The entire market has overlooked it! Morgan Stanley: Data center computing power costs are rapidly declining

As AI technology enters a new era of popularization and application, the demand for computing power is growing rapidly, while the cost of computing power is decreasing with technological advancements.

Morgan Stanley pointed out in its AI report on March 24 that with the continuous improvement of GPU technology, such as NVIDIA's development from the Hopper to the Blackwell GPU architecture, the cost of AI computing power is significantly decreasing.

Looking ahead, Morgan Stanley predicts that the cost of AI computing power will rapidly decline, the demand for AI computing power will grow vigorously, and it is estimated that by 2024/27, global data center electricity demand will account for approximately 2%/4% of the world's total, but the power infrastructure may not keep up with this growth trend.

Therefore, Morgan Stanley pointed out that companies that can meet the rapidly growing electricity demand for AI will benefit from this trend, especially those that provide power solutions capable of reducing power supply delays in data centers.

Rapid Decline in AI Computing Power Costs

Morgan Stanley pointed out that with the improvement in GPU efficiency, data center computing costs are rapidly decreasing. Taking NVIDIA's latest release of Blackwell as an example, it only consumes half the energy of Hopper. Specifically:

Data center models show that the capital cost per teraFLOPs (trillions of floating-point operations per second) for data centers using Hopper GPUs is $14.26, while for Blackwell data centers, the cost is $7.48. This means that when upgrading from Hopper GPUs to Blackwell GPUs, the capital cost per teraFLOPs decreases by approximately 50%.

In terms of data center construction costs, the total facility capital cost for a Blackwell data center (including GPUs) is $3.349 billion, while for a Hopper data center, the total facility capital cost is $3.583 billion. These costs are based on assumptions for a 100MW data center, including 8 chips/servers and 70% server utilization.

Assuming an electricity cost of $100 per megawatt-hour, the annual electricity cost for a Hopper data center is approximately $0.24 per teraFLOPs, while the annual electricity cost for a Blackwell data center is approximately $0.14 per teraFLOPs.

This decline is mainly due to the improvement in GPU energy efficiency with technological progress, allowing for more computing power to be provided under the same power consumption

AI computing power demand surges, posing challenges to power infrastructure

Meanwhile, Morgan Stanley predicts that AI power demand will grow rapidly. In the base case scenario, it is estimated that the global data center power demand (including GAI) for 2023-27 will be 430-748 terawatt-hours (TWh), equivalent to 2%-4% of global power demand in 2024/27.

Specifically:

In the base case scenario (GPU utilization increasing from 60% to 70%), the total power capacity of data centers for 2023-27 will be 70/122 gigawatts. The compound annual growth rate of GAI power demand for 2023-27 is ~105%, while the compound annual growth rate of global data center power demand (including GAI) during the same period will be ~20%.

In the bull market scenario (reflecting 90% chip utilization), the estimated global data center power demand for 2023-27 will be ~446/820 terawatt-hours.

And in the bear market scenario (reflecting 50% utilization), the estimated data center power demand for 2023-27 will be ~415/677 terawatt-hours.

Furthermore, Morgan Stanley also mentioned reducing the percentage of renewable energy services in new data center power, as data (such as Amazon/Talen transactions) indicates that non-renewable energy will play a greater role.

Looking further ahead, Morgan Stanley points out that according to forecasts for global data center power demand, it is expected to pose challenges to power infrastructure, including transmission line capacity constraints, planning and permitting delays, and supply chain bottlenecks. Therefore, looking to the future, companies that can meet the rapidly growing energy demand for AI, especially power solution providers that can reduce power delays in data centers, will be welcomed in the market.

Morgan Stanley states that data center growth varies globally, but similarities exist in that data center companies and hyperscale companies often collaborate with power developers to minimize power delays, reduce costs, and emissions.

In the United States, we see initial signs of partnerships, with hyperscale companies and nuclear plant owners building new large data centers at nuclear plants, benefiting power utilities like Constellation Energy and independent power producers like Vistra. We also see the potential for on-site generation, which may benefit many power stocks.

In Europe, by 2035, data center demand in European countries will drive a fivefold increase in power, undervalued European power stocks will benefit, including private multinational power companies like Iberdrola in Spain and energy firm Fortum in Finland.

In ASEAN, Malaysia, Singapore, and Thailand have seen strong growth in power demand. To minimize "time to power," we see data center developers partnering with local power companies in ASEAN: NVIDIA and Malaysia's YTL Group are collaborating to build a $4.3 billion artificial intelligence data center in Malaysia. Thailand's largest private power company, Gulf Energy, and Thailand's telecom company, AIS, have started constructing a new data center in Thailand.

Singapore's telecom company, Singtel, Indonesia's telecom company, Telkom, and Indonesia's energy company, Medco Power, have established a strategic partnership for their first data center project in Indonesia