AI Power Demand Crisis Forces Data Centers to Buy Ship Engines for Off-Grid Solutions

AI Power Demand Pushes Tech Giants Off the Grid

AI power demand has reached a breaking point that is forcing tech companies to abandon traditional grid connections entirely. HD Hyundai Heavy Industries closed a $673.8 million contract with Corban Energy Group in August to supply 1,000 megawatts of natural-gas engines for data centers run by an unnamed major US technology company1

1

. The order, built around roughly 100 of HD Hyundai's 9.6-MW HiMSEN engines, represents prime power generation designed to run facilities around the clock rather than serve as backup1

1

. This marks the largest power-generation engine contract in the company's history and its second US data center deal in four months, following a 684 MW order worth $425 million with Aperion Energy Group in April1

1

.

The bottleneck driving these purchases is clear. Grid interconnection queues now stretch into years, leaving operators with no choice but to contract for their own generation instead of waiting for utility connections1

1

. This mirrors Amazon's off-grid gas plant strategy in Texas and reflects the same logic behind state grid audits that are now slowing approvals1

1

. When a tech company needs a gigawatt and the grid cannot promise a delivery date, buying ship engines stops being an edge case and becomes standard operating procedure.

Electricity Replaces Chips as AI's Primary Constraint

The AI data center power consumption crisis has fundamentally shifted what limits growth in the sector. Electricity, not chips, now represents the binding constraint on AI infrastructure2

2

. The four largest cloud firms are set to spend around $725 billion on AI infrastructure this year alone, with single sites carrying price tags like Meta's $13 billion Texas data center2

2

. Server racks that drew about 3 kilowatts for ordinary computing now pull up to 150 kilowatts for AI workloads2

2

, and global data center capacity is expected to nearly double to roughly 200 gigawatts by 20302

2

.

This power crunch is sending data center builders to banks for billions in financing commitments before they commit to power deals and construction contracts2

2

. The borrowing has already reshaped credit markets, from a $5.9 billion loan for one data center operator to Oracle's $16.3 billion raise that leaned on private credit after banks grew cautious2

2

. Analysts estimate the sector is carrying around $1.65 trillion in off-balance-sheet obligations, structured through special vehicles that keep debt off tech giants' accounts2

2

. The Bank for International Settlements has flagged this circular financing between cloud firms, suppliers, and construction lessors as one of the bigger risks to financial stability2

2

.

Volatile Power Demand Damages Critical Infrastructure

AI workload power requirements are creating physical damage that extends far beyond planning spreadsheets. Rapid swings in AI data centers' power demands are straining vital equipment, causing batteries, generators, and cooling systems to malfunction or wear out far sooner than expected4

4

. Power increments equivalent to the consumption of factories or towns can appear and disappear within seconds, creating repeated shocks that connected equipment struggles to absorb4

4

. AI at times sees power usage spike as much as 50% above design capacity, meaning a 1 gigawatt facility may use 1.5 gigawatts for a split second4

4

.

Cranks on small natural gas combustion engines used to generate power at data centers have broken off, and at xAI's Colossus computing facility in Memphis, Tennessee, gas-fired turbines developed cracks4

4

. Batteries installed to smooth out power swings have sometimes needed replacement within months or even weeks due to high strain4

4

. These issues are causing delays or curtailing operations at some AI computing facilities, directly hitting revenue4

4

. The problems span data centers worldwide, from the Middle East and Africa to Europe and the US4

4

.

Next-Generation AI Data Centers Face Supply Chain Bottlenecks

Electrical infrastructure now determines whether next-generation AI data centers open on schedule, not capital or silicon availability5

5

. A single rack now pulls 40 to 130 kilowatts, with vendor roadmaps discussing cabinets measured in the hundreds of kilowatts as a near-term expectation5

5

. Large power transformers have become the primary supply chain constraint, with lead times stretching into multiples of years due to specialized requirements for grain-oriented electrical steel, bushings, tap changers, and skilled winding labor5

5

. Switchgear, breakers, generators, and basic busway have stretched in sympathy, forcing developers to order transformers speculatively and warehouse them until layouts catch up5

5

.

Power and Cooling Demands Drive Infrastructure Rethink

Meeting AI data center power consumption while maintaining efficiency is pushing developers toward gigawatt-scale facilities3

3

. Electricity consumption from data centers has grown at 12% per year over the last five years, with expected demand growth in AI training facilities set to drive substantial increases3

3

. Grid interconnection delays can push timelines back years, forcing many large-scale developments to pay a premium for off-grid solutions or hybrid architectures that include battery storage and renewable integration through microgrids3

3

. Liquid cooling is gaining traction as air cooling struggles to keep pace with thermal loads, though it introduces new dependencies around infrastructure support3

3

.

The Electric Power Research Institute projects data centers rising from 4-5% of US electricity today to 9-17% by 20301

1

. As this plan scales, more of AI's power problem gets answered with natural gas, tying running costs and emissions to fossil fuels and driving local backlash where power bills and pollution top complaint lists1

1

. Grid instability risks are mounting as utilities strain to keep lights on while data centers compete for the same electrons, with households in some regions already paying more2

2

. Watch for how grid operators respond with ramp-rate limits and power quality conditions in interconnection agreements, and whether the circular financing flagged by regulators unwinds as AI revenue growth faces pressure to justify infrastructure spending.