 20 × 5 MW); the remaining 310 MW is under letters of intent · phased deliveries to multiple U.S. sites through 2027 · intended for continuous on-site baseload generation · paired with power-conditioning equipment.17
The same buyer is procuring both engines and turbines · not all announced capacity is a firm order.
The engine factories were busy with ships before data centers came calling. Merchant-ship contracts at Korean and Chinese yards reached 173 million DWT through end-August—about 90% of the 194 million DWT full-year record set in 2007.18 DS Investment & Securities estimates that, if the current pace holds, the full-year total could reach 260 million DWT and surpass 2007.18
When ships sell, engines follow. A large merchant ship typically carries one or two two-stroke propulsion engines plus several four-stroke generator engines for onboard electricity. Clarksons Research puts the global orderbook at 216.43 million CGT at end-August, with China at 67% and Korea at 18%. The Newbuilding Price Index stood at 186.34.19
According to DS Investment & Securities, Chinese yards expanded dock capacity quickly, but domestic engine production failed to keep up. Scarce Chinese engine slots were allocated first to state-owned yards, pushing private yards with delivery schedules to meet toward Korean suppliers.18 In 2025, Korean-engine orders from private Chinese yards topped KRW 900bn, up more than 45% year on year. DS discusses the trend alongside two-stroke propulsion-engine market share. The demand that directly touches four-stroke capacity is the generator-engine package—multiple units per vessel.18
| Shipyard | Korean-engine orders |
|---|---|
| Jiangsu New Yangzi | KRW 629.2bn |
| New Times | KRW 216.0bn |
| Zhoushan Changhong | KRW 84.2bn |
| Qidong Xiangyu | KRW 35.8bn |
Of the 15 contracts HD Hyundai Marine Engine won in the first half of this year, 12 came from Chinese shipyards,20and broker analysis published at end-August said China’s share of Hanwha Engine’s order backlog had risen from 33% to 49%.21 Marine orders for Korean engine makers are coming not only from Korean shipyards, but also from Chinese yards short of engine slots.
The International Maritime Organization (IMO) postponed adoption of the Net-Zero Framework (NZF), its carbon-pricing regime for international shipping, for one year in October 2025 by a 57–49 vote.22 The discussion is scheduled to run through Nov. 30–Dec. 3 MEPC 85 and resume in a special session on Dec. 4, with the date to be confirmed at MEPC 85.23 The September intersessional working group also ended without agreement, while the United States and Saudi Arabia continued to oppose the framework.24
DS Investment & Securities calculates that even if the charge were implemented at half the proposed level, cumulative carbon costs over 20 years could reach 20–30% of a new vessel’s price, assuming USD 100–130 per tonne and no discounting.18 Treat that as an order-of-magnitude gauge of the regulation, not as an investment-economics model for an individual ship. If the NZF is adopted, vessel classes where dual-fuel (DF) penetration remains below 1%—including VLCC and small-to-mid-sized container ships—are cited as the next candidates for conversion;18if the framework is delayed again, the incentive weakens.
Adoption strengthens the incentive to move to DF engines; another delay weakens it · the next swing factor for marine-engine demand.
PAGE 03 Korean and Chinese merchant-ship contracts had already reached 173 million DWT by August, and private Chinese yards are also sourcing Korean engines. The part that meets data-center demand is the four-stroke generator engine—several units on each ship.
NEXT04 Where the production bottleneck really overlapsWärtsilä manufactures every engine in its portfolio at the Sustainable Technology Hub in Vaasa, Finland. In two announcements, in February and May, the company said it would expand the plant and related supply-chain capacity by 65% versus 2025. Both announcements cited demand from Energy and Marine.25 That is a company-level example of marine and power-generation orders meeting on the same manufacturing base.
Wärtsilä has also described how it allocates that capacity. In a summary of its CEO strategy call published Sept. 29, the company said data centers account for less than half of its Energy equipment backlog. It does not sign slot-reservation agreements with individual customers; instead, it allocates capacity to firm orders based on commercial opportunity, strategic priority, customer relationships, and geographic diversification, while continuing to serve both Energy and Marine customers.26 Data centers have not taken over the factory. Firm orders from two end markets are sharing the production clock.
But “made by the same company” does not, by itself, prove that two products are fighting for the same production hours. There are levels of overlap.
| Level | What it shows | What it does not prove |
|---|---|---|
| Same company | The company has exposure to both markets | The two products compete for the same production time |
| Same site | They may share land, labor, and some infrastructure | They share assembly, testing, and critical machining equipment |
| Same process / equipment | Production allocation can genuinely collide | That specific process is currently full |
On this test, two-stroke marine propulsion engines and data-center engines overlap only at the company level. The real collision is in four-stroke medium-speed lines that serve both shipboard generation and power plants. Most data-center engine orders, meanwhile, are high-speed gensets from Caterpillar, INNIO, Cummins, and others; those do not share marine-engine lines.2
| Company · site | What it builds | Where data-center demand can overlap | Utilization / expansion signal |
|---|---|---|---|
| HD Hyundai Heavy Industries Ulsan · new Onsan plant | Large two-stroke marine engines · HiMSEN four-stroke engines | HiMSEN four-stroke assembly and testing, until the new plant comes online | FILING Engine utilization 113% (2Q · entire Engine & Machinery division)27 |
| HD Hyundai Engine Yeongam | Power-generation engines up to 10 MW, among others | Smaller-scale power generation | BROKER ESTIMATE Expansion to 1.0 GW by 2027 (DS)18 |
| HD Hyundai Marine Engine Changwon · former STX Heavy Industries | Two-stroke marine engines · turbochargers · crankshafts | Components such as turbochargers used in four-stroke engines | BROKER ESTIMATE Turbocharger utilization forecast at 100% in 2H from marine volume alone (DS)18 |
| Hanwha Engine Changwon · former HSD Engine | Two-stroke marine engines · four-stroke medium-speed engines (new 900 MW/year line) | Allocation of four-stroke line between shipboard generation and land-based power | BROKER RESEARCH Two-stroke slots sold out through 2028 · 1H26 utilization about 96% (DS chart)18 |
| (COMPARISON) Wärtsilä Vaasa, Finland | All engines in the portfolio | Marine and power-generation engines made at the same factory | COMPANY DISCLOSURE Planned +65% capacity versus 202525 |
HD Hyundai Heavy Industries signed power-engine contracts in April with U.S.-based Aperion Energy Group for 684 MW (KRW 627.1bn) and in August with Corban Energy for 1,000 MW (KRW 956.0bn).29, 30 Together, roughly 1.7 GW of contracted volume is far above the company’s historical annual land-based power sales, described by SK Securities as only 0.2–0.3 GW per year.2 But both contracts are spread across several years, and the new Onsan plant in Ulsan that is meant to absorb the growth is not scheduled for completion until May 2028.31 The new plant includes assembly and commissioning facilities plus crankshaft machining and block-casting equipment. Reports relaying company comments say the main Ulsan plant will focus on marine engines, while the new plant and the Yeongam site will serve land-based power.28, 32 To say contracts are outrunning expansion, annual delivery volumes and annual capacity have to be put on the same timeline.
| Item | 2027 | 2028 | 2029 | 2030 |
|---|---|---|---|---|
| Onsan new plant FILING | Construction starts in 1Q | Completion in May · 3 GW/year installed capacity comes online | Ramp-up (initial utilization about 30%, DS estimate) | HiMSEN total-capacity target: 7.2 GW |
| Land-based power capacity BROKER ESTIMATE | about 1.0 GW | about 1.45 GW | about 3.1 GW | 4.0 GW |
| AEG 684 MW deliveries BROKER RESEARCH | – | 11 units (about 0.23 GW) | 11 units | 11 units |
| Corban 1,000 MW FILING | Contract period 2026.08–2030.07 · annual delivery schedule not disclosed | |||
On the table, AEG alone requires only about 0.23 GW per year—a slice of estimated land-based capacity in those years. Whether a real bottleneck emerges depends on the undisclosed Corban delivery profile and additional orders.
Hanwha Engine completed a new four-stroke medium-speed line in Changwon in August with annual capacity of 900 MW, or 180 units at 5 MW each.33 SK Securities estimates that about 60% of the line will be used for marine applications in 2027, leaving roughly 360 MW of headroom for data centers. Data-center buyers are typically looking for projects in the 250–400 MW range.2 Annual headroom and total project capacity are not the same unit of time. The answer changes with how many years the deliveries are spread across.
| Delivery profile | Annual volume required | Versus annual headroom |
|---|---|---|
| All in one year | 400 MW | Above headroom · marine allocation would have to move |
| Split over two years | 200 MW per year | Within headroom · but leaves less room for other data-center orders |
| Split over three years | about 133 MW per year | Within headroom |
Demand for shipboard generator engines and data-center medium-speed four-stroke engines overlaps in parts of the manufacturing and testing base and in the component supply chain.
The annual delivery plan for this specific line matters more than a company-wide utilization figure. HD Hyundai is trying to separate the two demand pools with a new plant; Hanwha Engine has to divide a limited four-stroke capacity pool between marine and power-plant customers.
An engine maker’s reported earnings mix three different clocks. Today’s margins reflect orders won in the past and delivered now. Data-center orders booked this year will turn into revenue over several years according to each contract’s delivery schedule. Bergen–Crusoe deliveries run through 2027, while some large Korean contracts ramp more meaningfully from 2028 onward.2, 17 Service revenue comes later still—and sometimes lands at a different company.
Operating margin at HD Hyundai Heavy Industries’ Engine & Machinery division rose from 18.3% in 2025 to 24.8% in 2Q26,27while HD Hyundai Marine Engine rose from 18.9% to 24.4% over the same period.34 Hanwha Engine’s backlog grew from KRW 2.5472tn at end-2023 to KRW 5.9789tn at end-June 2026, with dual-fuel (DF) engines accounting for 83%.35 HiMSEN land-based power sales had been only 0.2–0.3 GW per year (SK Securities),2 so it is reasonable to read most of those margins as coming from marine-engine deliveries. The Korean data-center engine contracts signed this year are either not yet being delivered or are still in early stages, making them unlikely to be the main source of 2Q margins.
Hanwha Engine’s average selling price per engine rose from roughly KRW 7.0bn in 2Q24 to about KRW 10.4bn in 1Q26, based on a DS Investment & Securities chart.18 That figure also reflects a richer mix of higher-priced DF engines. Reading it as a like-for-like price increase of the same magnitude would overstate the evidence.
Wärtsilä provides the clearest direct datapoint on the economics of new orders. Since early 2025, its Energy backlog has more than doubled, while the gross margin embedded in the Energy equipment order book has improved by more than 500 basis points.36 That is margin on orders to be delivered in the future, not margin already realized. In the same release, Wärtsilä said equipment delivery periods continue to extend, pushing revenue from the existing backlog further into the future.36 That is why backlog can grow faster than near-term revenue. GE Vernova management has also said pricing on new gas-turbine orders is running above 4Q25 levels.37
New orders are also appearing outside the engine OEMs. On Oct. 1, Canada’s Enerflex won a contract from a North American data-center developer to design, engineer, fabricate, and assemble roughly 450 MW of behind-the-meter natural-gas generation. Deliveries start in 2027 and finish in 2028; Enerflex approved about USD 15m of 2026 capex and a further roughly USD 85m, mostly for 2027.38 This is a separate datapoint from engine-OEM production slots. What it does show is that on-site power demand continues to convert into real contracts—and that system integrators are also adding capacity, so the supply side has begun to respond.
SK Securities estimates that data-center engines run nearly year-round, shortening the maintenance cycle to about three years versus five years for marine engines, and that over a 25-year operating life service revenue could equal roughly 2.5× the original engine sale.2 Who owns that revenue differs by company. HD Hyundai Marine Solution signed an MoU with AEG in May to pursue long-term maintenance and operations for the 33 engines HD Hyundai Heavy Industries is supplying to AEG. It is not yet a signed long-term service agreement.39 For marine engines, Hanwha Engine offers LTSA coverage that can run for the life of the engine and has signed a five-year agreement with Pan Ocean.40 For licensed power-generation models, the split of service rights has to be checked contract by contract.
| Company | Role | Stage secured |
|---|---|---|
| HD Hyundai Heavy Industries | Builds HiMSEN engines · contracting party for equipment supply | AEG 684 MW (filing · 2026.04–2030.10) · Corban 1,000 MW (2026.08–2030.07)29, 30 |
| HD Hyundai Marine Solution | Service and operations (LTSA · O&M) | MoU for maintenance of 33 AEG engines (2026.05) · not a firm service contract39 |
| HD Hyundai Marine Engine | Two-stroke marine engines · components including turbochargers | No reported direct data-center contract |
| Hanwha Engine | Builds two- and four-stroke engines · pursuing a licensed power-generation product | No reported firm data-center order as of Oct. 233 |
| Wärtsilä | Engine manufacturing and service | Gross margin in the Energy equipment order book improved by more than 500 bp36 |
U.S. power-market rules are moving toward opening a path for data centers to bring their own generation (BYOG). What is widening is the ability to choose on-site power. There is no rule telling developers to use gas engines. The direction is clear; the rulebook is still moving.
In December 2025, the Federal Energy Regulatory Commission (FERC) directed PJM, the largest U.S. grid operator, to create a new transmission service for large loads co-located with generation.41 In January 2026, the PJM board outlined a path for large loads to bring new generation with them,42while Texas approved Batch Zero in June to review large loads of 75 MW or more in a coordinated process.1 More of the cost is shifting to the data center as well. In July, the Oregon Public Utility Commission approved a 29.7% rate increase for PGE data centers and other large loads.43
The latest move came on Sept. 29. FERC accepted PJM’s Reliability Backstop Procurement proposal, but suspended its effective date for five months, until Feb. 28, 2027, subject to refund conditions and the outcome of further proceedings. Questions including cost allocation and collateral moved into additional proceedings, and PJM did not launch the procurement it had planned to begin on Sept. 30.44 In separate concurrences, Commissioners Rosner and See emphasized a cost-causation principle: new large loads such as data centers should bear the costs they create rather than shifting them to existing customers.45 Those are individual commissioners’ concurrences, not a rate rule, and they do not require on-site generation. But they are a signal that large loads may be asked to shoulder power-supply costs more directly.
The path for large loads to bring their own new capacity is also becoming more concrete. On Aug. 13, PJM filed BYONC (Bring Your Own New Capacity) and an Interim Resource Adequacy Service (IRAS) with FERC, requesting an Oct. 12 effective date. The proposal would allow qualifying large new loads that bring new generation to avoid first-in-line curtailment during emergencies. FERC approval is still pending.46
| Constraint | Recent case |
|---|---|
| Air permitting | Virginia · data-center air-permit applications filed from July 1 onward must meet Tier 4-equivalent or tighter emissions limits for each generator set under Va. Code § 10.1-1322.6; the statutory gen-set definition does not restrict fuel type47 · A September executive order removed new data centers of 25 MW or more from fast-track permitting and state support and ordered a review of emergency-generation equipment; it is not a ban on gas generation48 |
| Gas supply | Oracle Project Jupiter, New Mexico · pipeline completion delayed to February 2027 · force-majeure notice issued11 |
| Fuel cells need permits too | Same project · air permit for fuel cells not yet issued · permitting was paused and later resumed4, 11 |
| State-level moratorium | New York · state environmental permitting for data centers of 50 MW or more paused for up to one year from July49 |
PAGE 06 An engine can be on site and the project can still miss its date if the gas pipeline, air permit, or grid-connection rule is late. Oracle’s Project Jupiter in New Mexico is the live example.
NEXT07 What changes first when supply catches up?Almost every major power-equipment supplier is adding capacity. The biggest engine expansions cluster in 2028–29, while the broader gas-turbine market is not expected to “normalize” on lead time until 2031–32. Some individual turbine expansions arrive in 2027–28. That means it is too simple to say engines will loosen first; the better frame is that different parts of the market clear on different clocks.
| Company | Plan | Timing |
|---|---|---|
| Wärtsilä | Vaasa plant and supply chain +65% versus 2025 | 1Q28 (+35%) · 1Q29 (+30%) |
| Caterpillar | Large reciprocating-engine capacity 3× versus 2024 · 1.5 GW from restarted 10 MW-class medium-speed line | Investment concentrated in 2027–29 · medium-speed shipments start 4Q26 |
| HD Hyundai Heavy Industries | 3 GW/year Onsan power-engine plant · HiMSEN 7.2 GW | Plant completes May 2028 · 7.2 GW target in 2030 |
| Hanwha Engine | 900 MW/year four-stroke line | Completed August 2026 |
| GE Vernova | Gas-turbine capacity 20 GW/year → 24 GW → 30 GW | 2028 · 2030 |
| Siemens Energy | Heavy-duty roughly 35 units/year → +15 units · medium-size 80/year → about 100/year | 2027 · 2028 |
| Mitsubishi Heavy Industries | Gas-turbine capacity to double versus 2024 | In progress |
| Doosan Enerbility | Gas turbines 8/year → 12/year | 2028 |
SK Securities argues that all major gas-turbine makers are effectively running full, and that even with expansion the market does not normalize until 2031–32.2 Rystad Energy likewise says gas-turbine and reciprocating-engine orders reached about 100 GW in 2025—30% more than OEMs could actually deliver—and that the gap has not yet closed. The same report, however, warns that oversupply risk is building in the 20–50 MW segment as new entrants pile in.52 Meanwhile, major engine expansions begin coming online earlier, in 2028–29. Engine-to-engine competition can intensify before turbine queues clear, changing pricing power sooner. Some competitive supply arrives even earlier: Caterpillar has restarted a 10 MW-class medium-speed gas engine it discontinued in 2022, secured its first order, and plans to ramp to roughly 1.5 GW over 18 months after first shipments in 4Q26.50 In other words, the engine bottleneck may start easing before Korean suppliers finish their own expansions. That is why the expiry date of the bottleneck cannot be read from gas-turbine lead times alone; engine-capacity additions have to be tracked in parallel.
| Clock | What to watch | When supply rises |
|---|---|---|
| New equipment orders | Lead time · pricing and economics on new orders · conversion or cancellation of reservations | This is where growth can slow first—or where pricing can crack first |
| Installed engines | Operating hours · primary power vs. backup duty · utilization | Whether they keep running after grid or turbine capacity arrives is a separate economic decision |
| Service business | Service-contract attach rate · duration · scope · who owns the revenue | Follows the installed base and operating hours · revenue is not automatic without a contract |
SpaceX is building a gas-turbine blade-and-vane foundry in Bastrop, Texas. Elon Musk says blade and vane casting is the limiting factor in turbine production and claims in-house casting could cut delivery delays by as much as 18 months · this is a company claim, not verified production capacity.53
Parallax is developing a 10 MW-class data-center turbine using 3D printing to reduce part count · it has raised USD 117m · prototype targeted for end-2026, testing in 2027, first customer deliveries in 2028 · not counted as supply before validation.54
SpaceX bringing turbine components in-house and new turbine entrants could shorten the bottleneck’s shelf life · if turbines loosen sooner than expected, some demand that spilled into engines can unwind sooner too · neither is verified production capacity yet.
When gas-turbine demand collapsed in 2017–18, manufacturers were caught with expansion underway and responded with large layoffs.55 Siemens Energy’s CEO has said the company is managing capacity very carefully to avoid creating an oversupply that could hurt it over the next decade.56 Doosan Enerbility faces the same strategic tension: capacity added during a shortage can dilute pricing power once competitors expand too.51
Kyobo Securities reviewed 150 data-center projects that faced local opposition. Among the cases where power was a stated issue, 91% were canceled or halted. That is not a cancellation rate for the U.S. data-center market as a whole.49 In 1Q26, 75 U.S. data-center projects worth USD 130bn were halted or delayed amid community opposition.2 The IEA notes that data centers fill gradually, making peak-load needs uncertain and creating a risk of over-connection.57 On the other side, Caterpillar’s CEO says no customer is slowing order pace yet. In the same answer, however, he stressed that capacity expansion is not based on data centers alone, citing oil and gas—especially gas compression—plus mining and marine demand, and adding that the oil-and-gas backlog is nearly twice the year-earlier level.50
If turbine lead times fall, growth in new engine orders can slow. Engine makers’ own capacity additions may move pricing before that happens.
Operation of the installed engine base and service revenue are separate questions. Track new orders and installed units as two different datasets.
The three variables to track are new orders, lead times, and capacity. The next round of earnings and regulatory milestones should tell us which way each is moving.
Does the Korea + China contract pace still point above the 2007 record?
Check for FERC approval · does the pathway for large loads to bring new generation become an effective rule?46
Energy-equipment backlog margin · delivery lead times · data-center orders.36
20:30 KST · path toward GE Vernova’s year-end target of 125 GW+ backlog plus reservations · conversion of reservations to firm orders · share of contracting now reaching 2031 deliveries.5
Engine-segment margin and backlog · new power-generation order disclosures.
How capacity is allocated after the Vaasa expansion · whether Wärtsilä restates priorities between Energy and Marine.36
A live test of whether permitting, rather than equipment, sets the commercial-operation date.12
Adoption and any dilution · another delay would weaken the incentive to shift marine orders toward DF engines.
Everllence 35/44G agreement and whether group-linked orders emerge.
UPDATE PLAN · Refresh FIG 01-1, 04-3, and 07-1 after GE Vernova’s Oct. 28 3Q results and the Dec. 4 IMO outcome.
Parts of the four-stroke medium-speed line—and the testing and component supply chain around it. That is where shipboard-generator orders meet data-center power-plant orders. Two-stroke propulsion engines and high-speed gensets from Caterpillar, INNIO, Cummins, and others are outside this scope.
Hanwha Engine · about 360 MW of estimated 2027 data-center headroom(SK Securities estimate) · typical buyer project size is 250–400 MW. At HD Hyundai, the two demand pools can still overlap on the existing HiMSEN base before the new Onsan plant—3 GW/year of installed capacity—is completed in May 2028.
Engine expansions: 2028–29 · gas-turbine lead-time normalization: 2031–32(SK Securities forecast). Caterpillar moves earlier: its restarted 10 MW-class medium-speed engine begins shipping in 4Q26 and is planned to ramp to about 1.5 GW over 18 months. Engine supply can loosen first, changing bargaining power before turbines normalize.
New orders · installed base · service contracts—count them separately. They become revenue at different times, and the revenue can accrue to different companies.
This report is research on the supply-demand structure of power-generation and marine engines and gas turbines · it is not a recommendation to buy or sell any security or asset · broker estimates and manufacturer claims are identified as such and are not treated as established fact · reporting periods differ across overseas companies · broker estimates—including the Korea/China order table, four-stroke headroom, service-revenue multiple, carbon-cost calculation, and AEG annual unit schedule—follow the assumptions in those source materials · public information checked through 2026.10.02 20:00 KST.