Photo from Business Insider.
The rush to construct data centers to meet the rapidly growing demand for artificial intelligence is leading to shortages of the components, real estate, and energy needed for these vast complexes of supercomputers.
Data center executives report that the time required to obtain custom cooling systems has increased fivefold compared to just a few years ago. Similarly, delivery times for backup generators have stretched from as short as one month to as long as two years.
A shortage of affordable land with convenient access to adequate power and data connectivity is prompting builders to search worldwide and innovate. Plans for new data centers include locations next to a volcano in El Salvador and within shipping containers situated in West Texas and Africa.
Earlier this year, data center operator Hydra Host encountered difficulties as it searched for 15 megawatts of power required to run a planned facility housing 10,000 AI chips.
The company searched for suitable space, traveling from Phoenix to Houston, then to Kansas City, Missouri, followed by New York and North Carolina. However, they have yet to find the ideal location and are still actively searching.
Locations with sufficient power lacked the necessary cooling systems to keep the servers operational. Due to a supply shortage, it would take six to eight months for new cooling systems to arrive. Conversely, buildings with adequate cooling lacked the transformers needed to handle additional power, which would take up to a year to arrive.
Hydra Host Chief Executive Aaron Ginn remarked that the current enthusiasm for construction is likely the highest since the first dot-com wave. He mentioned that the search for suitable parts and space has taken several months longer than anticipated.
Since late 2022, when OpenAI's ChatGPT demonstrated the potential of AI technology, there has been a significant increase in the demand for computational power. This surge in demand for computer servers, equipped with the latest generations of AI chips—primarily graphics processing units (GPUs) from Nvidia—is straining existing data centers.
Raul Martynek, CEO of data-center company DataBank, described the surge as a "tsunami" and predicted a shortage of data-center inventory.
Developing and deploying complex AI systems demands an unprecedented number of chips. Analysts estimate that training the 2022 version of ChatGPT required over 10,000 Nvidia GPUs, while subsequent updates have demanded even more, adding to the pressure on data centers. Large tech firms have faced challenges in obtaining sufficient supplies.
According to real estate firm CBRE, data-center space in the U.S. expanded by 26% last year, with a record amount under construction. Rising prices for available space and negligible vacancy rates indicate that supply is failing to meet demand.
Bill Vass, Amazon Web Services' Vice President of Engineering, stated that a new data center is established somewhere in the world approximately every three days.
Jon Lin, the general manager of data-center services at Equinix, one of the world’s largest data-center operators, stated that it typically takes a year and a half to two years to construct a large new data facility. He explained that the industry finds it challenging to rapidly scale up when demand surges due to the extensive planning and supply-chain management involved.
Lin emphasized that these challenges are not easily solved, and it's not possible to quickly pivot and triple capacity.
Major cloud providers such as Amazon Web Services, Microsoft, and Alphabet's Google are pouring billions of dollars into the construction of new data centers. Google's capital expenditures, with nearly half allocated to its data infrastructure, surged by 45% compared to the previous year, reaching $11 billion in the three months ending in December.
To manage costs during its data-center expansion, Microsoft has reduced spending in other areas. According to an estimate by market research firm Dell'Oro Group, the company invested over $30 billion in data centers in 2023.
The urgency to construct data centers has extended the time required to obtain certain critical components. Transceivers, used to connect various server networks, now take several months longer to arrive compared to before. Additionally, labor costs have become a concern due to a shortage of construction workers trained in these specialized installations. Acquiring smaller networking components, such as cables linking different server racks, can also take several months.
The AI chips essential for powering this technology are notoriously scarce. Nvidia, the largest producer, has been inundated with demand from various companies and cloud infrastructure providers. Initially, as the AI demand surged, the lead time to acquire graphics processing units (GPUs) for AI computation in data centers stretched to several months, although this waiting period has decreased in recent months.
Data-center executives report that the time required to obtain large backup power generators has increased from less than a year to nearly two years.
The particular needs of the new GPU-driven AI data centers are prompting builders to seek locations with abundant, dependable, and cost-effective electricity. Amazon recently acquired a data center near a nuclear power plant in Pennsylvania, while Meta Platforms is preparing to invest $800 million in computing infrastructure in El Paso, Texas.
Standard Power, a digital infrastructure company, intends to provide power to data centers in Ohio and Pennsylvania using modular nuclear reactors similar to those employed in submarines and aircraft carriers. NuScale Power, headquartered in Portland, Oregon, has received approval to supply these reactors to Standard Power.
Clayton Scott, NuScale’s chief commercial officer, emphasized the urgent need for power in the data-center market.
Armada, a San Francisco-based startup, constructs data centers within shipping containers. These portable facilities, equipped with Nvidia chip-powered servers, can be deployed in locations such as remote areas of Texas or Africa, close to cost-effective power sources like gas wells.
Following efforts to improve the business environment through measures such as cracking down on gangs, El Salvador has reduced taxes on AI. Marc Seal, who manages a company investing in AI data centers, is contemplating utilizing the country's volcanoes to power these data centers with environmentally friendly geothermal energy.