On August 10, HD Hyundai Heavy Industries said it had signed a 956 billion won ($674 million) contract with the US developer Corban Energy Group to supply 1,000 megawatts of generating capacity for American data centers run by a large technology company, built around its 9.6 MW HiMSEN medium-speed engines [UPI, citing Asia Today, August 10, 2026]. It is the largest power-generation engine order in the company’s history. Three days earlier, Energy Vault announced its largest contract to date: 1.25 GW of integrated off-grid power for hyperscaler AI data centers in Texas, pairing its batteries, grid-forming inverters and controls software with a national EPC contractor deploying Caterpillar gensets [Energy Vault, August 7, 2026].

Two gigawatt-scale engine deals in four days. The reciprocating engine has stopped being backup and become the power plant. The question is how long its one advantage lasts.

What’s happening

Hold those last two against each other. April delivers in 2028. June delivers in 2029 and 2030. This desk covered GE Vernova’s heavy-duty gas turbine backlog near 100 GW in June, with roughly three-year lead times and 2029 and 2030 slots still selling. So the engine class beat the turbine queue in April and matched it in June, on a sample of two disclosed orders from one manufacturer. That is not a closed market. It is the first sign that speed to power, the entire reason this architecture won, is a depleting asset rather than a property of engines.

Brazil angle

Brazil is the market where none of this should be necessary, and it is not converting the advantage. The ONS counts 22 access requests to the Basic Network with signed contracts, 18 of them with connection authorization, and projects associated load rising from 304 average MW in 2026 to 3,457 average MW in 2030, in the first four-month revision of the 2026 to 2030 load forecast prepared jointly by CCEE, ONS and EPE [Canal Solar, April 13, 2026]. Average megawatts and nameplate capacity do not compare directly, but even so, Brazil’s whole projected 2030 data center load is the same order of magnitude as what HD Hyundai and Energy Vault each sold into the US buildout in one week.

The Brazilian pitch has always been the matrix: hydro, wind and solar, competitive free-market power. If the engine farm becomes standard, the hyperscaler stops buying the grid and starts buying the plant, and a clean matrix stops deciding site selection.

US angle

Interconnection is the constraint that created this market. In Northern Virginia, home of the world’s largest data center cluster, the wait for interconnection may stretch to seven years [Oxford Institute for Energy Studies, February 2026]. That number is why a Texas campus will burn gas through 42 engines rather than wait in a queue.

Note the supplier geography. The marginal supplier of American generating capacity in 2026 is a Korean shipbuilder, a Finnish engine maker and an American one. HD Hyundai is selling the same medium-speed platform it builds for ships. US energy security policy has spent two years on minerals and almost none on the machines.

China angle

China is not buying engine farms, and the reason is not superior AI power planning. Its advantage is build speed: 39 ultra-high-voltage projects in operation as of 2025, and as much as 80 GW of new coal-fired and 15 GW of gas-fired capacity added that year [OIES, February 2026].

The same paper is careful, and so are we. Grid connection is easy in Ningxia or Gansu and hard in Shenzhen, Shanghai and Beijing, where approvals carry PUE caps and power-use limits. OIES concludes that abundant energy and faster infrastructure approval do set China apart from the West, but that on the wider question of a durable advantage the evidence is mixed at best. Read the engine story as a symptom of a US interconnection problem, not proof of a Chinese win.

What it means

As last published, TAI-P stood at 104.2, up 13.6 percent year to date, against TAI-M at 101.8, up 6.4 percent. Those values carry a May 22, 2026 timestamp, so treat them as directional. The direction has held all year: power inputs have outrun mineral inputs.

Engine architecture may pull some of that back toward the materials leg. A 1,000 MW block built from roughly 104 generating sets, plus switchgear, plus a grid-forming inverter and battery layer, is a different bill of materials from one large-frame turbine of the same output. More copper per megawatt, more transformer content. We have no published figure for the delta and will not invent one, so read that as the desk’s view, not a sourced number.

The harder point is lock-in. Each deal is a long-dated gas commitment made outside utility planning, outside the interconnection queue, and outside the state processes meant to price this load. Virginia and New Jersey spent June building mechanisms to charge data centers for grid power. The engine farm is how you stop being a ratepayer.

What to watch