When AI Meets Power Shortages: Why Tech Giants Are Building Their Own Power Plants?
Hello everyone, I'm your financial journalist. Today, we're going to discuss a topic that might seem a bit out of place, but it's actually profoundly changing the global tech landscape: AI companies are racing to secure electricity supplies and are even starting to build their own power plants.
In the past, when we talked about AI, we discussed chips (GPUs), the parameters of large models, and algorithms. But things have changed. Tech giants like Google, Microsoft, and NVIDIA are accelerating their efforts. They're no longer just buying servers; they're directly signing nuclear power agreements, purchasing natural gas turbines, and even investing in entire power grids.
Why? Because we're running out of computing power, but what's even more scarce is electricity.
Next, I'll break down this news into five key points in plain language to help you understand the logic, risks, and future implications of this "power war."
---
1. The Core Problem: Rapid Chip Iteration vs. Slow Grid Expansion
First, we need to understand why tech giants are so anxious.
Think of AI training as running a superfactory. In the past, the bottleneck was the chips (GPUs); the company that bought more chips could process data faster. But now, even with the latest chips, we're running into a problem: there's not enough electricity to keep things running.
There's a huge time mismatch:
- AI models and chips are updated monthly or even weekly. A new model is released today, and by next month, we need more powerful computing power to train the next generation.
- Grid expansion and new power sources take years to build. It can take 5 to 10 years to construct a nuclear power plant or lay out a high-voltage transmission line.
This leads to a tricky situation: tech giants have the most advanced chips, but they can only wait for the grid to expand. As Silicon Valley investor Zhang Lu put it, "Before we run out of GPUs, we'll run out of electricity first."
Previously, electricity was a supporting factor, like water and electricity in a factory—always available. Now, it's become a bottleneck, like a tollbooth on a highway; no matter how many cars there are, if the road isn't open, they can't pass. So, the giants can't wait. They've decided to take control of their own power supply instead of relying on the public grid.
---
2. Strategic Shift: From Buying Electricity to Bringing Your Own Power (BYOP)
In the past, data centers just needed to be plugged into the grid and pay for electricity. But now, the giants have adopted a new approach called BYOP (Bring Your Own Power).
In simple terms, they say, "I won't wait for the grid to supply me with electricity; I'll generate it myself or secure it in advance."
- Google's Nuclear Power Deal: Google invested 13 billion euros in Finland to build a data center and signed a 22-year nuclear power agreement, securing 50% of the power output. It's like signing a 22-year electricity contract before you even move in, ensuring a steady supply at a fixed price.
- Microsoft's Restart Plan: Microsoft signed a 20-year agreement to restart an idle nuclear power plant in Pennsylvania. Why nuclear power? It's stable and low-carbon, ideal for long-term operation.
- Meta and Amazon's Nuclear Alliance: They're not only signing agreements but also investing in small nuclear reactor companies. For example, Amazon invested $500 million in X-energy to build over 5GW of nuclear power by 2039.
Why this move?
Traditional public grids are too slow, and electricity prices are unpredictable. With long-term agreements (PPAs) or direct investments, the giants turn electricity into a predictable cost, rather than a variable that could increase or disappear at any time. It's like switching from paying as you use to having a fixed supply for the long term, providing greater security.
---
3. Aggressive Measures: The Rise of Natural Gas Turbines and Behind-the-Meter (BTM) Power
Although nuclear power is stable, it takes too long to build. For companies in urgent need of computing power, speed is more important than sustainability. Thus, a more aggressive model has emerged: Behind-the-Meter (BTM) power supply.
What is BTM power supply?
Basically, it means installing generators inside the data center walls.
- Traditional Model: Grid → Substation → Data Center.
- BTM Model: Natural gas turbines/energy storage batteries → Data Center.
A report by SemiAnalysis reveals that the global demand for BTM power for AI has reached 75GW (gigawatts, with 1GW = 1000MW, equivalent to 100 large nuclear power plants). In the second quarter alone, another 20GW was added.
Who's doing this?
- xAI (under Musk): To quickly expand its computing power, xAI deployed many portable natural gas turbines in Memphis. It's like setting up large fuel tanks near the construction site to generate power—noisy and polluting, but fast!
- OpenAI and SoftBank: In Ohio, they plan to build at least 10GW of new power capacity and invest $4.2 billion to upgrade the regional grid.
Why natural gas?
Natural gas turbines are quick to build (within a few months to a year) and relatively cost-effective. They're also more environmentally friendly than diesel power (although still a fossil fuel). In the AI race, being online half a year earlier means generating revenue half a year earlier. SemiAnalysis predicts that BTM power could account for half of the electricity needed for new AI data centers in the US by 2028.
---
4. The Economic Calculus: Electricity as a New Asset
Many might wonder: Isn't generating power on your own more expensive? Why do the giants do it?
Because it makes financial sense.
- Revenue Logic: AI services only generate revenue when the GPUs are running. A 200MW data center that's offline for half a year can result in a net present value (NPV) loss of $400 million to $500 million. Although the cost of self-generated power is higher, the benefits of seizing the time window outweigh the expenses.
- NVIDIA's New Definition: NVIDIA now calls AI data centers "AI factories" and considers them a new type of investable asset. Electricity plus computing power equals intelligence, a combination that generates stable cash flows and can be used for financing or debt issuance.
But there are risks:
- Residents Pay the Price? In the US, if data center expansions increase grid expansion costs, these costs could be passed on to consumers through higher electricity prices. This has sparked political debates; for example, the Finnish opposition party is calling for a licensing system to prevent AI from straining the local power supply.
- Debt Pressure: Goldman Sachs estimates that the global AI industry's debt issuance is approaching $500 billion, accounting for 25%-30% of investment-grade corporate debt. If AI profits fall short of expectations, this debt could become a financial risk.
- Environmental Concerns: Natural gas turbines produce emissions and air pollution. xAI has avoided emission permits by using temporary facilities, leading to legal disputes with local communities. Environmental pressures are becoming a hidden obstacle to AI expansion.
---
5. Future Outlook: Electricity as the Slowest Link in AI Expansion
Finally, let's look ahead:
- Electricity Demand Soars: Gartner predicts that global data center electricity consumption will reach 565 terawatt-hours in 2026, a 26% increase. The IEA is even more optimistic, expecting data center electricity to account for nearly 3% of global electricity consumption by 2030.
- Competition Escalates: AI competition has expanded from models, chips, and data centers to power plants, grids, and energy financing.
- The Core Conflict: Large models are updated monthly, GPU architectures annually, but power infrastructure takes decades to build. Electricity is becoming the most challenging and rigid aspect of AI expansion.
In summary:
Tech giants are making a significant shift from the digital world to the physical one. They realize that no matter how smart the algorithms or powerful the chips, without stable electricity, everything comes to a halt.
The future AI giants will not only be tech companies but also energy companies. Those who can obtain electricity more efficiently, cheaply, and reliably will gain an advantage in the next round of AI competition. For ordinary people, this could mean changes in electricity pricing, community disputes, and a tech era defined by electricity supply.
In one sentence: The second half of the AI race is about who has the most electricity, not who is the smartest.