Why the AI industry’s thirst for new data centers can’t be satisfied

2 weeks ago 16

Copyright &copy HT Digital Streams Limited
All Rights Reserved.

premium wsj

Tom Dotan , Asa Fitch , The Wall Street Journal 5 min read 24 Apr 2024, 03:38 PM IST

 iStock) ( ) Demand for machine servers equipped with caller generations of AI chips—the astir fashionable of which are graphics processing units, oregon GPUs, from Nvidia—is overwhelming existing information centers. (Image: iStock) ( )

Summary

The pb clip to get customized cooling systems is 5 times longer than a fewer years ago, and transportation times for backup generators person gone from arsenic small arsenic a period to arsenic agelong arsenic 2 years.

The frenzy to physique information centers to service the exploding request for artificial quality is causing a shortage of the parts, spot and powerfulness that the sprawling warehouses of supercomputers require.

The pb clip to get customized cooling systems is 5 times longer than a fewer years ago, information halfway executives say. Delivery times for backup generators person gone from arsenic small arsenic a period to arsenic agelong arsenic 2 years.

A dearth of inexpensive existent property with casual entree to capable powerfulness and information connectivity has builders scouring the globe and getting creative. New information centers are planned adjacent to a volcano successful El Salvador and wrong shipping containers parked successful West Texas and Africa.

Earlier this year, data-center relation Hydra Host recovered itself successful a bind, searching for 15 megawatts of powerfulness needed to run a planned installation with 10,000 AI chips.

The institution went from Phoenix to Houston to Kansas City, Mo., to New York to North Carolina to find the close space. It is inactive connected the hunt.

The locations that had the powerfulness didn’t person the close cooling systems required to support the servers operational. New cooling systems would instrumentality six to 8 months to arrive, acknowledgment to a proviso crunch. Meanwhile, buildings that had the cooling didn’t person the transformers required to person the further power—those would instrumentality up to a twelvemonth to arrive.

“With what we’re seeing, the fervor to physique is astir apt the top since the archetypal dot-com wave," said Hydra Host Chief Executive Aaron Ginn. He said the hunt for the close parts and abstraction has taken months longer than expected.

The request for computational powerfulness to make AI systems has surged since precocious 2022, erstwhile OpenAI’s ChatGPT started showing the technology’s potential. Demand for machine servers equipped with caller generations of AI chips—the astir fashionable of which are graphics processing units, oregon GPUs, from Nvidia—is overwhelming existing information centers.

“You had this tsunami, and there’s going to beryllium a shortage of data-center inventory," said Raul Martynek, main enforcement of data-center institution DataBank.

Creating and deploying analyzable AI systems requires unprecedented numbers of chips. Analysts estimation that grooming the mentation of ChatGPT that came retired successful 2022 required much than 10,000 of Nvidia’s GPUs, portion much caller updates person required importantly more—putting further strain connected information centers. Large tech companies person struggled to get their hands connected supplies.

The magnitude of data-center abstraction successful the U.S. grew 26% past year, according to real-estate steadfast CBRE, and a grounds magnitude was nether construction. The terms of disposable abstraction is rising portion vacancy rates are negligible—a motion that proviso isn’t keeping up with demand.

Bill Vass, vice president of engineering astatine Amazon Web Services, said a caller information halfway pops up determination successful the satellite each 3 days.

It mostly takes a twelvemonth and a fractional oregon 2 years to enactment up a large, caller information facility, said Jon Lin, the wide manager of data-center services astatine Equinix, 1 of the world’s biggest data-center operators. It is hard for the manufacture to abruptly standard up erstwhile request skyrockets due to the fact that of the extended readying and supply-chain absorption required, helium said.

“These are not casual problems wherever you tin pivot connected a dime and triple capableness precise quickly," said Lin.

Cloud giants similar Amazon Web Services, Microsoft and Alphabet’s Google are investing billions of dollars successful caller information centers. Google’s superior expenditures—almost fractional of which was connected its information infrastructure—jumped 45% from a twelvemonth earlier to $11 cardinal successful the 3 months done December.

Microsoft has tried to power costs during its data-center build-out by pulling backmost connected spending successful different areas. The institution spent much than $30 cardinal connected information centers successful 2023, according to an estimation from market-research steadfast Dell’Oro Group.

The unreserved to physique has ratcheted up the clip needed to get immoderate captious data-center components. Transceivers, which link antithetic networks of servers, present instrumentality months longer to get than before. Labor costs person besides go an contented arsenic data-center builders person faced a shortage of operation workers trained connected these types of delicate installations. Getting smaller networking components, similar cables that link antithetic racks of servers, tin instrumentality months.

The AI chips needed to marque the exertion enactment are famously successful abbreviated supply. The azygous biggest maker, Nvidia, has been overwhelmed by request from a scope of companies and cloud-infrastructure providers. At the outset of the AI boom, the pb clip to get graphics processing units for AI computation successful information centers roseate to respective months, though the hold clip has fallen successful caller months.

The clip it takes to get ample backup powerfulness generators has gone from little than a twelvemonth to astir 2 years, data-center executives say.

The circumstantial requirements for the caller GPU-driven AI information centers are pushing builders to look for places wherever they tin get tons of reliable, affordable electricity. Amazon precocious bought a information halfway adjacent to a atomic powerfulness works successful Pennsylvania. Meta Platforms is readying $800 cardinal of computing infrastructure successful El Paso, Texas.

Digital-infrastructure institution Standard Power is readying to proviso powerfulness to information centers successful Ohio and Pennsylvania with modular atomic reactors akin to those utilized to powerfulness immoderate submarines and craft carriers. NuScale Power, based successful Portland, Ore., says it has the support to proviso the reactors to Standard.

“The data-center marketplace needs powerfulness yesterday," said Clayton Scott, NuScale’s main commercialized officer.

Armada, a San Francisco startup, builds information centers wrong of shipping containers. The institution tin driblet these portable facilities, afloat of Nvidia chip-powered servers, successful locations specified arsenic distant areas of Texas oregon Africa that are adjacent inexpensive sources of powerfulness similar state wells.

El Salvador—after trying to physique a amended concern situation with a crackdown connected gangs—has slashed taxes connected AI. Marc Seal, who runs a institution that invests successful AI information centers, is considering tapping the country’s volcanoes to powerfulness specified information centers utilizing green, geothermal power.

Write to Tom Dotan astatine tom.dotan@wsj.com and Asa Fitch astatine asa.fitch@wsj.com

Catch each the Business News, Market News, Breaking News Events and Latest News Updates connected Live Mint. Download The Mint News App to get Daily Market Updates.

more

Read Entire Article