2026~2030: in just five years, the power demand of new AI data centers must be filled. Which generation sources can actually be switched on, and where is the real bottleneck in the transmission grid? A ranking of the top 5 energy sources, an infrastructure checklist, and a tiering of US and Korean listed beneficiaries.
* Composite of 5-year revenue/order visibility + barriers to entry + pricing power. Analysis opinion, not a recommendation. 30 stocks are tiered in the body.
Let's start with the near-term numbers. The IEA (International Energy Agency) projects that global data center electricity consumption will double from 485 TWh in 2025 to roughly 950 TWh in 2030. Goldman Sachs forecasts +165% growth in 2030 versus 2023, and in some scenarios as much as +220%. Under any scenario, data center power alone will need an additional +400 TWh+ within five years.
One thing this trend tells us: Big Tech no longer waits for "the public grid to take care of it." They are doing one of three things: buying power plants directly, building them on their own sites, or locking in 20-year direct-sale PPAs. And all three paths mean someone has to machine the gas turbines, wind the transformers, and plant the transmission towers.
When evaluating generation sources within the 5-year window, what matters more than LCOE (levelized cost) is "can GW-scale new capacity from this source actually be switched on starting from this point?" The comparison below uses the criterion of "if construction begins in January 2026, when does it come online?"
| Source | Minimum Lead Time | Realistic Lead Time | Possible Within 5 Years? | Notes |
|---|---|---|---|---|
| Modular gas (aeroderivative) | 12~18 months | 18~30 months | ✓ | Adopted by Stargate·Meta. But turbine slots near sell-out by 2030 |
| Large CCGT (combined-cycle) | 36~48 months | 48~60 months | ✓ (barely) | Groundbreak 2026 → online 2030 |
| Solar + ESS | 12~24 months | 18~30 months | ✓ | The fastest GW-scale option once transformers and interconnection are unblocked |
| Restarting existing nuclear | 24~48 months | 36~60 months | ✓ | TMI/Crane 2027~2028, Palisades, etc. |
| Nuclear uprate (MUR/EPU) | 24~36 months | 24~48 months | ✓ | +5~10% to existing units. Nuclear fuel·turbine replacement |
| Onshore wind | 24~36 months | 36~60 months | △ | Permitting·transmission interconnection are the real variables |
| Enhanced geothermal (EGS) | 24~36 months | 30~48 months | △ | Capped at 1~2 GW by 2030. Site-dependent |
| Offshore wind | 48~72 months | 72~96 months | ✗ | Barely registers within 5 years |
| SMR | 60~84 months | 72~120 months | ✗ (tail) | First startup possible 2030~2032. The very tail of the 5-year window |
| New large-scale nuclear | 96~144 months | 120~180 months | ✗ | A next-cycle (post-2032) story |
* Lead times are estimates based on EIA·DOE·IEA·industry averages. Subject to ±30% variation depending on site·permitting·transmission interconnection conditions.
Why is gas #1? The answer is simple: there is no way to switch on 24-hour, GW-scale capacity within five years with any other source. In an era when Big Tech is building single campuses of 1.2 GW (Stargate Abilene) or 1 GW (Meta El Paso), "drawing power from the surrounding grid" no longer works.
"If there are no gas turbines, use ship engines." This is a proposition that has been validated in earnest in the US data center market since the second half of 2025. Four-stroke medium-speed gas engines (5~20 MW reciprocating gas engines typically used as marine propulsion) have begun entering en masse as generators for Big Tech's new campuses. What differs from before is that they go in not as mere backup but as primary generation or behind-the-meter baseload.
| Company | Ticker | Features | US DC Orders | 5-Year Visibility |
|---|---|---|---|---|
| HD Hyundai Heavy Industries | KRX 329180 | HiMSEN 20 MW-class, the only Korean firm to own medium-speed engine IP in-house | 684 MW (2026.4) + further talks | Strongest |
| Hanwha Engine | KRX 082740 | Formerly HSD Engine. Four-stroke medium-speed engine line restarting in 2027 | Hanwha Energy BTM vertical structure taking shape | Strong from 2027~ |
| STX Engine | KRX listed | Medium-speed engines + diversification into military·power generation | Related expectations, specific order timing undisclosed | Anticipated |
"Let's build new nuclear" is not a 5-year solution. But nuclear that is already built is a 5-year solution. The US has 94 operating units, several closed units that can be restarted, and room for capacity uprates (MUR/EPU) on existing units. As Big Tech signs PPAs directly and underwrites the generators' loss risk, units once shut down for "insufficient profitability" are coming back to life.
The solar + battery (BESS) combination is in fact the cheapest on an LCOE basis. Per IRENA·NREL data, in some US regions firm LCOE (including storage and transmission costs) has come down to the $50~80/MWh level. And the plant itself can be built at GW scale within 12~18 months. The problem is just one thing — when the sun isn't out, it doesn't generate.
"Geothermal" used to be possible only in regions with natural heat sources in the crust (Iceland·New Zealand). Enhanced geothermal (EGS) borrows shale gas drilling technology to create artificial fractures at depths of 3~4 km and circulate water, extracting geothermal heat anywhere. Fervo Energy commercialized this first, reaching TRL 8 (ready for commercial adoption).
EGS, being 24/7 clean baseload, is faster than SMRs and more stable than solar·wind. It will reach the 1~2 GW level globally within five years — small, but a decisive card filling the "final 1~2%" of Big Tech's clean-power portfolio.
SMRs are modular reactors in 50~300 MWe units. The concept is "made in a factory, transported by truck, and assembled on site," but in reality, the fastest scenario has first commercial operation in 2030. That is, only in the last year of the 5-year window (2026~2030) might the very first unit barely come online.
This is the single most important chapter of the article. More clogged than building power plants is the transmission grid·transformers·switchgear that carry that power to the data center. This is not mere rhetoric but a measured fact.
| Equipment/Process | Lead Time | Major Suppliers | Resolution Outlook |
|---|---|---|---|
| Large transformers (standard) | 128 weeks (about 2.5 years) | Hitachi Energy, GE Vernova, Siemens Energy, HD Hyundai Electric, Hyosung Heavy Industries, LS Electric | 2028+ |
| Generator-side transformers (GSU) | 144 weeks (about 3 years) | Same as above | 2029+ |
| HVDC converters·subsea cables | 3~6 years | Hitachi Energy, GE Vernova, Siemens Energy, Prysmian, Nexans, LS Cable & System·LS Marine Solution | ~2030 |
| Gas turbines (H/J-class) | Slots sold out ~2030 | GE Vernova, Siemens Energy, Mitsubishi, Doosan (new entrant) | 2030+ |
| Transmission line permitting (US PJM) | 5~10 years | Local·federal FERC·environmental regulation | Structural |
| High-voltage circuit breakers·GIS | 40~80 weeks | Hitachi Energy, ABB, Siemens, HD Hyundai Electric | Improving |
| Switchgear·LV/MV distribution | 30~60 weeks | Eaton, Schneider, ABB, LS Electric | Improving |
| Grain-oriented electrical steel for transformers | ~2 years | POSCO·Nippon Steel·AK Steel (now Cleveland-Cliffs) | Expanding |
This isn't simple luck. Two cycles erupted at once: (i) a surge in new generators·transformers for AI data centers, and (ii) the replacement cycle for aging US grid transformers (average transformer lifespan of 30~40 years — the 1980s~1990s installations are now due for replacement). The two overlap, putting global transformer demand at 1.5~2x capacity.
Within each tier, stocks are ordered by a composite assessment of 5-year visibility·barriers to entry·pricing power. All stocks are analysis opinions, not buy recommendations.
Korean stocks are exposed to two markets at once: (i) US grid transformer·generation exports, and (ii) domestic Korean data center transmission·generation. Both have strong 5-year visibility.
If a DeepSeek-style training-efficiency breakthrough happens once more, power consumption per GPU could fall 30~50%. If the demand curve flattens, gas turbine slots could empty before the backlog fills.
If the US 10Y staying at 4.5%+ becomes prolonged, the cost of capital for IPPs·utilities rises. PPA pricing power shifts slightly toward Big Tech. But Big Tech capex is covered by its own cash flow, so a major retreat is unlikely.
A single major accident in the US or Europe within five years would set back both restart and SMR momentum. A tail risk: low probability but large impact.
If global transformer capacity normalizes between 2028~2030, the pricing power of the KR three·Hitachi gradually weakens. But it barely loosens within the 5-year window.
The current structure has purchasers bearing US transformer tariffs — if Trump administration policy changes, it directly affects the margins of the Korean three. But as long as US capacity is short, it's only volatility in the near term; the structure holds.
Betting money on AI data centers themselves is already largely priced in. NVIDIA·Microsoft·Meta·Amazon·Oracle have all set new highs. But the companies that can actually switch on the electricity they need to run are far fewer. Gas turbines are the Big 3, transformers are 5~6 firms globally, nuclear operation is 3~4 firms, and BESS cells are within 5 firms. This narrow group of companies takes 30~40% of Big Tech's $5 trillion in capex over the five years.
To restate the priorities of the 5-year window: