《AI Now研究所：2025年人工智能权力格局研究报告：权力集中化及其威胁（英文版）（122页）.pdf》由会员分享，可在线阅读，更多相关《AI Now研究所：2025年人工智能权力格局研究报告：权力集中化及其威胁（英文版）（122页）.pdf（122页珍藏版）》请在三个皮匠报告上搜索。

1、ARTIFICIAL POWER2025 Landscape ReportEXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT2With contributions and feedback from:Annette Bernhardt,UC Berkeley Labor Center Abeba Birhane,Artificial Intelligence Accountability Lab,Trinity College Dublin Brian Chen,Data&Society Jane Chung,Justice Speaks Stefani

2、e Coyle,New York Civil Liberties Union Andrea Dehlendorf,AI Now Kevin DeLiban,TechTonic Justice Becca Deutsch,Amazon Employees for Climate Justice Alix Dunn,The Maybe Ali Finn,AI Now Timnit Gebru,Distributed AI Research Center Ryan Gerety,Athena Coalition Lisa Gilbert,Public Citizen Sam Gordon,Tech

3、Equity Collaborative Janet Haven,Data&Society Nidhi Hegde,American Economic Liberties Project Ben Inskeep,Citizen Action Coalition of Indiana Taylor Jo Isenberg,Economic Security Project Heidy Khlaaf,AI Now Stephen Lerner,Action Center for Race and the Economy-Bargaining for the Common Good Barry Ly

4、nn,Open Markets Institute Varoon Mathur,Duke Institute for Health Innovation,Duke Health Jill McArdle,Beyond Fossil Fuels Michelle Meagher,SOMO Brian Merchant,Independent Journalist Erie Meyer,Georgetown University Sarah Miller,American Economic Liberties Project Stacy Mitchell,Institute for Local S

5、elf-Reliance Arvind Narayanan,Center for Information Technology Poli-cy,Princeton University Chris Nielsen,National Nurses United Teri Olle,Economic Security Project Britt Paris,Rutgers University Tekendra Parmar,Investigative Journalist Reshma Ramachandran,Yale School of Medicine Steven Renderos,Me

6、diaJustice Rashida Richardson,Worcester Polytechnic Institute&Northeastern University Hilary Ronen,Local Progress Leevi Saari,AI Now Hannah Sassaman,Peoples Tech Project Matt Scherer,Center for Democracy and Technology Paromita Shah,Just Futures Law Ganesh Sitaraman,Vanderbilt University Andrew Stra

7、it,UK AI Security Institute Kasia Tarczynska,Good Jobs First Jim Thomas,Scanthehorizon.org Max Von Thun,Open Markets Institute Jai Vipra,Cornell University Robert Weissman,Public Citizen Meredith Whittaker,Signal Savannah Wilson,Clean Virginia Boxi Wu,Oxford UniversityCite as:Kate Brennan,Amba Kak,a

8、nd Sarah Myers West,“Arti-ficial Power:AI Now 2025 Landscape”,AI Now Institute,June 3,2025,https:/ainowinstitute.org/2025-landscape.ACKNOWLEDGMENTSAuthored by Kate Brennan,Amba Kak,and Dr.Sarah Myers West.With research support from Mohammed Ali,Yasmine Chokrane,Madeline Kim,Tekendra Parmar,Tanya Raj

9、a,and Boxi Wu.Special thanks to Rhiana Gunn-Wright for her feedback and editorial contributions.Project management by Ellen Schwartz.Copyediting by Caren Litherland.Design by Partner&Partners.EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT3TABLE OF CONTENTSExecutive SummaryChapter 1:AIs False Gods1.1:

10、The AGI Mythology1.2:Too Big to Fail1.3:AI Arms Race 2.01.4:Recasting RegulationChapter 2:Heads I Win,Tails You LoseChapter 3:Consulting the RecordChapter 4:A Roadmap for ActionEndnotes4181924283236466793EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT4EXECUTIVE SUMMARYThose of us broadly engaged in ch

11、allenging corporate consolidation,economic injustice,tech oligarchy,and rising authoritarianism need to contend with the AI industry or we will lose the end game.Accepting the current trajectory of AI proselytized by Big Tech and its stenographers as“inevitable”is setting us up on a path to an unenv

12、iable eco-nomic and political futurea future that disenfranchises large sections of the public,renders systems more obscure to those it affects,devalues our crafts,undermines our security,and narrows our horizon for innovation.This is true whether or not the technology even works well,on its own ter

13、ms;it often doesnt.The good news is that the road offered by the tech industry is not the only one avail-able to us.This report explains why the fight against the industrys vision for AI is a fight worth having,even as we turn ourselves tirelessly toward the task of building out the shared project o

14、f a just,equitable,sustainable,and democratic society.EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT5Over the past decade,taming the power of big tech-nology platforms like Microsoft,Amazon,Google,and Meta has increasingly become a central question in American political and public discourse.Unless we

15、 contend with the power vested in these firms,we wont meaningfully be able to hold the industry ac-countable to the interests of the broader public,even as these companies reshape markets,institutions,and infrastructures core to public life.What does AI have to do with any of this?As we ar-gued in o

16、ur 2023 report,1 AI is fundamentally about concentration of power in the hands of Big Tech.At the start of the year,it seemed like the market was poised for disruption,with a new crop of Silicon Valley chal-lengers gaining prominence,like OpenAI,Anthropic,StabilityAI,and Inflection AI.But now,just t

17、wo years later,it is clear that the bench of key players in this market hasnt changed much:Microsoft,Google,Meta,Musks xAI,OpenAI(backed by Microsoft),and Anthrop-ic(backed by Amazon and Google).2 The new suite of LLM-powered AI products has pushed these firms into the spotlight,dominating headlines

18、 and,increasingly,becoming the subject of dinner-table conversation.Amid the frenzy,theres been a misplaced focus on blinkered questions of whether one AI system or application is good or bad,or evaluating the moral quandaries of hypothetical worlds.Instead we need to redirect attention to the AI ec

19、osystem,and its depen-dencies and risks,as a whole.The question we should be asking is not if ChatGPT is useful or not,but if OpenAIs unaccountable power,linked to Microsofts monopoly and the business model of the tech econo-my,is good for society.Looking beyond individual use cases allows for a mor

20、e comprehensive look into the centers of power that drive our current tech landscape.AI as a field has been not just co-opted but constituted by the logics of a few dominant tech firms.It is no coincidence that the“bigger-is-better”paradigm that dominates the field today,where the scale of compute a

21、nd data resources are generally used as a proxy for perfor-mance,lines up neatly with the incentives of Big Tech,which disproportionately controls these resources,the talent to leverage them,and the pathways to mon-etization.Around 2012,as it became apparent that substantial gains in model performan

22、ce could come simply from applying larger and larger scale data and computational resources to existing algorithms,tech giants moved quickly to shore up their existing advan-tages and hire talent.3 Corporate influence over AIs research trajectory has been cemented through tech firms AI labs dominant

23、 presence at prestigious ma-chine learning conferences,further shaping the field of research in ways that align with industry.4 This is in part because building AI bigger requires enormous resources,both financial and social,to achieve unre-stricted growth at breakneck speedresources that AI compani

24、es own and control.5But its not just market power we need to be con-cerned with:These tech oligarchs are counting on a wholesale rewriting of our social and economic foundations,using AI as the justification.From break-ing apart the US federal government and raiding citizen data under the guise of e

25、fficiency,to redesign-ing workflows to devalue human labor and creativity so they are AI-ready,to redirecting our entire ener-gy infrastructure to prioritize their technology over peoples basic needs,the vision promulgated by tech oligarchs requires,as a foundation,the unraveling of core social,poli

26、tical,and economic fabrics.Across our information ecosystem,from science to education,healthcare,culture,and art,AI is being positioned as a disruptive new infrastructure and a mediating force.In truth,though,it rehashes an old playbook,helicoptering in solutions built on the extraction of expertise

27、 and value from all corners of societysolutions that always,eventually,amount to the further degradation of life for the most EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT6marginalized among us.While generative AI and AI agents are the buzzwords that splash across the headlines,the same dynamics are

28、 true of precursors to contemporary AI systems like automated decision-making technologies used in banking,hiring,and criminal justice.The techniques and vendor names vary,but the industry incentives powering proliferation,as well as the failure modes across these systems,share much in common.More t

29、han a decade of evidence demonstrates how it goes:The introduction of these systems concentrates power among the deployers of the tech,leaving those on the receiving end more insecure,vulnerable,and unable to contest the determinations made by the“smart machine”at the expense of the broader pub-lic.

30、These tools are often invisible to those judged by them,and inscrutable even when they are visible.Why society would ever accept this bargain is the critical question at hand.Amid the excitement over AIs(speculative)potential,the sobering reality of its present and recent past is obscured.When we co

31、nsult the record on how AI is already intermediating criti-cal social infrastructures,we see that it is materially reshaping our institutions in ways that ratchet up in-equality,render institutions opaque to those they are meant to serve,and concentrate power in the hands of the already powerful.(Se

32、e Chapter 3:Consulting the Record.)It makes clear that for all the whiz-bang demos and bold Davos proclamations,on the ground AI is consistently deployed in ways that make everyday peoples lives,material conditions,and access to opportunities worse and the systems that incorporate them stronger.This

33、 reports title,Artificial Power,captures the critical,and at times contradictory,moment we find ourselves in.On one hand,the tech oligarchy has successfully deployed“AI”as a strategic marketing term and as a set of automation technologiesto cement and grow its power.At the same time,this power is va

34、stly inflated,contingent,and poised for disruption.Con-tending with the dual reality of how those with power have deployed AI systems to enact significant harm while exposing the ways this power can and must be disrupted is the central work of this moment.THE ELEPHANTS IN THE AI ROOM:OF BUSINESS MOD

35、EL(S)AND FATAL TECHNICAL FLAWSThe AI industry is on shakier ground than it may seem.Valuations are sky high,while the business model hinges on an intensely expensive technology that lacks a consistent revenue stream.AI compa-nies bleed money for every user they gain:Anthropic burned through$5.6 bill

36、ion6 this year but was valued at$61.5 billion.7 OpenAI lost$5 billion8 but is valued at$300 billion.9 No profit-making use cases exist yet,or are even on the horizon.This may seem like business as usual for the move-fast-and-break-things ethos of Big Tech,but we are in a profoundly different mon-eta

37、ry environment nowits no longer the 2000s or 2010s.Markets are saturated,market dominance has been established among the platform and infrastruc-ture winners that emerged from those decades,and,put simply,the cost of large-scale AI is eye-watering at a level not seen before in tech.EXECUTIVE SUMMARY

38、AI NOW 2025 LANDSCAPE REPORT7The question now circulating more,and more open-ly,is this:When will the AI bubble burst and who will be impacted by it?Because this is a bubble.For their part,companies are pushing out shiny objects to detract from the business reality while they desper-ately try to der

39、isk their portfolios through government subsidies and steady public-sector(often carceral or military)contracts.While its very clear how tech companies benefit from claims to“public-interest AI”used to justify the pouring of taxpayer dollars into this industry,it is not at all clear how this benefit

40、s the rest of society.(See Chapter 2:Heads I Win,Tails You Lose)Chapter SnapshotHeads I Win,Tails You LoseThis section maps the drivers that are securing Big Tech firms advantage in the AI market,before turning to the question of who loses in the end:Cloud infrastructure providers benefit from cycle

41、s of AI dependence Big Tech firms benefit from leveraging control over the tech ecosystem Big Tech benefits from the data center boom.even if the AI boom doesnt pan out With generative AI,in particular,the hyped claims stand in stark contrast to the largely mundane use cases that are being shoved in

42、to nearly every app and service.In contrast to the claims of world-changing tech,Meta is investing heavily in AI advertising infra-structure.10 OpenAI is creating AI agents that fill out forms and call web browsing“research”sucking up your data and requesting invasive permissions as it does so.11 Pr

43、essured by their employers,software engineers are using Microsofts Copilot to produce more code,more rapidly,undercutting their skills and trade.12 And cloud companies are happily lock-ing enterprise users into their software-as-a-service(SaaS)ecosystems by automatically upgrading them to new AI fea

44、turesand raising the price.13 These are not examples of a technology being embraced by a society glad of its utility.Despite being positioned as critical infrastructure,AI systems in their current form have fundamental flaws:there is an intractable problem of“hallucinations”with LLMs reliant on rand

45、omly generated coordination,leav-ing the humans in this technology in the unenviable position of fact-checking the tech meant to make their lives easier.14 Peer-reviewed research indicates that in many cases,AI systems fail profoundly at even basic tasks when applied in real-life contexts.15 Theyre

46、also far from resilient,prone to cybersecurity vulnerabilities like web poisoning attacks and new jailbreaking meth-ods that enable the persistent unauthorized disclosure of training data and other sensitive information.16 And its not that the trade-offs are weighted and deemed worth it:In many use

47、cases,AI is deployed by thwose with power against those who lack it,and who have no opportunity to opt out or seek remedy when mistakes are made.(See Chapter 3:Consulting the Record.)EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT8Chapter SnapshotConsulting the RecordThis section compiles over a decad

48、e of evidence showing how the tech industry has sought to reshape society to enable more widespread deployment of the technologies it builds and profits from,often contributing to the deg-radation of our social,political,and economic lives.The section aligns on five key takeaways:AIs benefits are ov

49、erstated and underproven,from cancer cures to hypothetical economic growth-while some of its flaws are real,immediate,and growing.AI-sized solutions to entrenched social problems displace grounded expertise,in disparate domains like higher education,healthcare,and agriculture.AI solutionism obscures

50、 systemic issues facing our economy obscuring economic concentration and acting as a conduit for deploying austerity man-dates by another name.The DOGE power grab is instructive,though New Yorks MyCity offers another example where millions of taxpayer dollars were invested into flawed AI solutions t

51、hat failed to deliver tangible benefits to the public.The productivity myth obscures a foundational truth-the benefits of AI accrue to companies,not to workers or the public at large,even as algorithmic manage-ment tools make work unstable and unsafe.Agentic AI will make workplaces even more bureauc

52、ratic and surveillant,reducing not increasing autonomy.AI use is frequently coercive,violating rights and undermining due process.This is nowhere more clear than the rise of AI usage in immigration en-forcement,where human rights abuses are com-mon and legal norms are routinely violated-even before

53、AI is in the mix.with products that are patently inaccurate,insecure,and compromise the safety of consumers;engage in anti-competitive practices that shore up their advan-tages to shut the door behind them;and deploy larger than life narratives around AGI and innovation to quell any form of interrog

54、ation and critique.It doesnt help that the prevailing deregulatory current is an industry that continually acts above the law and is driven narrowly by its bottom line(See 1.4:Reg-ulation):in 2024,we saw companies rush to market Chapter SnapshotAIs False GodsThis section interrogates narratives that

55、 advance AI indus-try dominance and make the current trajectory of AI seem inevitable:The AGI Mythology:The Argument to End All Arguments unpacks the nebulous claims surround-ing“artificial general intelligence,”arguing that the term collapses complex technical realities into a singular,imminent,and

56、 inevitable future that con-veniently advances the interests of the companies claiming to build it.Too Big to Fail:Infrastructure and Capital Push explores how tech firms are deploying unprecedent-ed amounts of capital to perpetuate a“bigger-is-bet-ter”AI paradigm,shoring up their continued market d

57、ominance through government and taxpayer support.AI Arms Race 2.0:From Deregulation to Industrial Policy details how the US-China AI arms race has heightened,and is now used to brand a slate of in-dustrial policy initiatives designed to boost the tech industry and avert regulatory scrutiny.Recasting

58、 Regulation as a Barrier to Innovation shows how the AI industry has strategically pitted regulation against innovation,leading to a global deregulatory posture that ignores the role regulation plays in enhancing innovation and competition.EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT9technological

59、infrastructures in ways that are robust,accountable,and protected from systemic shocks.What weve seen play out within the AI industry is not unique to this industry,of course.The dynamics of“gain for me,loss for thee”have been examined in many critiques of shareholder capitalism broadly,which emphas

60、ize the corporate willingness to specu-late in ways beneficial to shareholders but not to so-ciety at large,alongside the perverse incentives that lead firms to act against their own business interests,and develop an orientation toward monopolization and sclerosis.If anything,the AI market is the pe

61、ak exem-plar of overreliance on venture-based investment.18But AI introduces new dynamics and accelerants.As designed,developed,and deployed currently,AI works to entrench existing power asymmetries,and to ratchet them up.It naturalizes inequity as destiny and deservednesssimply the classification g

62、iven by the intelligent systemwhile rendering these underlying patterns,judgments,and self-interested drivers inscrutable to those affected by AIs judg-ments and instructions.IT IS TIME TO BUILD JUST NOT AI.AI is predominantly used on us,not by us,to shape ac-cess to resources and life chances.But w

63、hile there is a clear path dependency within this narrow trajectory for AI proselytized by big tech and its stenographers,the good news is that its not the only road available to us.17 Not by a long shot.AI hype has tapped into a sentiment that is real and widespread:genuine enthusiasm to build a fu

64、ture where all people can thrive,a future that will likely look radically different from the present.It is a catalyzing goal we should unite around;most of us want a future that frees us from the endless cycle of war,pandemics,and environmental and financial crises that charac-terize our present.The

65、 2024 US presidential election brought the need to create social and political insti-tutions connected to the needs and lived realities of people even closer to home,across the country and across the world.But AI doesnt create any of theseand pegging our shared future on AI makes that future harder,

66、not easier,to achieve because it binds us to a decidedly bleak path,stripping us not only of the ability to choose what to build and how to build it,but also stealing what joy we might take in that building.This hype-prescribed AI future further distances us from a life with dignity,one where we hav

67、e the autonomy to make our own decisions and where democratically accountable structures work to distribute power and EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT10 ANOTHER AI IS POSSIBLE.HOW DO WE GET THERE?Although real ways in which the AI market could be structured to benefit the public may exi

68、st,the path charted by the companies controlling AI,and those wanting a piece of the control AI could give them over our lives and institutions,wont lead us there.One thing is clear:we cant fight tech oligarchy with-out rejecting the current industry trajectory around large scale AI.Its a crucial in

69、flection point and how American policymakers and movement leaders choose to respond to the AI industry will write the coming chapters of the story of tech power.AI com-panies,and those who lead them,have positioned themselves to reshape broad swaths of societywithin and beyond the USnot only to work

70、 in their interest,but to do so in ways that allow their firms to capture the lions share of the value.This isnt inevitable.In fact,the tide of public opinion is moving decisively against the entrenched power of tech firms.And weve seen major legal wins in the landmark antitrust cases filed by the D

71、OJ and FTC against Google and Meta.After successfully proving that Google maintains an illegal monopoly in search and advertising markets,19 the DOJ is now requesting bold,structural remedies that were all but inconceiv-able a few years ago.20 These remedies,which include breaking up Googles adverti

72、sing technology business and spinning off Chrome,strike directly at the heart of Googles business model.But the remedy trials have revealed a larger truth:AI startups cant scale or achieve distribution without Big Tech firms infrastruc-ture.Thats why OpenAI offered to buy Chrome.Its why Perplexitys

73、CEO said hed want to buy itthen pandered so as not to aggravate a company hes de-pendent on.21 This is why its especially important that we not cede the momentous ground these regulatory actions have pushed us towards when Big Tech com-panies use AI as cover for staying unregulated.In this report,we

74、 lay out another path forward.First,we map what we mean by AI in the first place,pro-vide an accounting of the false promises and myths surrounding AI,and examine whom its working for and whom its working against.Then,given that AI consistently fails the average public,even as it en-riches a sliver,

75、we ask what we lose if we accept the current vision of AI peddled by the industry.Finally,we identify leverage points that we can latch on to as we mobilize to build a world with collective thriving at its centerwith or without AI.We are not naive.The headwinds against sensible AI alternatives have

76、never been stronger.The tech indus-try is better resourced and the political environment more bleak than ever before.Indeed,considering the tech industry to be simply a collection of firms itself misses key sources of their power,from their surveil-lance apparatus to their control over global digita

77、l infrastructures that shape our states,our institutions,our economies,and,most importantly,our lives.But the battle has never been more important.Contest-ing AI links the movements we must build not only to create meaningful public power,but also to seed a new path defined by autonomy,dignity,respe

78、ct,and justice for all of us.In the following sections,we set out a tool kit for reasserting public power amid a takeover by AI firms.EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT11Chapter SnapshotA Roadmap for Action:Make AI a Fight About Power,Not Progress 1.Target how the AI industry works agains

79、t the interests of everyday people.In the aftermath of the 2024 election,theres growing consensus across the ideological spectrum that focusing on the material conditions and economic interests of working people is key to building political power.We need to not only make AI-related issues more relev

80、ant to movements fighting for economic populism and against tech oli-garchy;we also need to better target the AI industry as a key actor working against the interests of the working public.The pushback against the Depart-ment of Government Efficiency(DOGE),the buildout of AI data centers,and algorit

81、hmic prices and wages constitutes fertile ground for building a broader movement unified in its focus on rejecting AIs unac-countable tech-enabled social and political control.2.Advancing worker organizing is the clearest path to protecting us and our institutions from AI-en-abled capture.Labor camp

82、aigns have demonstrat-ed that working people have a particular form of power to wield,power that can determine how their employers deploy AI and digital systems.The deeper opportunity for labor,and a more transfor-mative ambition,however,would be to direct labors power not just toward whether and ho

83、w AI is used in the workplace,but also toward recalibrating the technology sectors power overall and shaping the trajectory of AI in the public interest and common good.3.Enact a zero trust policy agenda for AI.Trust in AI firms benevolence is not a smart,informed,or credible optionnot if were going

84、 to proceed with serious work.Enacting a policy agenda built on bright-line rules that restrict the most harmful uses of AI,regulate the AI life cycle from nose to tail,and ensure that the industry that currently creates and profits from AI isnt left to regulate and evaluate itselfessentially gradin

85、g its own homeworkmust be a priority at the state and federal levels in the US,and internationally.4.Bridge networks of expertise,policy,and narrative to strengthen AI advocacy.AI advocacy and policy has often been undermined too often by blinkered views that fail to see the different components of

86、the AI supply chain materially,are often single issue-focused,and it is easy to miss the ways in which big picture narratives manifest to limit possibilities in policy fights.From national security logics that can be a vector both for,and cutting against,moves towards industry accountability;to refr

87、aming traditional data privacy levers as key tools in the fight against automation and addressing market power.5.Reclaim a positive agenda for public-centered innovation without AI at the center.The current trajectory of AI puts the public under the heel of un-accountable tech oligarchs.But their su

88、ccess is not inevitable.By moving out from under the shadow of the idea that large-scale AI is inevitable,we can reclaim the space to conduct real innovation and to push for exciting and novel alternative pathways that leverage tech to shape a world that serves the public and is governed by our coll

89、ective will.A more fulsome treatment appears in Chapter 4:A Roadmap for Action,but here are the highlights:EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT12EPILOGUE:THE WORLD WE WANT (AND WHY THE CURRENT TRAJECTORY OF AI WONT GET US THERE)The AI hype of the past year has sucked the air out of an alrea

90、dy stuffy room,making it feel futileat times,impossibleto imagine anything other than a steady march toward the inevitable supremacy of AI.But no matter how true that may feel,it is only that:a feeling.It is not realitynot yet,at least.There are,in fact,many alternatives to this version of AI,many w

91、ays to shape new worlds.Like AI,though,these are not in-evitable either.Making them possible starts by asking and answering a single question:Is this the world we want?At AI Now,we want to see a world that has:Good JobsEveryone deserves access to the resources to live a happy,fulfilled life,and a di

92、gnified job that will provide them with these resources.Under the right conditionslike policies intent on uplifting,rather than exploiting,workersnew technologies have the potential to make everyones working lives better.Yet AIs current trajectory is fundamentally incom-patible with the proliferatio

93、n of good jobs rooted in human flourishing.As the AI industry embeds itself into nearly every sector of the economy,firms are increasingly shaping a job market contingent on EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT13worker displacement and exploitation.Overwhelming-ly,AI companies are embedding

94、“productivity”tools designed to help businesses optimize their bottom line across the entire labor supply chain.This requires work itself to become legible to AI systems,making working life more routinized,surveilled,and hierar-chical.Furthermore,instead of working to protect workers from the uncert

95、ainty coming from this new market,AI companies are undermining hard-won labor protections,exploiting legal loopholes to avoid corporate accountability,and lobbying governments to support policies that prioritize corporate profits over fair and just treatment for workers.As the current vision of AI t

96、akes hold,we lose a future where AI technology works in support of stable,digni-fied,and meaningful jobs.We lose a future where AI supports fair and livable wages,instead of wage depreciation;where AI ensures that workers have the control to decide how new technology affecting their careers is deplo

97、yed,instead of undermining their expertise and knowledge of their own work;where we have strong policies to support workers if and when new technologies automate existing rolesincluding laws that broaden the social safety netinstead of AI boosters who brag to shareholders about cost savings from aut

98、omation;where robust public benefits and time-off policies ensure the long-term wellness of em-ployees,instead of AI being used to surveil and nickel-and-dime workers at every turn;where AI helps protect employees from health and safety risks on the job,instead of perpetu-ating conditions that make

99、work dangerous and celebrating employers who exploit labor loop-holes to avoid responsibility;and where AI fosters meaningful connection through work,instead of driving cultures of fear and alienation.Shared ProsperityThe proliferation of any new technology has the po-tential to increase economic op

100、portunity and lead to widespread shared prosperity.But shared prosperity is incompatible with AIs current path toward maximiz-ing shareholder profit.The insidious myth that AI will lead to“productivity”for everyone when it really means productivity for a select number of corporate firms propels us f

101、urther down the path of shareholder profit as the singular economic goal.Even well-intentioned government policies designed to boost the AI industry steal from the pockets of workers.For example,government incentives meant to revitalize the chip manufacturing industry were thwarted by corporate buy-

102、back provi-sions,sending millions of dollars to companies,not to workers or job creation.And despite some meaningful moves to investigate the AI industry under the Biden Administration,companies have still gone largely unchecked,meaning new entrants cannot come in to challenge these practices.By pro

103、liferating the myth that AI will inherently lead to shared prosperity(or that“a rising tide lifts all boats”),we lose the economic policies that could meaningfully lead us into a period of shared prosperity,including pro-enforcement policies to break up the concentration of corporate power,a strong

104、pro-labor agenda to center the needs of workers,and industrial policy EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT14strategies designed to put workers and communities before the bottom line of large corporations.Crucially,we lose a thriving and competitive economy,where innovators and entrepreneurs

105、 are incentivized to launch sustainable and prosperous businesses that need not rely on surveillance mechanisms,hyper-growth venture capital funding,and extractive business models to succeed.Freedom&AutonomyWe all deserve to live in a world where our personal,political,and economic lives are free fr

106、om coercion.But the current trajectory of AI is grounded in coer-cion and opacity.By amassing such concentrated power,AI companies have assumed control over many aspects of our lives,subjecting us to coercive practic-es designed to maximize their own profit-making po-tential.Nowhere is this more exp

107、licit than the regime of surveillance pricing algorithms.These algorithms collect extraordinary amounts of consumer data to set individualized prices for goods and services,such that important aspects of daily lifefrom buying gro-ceries,taking an Uber,or buying an apartmentare controlled by companie

108、s looking to squeeze consum-ers to pad their own bottom lines.As an inherently centralizing technology,AI is consistently deployed to support and benefit from surveillance states,carceral systems,and military techniques,embedding corpo-rate interests into state apparatuses and vice versa,making life

109、 more coercive and violent.AI could contribute to a future built on autonomy and transparency.But right now,we stand to lose a robust ecosystem of public resources that are not depen-dent on private industry players or priced for profit;pricing for goods and services driven by principles of fair com

110、petition;and changing life circumstances not being treated as grounds for companies to profit.A public life invested in freedom would be divested from the proliferation of surveillance systems,carceral logics,and military state apparatuses,enabling people and communities to thrive.Sustainable Future

111、Technological progress does not have to come at the expense of our natural environment.A publicly bene-ficial AI landscape places principles of sustainability and environmental justice at its center,recognizing that the health and safety of our planet and communi-ties is of paramount importance.AIs

112、current trajectory is not merely incompatible with sustainability;it is fueling climate degradation.The focus on scale at all costs within the AI industry makes it dependent on climate extraction and energy dominance.These include a deregulatory policy intent on accelerating the AI industry(meaning

113、more ener-gy,more infrastructure,more natural resources going into AI)as well as the insidious belief that AI is going to help solve the climate crisis,“greenwashing”the environmental harm it is already enacting.As a result,what we lose is a world where govern-ments and companies work together to ad

114、vance principles of sustainability and environmental justice,investing in green and renewable energy infrastruc-ture to support the additional energy usage that new EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT15technologies require.Decisions on siting,permitting,and constructing new technology infr

115、astructure should be made in relationship with local communi-ties,with particular attention to those communities who are most significantly affected by technologys industrial evolution.Cities and states should not be coerced into providing subsidies for infrastructure that come at the expense of loc

116、al communities needs,like funding for schools and healthcare.We also lose a future where industry takes its climate commitments seriously and works to mitigate the harmful effects of its industrial processes,recognizing that the health and safety of our planet and communities is key to everyones sur

117、vivalincluding industrys.Strong Social Safety NetWe should live in a world where everyone has access to and community control over a robust system of public resources.Yet as AI rapidly shifts our social landscape,govern-ments are increasingly driven by privatization and austerity measures.AI is ushe

118、ring in a policy agen-da designed to enrich private interests rather than provide robust public benefits.AI firms are pushing AI integration in local and federal government agen-cies driven by austerity,restricting peoples access to needed resources and the social safety net.More-over,the purported

119、AI arms race with China is being leveraged to convince governments that domestic infrastructure is an imperative for national competi-tiveness and security,encouraging public agencies to throw generous tax exemptions at private companies in order to build massive data centers in communities that may

120、 not want them there.These exemptions totaling billions of dollarssteal investment from strong public resources that benefit everyone,like investments in more teachers,roads,and libraries.We could have a world where technology works in service of the broader public,like algorithms designed to maximi

121、ze peoples options,not reduce them;and where companies are required to pay their fair share of taxes to local communities,rather than fighting efforts to make them pay their fair share.SecurityIn an increasingly complex world,security and resil-ience are more important than ever.Our infrastructures

122、are largely invisible to us until the moment they break down-and weve felt the shock of infrastructural failure frequently over recent years,from supply chain issues during the pandemic to power outages to bank closures.Layering AI into our critical infrastructures,particular-ly when AI systems are

123、highly concentrated,creates a real and present security risk.These risks are mani-fold:there are cybersecurity risks that are inherent to many AI systems that make them vulnerable to hack-ing,and some of these cannot be remedied.There are systemic risks introduced by overreliance on a single technol

124、ogy:for example,if banks,hospitals,and schools all use the same cloud infrastructure provider,an outage could affect all of these at one fell swoop.These risks are at their highest when AI is used in life or death settings-and from healthcare to defense,these are some of the industrys prime markets.

125、And there are risks emanating from decisions made by EXECUTIVE SUMMARYAI NOW 2025 LANDSCAPE REPORT16the companies themselves to experiment in the wild,bringing to market technologies that have not been adequately tested or validated and with little certainty that they will work as intended,let alone

126、 cause harm.We could have a world where our safety and well-be-ing are not vulnerable and exposed to an industry that is scaling at an unprecedented rate with little regard for safety and security,let alone compliance with the law;where AI is validated,tested and built safe by de-sign and used with

127、prudence rather than impunity.Innovative Tech EcosystemTechnology has the potential to solve important soci-etal challenges and push the frontiers of innovation forward.In a thriving tech ecosystem,companies big and small are able to succeed,not by amassing con-centrated power but by engaging in fai

128、r competition.Society benefits from the distribution of a diverse set of products,services,and technologies that result from such competition.The current AI industry is defined by concentration,precluding a truly diverse and innovative tech ecosys-tem from flourishing.There is no AI without Big Tech

129、 firms,which have spent decades amassing unre-strained data access and economic power and then used those advantages to control key inputs at all lev-els of the AI stack.Even where new entrants are able to enter the AI market,they are still dependent on the cloud and computing infrastructures of Big

130、 Tech firms in order to succeed,creating an ecosystem of depen-dence rather than competition.Where consolidation has been averted-as was the case with the proposed Nvidia-ARM merger-firms have been able to thrive.ARM went on to IPO and beat quarterly estimates,all after its acquisition was blocked.F

131、urthermore,Big Tech companies control most pathways to consumers and enterprise businesses at scale.This centralized power is also driving a crisis of innovation,where Big Tech companies are bloated,stalled by legal reviews,and are stuck repackaging their existing technologies in order to revive and

132、 boost their bottom lines.What we lose in this bloated and stale tech ecosystem is a truly diverse horizon of possibility,filled with inno-vations that tackle peoples real-life needs,rather than an endless soup of enterprise software and AI agents.We can inspire an ecosystem where people can use tec

133、hnology to build companies slowly and sustain-ably,without the need to grow rapidly and amorally in order to stay in business.Our entire tech ecosystem is in need of a paradigm shift,one that tears down existing structures to make room for complexity and emergence.22 This includes breaking up big co

134、mpa-nies,overhauling the VC-backed funding structure so more companies can thrive,investing in public goods to ensure tech resources are not dependent on large private companies,and increasing institutional invest-ment to bring more diverse peopleand thus ideasinto the tech workforce.Vibrant,Democra

135、tic StateWe deserve a technological future that works to sup-port strong democratic values and institutions.What we have now is a society captured by the tech billionaire class.Over the past few decades,a handful of billionaires degraded our entire information system EXECUTIVE SUMMARYAI NOW 2025 LAN

136、DSCAPE REPORT17under the guise of disruption,killing the business model of newsrooms and replacing it with ad-based products required to keep our constant attention.And despite rhetoric that AI has the potential to“de-mocratize”the world,the inherent pathologies of AI make it a centralizing force,23

137、 contingent on the mass accumulation of data and compute resources in the hands of a few big players.24 Now the tech billionaire class threatens to destroy our creative industries,transforming hard-earned craft into“content”that is then fed to AI models intended to churn out lossy xerox copies of ou

138、r masterworks.And,as if thats not enough,these same billionaires have begun to destroy our institutions,purchasing newspapers and taking over the opinion pages,buying elections,and hollowing out our social services.Contesting the economic and political capital amassed within the tech industry is nec

139、essary to create the conditions for a thriving democracy.We urgently need to restore the institutional structures that protect the interests of the public against oli-garchy.This will require confronting tech power on multiple fronts,from enacting corporate account-ability measures that keep tech ol

140、igarchs in check,to staving off efforts to use AI to hollow out our institu-tions,to bolstering work happening at the community level among local government officials,organizers,and workers devoted to rebuilding a democracy that serves the broader public.CHAPTER 1:AIS FALSE GODSWHATS PROPPING UP THI

141、S BUBBLE AND WHY IS IT SO HARD TO NAME?AI NOW 2025 LANDSCAPE REPORT19CHAPTER 1:AIS FALSE GODS1.1:THE AGI MYTHOLOGY:THE ARGUMENT TO END ALL ARGUMENTSThe promise that artificial general intelligence,or“AGI,”is hovering just over the horizon is tilting the scales for many of the debates about how AI is

142、 affect-ing society.AI firms investing in the development of very large models at scale constantly assert that AGI is months or weeks away1,poised to have transforma-tive effects on society at largemaking this central to their sales pitch for investment.2 The discourse around AGI adds a veneer of in

143、evitability to conversa-tions about AI;if one company doesnt achieve it,an-other will.This also gives governments an excuse to sit on their hands even as current versions of AI have profound effects on their constituents,as though the race to create AGI has its own momentum.If anything,under both th

144、e Biden and Trump admin-istrations,the US government has instead positioned itself as chief enabler:ready to wield every tool at its disposalincluding investment,executive author-ity,and regulatory inactionto push American AI firms ahead of their competitors in this race to AGI.3 Its worth noting th

145、at those most vocal about their The“common sense”around artificial intelligence has become potent over the past two years,imbuing the technology with a sense of agency and momentum that make the current trajectory of AI appear inevitable,and certainly essential for economic prosperity and global dom

146、inance for the US.In this section,we break down the narratives propping up this“inevitability,”explaining why it is particularly challengingbut still necessaryto contest the current trajectory of AI,especially at this moment in global history.CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT20fea

147、rs about the so-called“existential risks”posed by AGI have done as much to prop up and speed along industry development as anything or anyone else.4 OpenAIs assertion that“its hard to fathom how much human-level AI could benefit society,and its equally hard to imagine how much it could damage societ

148、y if built or used incorrectly”5 drives home that the AI boosters and the existential(“x-risk”)fearmongering both play a role in propping up this vision of AI with supreme capabilities.WHAT IS AGI?THE HISTORY OF ARTIFICIAL GENERAL INTELLIGENCE As Brian Merchant chronicles in his report“AI Gener-ated

149、 Business,”the term AGI,coined in 1997,cap-tured the notion of a“general intelligence”as a coun-terpoint to the then-dominant current in AI research,“expert systems,”which operated on rule-based logic designed as a formalized representation of how humans think.6 Where expert systems only worked in t

150、he narrowest of applications,AGI would operate broadly across a wide range of domains.But develop-ers in the field largely ditched these ways of thinking about AI,turning instead to deep-learning techniques that proved more effective and that form the basis of todays automated decision-making system

151、s,among others.Interest in AGI was revived in the 2010s when com-panies like OpenAI seized on the term,first as short-hand for a form of machine intelligence intended to rival and eventually surpass human intelligence,and later as a term“central to their marketing efforts.”7 The images invoked by AI

152、 firms is instructive,from Anthropic founder Dario Amodeis use of the sublime imagery of“geniuses in a data center”capable of par-adigm-changing scientific leaps like“designing new weapons or curing diseases,”8 to the straightforwardly commercial logic underpinning OpenAIs agreement with Microsoft:A

153、GI is when AI can create$100 billion in profits.9In this sense,ChatGPT walked so that AGI could run;the current crop of LLMs in the consumer market are examples of brilliant marketingproof,as AI firms argue,that big,unexpected advancements in AI were not only possible but“just around the corner.”10

154、AGI has since been positioned as the next big step in the LLM advancement trajectory,albeit with little proof,beyond speculation,of how far or wide this leap will have to be.11 However,while this belief seems to be spreading among the general public,it is widely con-tradicted by many within the AI r

155、esearch community.For instance,in a recent survey of members of the Association for the Advancement of AI,84 percent of respondents said that the neural net architectures that large models rely on are“insufficient to achieve AGI.”12 In another,more fundamental,debunking of AGI claims,scholars like E

156、mily Bender13 and Henry Farrell,14 among others,have contested the basis of claims to AGI,arguing instead that large models can“never be intelligent in the way that humans,or even bumble-bees,”15 are because AI cannot,in fact,create.It can only reflect,compress,even remix content that humans have al

157、ready created to help people to coor-dinate or solve problems.16 While current AI models make the promise of AGI more tangible for policymakers and the general pub-CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT21lic,AGI is conveniently distanced from the fundamen-tal and persistent limitations

158、 of LLMs on the ground that AGI,by definition,will be a wholly new paradigm that leapfrogs these material concerns.The mythol-ogy around AGI masks the shallowness of todays AI models,providing substance and imagination that innovations are just around the corner.IF AGI WERE HERE,HOW WOULD WE EVEN KN

159、OW?Despite bold public claims from the tech industry that AGI is“as little as two years”18 away,the research community has yet to agree.19 A recent survey by the Association for the Advancement of AI(AAAI)of nearly five hundred AI researchers found that 76 percent of respondents assert that scaling

160、up current approaches to yield AGI is unlikely or very unlikely to succeed.20 So how will we even know when AGI is here?The metrics currently on offer are largely narrow,vague,and self-serving benchmarks21and some research-ers have argued that the preoccupation with AGI is“supercharging bad science.

161、”22 In place of scientific breakthroughs,industry labs are hinging claims to proximity to AGI on grandiosely named tests like“Humanitys Last Exam”23 and“Frontier Math”24 that gauge only a very narrow ability to answer clear,closed-ended questions25poor proxies for the promises these companies make a

162、bout the capabil-ity of this technology like inventing cures to cancer or solving for climate change.AI company Hugging Faces Chief Science Officer Thomas Wolf has argued were currently testing systems for their ability to be obedient students,rather than for their mastery of bold counterfactual app

163、roaches or the ability to chal-lenge their own training data,which might show more promise for solving complex,intractable problems.26 In 2025,a group of AI researchers from across aca-demia and industry pointed to an endemic challenge within the current field of AI evaluations that is more preoccup

164、ied with“coarse-grained claims of general intelligence”than“real-world relevant measures of progress and performance.”27 In sum,there is a widespread and endemic lack of clarity on both the definition and time scales of the AGI conversation,which makes it hard to contest or reason its merits.The mor

165、e urgent inquiry,however,is who and what does this disproportionate focus on AGI work in service of?How will it shape the current trajectory of AI?WHO BENEFITS FROM AGI DISCOURSE?AGI has become the argument to end all other argu-ments,a technological milestone that is both so ab-stract and absolute

166、that it gains default priority over other means,and indeed,all other ends.It is routinely cast as a technology so all-powerful that it will over-come some of the most intractable challenges of our timeand that both investment into the sector and CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT22

167、ancillary costs are justified by the future solutions it will offer us.For example,Eric Schmidt recently dismissed the climate costs imposed by AI by assert-ing that humans arent set up to coordinate to solve climate change.Thus,the reasoning goes,we need to supercharge data centersbecause in the lo

168、ng term,AGI has the best shot at solving for it.28 This not only reflects abstract AI solutionism at its peak;it also serves to flatten and disguise the problem of climate change itself as waiting for its technical silver bullet,rendering the challenges of political will,internation-al cooperation,a

169、nd material support for people to rebuild homes or house climate refugeeseverything it will take to meaningfully“solve”climate changeinvisible.29 Presenting AI as a quick technical fix to long-stand-ing,structurally hard problems has been a consistent theme over the past decade(as we explore in our

170、chapter on Consulting the Record),but past variants of technosolutionism at least had to demonstrate how the technology would solve the problem at hand.With AGI,though,were not clear how this transformation will happen beyond the assertion that the current state of affairs will be overhauled.The deb

171、ates around DOGE transforming government using AI have this flavor:In his interview with Ben Buchanan,Ezra Klein speaks of the general sentiment that with superintelli-gent AI potentially around the corner,the government will inevitably need to be taken apart and rebuilt in the age of AGI.30 Its the

172、 same logic that dictates that if AGI is truly going to propel scientific discoveries of the kind that Amodei promises,then perhaps there will be no need for federal funding for science at all.AGIS MARKET-BOOSTING FUNCTION Asserting that AGI is always on the horizon also has a crucial market-preserv

173、ing function for large-scale AI:keeping the gas on investment in the resources and computing infrastructure that key industry players need to sustain this paradigm.As weve argued,this current avatar of large-scale AI was set in motion by the simple rule that scaling up data and compute would lead to

174、 performance advancementsa logic that sedimented the dominance of the handful of companies that already controlled access to these inputs,along with pathways to market,31 and in whose hands power would be further concentrated if AGI ever were achieved.32 The quest for the ever-shifting goalpost of A

175、GI only reinforces this“bitter lesson”(as Anthropic CEO Amodei calls it).33 Theres a lesson here from the 1980s,when,even before the term AGI was in vogue,the Reagan admin-istration pushed for a wildly ambitious(for the time)“Strategic Computing Initiative”that was focused on propelling general adva

176、ncements in“AI”along the lines of the AGI promise.34 It was propelled by the promise of new military capabilities,anxieties about Japanese domination on AI,and the potential of pri-vate-sector opportunities.A billion dollars in taxpayer money was spent then on a program,now universally acknowledged

177、as a failure,that didnt yield results even on the terms it set for itself.The postmortem of why it failed yields varied conclusions,but one is CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT23worth underscoring:Then,as now,these advance-ments hinged not on revolutionary feats in science,but on

178、scaling up computing power and data.Coincidentally,existential risk arguments often have the same effect:painting AI systems as all-powerful(when in reality theyre flawed)and feeding into the idea of an arms race in which the US must prevent China from getting access to these purportedly dan-gerous

179、tools.35 Weve seen these logics instrumented into increasingly aggressive export-control regimes.By drawing attention to the very systems they pur-portedly aim to contest,x-risk narratives create a Streisand effect:encouraging more people to see the AI dystopia in their present,fueling adoption and

180、bolstering industry players rather than curbing their power.They have also narrowed the scope for policy intervention,bolstering a debate centered around the two poles of accelerationism and deceleration rather than facilitating a broad dialogue about AI develop-ment and its societal implications.Ul

181、timately,these twin myths around AGI position AI as powerful and worthy of investment,and draw atten-tion away from the evidence to the contrary.DISPLACING GROUNDED EXPERTISE:WHO IS DISEMPOWERED BY THE AGI DISCOURSE?Elevating AGI over other paths to solving hard prob-lems is just a supercharged form

182、 of technosolution-ism,36 but it also means that those with technical expertisenot only those driving the tech develop-ment but also those fluent in using this new suite of toolsare normalized as primary experts across broad areas of society and science in which they lack domain-specific context and

183、 experience.37 This has been a familiar fight over the past decade of AI devel-opment:Those with lived experience and sector-spe-cific knowledge have had to advocate for a determin-ing role in questions around whether,and how,AI is deployed.Whether that means nurses having a say in how AI is integra

184、ted in patient care,or parent groups fighting against the use of facial recognition on their children in the classroom,there has been a consistent push to recenter who is counted as an expert on baseline questions about AI integration.(Notably,some of this has often resulted in tokenistic approaches

185、 that pro-vide nominal seats at the table to impacted commu-nitiestoo little,too late.)AGI presents a more formi-dable version of this challenge given its abstract and absolutist form.For example,narratives around AGI upending the world of work routinely position work-ers across industries as being

186、subjectsor worse,collateral damageof a great transformation,rather than as participants and indeed experts in the ways in which these transitions will take place.38AI NOW 2025 LANDSCAPE REPORT24CHAPTER 1:AIS FALSE GODS1.2:TOO BIG TO FAIL:INFRASTRUCTURE AND CAPITAL PUSHThe AI industrys growth model,f

187、ueled by the asser-tion that infinitely increasing scale leads to superior products,has spawned AI firms that are positioned to be too big to fail.Americans are actively subsidizing this unstable system under the premise that the adop-tion of AI is a“national strategic priority.”As we illus-trate in

188、 this chapter and in Chapter 1.4,though,this has enabled an industrial-policy approach that will ultimately undermine,rather than strengthen,our na-tional security.Finally,we discuss how the abundance agenda,with its seemingly benign focus on what it calls“supply-side progressivism,”is a very conven

189、ient tool for big AI to justify expanding its energy needs.Tech firms are deploying unprecedented amounts of capital to maintain their lead and advance in the current paradigm of“scale is all you need”AI,dou-bling down on infrastructure build-out and seeking federal funding and regulatory support ac

190、ross several dimensions:access to chips and associated hardware to equip data centers,approvals for the construction of the data centers themselves,and the energy nec-essary to power them.The stock market is riding this hype wave,and the“Magnificent Seven”stocks(Al-phabet,Amazon,Apple,Meta,Microsoft

191、,Nvidia,and Tesla)now represent more than 30 percent of the S&P 500,the largest sector of the indexin prominent part because of the AI boom.1 Its important to remember that the pursuit of scale was a choice that locked us into a future where a handful of Big Tech firms retained control of the market

192、(see the Introduction).This is not the only way for AI to develop,nor is advancement measured on a narrow set of self-serving benchmarks2 a meaningful proxy for evaluating the societal utility of these sys-tems.3 But because it is what these key market players have doubled down on,and because of the

193、ir central-ity to market indices,the success or failure of the AI bubble will now have a profound effect on the stock CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT25market as a whole.4 This raises the stakes around the push for public investment in AI infrastructurea move that is at best a he

194、dge,and at worst a subsidy,for the profoundly risky and self-interested set of bets by AI firms.If successful,this effort will lock in infrastructures that the public will pay dividends on for years to come,in the form of financial and mate-rial costs(see Chapter 2:Heads I Win,Tails You Lose),creati

195、ng a path dependency toward continued domi-nance by large AI firms.TECHS CAPEX FRENZYFirms like Microsoft,Google,and Meta need AI to be profitable because they have funded the AI boomat many orders of magnitude more than traditional venture capital5boosting the valuations of startups that are far fr

196、om demonstrating the kind of profitabil-ity that traditional investors would seek.They have gone all in on the most capital-and resource-intensive version of AI by adopting the“scale is all you need”paradigm as canon.This is not the only way to ap-proach building AI models,and the companies leading

197、AI development have occasionally gestured toward the need for model efficiency to address compute infrastructure bottlenecks.This was brought home especially by the release of DeepSeeks R1,which demonstrated model capabilities on par with the lead-ing-edge models of US firms,without anything like th

198、e scale US firms rely on.6But rather than make concerted efforts to build models differently,many dominant firms are doubling down on this approach by seeking public investment and the rollback of regulation to de-risk the expan-sion of the AI market.For example,within weeks of the DeepSeek announce

199、ment,OpenAI announced its Stargate investment with SoftBank,which will allocate a$100 billion investment into data center infrastruc-tures for model training.7GETTING HIGH ON AI SUPPLYThe US has adopted a position over the past two years that treats AI as an exceptional sector core to the nations ec

200、onomic and national security interests.This stance exists in tension with growing friction with Big Tech firms,most clearly articulated in the Biden ad-ministrations Executive Order on Competition,which articulated the perpetuation of national monopolies as antithetical to the national interest.8 Th

201、e Trump administration has likewise bought into AI boosterism even as it has gestured toward the need for antitrust,mostly as a political tool for addressing firms it sees as adversarial to its interests.9 As chief case in point,Trumps pick to head the FTC,Andrew Ferguson,vowed to go after tech mono

202、polies while taking a hands-off approach to AI regulation,proving that at-tacks on corporate tech power reach their limit when it comes to AI.10 In tandem,a cadre of appointments related to the environment and energyincluding Lee Zeldin as head of the EPA;Jacob Helberg as under secretary for economi

203、c growth,energy and the envi-ronment;Doug Burgum as dual interior secretary and CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT26“energy czar”;and David Sacks as a newly created“AI czar”have inextricably tied support for a strong national AI industry to achieving energy dominance,positioning en

204、ergy expansionism as the essential tool to achieve the administrations economic nationalism agenda.11Recent movements from within the federal govern-ment have backed this stance:The Department of Energy recently announced it had identified sixteen federal sites across the country positioned for rapi

205、d data center construction,12 and in April the Trump Administration signed an executive order ramping up domestic coal mining using growth in demand from AI data centers as justification.13 AI FIRMS WANT TO BE TOO BIG TO FAILThese infrastructure investments function to lock us into a world where US

206、continued dominance in the AI market is guaranteed by the government,and,for now,largely supported by investors in the stock mar-ket seeking to avoid an end to the AI bubblewhile taxpayers foot the bill(whether by taxes that contrib-Small(AI)Is Beautiful?Differentiating to Avoid Industry Co-OptionA

207、growing number of technologists and civil society organizations advocate for smaller models as the alter-native trajectory to the bigger-is-better paradigm.14 This makes sense,because many of the clearest pathologies within the AI industry are driven by scale:from climate impacts;to risks of contagi

208、on effects from privacy,security,and accuracy failures;to the ways in which scale breeds ultra-concentrated markets in AI.The dangerous impacts of the vague and all-encompassing“AGI”(see Chapter 1.1)also demonstrate the scale thesis taken to its logical end:a system that exists at a scale and level

209、of universality that,hypothetically,displaces all other forms of expertise and value.But industry is flocking to a version of the“small is beauti-ful”thesis,too,as part of their plans for market expansion,creating a familiar risk of co-option of the alternative by the same players who have driven an

210、d shaped this current paradigm.In the summer of 2024,Microsoft heralded“tiny but mighty”smaller language models that would provide impressive performance despite a reduced number of pa-rameters.15 Apple,Meta,and Google also released AI mod-els with many fewer parameters,signaling that industry is in

211、centivized to move away from simply bigger-is-better in pursuit of compute-efficient methods.16 DeepSeek only propelled this trend,making it clear that frugality would be a key competitive advantage in this market.17This is only superficial common ground.Positioning“smaller”models as one of the opti

212、ons in an“all of the above”approach for the biggest AI companies should not be confused with a rejection of the bigger-is-bet-ter paradigm.As Satya Nadella said after the DeepSeek announcements,signaling that these efficiencies only consolidate benefits for the tech giants best placed to capture dem

213、and(see Chapter 2:Heads I Win,Tails You Lose):“As AI becomes more efficient and accessible,we will see exponentially more demand.”18 It also ignores that pushing advancements at the“frontier”of this tech is still dictated by scale,even as firms play around with a mix of approaches across their portf

214、olio to target different types of consumers.Most importantly,the large-scale version of this tech is what drives these firms policy lobbying around infrastructure expansion with deleterious impacts on the public.Movements that aim to disrupt the consensus around scale-driven AI must explicitly name

215、and distance themselves from this industry-driven discourse.CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT27ute to these investments,or more directly through increased energy bills,as we unpack in Chapter 2:Heads I Win,Tails You Lose).AI industrial policy serves either to secure demand via pro

216、curement policies19 or to underwrite and attract continued investment(as The Abundance Agenda:AIs Fundamental Incompatibility with Supply-Side ProgressivismThe emergence of“abundance”as a narrative strategy and pol-icy platform is being used by tech firms to get access to scarce public subsidies and

217、 energy.This stance has formed around a constellation of thinkers and organizations working across party lines to articulate a policy agenda premised on building a policy apparatus in support of more,and more efficient,con-struction of critical resources with low supply and high demand,including hou

218、sing,healthcare,and energy.It operates under the presumption that(1)government regulation makes building too burdensome in these sectors,leading to cost inflation;and(2)progressives have focused too intently on subsidy programs that cut or block access,rather than on the underlying reasons for cost

219、inflation.The solution,abundance movement advocates posit,is to push forward“supply-side progressivism,”or,as Ezra Klein puts it,“to take innovation as seriously as they take affordability”20 by implementing regulatory reforms that speed development and solve scarcity.Abundance proponents centrally

220、contend with energy markets,in that they argue in favor of cutting regulation to enable an increase in energy production.For example,Jerusalem Demsas wrote in the Atlantic that the ability for NIMBY-minded commu-nity organizations and climate groups to shut down renewable development is hindering th

221、e USs ability to meet its climate goals.21 Klein and Derek Thompson argue that overhauling ener-gy infrastructure is crucial to mitigating climate change,empha-sizing that the first step toward an abundant clean-energy future is reducing the current fossil fuel reliance from 60 percent as of 2022 to

222、 nearly 0 percent.22As a growing number of AI companies prioritize building and opening new data centers,more energy is needed to meet the staggering demand.One might think that AI-driven demand would concern abundance advocates,because AI firms soak up the available supply of renewable energy.Data

223、centers already account for 4.59 percent of all energy used in the US.That number has doubled since 2018.23 Goldman Sachs estimates that data center power demand will grow 160 percent by 2030.24 These are staggering numbers wreaking havoc on an already fragile energy grid.Instead,we see a more uneas

224、y alliance,where the abundance agenda potentially converges with the energy deregulation camp for whom the“urgent”need to advance AI is being used as a justification to fast-track and expand fossil fuel production and use.At the House Oversight Committee hearing on data centers,AI,and energy,legisla

225、tors repeatedly threw renewables under the bus,even touting that China is powering their AI systems with coal-fired plants.25 The fossil fuel company talking point that wind and solar are not a reliable source of energy to meet data centers 24/7 demands is deeply ingrained,26 with legislators and da

226、ta center trade groups pivoting toward the expansion of nuclearrather than renewableenergy to provide“reliable”and sturdy energy for AI.Despite the substan-tial evidence on hand,this sustainability critique has not been taken seriously by abundance advocates skeptical of the climate movement.is the

227、case with the Stargate deal).This approach to AI is akin to industry bailoutsrarely a popular pol-icy stancebut compared to the auto industry and banking,the AI market is much more speculative and its value to the public is unproven.AI NOW 2025 LANDSCAPE REPORT28CHAPTER 1:AIS FALSE GODS1.3:AI ARMS R

228、ACE 2.0:FROM DEREGULATION TO INDUSTRIAL POLICYThe fusing of economic and national security goal-posts under the banner of the US-China AI arms race is a critical asset for US AI firms:It affords them patronage not just from their own government,but potentially from the many other nation-states vying

229、 for a fighting chance at national competitiveness in this market;it insulates them from regulatory friction by framing any calls for accountability as not just an-ti-innovation but harming national interests;andas we explore in Chapter 1.2:Too Big To Failis a key factor in positioning them as not j

230、ust too big,but too strategically important,to fail.Nation-states have developed their own flavors of“AI Nationalisms,”embarking on initiatives designed simultaneously to support homegrown development and sovereign infrastructures free of dependency on US tech firms,and to attract AI investment.1 Bu

231、t though AI nationalism is on the rise globally,the rhetoric around the AI arms race remains centered around two poles:the US and China.Since the mid 2010s,the notion of a US-China AI arms race has been primarily deployed by industry-motivated actors to push back against regulatory friction.A freque

232、nt motif in policy discussions at moments where the industry has sought to stem the tide of regulation,the notion of an arms race was one of the key arguments made against the introduction of a federal data protection law,a package of antitrust reforms targeting the tech industry in 2022,and an omni

233、bus AI Accountability Bill that was considered before Congress.2 In the past two years,this so-called race has taken on a new character(lets call it the“AI arms race 2.0”),taking shape as a slate of measures that go far be-yond deregulation to incorporate direct investment,subsidies,and export contr

234、ols in order to boost the interests of dominant AI firms under the argument that their advancement is in the national interest(what we refer to as AI industrial policy3).Such an CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT29approach predates the Trump administration.Argu-ably a number of the

235、 core measures propping up the AI arms race 2.0 were outlined under the Biden Administration;Jake Sullivan,in particular,was a vocal proponent of the logics of economic security.4 The Biden administrations AI Executive Order,5 National Security memo,6 and export controls7 all established an intent f

236、or the US government to widely adopt AI and to clear the pathway for the industry to expand through infrastructure build-out,while simultaneously hindering the advancement of strategic adversaries like China by limiting the export of leading-node chips.Unsurprisingly,this stance ran parallel to the

237、lobby-ing platforms of firms like OpenAI that have sought government cooperation,with a narrow list of condi-tionalities such as the use of renewable energy and compliance with security measures.8 OpenAI specifi-cally has made threats that it will relocate its business absent commensurate support fr

238、om the US govern-ment.9 Since inauguration,the Trump administration has escalated support for the AI industry,rolling back the conditionalities articulated by the Biden adminis-tration by repealing the AI Executive Order and replac-ing it with a blanket assertion:“It is the policy of the United Stat

239、es to sustain and enhance Americas global AI dominance in order to promote human flourishing,economic competitiveness,and national security.”10 A NEW SILICON VALLEY CONSENSUS BEYOND TARGETED ADS TO TARGETED AI WEAPONS?While the Trump administration has firmly asserted AI as a strategic national asse

240、t,they are likely to expect the industry to act in ways that align more closely with state interest.The specifics of what that means is left deliberately hazy,but a popular refrain has been that companies should be devoted less to targeted advertising,and more to AI that would bolster nation-al secu

241、rityand defense tech is increasingly front and center of events like the Hill&Valley Forum,11 an annual consortium of Silicon Valley elites and DC law-makers that first convened in March 2023 to combat Chinas influence on the American tech industry.12 Cofounded by Palantirs Jacob Helberg,the Hill&Va

242、l-ley Forum is more aligned than ever before with state national security interests,13 as Helberg,14 like Michael Kratsios and David Sacks,is one of many industry rep-resentatives who find themselves in key policy roles under the Trump administration.15 So far,the industry seems to support this visi

243、on.This is best seen in the rhetoric of Palantirs CEO Alex Karp,who has long framed the companys mission as addressing a civilizational need to support democratic and Western supremacy through leading-edge tech-nology.But emboldened by Trumps intent to scale CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDS

244、CAPE REPORT30up mass deportations and police surveillance,Karp has escalated,saying in an investor call in early 2025:“We are dedicating our company to the service of the West and the United States of America,and were su-per-proud of the role we play,especially in places we cant talk about.Palantir

245、is here to disrupt.And,when its necessary,to scare our enemies and,on occasion,kill them.”16Karp isnt alone.Since the Biden administrations shift toward securitization of AI in 2024,companies that have historically distanced themselves from the military have also doubled down on national secu-rity:A

246、fter making an amendment to its permissible use policy enabling its tools to be used by militaries,17 OpenAI has increasingly leaned in to making policy arguments on security grounds,18 going so far as to assert that expanding fair use under copyright law to include AI development is a security impe

247、rative.19 In February 2025,Google amended its guidelines to allow its AI technologies to be used for military weap-ons and surveillance,despite ongoing protests by its employees and a long-standing ban on use of its technology for weapons following the Project Maven protests of 2018.20 And Meta made

248、 an announce-ment in November 2024 that it would make its Llama models available to the US government for national security use.21Meanwhile,Anthropics CEO Dario Amodei recently wrote about the threat of authoritarian governments establishing military dominance on AI as a reason to accelerate US lead

249、ership22 and the VC firm Andrees-sen Horowitz operates an“American Dynamism”practice expressly designed to support the national interest in strategically important sectors:aerospace,defense,public safety,education,housing,supply chain,industrials,and manufacturing.23 A DOUBLE-EDGED SWORD:CHIP DIFFUS

250、ION AND“SOVEREIGN AI”Its worth noting that the AI arms race 2.0 has shifted from being an absolute policy advantage for the tech industry writ large to being a double-edged sword for some:Aggressive restrictions on the export of chips are closing off a huge market for US AI hardware com-panies and d

251、ata center products,which has left firms like Nvidia and Oracle deeply unhappy.24 During the Biden administration,the implementation of export controls restricting the sale of semiconductors to certain countries through the“diffusion framework”received the bulk of the criticism,with a number of firm

252、s invested in the global chip market particularly up in arms about the impact to their businesses.25 The Trump Administration may make changes to the diffusion rule,26 and is internally fragmented between factions that are supportive of tariffs and hawkish toward China,and those that are interested

253、in global expansion of the AI market.27For its part,Nvidiathe leading semiconductor firm,which is most directly affected by the export con-trolshas embarked on a push for“sovereign AI,”a term coined by the company to refer to nations abili-ties to produce their own AI using some combination of homeg

254、rown infrastructures,data,workforces,and business networks.28 CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT31Nvidias stance is an example of a play at market expansion.As the provider of computing chips for the data center infrastructures central to sovereignty initiatives,the company stands

255、to benefit from na-tion-states growing interest in building out their own homegrown industries and attracting AI investment.For chip manufacturers,the push toward sovereign AI can be seen as a way of diversifying their customer base away from the hyperscalers and hedging their business against the p

256、otential slump in the demand from these companies.29 The European Union and its member states have also espoused interest in sovereign investment into AI in a bid to compete at the frontier.The European Commission has gradually repurposed its existing European high-performance supercomputing ca-paci

257、ty toward training large-scale AI models.30 To further up the ante,the Commission announced a 20 billion InvestAI initiative to establish European“gigafactories”that would house one hundred thou-sand GPUs with the objective to facilitate training of models with“hundreds of trillions”of parameters.31

258、 Investment has also picked up in the member states.In February 2025,France hosted the Paris AI Action Summit,during which president Emmanuel Macron announced around 110 billion in investment pledges to boost Frances AI sector,with a focus on infrastruc-ture investments.32,33 In Germany,the new gove

259、rn-ment coalition has agreed to house at least one of the gigafactories,complemented with commitment to develop a sovereign tech stack,as well as support for a budding“Eurostack”movement,an informal coalition34 at the European level that aims to reduce European tech dependencies by developing domest

260、ic alternatives.35These investments at the level of the EU and its mem-ber states still pale in comparison to the scale of the private investment plans in the US,with the$500 bil-lion joint venture fund Stargate announced in January 2025;the fund arguably cements monopoly domi-nance by a cartel of U

261、S-based firms.36 Meanwhile,the UAE and Saudi Arabia are geopolitical swing states,given their financial capital to sustain infrastructural build-out,and have been flooding the market with money via the AI funds MGX,G42,and the Saudi Public Investment Fund(PIF)for AI,37 money that the leaders of AI f

262、irms are avidly seeking.38Nationalism thus still remains a critical shaping force in AI policymaking:The“AI arms race”has if anything become increasingly complex in a moment of geo-political uncertainty,and is wielded by firms both to avert regulation and to court investment.AI NOW 2025 LANDSCAPE RE

263、PORT32CHAPTER 1:AIS FALSE GODS1.4:RECASTING REGULATION AS A BARRIER TO INNOVATIONThere has been a swift and aggressive narrative at-tack on AI regulation as anti-innovation,superfluous bureaucracy,and unnecessary friction.Weve seen a total reversal in the US federal stance and,increasing-ly,a regula

264、tory chill reverberating across quarters in the EU.We saw early signs towards the end of Bidens term setting the governments primary role as enabler of the AI industry,1 and with the Trump Administration it is the headlining message.The headwinds against baseline accountability against the tech sect

265、or in general,and AI companies in particular,are greater than ever.The tech industrys fickle policy promises have also revealed their true colors.Companies spent 2023 insisting they were extremely concerned about safety and were firmly“pro-regulation.”2 But as the center of power has shifted towards

266、 a deregulatory current,any superficial consensus on guardrails has just as quickly fallen away.OpenAIs CEO Sam Altman,for instance,went from testifying in a Congressional hearing that regulation is“essential”to lobbying against a minor safety provision in just fifteen months.The governments narrati

267、ve change has been just as swift.In 2023,future-looking existential(“x-risk”)concerns took center stage.In policy fights these x-risk safety concerns have often eclipsed the long list of material harms arising from corporate AI con-trol,often moving public and policy attention away from enacting pol

268、icy and enforcing existing laws on the books to hold companies accountable.3 Nota-bly,Vice President Harriss speech on the sidelines of the UK AI Safety Summit called out this tension explicitly,and set up an(implicit)counterpoint to the x-risk-dominated agenda at the rest of the summit led by forme

269、r prime minister Rishi Sunak:“These exis-tential threats,without question,are profound,and they demand global action.But let us be clear.There are additional threats that also demand our actionCHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT33threats that are currently causing harm and which,to

270、many people,also feel existential.”4 Harris went on to describe the ways in which ordinary people have already been harmed from faulty,discriminatory,and inaccurate AI systems.Unlike other regulatory conversations,the broad philanthropic and government interest in addressing x-risk safety concerns e

271、ventually served to further ce-ment government relationships with the tech industry.The vast majority of efforts under the safety umbrella have been voluntary and industry-ledfor example,numerous safety validation standards within the UK and US AI Safety Institutes were set by or done in collaborati

272、on with industry players like Scale AI5 and Anthropic6revealing that the government had been successfully convinced to regulate AI in lockstep with and led by industry-centered expertise.On the other hand,when the rubber met the road with SB 1047,the California bill that sought to impose baseline do

273、cu-mentation and review requirements on the largest AI companies for a very narrow class of advanced mod-els,large parts of the tech industry pulled out the rug and pushed against even this narrow regulatory inter-vention with all their might.7 Even Anthropicwhich positions itself as a company respo

274、nsive to safety and the risks of AIwaffled on SB 1047 support,first coming out against the bill before dragging their feet into a hedged statement of support,saying the“ben-efits likely outweigh its costs,”but“we are not certain of this.”8 Government players fell in line,with key Democratic legislat

275、ors9 framing the bill as detrimental to innovation.10 In a letter to Governor Newsom,eight Democratic members of Congress succinctly summed up this position:“In short,we are very concerned about the effect this legislation could have on the innovation economy of California.”11 Facing immense pressur

276、e,Governor Newsom ultimately vetoed the bill.The fight for SB 1047 opened the floodgates for pit-ting regulation against innovation.A recent one-two punch has shifted the terrain entirely:Groups advo-cating for legislation mirroring SB 1047s provisions are being politically targeted by Republicans12

277、 and a new troubling bill,SB 813,13 is gaining support in Cal-ifornia that allows AI firms to self-certify their models as safe and then use that certification as a legal shield to avoid liability in a civil action for harm.14At the federal level,there was vanishingly little prog-ress legislatively,

278、leaving large swaths of industry use entirely outside of regulatory constraints.Bidens now-repealed EO15 and the OMB memo16 were bright spots,making strong progress in terms of hooks for actionable accountability via targeting government use of and procurement of AI.Even public investment proposals

279、such as the National AI Research Resource pilot,originally positioned as a counterforce to con-centrated power and resources in the AI industry,was recast under Bidens 2024 National Security Memo as a national competitiveness project.Former National Security Advisor Jake Sullivans October 2024 speec

280、h before the National Defense University also firmly po-sitioned the US government as an enabler of frontier AI companies and emphasized the need for US invest-ment in the AI sector to go full steam ahead in order to shore up the countrys strategic positioning against China.17Still,despite a far-fro

281、m-coherent policy stance on AI under Biden,the attack on regulation ushered in by the Trump administration cannot be overstated.18 Since being elected,President Trump has positioned regulation as a clear-cut way for the US to“lose”the global arms race,and his allies have propagated fears of Chinese

282、control of global AI infrastructure as a threat to American security and democracy.On his first day in office,Trump gutted Bidens Executive Order on AI,replacing it with his own Executive Order set to revoke existing federal AI policies that“act as barriers to American AI innovation.”19 At a series

283、of high-profile events including Davos,the French AI CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT34Action Summit,and the Munich Security Conference,the Trump administrations message rang loud and clear:Global regulation is a targeted economic attack on US companies,and the antithesis to inno

284、vation.Meanwhile,the administration has expressly target-ed the administrative state,calling into question the independent status of enforcement agencies and gutting the federal workforce,including key employ-ees tasked with enforcing existing laws to rein in corporate dominance(this included unlawf

285、ully firing key Democratic FTC Commissioners with a record on tech enforcement).The Trump administrations recent OMB memos do little to impose accountability on AI systems,and are instead designed to fast-track the procurement of AI across the federal government.20 Meanwhile,AI Industrial policyor f

286、inancial and reg-ulatory support for expanding the national AI indus-tryis being positioned as the counterpoint to reg-ulation,and a more appropriate role for government intervention.Unsurprisingly,Silicon Valley tech and AI executives have fallen21 quickly22 into23 line,shoring up their seats at th

287、e table.Because,while Trumps tangible industrial AI policy moves remain to be seen,the dominos set in motion by the Biden administra-tion are poised to rapidly accelerate under Trump.Trumps agenda for global AI dominance is mutually reinforced by an expansive energy dominance agen-da,and his adminis

288、tration has repeatedly highlighted the need to expand US energy resources24 to remain competitive in AI.25 Debates about permitting require-ments for infrastructure build-out had already taken center stage during the Biden administration.Senator Joe Manchins Energy Permitting Reform Act of 2024 expe

289、diting review procedures for energy and mineral projects advanced out of committee with a bipartisan vote.26 The bill is supported by a coalition of fossil fuel companies and tech lobbyists,who claim that AI tech innovation is tied to energy expansion.As they wrote in a letter to Congress:“Americas

290、leadership in global innovation depends on the passage of permit-ting reforms that allow the US to build critical energy infrastructure.”27 In some ways,the Trump administrations pro-enforce-ment posture toward Big Tech companiesseen in the continuation of the DOJs case against Google and the FTCs r

291、ecent trial against Metais consistent with the Biden administrations antitrust policies,and runs orthogonal to the otherwise deregulatory headwinds and hands-off approach to the tech industry.At the same time,these cases are not designed to strike at the root of power facing the AI industry,which ha

292、s received an“all systems go”message from the Trump White House,but rather to curtail Big Tech censorship and undermine platform authority over state pow-er.Already we see tech companies attempt to wield political favor to end the trials.28 And Google is set to argue that structural separation will

293、undermine US national security issues,29 potentially derailing bold antitrust remedies from the court.Despite these cas-es,it is unlikely that the Trump DOJ and FTC are set to broadly undermine the AI industrys market power as a matter of policy,no matter how the antitrust suits are decided.30The dr

294、ift toward deregulation has begun even in the European Union,traditionally seen as a staunch regulatory power.Driven by rightward electoral shifts,increasing securitization of AI,and new geo-political realities driven by Trump,the once proudly proclaimed digital regulation agenda is now seen as a li

295、ability by European policymakers.In addition to scrapping planned bills,such as the AI Liability Direc-tive that created a product liability framework for AI,31 there is appetite in the high halls of EU policymaking to walk back on rules already agreed to.While back-tracking is constrained by the em

296、barrassing optics of bending under US pressureat least thus farwhen it comes to implementation,there is growing pressure to create as much flexibility as possible so as to mute CHAPTER 1:AIS FALSE GODSAI NOW 2025 LANDSCAPE REPORT35the impact of the laws without changing their letter.32 This push to

297、create flexibility for domestic companies is complicated by the importance of these rules as a rare source of leverage in the nascent trade war between the EU and the US.33 The extent to which European digital regulation becomes a pawn in this debate remains to be seen.More generally,the tone in the

298、 European Union and member states has become more enabling,parallel-ing the developments elsewhere.French President Emmanuel Macrons“plug,baby,plug”quip at the Paris Action Summit crystallized this shift in senti-ment.34 Leveraging the tools of statecraft and existing infrastructures(such as abundan

299、t nuclear energy in France)toward promoting the development of AI is in-creasingly central to the broader push toward Europe-an sovereignty.In addition to new public investments in AI infrastructures,new political coalitions and power players are also emerging in the background to facilitate this ch

300、ange.A recent large public-private partnership with an investment pledge of 150 billion by a collective of leading European industrial giants and tech companies,complemented by direct access to heads of European states to discuss a“drastically simplified regulatory framework for AI,”is one exam-ple

301、of these changing winds.35 Absent from this discussion is the role regulation can play in fostering innovation within markets,particular-ly given the dynamism and complexity that AI exhib-its.By creating a stable regulatory environment with robust competition among firms and an equal playing field t

302、hat enables new entrants to thrive,well-crafted regulation can act as an enabler rather than an adver-sary to innovation in emerging markets(See Chapter 4:A Roadmap for Action).CHAPTER 2:HEADS I WIN,TAILS YOU LOSEHOW TECH COMPANIES HAVE RIGGED THE AI MARKET AI NOW 2025 LANDSCAPE REPORT37CHAPTER 2:HE

303、ADS I WIN,TAILS YOU LOSEIn some regards,the current market behavior of AI firms appears wholly irrational:Tech companies are pumping billions of dollars into an unproven tech-nology with little market demand,firing their own workers,1 and acquiescing to the political demands of an administration def

304、ined by its tech factionalism and personal vendettas.2 On its face,the AI market appears to be driven more by AI“FOMO”a fear of missing outthan sound business decisions,with AI firms throwing product use cases at the wall to see what sticks,and firms across the economy force-fit-ting AI solutions in

305、to their workflows,buckling un-der the generalized pressure that any competitive company must today have an“AI strategy.”3 Big Tech firms have guaranteed their own success by making the wall as sticky as possible,gaming the market to ensure they benefit if and when the returns come rushing in.Whethe

306、r by locking customers into existing ecosys-tems,bending the law to work in their favor,co-opting political processes and media narratives,or pegging their own futures to an industrial strategy of national dominance and government investment,Big Tech firms are shaping the market to consolidate their

307、 own power and to hedge against the considerable risks theyre exposed to.The reality is that Big Tech firms and AI developers(propped up by Big Tech firms)can successfully gamble on AIs future because the house always wins.Their deep pockets allow them to suffer short-term losses as they shuffle thr

308、ough product use cases and burn money,AI chips,and energy at an alarming rate,but ultimately theyand power players in adjacent in-dustries that hinge on AI infrastructure build-outare best positioned to net long-term gains in this market.This section maps the drivers that are securing Big Tech firms

309、 advantage in the AI market,before turn-ing to the question of who loses in the end.CLOUD INFRASTRUCTURE PROVIDERS BENEFIT FROM CYCLES OF AI DEPENDENCEBecause the quickest path to AI profit is through the increased demand in cloud services this market drives,Big Tech firms that offer cloud computing

310、 ser-vices and control cloud infrastructure(like Amazon,Microsoft,and Google)are best positioned to win the AI race.Governments and investors are funneling billions of dollars into a speculative AI industry without a clear business model or pathway to profitability.In Chapter 1,we identified the myt

311、hs undergirding the hype despite obvious red flags and warning signs.But the reality on the ground is far less distributive;here,we explain how a handful of firms are poised to capture the AI market.CHAPTER 2:HEADS I WIN,TAILS YOU LOSEAI NOW 2025 LANDSCAPE REPORT38Because of the“bigger-is-better”par

312、adigm,AI devel-opers require more and more compute resources to effectively train their larger models and run“infer-ence,”such as the queries returned each time you enter a prompt into ChatGPT.This dependency on compute has made large-scale AI development con-tingent on access to compute resources,w

313、hich has led AI developers like OpenAI and Anthropic to secure partnerships with cloud companies like Microsoft and Amazon in order to successfully train and run their models.The early exclusive partnership between OpenAI and Microsoft has received the most atten-tion among these:OpenAI received Mic

314、rosofts cloud resources at a fraction of the cost;in return,Microsoft locked OpenAI into billions of dollars in cloud commit-ments and a share of OpenAIs future revenue.4 Ope-nAI wasnt alone:Anthropic developed arrangements with Google5 and Amazon,6 Deepmind solidified its cloud partnership as Googl

315、e DeepMind,7 and Mistral struck a deal with Microsoft,8 for example.But the advantage these cloud firms hold is multifac-eted:Unlike other cloud companies like Oracle and Coreweave,Amazon,Microsoft and Google also hold a dominant advantage along the AI supply chain,with advantages in access to data,

316、paths to market,and talent.The partnership model between hyperscalers and AI developers is evolving from being predicatedas was the case in the 2018 deal between Microsoft and OpenAIon exclusivity to being predicated on mutual dependence.9 For example,even though OpenAI is no longer locked into an e

317、xclusive partnership with Microsoft,Microsoft remains able to secure a market advantage where it matters mostAI model deploy-mentwhile ensuring their investment is recouped through circular spending agreements and revenue shares.10 Under the new partnership,Microsoft retains access to OpenAIs IP(inc

318、luding insight into how Ope-nAI and Oracle will manage the new Stargate servers);OpenAI API is still exclusive on Azure(and pays more than$1 billion per year on Microsoft services);and revenue-sharing commitments are still in place(Mic-rosoft retains a 20 percent share of OpenAIs revenue and future

319、profits up to$92 billion).11 Microsoft is also positioned to effectively block OpenAIs effort to con-vert into a for-profit company,while OpenAIs board can trigger a clause that prevents Microsoft from accessing its most cutting-edge tech,which OpenAI officials have reportedly proposed doing.12In Ja

320、nuary 2025,markets were temporarily rattled by the announcement that Chinese startup DeepSeek was able to launch an AI model comparable to Ope-nAIs latest release at a fraction of the compute cost.13 For some,DeepSeek cast doubt on the self-serving,bigger-is-better paradigm advanced by companies lik

321、e OpenAI,projecting future efficiencies in compute resources.But DeepSeeks release does not change the current paradigm of cloud company dominance:Despite the models smaller use of compute in the final training run,the technical advancements in advanced reasoning driven by the inference-time compute

322、 approach are still reliant on scale for their performance advantages.And any efficiency gains would likely be overridden by growth in demand,a phenomenon known as the Jevons paradox.As Satya Nadella declared in the wake of DeepSeeks release,“As AI gets more efficient and accessible,we will see its

323、use skyrocket,turning it into a commodity we just cant get enough of.”14 DeepSeek thus solved one pressing business problem for Microsofthow to deal with its escalating expenditures on data centerswithout disrupting the overall business proposition for the company:capturing the market through its co

324、n-trol over the cloud ecosystem.Efficiency gains through models like DeepSeeks also dont necessarily undercut the advantage that Big Tech companies hold from their access to compute.For one,pushing ahead on performance gains at the CHAPTER 2:HEADS I WIN,TAILS YOU LOSEAI NOW 2025 LANDSCAPE REPORT39cu

325、tting edge of the technology is still extremely com-pute intensive.15(It is also widely believed that Deep-Seek essentially“distilled”its model building off of OpenAIs o1.16)Moreover,cloud companies reap con-sistent gains even as weconsumers and companies alikefigure out whether AI delivers what it

326、promises:Every response generated on ChatGPT,every query run on Gemini,and every customer service chatbot integration incurs a cost that customers pay back to the hyperscalers.Now,if compute remains a scarce resource,Big Tech companies with cloud businesses win by controlling limited supply.Similarl

327、y,if models become more efficient,these firms still win because the efficiencies will lead to overall reduced infrastruc-ture costs,allowing them to deliver more product at cheaper cost.This means that cloud firms are incen-tivized to boost AI demand either way,ensuring that AI demand balloons to fi

328、t a growing market for infra-structure that depends on its success.This relationship of dependence extends not just to AI developers but to cloud startups,too.For instance,CoreWeave is a new entrant into cloud computing that chipmaker Nvidia has invested in,and has marketed itself as a solution to c

329、ompute bottlenecks in the AI market.But the company recently went up for pub-lic offering,17 and financial documents revealed that CoreWeave is saddled with debt and almost entirely dependent on Big Tech companies like Microsoft that need to offload their excess demandthe very com-panies it is attem

330、pting to compete with.18 BIG TECH FIRMS BENEFIT FROM LEVERAGING CONTROL OVER THE TECH ECOSYSTEM There is increasing consensus that AI models are be-coming“commoditized,”meaning that gains in model efficiency decrease costs,and more large-scale mod-els will emerge to compete.In response,firms like Mi

331、crosoft are advising those in the market to“focus more on how they integrate these models with their own data and workflows.”19This advice reflects Microsofts position in the mar-ket:It,like Google and Meta,has an advantageous position due to its dominant role in enterprise and consumer-facing softw

332、are.This is precisely why,on the day that the chipmaking firm Nvidias stock fell nearly 17 percent following the DeepSeek news,Am-azon,Meta,and Apples stock went up.20 Because if AI models become cheap to integrateand compute becomes significantly cheaperthe firms who own AI products,distribution,an

333、d data centers are at an ad-vantage.This makes ecosystem powercontrol over the paths to marketan important element in the AI market.Ads all the way down:Metas advertising ecosystem positions it well in the generative AI market.Because one of the main use cases for generative AI technolo-CHAPTER 2:HEADS I WIN,TAILS YOU LOSEAI NOW 2025 LANDSCAPE REPORT40prise products,from Workspace to Email,so comp