Earth Day has passed. Companies’ ESP/social responsibility reports are beginning to appear and they have been commenting on their progress in their Q1 earnings report. The early data in the semiconductor and microelectronics segment looks encouraging.
The percentage of renewables in their energy portfolio is increasing. Water conservation and water-positive projects are showing positive results, and waste sent to landfills is decreasing. The majority of reports are being released in May and will continue to trickle out until the September time frame.
The sustainability news was positive for 2023, but what does 2024 look like?
The mainstream media, Wall Street Journal, and Bloomberg have finally picked up on the fact that data center growth and the energy growth that comes with it is not sustainable.
While renewables are increasing rapidly, data centers are growing more rapidly, driven by the demand for AI and the massive amount of computing power it takes to run large language training models and store data. It is unlikely that the new renewable projects will come online fast enough to meet the demand in 2024, and companies like Amazon, Meta, and Alphabet may likely need to reset their 100% renewables and net zero goals.
Capital expenditures according to earnings reports summarized in the Wall Street Journal CIO Journal on April 26, 2024, have Meta, Alphabet, and Microsoft all increasing capex for data centers. Meta is projected to spend $40 billion in capex in 2024. Microsoft spent $14 billion and Alphabet/Google spent $12 billion in Q1 2024. If you project that out for the year you are looking at $52 billion and $48 billion respectively.
The spending for data centers is considerable, but these companies are also trying to finance renewables to help them achieve their net zero goals. Figure 2 shows the increase in electricity use at Meta, Alphabet, Oracle, and Microsoft taken from their 2022 ESG reports.
From 2017, Microsoft and Meta have increased electricity consumption by over 60%. Oracle and Alphabet are slightly over 40%. News reports out of Georgia and West Virginia are showing rapid growth of data centers. The concern in Georgia is they will not be able to build or access renewables fast enough, so natural gas will supply the energy required for data centers. Georgia Power estimates that the demand growth will be 6600MWh in the next 7 years an increase of 30% from 2023. In Northern Virginia’s data center alley, there is a similar problem, except instead of natural gas coal is the fuel of choice (Figure 3).
Early in the data center expansion, I heard multiple times, “It’s okay, all of our data centers will run on renewables”. I’m interested in seeing how the ESG reports for 2023 will address this disconnect, and what 2024 ESG reports will look like.
The AI companies are not standing still. Microsoft signed a 10.5 GW deal to support its data center expansion. This will start coming online in 2026. Alphabet purchases more renewable power than they need to offset the lack of renewables in other countries. It’s possible that some wind power might come to support the North Virginia data centers, as solar in that region is challenging. it will be a good case study to see how the AI and the power suppliers manage this.
Compounding the situation is that the chips used for AI are getting bigger and will use more energy. At TSMC’s recent technology symposium, they discussed how advanced packaging technology will enable putting together chip systems that consume significantly more power than current technology.
Using the CoW-SoW, technology TSMC can grow the amount of silicon in a package from three reticle sizes worth of silicon in 2024 to 5.5 reticles worth of silicon in 2027. This increases the computing power by 3.5 times so that you are now drawing thousands of watts of power instead of hundreds of watts of power. The compute performance will also go up considerably, but does it offset the increase in power, and more importantly, will the industry continue up the power curve, or find ways to reduce the amount of power needed for computing?
Tesla’s Dojo training in Figure 5 shows how these systems are put together and the complexity of the system. It will be interesting to see if the new tile helps to reduce data center power, or as the industry has done historically it keeps ramping up the power consumption.
McKinsey reports that data center energy demand was nearly 17 gigawatts in 2022, and is projected to reach 25 gigawatts in 2026. A large nuclear generates about 1 gigawatt of power. So a lot of renewables will need to be added to meet ESG goals.
To draw attention to the AI power demand, Salesforce is calling for companies to disclose their AI emissions publicly. What might be a better measurement is a power-to-learning output. The large language models consume most of the power in training, the chips are increasing in size and power. We are told that along with the increase in power also comes a performance improvement, but does that guarantee that there will be a decline in power consumption?
History and McKinsey tell us otherwise. Even though the power performance equation has improved with the introduction of new technology nodes. The users have managed to take full advantage of both and have increased power consumption over time. This time doesn’t appear any different. The end users are taking full advantage of both the power and performance improvements we will see from the next-generation packaging. The only real question is will it be powered by renewables, as they have promised, or will they need to fudge a little bit.
