《ORF：2025全球人工智能“DeepSeek 时刻”：影响及启示（英文版）.pdf》由会员分享，可在线阅读，更多相关《ORF：2025全球人工智能“DeepSeek 时刻”：影响及启示（英文版）.pdf（29页珍藏版）》请在三个皮匠报告上搜索。

1、 2025 Observer Research Foundation.All rights reserved.No part of this publication may be reproduced,copied,archived,retained or transmitted through print,speech or electronic media without prior written approval from ORF.SPECIALREPORTno.260Global AIs DeepSeek Moment:Impact and Implications Rahul Ba

2、traM A Y 2 0 2 52The United States(US)is the current world leader in the race to develop the most powerful advances and applications of Artificial Intelligence(AI),but China has steadily caught up in recent years.The launch in January 2025 of DeepSeek-R1,a powerful AI(reasoning)model developed and o

3、ffered by a Chinese startup under open-source technology norms at a fraction of the current market costs,is being considered a milestone moment in this race.Given global AIs underwhelming socio-economic impact and commercial returns thus far,DeepSeek offers a paradigm shift in its accessibility and

4、influence.This report assesses the weight of this moment across pillars of the international system,and reviews how countries like India can respond to it.AbstractAttribution:Rahul Batra,“Global AIs DeepSeek Moment:Impact and Implications,”ORF Special Report No.260,May 2025,Observer Research Foundat

5、ion.3The theme of the Global Technology Summit held in April in New Delhi under the auspices of Indias Ministry of External Affairs1 was Sambhavnaa Hindi-language word meaning“possibility”.It sums up the progressive,albeit cautious,sentiment expressed by Foreign Minister S.Jaishankar at the event2th

6、at in a talent-and data-centric world,disruptions to the status quo can lead to opportunities for exponential leaps,given there is political will.With this in mind,it is worth recalling Chinese President Xi Jinpings claim sometime in March 2023 that the West,led by the United States(US),was suppress

7、ing Chinas development and growth.3 Besides the realm of Chinas foreign trade,Xi referred to US crackdowns on Chinese telecommunications giant Huawei and social media platform TikTok.By January 2025,however,then IntroductionaTheprojectwillexpanddatacentresandcomputationalinfrastructureintheUS.bCompu

8、tecapacityreferstotheamountofcomputationalpowerasystem,server,ornetworkcanprovidetoperformtasks,processdata,orrunapplications.newly elected President Donald Trumps“Xi or me”argument4 faced a more difficult challenge.Shortly after Trumps announcement of the Stargate Project,5 a new US$500-billion pub

9、lic-private commitment to the USs Artificial Intelligence(AI)future,a China responded with DeepSeek.6 DeepSeek is a startup with a“hacker spirit”that has disrupted7 the most advanced practices in global AIs approach to compute capacityb and algorithmic output.Essentially,it promises a more efficient

10、,affordable,and adaptable approach to future AI development,undercutting8 the USs“moat”comprising export controls9 and the CHIPS and Science Act.10The rest of this report analyses the impact of this technological milestone across pillars of the global system.It ponders how a rising power like India

11、can respond.4technological competition.13 The launch led to a slide in US stock markets,directly impacting Big Tech companies such as Nvidia,OpenAI,Meta,Microsoft,and Google.The Nasdaq Composite index lost more than US$1 trillion in a single day,on 27 January 2025,a week after the announcement.14 De

12、epSeeks cheaper but capable open-source developments promise efficiency and productivity gains,15 especially comparing OpenAIs resource-heavy approach to scalability and its“closed”approach16 in achieving commercial goals,strategy,and dealings.c Various experts have termed Chinas launch of DeepSeeks

13、 latest model R111 as AIs“Sputnik moment”,12 drawing a parallel to the Soviet Union launching the first artificial satellite from Earth in 1957,which stunned the US government at the time and pushed it to boost its own Apollo space programme.The launch of DeepSeek-R1 on the day of Trumps inauguratio

14、n has added to speculations that it is linked to the US-China Geopolitics,Geoeconomics,and GeotechnologycInFebruary2023,ElonMusk,oneoftheco-foundersofOpenAI,said,“OpenAIwascreatedasanopensource(whichiswhyInameditOpenAI,)non-profitcompanytoserveasacounterweighttoGoogle,butnowithasbecomeaclosedsource,

15、maximum-profitcompanyeffectivelycontrolledbyMicrosoft.”See:https:/ analysis suggests17 that DeepSeek is an expected outcome of the trade restrictions being imposed by the US on China18(directly and through partner nations/networks)for Nvidias top-grade H100 chips.d This move forced the Chinese start

16、up to innovate to overcome constraints.While some reports claim that the founder and CEO of DeepSeek,Liang Wenfeng,had amassed a large stockpile(10,000-50,000)of the chips before the restrictions on acquiring them came into effect,19 other accounts suggest that DeepSeek acquired these chips through

17、illegal channels via intermediary nations like Singapore and the United Arab Emirates(UAE),even using OpenAIs proprietary technology to develop its models.20 Therefore,assessing this development as a“Sputnik moment”may be premature,if not altogether misleading.21 The long-term sustainability of faci

18、litating the above-mentioned computational infrastructure to develop these frontier AI(large-language)models is also a subject of global debate.22 Advanced data centres have a steep cost.The financial and infrastructure investments impact the environment(Figure 1),such as by distorting the power sup

19、ply(Figure 2)and worsening air quality.dTheNvidiaH100iscurrentlythemostpowerfulgraphicsprocessingunit(GPU)chiponthemarket.Capableofprocessinglargeamountsofdata,itisidealforAIapplicationsthatrunonlargelanguagemodels(LLMs).See:https:/www.weforum.org/videos/what-is-h100-gpu-chip-ai-nvidia/Figure 1:The

20、Environmental Cost of AISource:Stanford AI Index Report 2025236Figure 2:The Environmental Cost of AISource:Energize Capital24The question that arises is whether Chinas evolving worldview can be considered as more“open”than the USs increasingly“closed”,America-first stance.As Chinas Foreign Minister

21、Wang Yi has said,open-source AI is an opportunity for China to support Global South countries in developing their own AI ecosystems.25 According to a white paper26 by the China Academy of Information and Communications Technology,as of July 2024,the number of AI large language models(LLMs)worldwide

22、was 1,328,with 36 percent originating in China.China is also the global leader in publishing research papers on AI(Figure 3).7Figure 3:U.S.-China Rivalry in AI ResearchSource:Allen Institute for Artificial Intelligence278Less hawkish observers are of the view that DeepSeek is just a Chinese upstart

23、with young Chinese talent harbouring“open source zeal”,coming at a global opportunity in their own culturally and organisationally unique manner.28 Their reputation precedes them,from years of being seen as followers of western technology trends(servicing them at best)incapable of path-breaking inno

24、vation in software engineering on account of poor organisational and incentive structures;seeking global validation and approval;landing a milestone moment forced by constraints(known historical trend in deep-tech research and development endeavors).The timing of the launch of DeepSeeks first-genera

25、tion reasoning model,R1,likely had less to do with Trumps inauguration and more with wrapping up a work milestone before the slowdown leading up to the Chinese New Year.Rather than the alarmist“Sputnik moment”comparison,the DeepSeek launch is being likened to Googles 2004 S-1 filing,which was a publ

26、ic acknowledgment and sharing of the leaps in search-engine technology research and development at the time.29 DeepSeek R1,trained via large-scale reinforcement learning(RL),draws on American company Metas open-source LLM Llama.30 Training the model reportedly cost the startup less than US$6 million

27、.31 In April 2023,OpenAI CEO Sam Altman claimed that they had spent over US$100 million just in training the GPT-4 model.32 Alternative Perspectives:Minimising Politics,Maximising Potential9Indeed,analysts surmise that DeepSeeks low-cost claims do not consider factors such as training and research o

28、n the model;33 nevertheless,the 30 x difference in pricing34(Figure 4)between OpenAI and DeepSeekwith OpenAIs o1 model costing US$60 per million output tokens compared to DeepSeek-R1s US$2.19has drawn global attention.Figure 4:DeepSeeks Pricing DisruptionSource:Stanford AI Index Report 20253510Compu

29、ter scientist and entrepreneur Andrew Ng,who has played a central role in both the American and Chinese deep-tech ecosystems,highlighted some key points about the development.36 These include the implications for the AI supply chain of China catching up to the US in generative AI;open-source AI mode

30、ls commodifying the foundation-model layer,creating opportunities for application builders;and reducing training costs due to algorithmic innovations,proving that scaling up processing power is not the only path to AI progress.Economies of scale aside,from a purely technical perspective,DeepSeek has

31、 cut down the processing power required for LLM computation by scheduling and batching their foundational-model training tasks.37 The model does this by using a“mixture-of-experts”approach,which is similar to the human brain,where numerous smaller neural networks are used and activated independently

32、 for specialised tasks.This approach results in less memory being required to train the overall model.Additionally,DeepSeek switched from the standard,time-consuming,and expensive supervised-fine-tuning(SFT)technique for LLM training to a rules-based reward system for learning to reason,called reinf

33、orcement learning(RL),and then combined both techniques to remove difficult-to-read mixed language responses.38 Ng also warns against US companies pushing for regulation to stifle open-source development by inflating hypothetical AI dangers such as superintelligencee and human extinction.39 Open-sou

34、rce/weight models have become a key part of the global AI supply chain,and many companies will use them.If the US continues to stymie the development of open-source models,China will begin to dominate this segment,which might result in many businesses using models that reflect Chinas values rather t

35、han the USs.A 2025 Stanford University study positioned China as the top nation for AI patent applications and the second-largest contributor to global AI behind the US(see Figure 5).eArtificialsuperintelligence(ASI)isahypotheticalentityoragent(AI)thatpossessesintelligencesurpassingthatofthehumanmin

36、dinsomeorallmeasures.See:https:/ 5:Chinas Global AI Leadership Source:Stanford AI Index Report 202540 12Historically,as in the case of energy consumption via fossil fuels(e.g.,industrial coal and transportation gas),41 increased energy efficiency results in lower fuel consumption,thereby reducing ov

37、erall dependence on and demand for the fuel.This results in better resource planning and a more sustainable future.However,a paradoxical theory,known as Jevons Paradox,42 states that the“rebound effect”of this energy efficiency can lead to increased consumption of a technology.This theory suggests t

38、hat wider adoption of the DeepSeek approach and model could democratise AIs capability and efficiency,driving up the global demand for AI production and the financial and infrastructural investments required for it.It is worth examining whether this global democratisation of AI(and the resulting inc

39、rease in demand)could lead to a review of the USs export control policy.Possibilities include additional restrictions geared towards self-interest or a relaxation to serve the increased global demand for the H100 Nvidia semiconductors to participate in and control valuable supply chains.Early signs

40、suggest that the 2022 chips ban43 passed by the Biden administration might have backfired.While its intention was to delay Chinas rapid progress in advanced AI using high-end US silicon chips,the policy might have spurred China to invest in chips research and development(R&D)of its own.More Competit

41、ion and Collaboration13In April 2025,the Trump administration extended the ban on US AI chip exports to China by including the Nvidia H20 processor(scaled down version of the already banned H100).In response,Huawei immediately announced44 its next-generation processor called the Ascend 920,f expecte

42、d to compare with the Nvidia H100s performance.It is due to hit mass production later this year.However,it is clear that DeepSeek is not an outcome of any major technological advance in the Chinese semiconductor value chain(see Figure 6),but rather the result of innovation driven by constraints in c

43、omputational capacity and technical brilliance in AI model R&D techniquesboth despite the lack of access to high-end US chips due to trade restrictions.Figure 6:The Global Semiconductor IndustrySource:McKinsey&Company45fTheAscend920isanupgradeonHuaweismostpowerfulcurrentoffering,theAscend910Cwhichse

44、eminglydeliversabout60percentoftheUS-basedglobalGPUmarket-leaderNvidiaH100sinferenceperformance.See:https:/ for the US,there is a potential scenario in which the universal success of DeepSeek causes a democratisation of AI across labs,but defies Jevons Paradox.This could cause a supply-demand imbala

45、nce within the USs resource-heavy,Stargate-led ambitions.This imbalance can happen in two ways:Widespread access and use of a resource-light DeepSeek model will reduce the workload of cloud-computing providers46 like Microsoft,Amazon,and Google,thus reducing the demand for top-grade Nvidia-chip-depe

46、ndent graphics processing units(GPU)and their computational power across Big Tech.Technology-wise,most consumers,especially in Global South countries like India,do not need47 a supercharged AI model(e.g.,LLMs)for everyday needs.Kompact AI,a framework developed by Indian startup Ziroh Labs in collabo

47、ration with the Indian Institute of Technology,Madras,confirmed this recently.48 Smaller models emerging from the DeepSeek innovation,or post-training,inference-stage LLMs in Kompact AIs case,can run on edge-computing devices with CPUs(like PCs,mobile phones,and smart wearables)without the need for

48、the security,privacy,and latency of cloud-computing providers.This will further depreciate the demand and price of heavy-duty GPUs.Since the start of 2025,researchers at UC Berkeley have trained“reasoning”models comparable to OpenAIs o1 for just US$450;49 since the release of DeepSeeks open-source m

49、odel R1,this figure has dropped to US$50.50 As with supercomputers of the past,advanced research labs working on long-term goals and improving upon current advances will remain a notable market for current computing architecture and language models.For many,the biggest concern with the DeepSeek mome

50、nt is a scenario51 where open-source models become too powerful in their functional capabilities,which could have implications in both commercial competition and national security.15In such a context,how would the Chinese government perceive its firms pushing the AI frontier?The Chinese Communist Pa

51、rty is almost certainly following DeepSeek and its technology more closely52 and is highly likely to have a say on its open-versus closed-source nature.In a few years,it might be possible to access sensitive information or perform cyberattacks by using the capabilities of the DeepSeek model.Once the

52、se open-source foundational models begin to allow anyone to modify,study,build on,and use them,offering advantages such as external oversight,accelerating progress,and decentralised control,there is likely to be increased discretion around what is and is not allowed for international release.Open-so

53、urcing basic software is more conducive to defensive activities that guard against misuse(e.g.,transparency over biases,community auditing),than allowing offensive misuse by malicious actors.On the other hand,open-sourcing advanced LLMs shifts the balance in favour of offensive use cases via the pot

54、ential disabling of safeguards,increased public knowledge of vulnerabilities,likely perpetuation of flaws,and safety issues.53 Countries like Italy,Taiwan,Australia,South Korea,the US,and India have already taken measures to restrict or prohibit the use of DeepSeeks“application layer”,54 if not the

55、actual AI model,g especially for sensitive functions like government and security.h According to OpenAIs former Head of Policy Research,Miles Brundage,the US benefits from having a stronger AI sector than Chinas,including through direct military applications,economic growth,speed of innovation,and o

56、verall dynamism.55 However,the US should not be stuck between letting Nvidia sell their entire offering and completely cutting off China.Instead,its policy approach needs to be more nuanced than a zero-sum strategy and involve more fine-grained controls.Despite the USs interest in slowing down Chine

57、se AI development,the US needs to have a policy vision that allows China to grow its own economy and benefit from the use of AI.56 However,such choices will require proper staffing in the Department of Commercewhich has the mandate to implement detailed agreements around retaining certain technologi

58、es for civilian purposes and restricting others from military useswhich may be challenging under the Trump administrations“lean government”approach.57 g“Applicationlayer”referstoafront-endtechnologyinterfacethatisaccessedbygeneralconsumerse.g.,theDeepSeekAIchatbotthatisavailableonsigningupunliketheb

59、ack-endtechnologysoftwarethatisusedtopowertheapplicationinthiscase,theDeepSeekLLMs.AnyandalldatacollectedfromtheusageofDeepSeeksapplicationlayerisautomaticallystoredonserversinChina,whereasdownloadingtheAImodellocallytopowerapplicationsdevelopedinacertaingeographywouldleadtothedatacollectedfromusers

60、inthoseapplicationsbeingstoredonserversinthatgeography.hAccordingtoAIsecurityandcompliancefirmEnkryptAI,“Inourevaluations,theDeepSeekmodelwasfoundtobehighlybiasedaswellashighlyvulnerabletogenerateinsecurecode,toxic,harmfulandCBRN(Chemical,Biological,Radiological,andNuclear)content.Wealsocompareditsp

61、erformancewithgpt-4o,o1andclaude-3-opus.”See:https:/ we therefore fundamentally misunderstood what this“AI race”is about?According to Yann LeCun,Metas Chief AI Scientist,neither the US nor China is“winning”as both are constrained by the same conceptual framework regarding the foundations of modern-d

62、ay AI.58 This framework prevents truly“intelligent”behaviour in AI modelsi due to a lack of factors that are considered natural to human intelligence,like understanding of the physical world,persistent memory,reasoning,and complex planning capabilities.Like other leading experts backing the exponent

63、ial power of AI to become the definitive technology of the 21st century,59 LeCun is of the view that real breakthroughs will not come from marginally better language models but from entirely new paradigms of AI architecture.In a potential“second DeepSeek moment”60 the March 2025 launch of the worlds

64、 first fully autonomous AI-agent,Manus,from Shenzhen,Chinahas led to the view that the world is nearing truly“intelligent”AI.In revolutionary style,Manus can think,plan,and execute real-world tasks independently,without human intervention.iAI,asatechnology,translates“intelligent”behaviourtoeverydaya

65、pplicationsservingsocietalneedsviaAImodelslikeDeepSeek.17Experts believe that DeepSeek gives hope to middle powers like India and France.61,62 Both countries emphasise“sovereign AI”63 and will need to develop strategies that ensure cooperation with the US and independent development of AI.Even if th

66、e AI world never becomes truly multipolar,there will be enough room for other players to conduct research,development,and innovation.As of 2024,India had 863 million internet users,US$568 billion in monthly Unified Payment Interface(UPI)transactions,over 100 unicorns,and handled 49 percent of global

67、 real-time payments.64 The IndiaAI Mission,launched in March 2024,65 aims to build a comprehensive ecosystem for the growth of responsible and inclusive AI in India via seven pillars:computing infrastructure,R&D,data access,impactful application,skills enhancement,financing,and trust and safety.Howe

68、ver,in order to realise these goals and catch up with global AI powerhouses like the US and China,India needs to address existing challenges via four main approaches.66The Requirements for India:Strong Foundations and Partnerships18Set Up Sufficient High-End Hardware to Enhance National Computationa

69、l Capacity InfrastructureDespite being a Quad partner,India was excluded from the list of 18 countries granted unchecked access to the USs high-end GPU chips in the January 2025 Framework for Artificial Intelligence Diffusion67 under the Biden administration.j The framework comes into effect on 15 M

70、ay 2025.In response,the Indian government accelerated its pursuit of developing an indigenous,affordable,safe,and secure AI model by announcing the empanelment of 18,693 GPUs,including 12,896 Nvidia H100 chips,under a Common Computing Facilityk68 at the end of January 2025.69 At the time of writing,

71、with 14,000 of these GPUs already accessible,reports suggest70 that the Indian government is looking to double this capacity by acquiring another 15,000 chips.Invest in Foundational R&D Across Academia and Industry,Promoting Long-Term Innovation Alongside Short-Term Optimisation Indias gross domesti

72、c expenditure on R&D(GERD)is below 1 percent of Gross Domestic jThisexclusioncreatedhurdlesforIndiasAIsector.AccordingtoAshokChandak,PresidentoftheIndiaElectronicsandSemiconductorAssociation,“RestrictedaccesstoadvancedAIchipsmayslowinnovationanddevelopmentandscalingupofinstallations.Second,licensing

73、requirementscouldraisecostsandintroducedelaysduetoauthorisations.IndiancompaniesmayrelyheavilyonglobalcorporationsforAIinfrastructure,suchasdatacentres.”See:https:/ which is far less than other major economies.China,despite a slowing economy,invested 2.68 percent72 of GDP to R&D in 2024.AI industry

74、stakeholders in India have echoed these concerns and sought commitments similar to the Common Computing Facility in core AI R&D in universities and labs across the country.73 They also highlight the need for public-private support to PhD programmes and scientific research papers on foundational/sove

75、reign AI models(see Figure 7).However,compared to the Central Governments efforts,Indias private sector is investing far less in R&D compared to its counterparts in global hubs like the US,China,Japan,and South Korea.74 Even governments of states like Tamil Nadu(0.01 percent)and Karnataka(0.07 perce

76、nt),among the top 5 wealthiest states in India,are faring below the national average on this commitment,having allocated less than 0.1 percent of their Gross State Domestic Product(GSDP)to R&D.75 Given Indias diverse and complex ecosystem,developing a one-size-fits-all national LLM need not be the o

77、nly goal.Instead,building Small Language Models focused on specific languages,industry domains,and organisational needs may be more effective and,when layered on top of horizontal LLMs,could pose the ideal multi-model approach.76 19Organise Fragmented,Unstructured DataDespite the scale of Indias dig

78、ital ecosystem and the countrys growing digital public infrastructure,78 AI researchers,innovators,and startups in the country struggle to access good-quality,India-Figure 7:Global Leaders in Foundational/Sovereign AISource:Stanford AI Index 202477specific data(public and proprietary)in the volumes

79、necessary to train foundational LLMs.79 This is primarily because most of the digital consumption in India occurs on global platforms,where user data is locked in.80 20Critical,anonymised digital data from Indian public services like the UPI and the Unique Identification Authority(Aadhaar)is unavail

80、able or available in outdated,AI-unfriendly formats.This hinders the development of AI language models and applications for Indian public sectors like telecom,education,healthcare,finance,agriculture,and transport.81 India needs a long-term approach to AI-friendly data generation,focusing on quality

81、 and multilingual/multi-modal(text,image,audio,video)formats.82 Access to globally available public and licensed data needs to be pursued via partnerships.In March 2025,the Indian government launched its IndiaAI Datasets Platform,AIKosha,to start addressing some of these needs.Housed under the India

82、AI Mission,the IndiaAI Datasets Platform currently hosts 315 datasets and 84 models from 12 organisations.83 It has also partnered with various central and state government ministries,including the Digital India Bhashini division,and offers models from private players like Sarvam AI and Ola Krutrim.

83、Eventually,to supercharge its AI industry,India must aim to truly open-source its national data on a large scale.Much in the manner that the worlds largest free,open repository of datasets from the global webCommon Crawl84does,but with an Indian perspective on choosing and defining these datasets.Le

84、verage Soft-Power-Led Strategic Partnerships India must deepen its strategic partnerships,leveraging frameworks like TRUST(Transforming Relations Utilising Strategic Technologies)85(formerly the Initiative on Critical and Emerging Technologies)and the Indo-Pacific Economic Framework86to co-develop c

85、utting-edge technologies and scale its domestic capabilities.Most importantly,as the USs strategic partner in the Indo-Pacific,India should look to renegotiate its position87 to secure a place among the 18 Tier-1 countries in the AI diffusion framework to avoid being subjected to restrictions or con

86、trols going forward.De-risking against any misalignments in US investments and operations in Indiaor potential knowledge transfer and technological innovation away from Indiais imperative.88 21ConclusionIn 2024,AI and machine learning startups received one-third of total venture capital funding glob

87、ally.89 This raised concerns that too much money is being pumped into foundational models and infrastructure while not enough is being done towards the real-world application of AI.90 Through DeepSeek,China claims to have achieved in weeks what Silicon Valley had assumed would take years.DeepSeek de

88、monstrated its global competitiveness without endless warehouses of US$40,000 chips,achieving remarkable results with 1/50th of the resources.Beyond clever engineering,this is a complete reimagining of what is possible.It is time to stop treating the future of AI like a nuclear arms race from the 20

89、th-century,post-Second World War era and start seeing it as a global public good in the 21st century.91 Many of the worlds nuclear weapons sit as inactive stockpiles,used as a deterrent to large-scale military action and for occasional sabre-rattling.AI,on the other hand,permeates our everyday lives

90、 and has the potential to impact radically the way societies and nations function.What is required is collective thinking and a new“technological quotient”92 that considers not only the technology but also the skills and processes to develop it.In this context,the future will belong not to whoever h

91、as the most physical infrastructure(e.g.,chips and data centers)but to those who invest and build beyond.221 Rudra Chaudhari,“At the Global Technology Summit,Realism Meets Imaginationand India Makes it Possible,”The Print,April 7,2025,https:/theprint.in/opinion/global-technology-summit-2025-sambhavn

92、a-geopolitics-tech/2580575/2 The Print,“LIVE:EAM S Jaishankars Keynote Address at Global Technology Summit,”Youtube video,59:25 min,April 11,2025,https:/ Xi Jinping Accuses US of Leading Western Nations to Suppress China,”The Indian Express,March 7,2023,https:/ Rahul Batra,“TikToks Geopolitical Dile

93、mma:Which Values,Whose Interests?,”Observer Research Foundation,May 1,2024,https:/www.orfonline.org/expert-speak/tiktoks-geopolitical-dilemma-which-values-whose-interests5 Karen Hao,“OpenAI Goes MAGA,”The Atlantic,January 23,2025,https:/ Ananya Bhattacharya,“Non-Western Founders Say DeepSeek is Proo

94、f That Innovation Need Not Cost Billions of Dollars,”Rest of World,January 30,2025,https:/restofworld.org/2025/deepseek-ai-model-openai-dominance-challenge/7 Mathew S.Smith,“What DeepSeek Means for Open-Source AI,”IEEE Spectrum,January 31,2025,https:/spectrum.ieee.org/deepseek8 Sameer Patil and Saur

95、abdeep Bag,“DeepSeek and the Shifting Tides of the US-China AI race,”Observer Research Foundation,February 12,2025,https:/www.orfonline.org/expert-speak/deepseek-and-the-shifting-tides-of-the-us-china-ai-race9 Rahul Rao,“The U.S.-China Chip Ban,Explained,”IEEE Spectrum,November 21,2022,https:/spectr

96、um.ieee.org/chip-ban10 McKinsey&Company,“The CHIPS and Science Act:Heres Whats In It,”October 4,2022,https:/ DeepSeek-AI,DeepSeek-R1:Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,January 2025,https:/arxiv.org/pdf/2501.1294812 Dan Milmo et al.,“Sputnik Moment:$1tn Wiped Off US

97、 Stocks After Chinese Firm Unveils AI Chatbot,”The Guardian,January 28,2025,https:/ Erin Watson,“Chips,Clouds,and Checkpoints:The New AI Export Battlefield Under Trump 2.0,”Observer Research Foundation,March 17,2025,https:/www.orfonline.org/expert-speak/chips-clouds-and-checkpoints-the-new-ai-export

98、-battlefield-under-trump-2-014 Milmo et al.,“Sputnik Moment:$1tn Wiped Off US Stocks After Chinese Firm Unveils AI Chatbot”15 Caiwei Chen,“How a Top Chinese AI Model Overcame US Sanctions,”MIT Technology Review,January 24,2025,https:/ David Gray Widder et al.,“Why Open AI Systems Are Actually Closed

99、,and Why This Matters,”Nature(635)2024:827833,https:/ Michael Froman,“DeepSeek:Making Sense of the Reactionand Overreaction,”Council on Foreign Relations,January 31,2025,https:/www.cfr.org/article/deepseek-making-sense-reaction-and-overreaction18 Rao,“The U.S.-China Chip Ban,Explained”19 Dylan Patel

100、 et al.,“DeepSeek Debates:Chinese Leadership on Cost,True Training Cost,Closed Model Margin Impacts,”SemiAnalysis,January 31,2025,https:/ Swagath Bandhakavi,“DeepSeek Being Investigated for Allegedly Using Restricted US-Made AI Chips,”Tech Monitor,January 31,2025,https:/www.techmonitor.ai/hardware/s

101、ilicon/deepseek-under-scanner-restricted-us-made-ai-chips?cf-view21 Sebastien Laye,“No,DeepSeek Isnt Americas AI Sputnik Moment,”The Hill,January 30,2025,https:/ James ODonnell,“AIs Energy Obsession Just Got a Reality Check,”MIT Technology Review,January 28,2025,https:/ Yolanda Gil and Raymond Perra

102、ult,The 2025 AI Index Report,Stanford HAI,April 2025,https:/hai- Alexandra Lum,“Powering the Future of AI:The Data Center Energy Crunch,”Energize Capital,November 19,2024,https:/ Alex Colville,“AI for All,”China Media Project,December 20,2024,https:/chinamediaproject.org/2024/12/20/ai-for-all/26 Sta

103、te Council Information Office,The Peoples Republic of China,China Home to Over One-Third of Worlds AI Large Language Models,July 3,2024,http:/ Field Cady and Oren Etzioni,“China May Overtake US in AI Research,”Allen Institute for Artificial Intelligence,March 13,2019,https:/ Lily Ottinger and Jordan

104、 Schneider,“DeepSeek:What It Means and What Happens Next,”China Talk,February 1,2025,https:/www.chinatalk.media/p/deepseek-what-it-means-and-what-happens29 Yishan(yishan),“I Think the Deepseek Moment Is Not Really the Sputnik Moment,but More Like the Google Moment,”X post,January 28,2025,https:/ Dan

105、 Milmo,“Llama 2:Why is Meta Releasing Open-Source AI Model and Are There Any Risks?,”The Guardian,July 20,2023,https:/ Ananya Bhattacharya,“Non-Western Founders Say DeepSeek Is Proof That Innovation Need Not Cost Billions of Dollars,”Rest of World,January 30,2025,https:/restofworld.org/2025/deepseek

106、-ai-model-openai-dominance-challenge/32 Will Knight,“OpenAIs CEO Says the Age of Giant AI Models is Already Over,”Wired,April 17,2023,https:/ Patel et al.,“DeepSeek Debates:Chinese Leadership on Cost,True Training Cost,Closed Model Margin Impacts”34 Vinod Chugani,“DeepSeek vs.OpenAI:Comparing the Ne

107、w AI Titans,”datacamp,February 6,2025,https:/ Yolanda Gil and Raymond Perrault,The 2025 AI Index Report,Stanford HAI,April 2025,https:/hai- Learning Heats Up,White House Orders Muscular AI Policy,and More.,”deeplearning.ai,January 2025,https:/www.deeplearning.ai/the-batch/issue-286/37 Edd Gent,“When

108、 AIs Large Language Models Shrink,”IEEE Spectrum,March 31,2023,https:/spectrum.ieee.org/large-language-models-size38“DeepSeek-R1:Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,DeepSeek-AI”39“Reinforcement Learning Heats Up,White House Orders Muscular AI Policy,and More.”40 Yol

109、anda Gil and Raymond Perrault,The 2025 AI Index Report,Stanford HAI,April 2025,https:/hai- Greg Rosalsky,“Why the AI World Is Suddenly Obsessed with a 160-Year-Old Economics Paradox,”NPR,February 4,2025,https:/www.npr.org/sections/planet-money/2025/02/04/g-s1-46018/ai-deepseek-economics-jevons-parad

110、ox42 Rosalsky,“Why the AI World Is Suddenly Obsessed with a 160-year-old Economics Paradox”43 Rao,“The U.S.-China Chip Ban,Explained”44 Patrick Seitz,“Chinas Huawei Launches AI Chip To Fill Nvidia Void,”Investors Business Daily,April 21,2025,https:/ Ondrej Burkacky et al.,Exploring New Regions:The G

111、reenfield Opportunity In Semiconductors,McKinsey&Company,January 29,2024,https:/ Alvin Wang Graylin(/agraylin),“#Misunderstandings of Applying Jevons Paradox to AI,”facebook post,January 31,2025,https:/ Chitra Narayanan,“Jio-Backed Two AIs Sutra-R0 Bets on Language Capability to Power Ahead,”The Hin

112、du Business Line,February 7,2025,https:/ Startup Unveils System to Run AI Without Advanced Chips,”The Economic Times,April 11,2025,https:/ Kyle Wiggers,“Researchers Open-Source Sky-T1,a Reasoning AI Model That Can Be Trained for Less Than$450,”Techcrunch,January 11,2025,https:/ Maxwell Zeff,“Researc

113、hers Created an Open Rival to OpenAIs o1 Reasoning Model for Under$50,”Techcrunch,February 5,2025,https:/ Ottinger and Schneider,“DeepSeek:What it Means and What Happens Next”2552 Ottinger and Schneider,“DeepSeek:What it Means and What Happens Next”53 Elizabeth Seger et al.,Open-Sourcing Highly Capa

114、ble Foundation Models:An Evaluation of Risks,Benefits,and Alternative Methods for Pursuing Open-Source Objectives,Center for the Governance of AI,September 29,2023,https:/cdn.governance.ai/Open-Sourcing_Highly_Capable_Foundation_Models_2023_GovAI.pdf54 Natasha Ganesan,“CNA Explains:Are Countries Ban

115、ning DeepSeek for Legitimate Reasons?,”Channel News Asia,February 12,2025,https:/ Jordan Schneider,“DeepSeek and the Future of AI Competition with Miles Brundage,”China Talk,January 25,2025,https:/www.chinatalk.media/cp/15570418656 Schneider,“DeepSeek and the Future of AI Competition with Miles Brun

116、dage”57 James Fitzgerald and Holly Honderich,“What is Doge and Why is Musk Cutting So Many Jobs?,”BBC,February 19,2025,https:/ Paul Sawers,“Metas Yann LeCun Predicts New Paradigm of AI Architectures Within 5 Years and Decade of Robotics,”Techcrunch,January 23,2025,https:/ Gary Marcus,“The Race for A

117、I Supremacy Is OverAt Least for Now,”Marcus on AI,January 26,2025,https:/ Craig S.Smith,“Chinas Autonomous Agent,Manus,Changes Everything,”Forbes,March 8,2025,https:/ C.Raja Mohan,“DeepSeeks Sputnik Moment:5 Ways in Which US-China AI Race Plays Out for India,”Indian Express,January 29,2025,https:/ S

118、ameer Patil and Sauradeep Bag,“US-China AI Rivalry:Lessons for India,”Deccan Herald,February 8,2025,https:/ Trisha Ray,“The Future of AI Is Sovereign:How it Evolves Is Up to Us”,Observer Research Foundation,April 16,2025,https:/www.orfonline.org/expert-speak/the-future-of-ai-is-sovereign-how-it-evol

119、ves-is-up-to-us#_ednref364 Amitabh Kant,“In the AI Race,What India Needs to Do,”The Indian Express,February 9,2025,https:/ Ministry of Electronics and IT,Government of India,https:/pib.gov.in/PressReleasePage.aspx?PRID=201237566 Bhuvana Kamath,“As US Launches$500B Stargate,India Cant Afford to Be Id

120、le in AI race,”Yourstory,January 27,2025,https:/ Prateek Tripathi,“The US AI Diffusion Framework:Global and Indian Implications,”Observer Research Foundation,April 10,2025,https:/www.orfonline.org/expert-speak/the-us-ai-diffusion-framework-global-and-indian-implications68 Ministry of Electronics&IT,

121、Government of India,https:/pib.gov.in/PressReleasePage.aspx?PRID=20977092669“DeepSeek to Be Hosted on Indian Servers Soon to Address Data Privacy Concerns:Vaishnaw,”Yourstory,January 31,2025,https:/ Aashish Aryan,“Govt Hopes to Add Another 15K GPUs in Second Round of IndiaAI Mission”,Business Standa

122、rd,April 25,2025,https:/www.business- Gopal Krishna Roy,“DeepSeeks Market Disruption Must Awaken India,”The Hindu,March 6,2025,https:/ Abhijeet Kumar,“Chinas R&D Investment Grows Despite Slower Funding,US Tops Global List,”Business Standard,January 24,2025,https:/www.business- S Ronendra Singh,“Buil

123、ding Indias Foundational AI Model:Start-Ups Community Says Research Ecosystem Needs Strengthening,”The Hindu Business Line,February 10,2025,https:/ Sriram Iyer,“Budget 2025:Government Spends Half of All R&D Expense in India,”CNBC TV18,February 1,2025,https:/ Aprameya Rao,“Are Indian States Doing Eno

124、ugh to Fund R&Dan RBI Report Paints a Bleak Picture,”CNBC TV18,December 22,2025,https:/ K V Kurmanath,“Beyond LLMs:A Multi-Model Approach to AI,”The Hindu Business Line,February 12,2025,https:/ Maslej,“Inside the New AI Index:Expensive New Models,Targeted Investments,and More”78 Kay McGowan et al.,“

125、Indias G20 Presidency Rallied Consensus on DPI.Brazil Has the Opportunity to Go Even Further,”Digital Impact Alliance,December 1,2023,https:/dial.global/india-g20-presidency-consensus-dpi-brazil-opportunity/79 Rudraksh Lakra and Rutuja Pol,“Charting Indias AI Future:Overcoming Data Challenges for Re

126、sponsible Innovation,”The Secretariat,July 9,2024,https:/thesecretariat.in/article/charting-india-s-ai-future-overcoming-data-challenges-for-responsible-innovation80 Subhayan Chakraborty,“Indian Data Should Remain in Indian Data Centers,Says Akash Ambani,”Business Standard,October 15,2024,https:/www

127、.business- Sam Neufeld,“Deploying Open Government Data for AI-Enabled Public Interest Technologies,”Observer Research Foundation,July 21,2021,https:/www.orfonline.org/expert-speak/ai-driven-public-interest-technologies-employing-open-government-data-achieve82 Rohan Pai,“Indias AI Development Faces C

128、hallenges,Calling MeitYs Attention,”Deccan Herald,June 17,2024,https:/ Aditi Agarwal,“Centre Opens Portals to Fuel Indias AI Mission,”Hindustan Times,March 7,2025,https:/ Alejandro Alija,“Open Data for the World:Common Crawl Frees Petabytes of Web Data”,datos.gob.es,May 22,2024,https:/datos.gob.es/e

129、n/blog/open-data-world-common-crawl-frees-petabytes-web-data85 Prashant Jha,“TRUST Replaces iCET:New Brand,Old Spirit on Tech Partnership,”Hindustan Times,February 14,2025,https:/ Office of the United States Trade Representative,Executive Office of the President,Indo-Pacific Economic Framework for P

130、rosperity(IPEF),https:/ustr.gov/trade-agreements/agreements-under-negotiation/indo-pacific-economic-framework-prosperity-ipef87“India to Press for Technology Access Parity with US Allies in New Trade Agreement Talks,”Financial Express,April 27,2025,https:/ Ashwin Prasad,“Implications of the AI Diffu

131、sion Framework,”The Hindu,February 13,2025,https:/ Jacob Robbins,“AI Startups Grabbed a Third of Global VC Dollars in 2024,”Pitchbook,January 9,2025,https:/ Allison Nathan et al.,Gen Ai:Too Much Spend,Too Little Benefit?,Goldman Sachs,June 25,2024,https:/ Dan Hendrycks and Eric Schmidt,“The Nuclear-

132、Level Risk of Superintelligent AI,”Time,March 6,2025,https:/ Ayesha Khanna and Parag Khanna,“Which Nation Has the Best Technik?,”Harvard Business Review,June 12,2012,https:/hbr.org/2012/06/which-nation-has-the-best-tech28About the AuthorRahul Batra is an independent consultant with extensive experie

133、nce at the intersection of digital technology and international affairs.Cover photo:Getty Images/Yuriko NakaoBack cover image:Getty Images/Andriy OnufriyenkoIdeas.Forums.Leadership.Impact20,Rouse Avenue Institutional Area,New Delhi-110 002,INDIAPh.:+91-11-35332000.Fax:+91-11-35332005 E-mail:contactusorfonline.org Website:www.orfonline.org