- AI data centers are expected to consume up to 9% of all electricity generated in the United States by 2030.
- AI GPUs have already tripled their power consumption, with Nvidia Blackwell raising the bar as the GB200 is expected to use up to 2,700 watts of power, up from 250 watts for the earlier A100.
- Power consumption is the chokepoint for AI data centers that are desperate to lock in power purchase agreements (PPAs) to procure long-term power.
The artificial intelligence (AI) revolution is driving an intensely competitive race across various facets, from GPUs and ASICs to CPUs, storage, LLM models, and beyond. However, one critical component stands out as the key enabler for AI’s future: power.
Simply put, AI cannot exist without the electricity that powers its applications, making energy the chokepoint for AI data centers. Electricity must be generated to keep these data centers running at full capacity, and there are five primary types of fuel sources: coal, natural gas, solar, nuclear and fuel cells.
AI and Data Centers Are Driving Up Power Consumption:
Power consumption is rising with each generation of GPUs. Nvidia’s A100 max power consumption was 250W. Its H100 GPU consumes 350 to 350W and up to 700W with SXM. IO Fund wrote, “Nvidia’s upcoming Blackwell generation boosts power consumption even further, with the B200 consuming up to 1,200W, and the GB200 (which combines two B200 GPUs and one Grace CPU) expected to consume 2,700W. This represents up to a 300% increase in power consumption across one generation of GPUs with AI systems increasing power consumption at a higher rate.”
Join thousands of investors who trust I/O Fund’s expert stock analysis on AI, semiconductors, cryptocurrency, and adtech — sign up for free! Click here!
AI data centers need to meet this power demand, which is expected to surge with each generation of GPUs, which are now expected to roll out annually (rather than bi-annually). A single rack is used to consume 10 to 15 kilowatts (KW) of power, but current AI racks are now averaging 80 KW. The next generation is expected to draw 120 KW to 150 KW towards 200 KW within the next few years. This will drive data center consumption by 160% by 2030, where data centers could draw 9% of all the electricity produced in the United States.
Using Thermal Efficiency to Rank Fuel Sources
AI data centers are actively trying to secure long-term power purchase agreements (PPA) and are even exploring nuclear energy options, as evidenced by the 20-year PPA Microsoft signed with Constellation Energy. Hyperscalers are also co-locating data centers near power sources to ensure reliable electricity.
To ensure reliable power, hyperscalers are increasingly seeking long-term power purchase agreements (PPA) and exploring various energy options, including nuclear power. This was underscored by the 20-year PPA that Microsoft signed with Constellation Energy marking the largest-ever PPA in its history. Constellation Energy will launch the Crane Clean Energy Center, restoring the Three Mile Island Unit 1 nuclear reactor, which will add 835 MW of carbon-free energy to the electrical grid.
While electricity is generated from various energy fuels, not all fuels produce the same amount of energy. Thermal efficiency measures how effectively a fuel is converted into electricity. For example, a thermal efficiency of 25% means that 75% of the fuel is lost as heat. Higher thermal efficiency means more energy is converted into electricity, while lower efficiency results in greater energy loss. Let’s see how the energy fuel sources size up.
Premium members receive real-time trade alerts on Nvidia (including 9 we sent under $20 in 2022), as well as deep-dive research on numerous AI stocks and crypto that we track closely. Learn more here.
Sizing Up the Five Types of Energy Fuel
These are the five energy fuel sources and their general thermal efficiency. These are general thermal efficiencies, which can vary based on location, technology and method:
- Solar PV: While solar photovoltaic (PV) panels produce renewable clean energy, they are also the least efficient, averaging 15% to 20%, or 17.5% midpoint. They generate electricity directly from the sunlight. Of course, sunlight is not available 24/7, and efficiencies can drop during overcast and precipitation.
