Topics
former
AI
Amazon
Image Credits:Romilly Lockyer / Getty Images
Apps
Biotech & Health
Climate
Image Credits:Romilly Lockyer / Getty Images
Cloud Computing
Commerce Department
Crypto
Enterprise
EVs
Fintech
fund-raise
appliance
Gaming
Government & Policy
ironware
Layoffs
Media & Entertainment
Meta
Microsoft
Privacy
Robotics
Security
Social
Space
startup
TikTok
shipping
Venture
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
adjoin Us
Chinese AI startup DeepSeek stunned the world with the release of its R1 model , which appears to performnearly as wellas lead mannequin from Google and OpenAI , despite the company ’s title that it used a comparatively meek number of GPUs to prepare it .
DeepSeek ’s comparative efficiency hasexperts and investorsquestioning whether AI really need the monumental hardware outlay everyone had been forebode . And that could change information nerve centre demand — and the energy needed to power them .
The company claims it ran 2,048 Nvidia H800 GPUs for two month to train aslightly older framework , a fraction of the compute that OpenAI is rumour to use .
Few companies are as unwrap as Nvidia , the part price of which was down 16 % at the time of publication . Perhaps even more vulnerable are the startups and magnate producers that are betting big on new nuclear and natural gas capacity .
Nuclear power , in peculiar , has been on the cusp of a renaissance for years , driven by betterment in fuel and reactor design that promise to make a novel generation of tycoon plants secure and tawdry to build and operate . Until now , there was short reason to blaze out ahead . Nuclear is still expensive relation to idle words , solar , and lifelike gas . Plus , next - genesis nuclear has yet to be test at commercial scale .
The surge in power demand from AI changed the equation . With datum centers predicted to consume as much as 12 % of all electricity in the U.S. — more than ternary their share in 2023 — and prognosis of underpowered AI data point centers by 2027 , tech company have been racing to secure newfangled supply , and throwing billions of dollars at the problem . Google has pledged tobuy 500 megawatt of capacityfrom atomic inauguration Kairos , Amazon led a$500 million investmentin another nuclear startup , X - Energy , and Microsoft is working with Constellation Energy on a $ 1.6 billionrenovation of a reactorat Three Mile Island .
But what if the problem has been grandiloquent ?
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
There is no hard and fast pattern suggesting that the only way to improve AI performance is to practice more compute . For a while , that tactic process well , but more of late , more computehasn’t yielded the same results . AI researchers have been drop about for solutions , and it ’s potential that DeepSeek found one for its R1 manikin .
Not everyone is convince , of class .
“ While DeepSeek ’s accomplishment could be groundbreaking , we question the impression that its feats were done without the use of modern GPUs , ” Citigroup analyst Atif Malikwrote .
Still , history suggests that even if DeepSeek is veil something , someone else will belike find a way to make AI cheaper and more efficient . After all , it ’s easier and potentially quicker to tax some PhDs with developing good models than it is to progress new power industrial plant .
The current wave of new reactors are n’t scheduled to add up online until 2030 , and young natural gas power plants wo n’t be available until the end of the decade at the soonest . In that context , technical school companies ’ power investments appear to be hedge in case their software system wager do n’t pan out .
If they do , expect tech companies to scale back their power ambitions . When chip in the pick between spend billions on physical assets or software , tech companies almost always chose the latter .
Where will that entrust atomic inauguration and energy companies ? It depend . Some might be able-bodied to produce power at a humble enough cost that it wo n’t matter if AI ’s business leader needs wane . The globe is electrifying , and even before the AI bubble start inflate , demand for electricity wasexpected to grow .
But absent need from AI , those toll pressing are probably break to increase . Wind , solar , and battery are gaudy and getting cheap , and they ’re inherently modular and mass - bring forth . Developers can tramp out new renewable plants in phases , delivering electricity ( and tax revenue ) before the intact project is consummate while offering some control over their future tense in the fount of incertain need . The same ca n’t be allege of a atomic reactor or a flatulency turbine . Tech companies know this , which is whythey’vebeenquietlyinvestinginrenewablesto power their datum centers .
Few mass predict the current AI boom , and it ’s unlikely that anyone know how the next five years will play out . As a result , the safer wager in energy will belike flow to proven technologies that can be apace deployed and scaled consort to a rapidly evolving marketplace . Today , renewables fit that posting .
