Big Tech’s Debate Over “Open” and “Safe” AI Leaves Corporate Power Unquestioned

On July 24, 2026, an initial coalition of 25 technology companies and organizations, including NVIDIA, Microsoft, and Meta, released the Open Weights and American AI Leadership letter, calling for wider adoption of AI models that institutions can download and modify. By August 3, more than 270 companies and organizations had signed, spanning Big Tech, venture capital, investment management, and the defense and security industries. Four days after Open Weights’ release, another letter, Pacing the Frontier, organized by employees of frontier AI companies with support from two AI-policy nonprofits, drew more than 1,300 signatures. It warned that leading companies may be close to automating AI research, accelerating development beyond society’s ability to understand or control the resulting AI systems.

The letters seem to argue for opposite futures. One presents wider access and fewer restrictions as the route to security and prosperity. The other asks the U.S. government to support an international effort to slow frontier development before competitive pressures make coordinated restraint impossible.

But they address different levels of the industry. Open Weights seeks to deepen AI’s already pervasive presence by making customizable models easier to embed throughout the economy and public life. Pacing the Frontier proposes government-backed management of frontier AI, the industry’s term for its most advanced and broadly capable models. That framework could make the small circle of companies building these systems indispensable to defining the dangers and writing the rules meant to restrain them.

Read together, the letters point toward a two-tier arrangement. AI becomes entrenched across society while ownership of its most powerful systems, infrastructure, and governance remains concentrated at the top. Their connection requires no secret agreement. It lies in what both take for granted: AI development will continue, public institutions will finance and adopt it, and private ownership and corporate power will remain intact.

The letters leave out harder questions: Who will own AI and decide its direction? How will it be used in surveillance and war, and how will it reshape work and deepen inequality? Whose lives and futures will be valued, and does the public have any right to limit or refuse a technology presented as inevitable? Beneath the language of openness and safety lie deeper a general agreement over corporate power, military dominance, elite rule, and the authoritarian currents shaping Silicon Valley’s vision of the future.

Open Weights Are Not Open Source

An AI model is the underlying system that generates answers, predictions, images, or other outputs. Its “weights” are the billions of numerical settings developed from patterns in training data. Releasing those weights allows organizations with computing power and expertise to download, run, and modify the model rather than use it only through a company-controlled service such as ChatGPT, Gemini, or Claude.

Researchers can examine these models, test their weaknesses, and build specialized tools, while organizations can run them locally rather than send sensitive information to a commercial service. Yet proprietary, company-controlled services such as ChatGPT, Gemini, and Claude are not inherently safer: their operators can keep flaws from public scrutiny, monitor how people use them, impose restrictions, raise prices, or withdraw access.

But releasing a model’s weights does not by itself make the AI system open source. Under the Open Source Initiative’s definition, open-source AI must provide the necessary materials and the legal freedom to use, study, modify, and share the system for any purpose. Those materials include the model’s parameters, the complete code used to train and run it, and detailed information about its training data and methods. Most open-weight releases, by contrast, provide the finished numerical settings and enough supporting files to run or adapt the model, but not the complete training code or the required details about its training data and methods. The training data itself and much of the development process often remain proprietary or unavailable. Beyond the definition’s formal requirements, the labor arrangements and corporate decisions behind these systems also remain undisclosed. Users of open-weight models can therefore operate and adapt them without gaining a full account of how they were built.

The Open Weights letter borrows the democratic reputation of open source to present open-weight access as a path to competition, independence, and shared prosperity. Yet an open-weight release gives the public access to a corporate product without transferring ownership or giving workers, affected communities, and the broader public meaningful power to govern the data, computing infrastructure, labor conditions, and deployment decisions behind it. A genuinely open-source release would provide broader materials and legal freedoms, but even that would not place the resources, infrastructure, or institutional uses of AI under public, community, or worker governance. In either case, access remains unequal. Individuals may run smaller models, but corporations and agencies possess the data centers, private datasets, institutional records, and specialists needed to customize them, connect them to institutional data systems, and deploy them at scale. An open-weight model that supports independent research can therefore also expand institutional surveillance and coercion by enabling governments and corporations to monitor and classify people, inform or automate consequential decisions, and control access to services.

That unequal distribution of power is only one danger. Open weights also create an irreversible loss-of-control problem. Once copied, a model cannot be fully retrieved, monitored, or patched across all distributed copies, nor can its safeguards be guaranteed to remain intact. Those with sufficient technical expertise and computing resources can modify open-weight models, remove their refusal mechanisms, introduce discriminatory behavior or hidden backdoors, and redistribute the results. Modified models could facilitate cyberattacks, fraud, impersonation, targeted disinformation, nonconsensual sexual imagery, harassment, or guidance involving weapons, abuse, and self-harm. Connected to institutional records, they could expose private information, intensify surveillance, discriminate, restrict services, or help target communities. Responsibility then fragments as developers, vendors, and institutions blame one another.

Even as the weights circulate and open the door to still more harmful uses, the infrastructure remains firmly closed. AI depends on specialized chips, enormous data centers, electricity, water, and cloud platforms. Open weights may reduce an institution’s dependence on a particular model provider, but training, adapting, and deploying AI at scale still depend on highly concentrated infrastructure markets, led by NVIDIA in advanced AI accelerators and by Amazon Web Services, Microsoft Azure, and Google Cloud in cloud computing. All four companies now appear among the letter’s signatories. The model may be downloadable, but control over the infrastructure surrounding it remains concentrated in a small group of corporations.

For researchers and organizations in the Global South, open weights may provide some genuine independence from closed commercial services and make it easier to adapt models to local languages and needs. But that limited autonomy does not alter the global labor system through which AI is produced, requiring a vast workforce largely across the Global South that gathers, labels, cleans, translates, moderates, and evaluates data and model outputs. These workers are often monitored and managed algorithmically, bound by nondisclosure agreements, and required to endure prolonged exposure to traumatic material. Releasing a model’s weights reveals none of these labor conditions and gives the workers neither an ownership stake nor decision-making power over the systems their labor helped create.

Open Weights asks policymakers to expand computing access, fund datasets and testing tools, avoid restrictions that might slow development, and spread AI throughout the economy. This is a familiar Big Tech strategy: use public resources to build markets for privately controlled technology and make everyday life dependent on products the public neither owns nor governs. The costs are distributed across workers and the public, while private companies retain control over much of the infrastructure, intellectual property, and resulting wealth.

Pacing the Frontier — But Under Whose Authority?

Building frontier AI requires extraordinary amounts of computing power, energy, data, and capital, largely confining its development to a small number of Big Tech corporations, heavily financed AI laboratories, and state-backed institutions. Pacing the Frontier warns that if AI systems begin automating the research and engineering needed to create more powerful successors, development could outpace researchers’ and governments’ ability to test, understand, or control them. These systems could facilitate cyberattacks, weapons development or deployment, mass manipulation, and other catastrophic harms before effective safeguards are in place. Because no company or country wants to surrender its competitive position while others continue racing ahead, the statement calls on the U.S. government to support an international effort to develop technical and governance mechanisms for deliberately pacing frontier development.

Many of the letter’s signatories may sincerely seek to prevent catastrophe, and coordinated restraint may indeed be necessary. Yet the letter calls for no immediate pause and leaves the decisive questions unanswered: Who decides when development must slow? What capability threshold or kind of security breach would trigger intervention? Who oversees the companies, and what prevents the largest laboratories from shaping rules that entrench their dominance by imposing barriers smaller competitors cannot meet?

Many of the signatories work for companies that would be subject to any resulting rules, yet these same companies possess enormous power to shape those rules. These firms are private centers of power and rank among the world’s wealthiest and most politically connected institutions. Collectively, they own or control much of the frontier-model, cloud, and data-center infrastructure, as well as the proprietary information and technical expertise on which regulators depend. Their wealth funds lobbying, research partnerships, political access, and policy campaigns. Through search engines, social platforms, AI assistants, media partnerships, and sponsored research, they also shape public perceptions of AI: what it promises, which harms receive attention, and which alternatives appear possible.

Yet this power is subject to little democratic accountability. Major decisions are made without public scrutiny, while the people and communities affected by these systems have little power to inspect them, challenge their decisions, or refuse their use. The companies can help define what constitutes a danger, which responses are acceptable, and which restrictions remain outside serious debate. Those creating the risks thereby become indispensable government partners and coauthors of the rules meant to restrain them.

The emphasis on future catastrophe can push present harms out of view. AI is already intensifying racist systems of mass surveillance, policing, detention, and deportation; supporting military targeting and the rapid development of increasingly autonomous weapons systems; automating work and displacing workers; deepening data extraction; and accelerating environmental destruction. A Brennan Center report details racial disparities in facial recognition and notes that Black people account for at least 8 of 10 known wrongful arrests based on faulty matches. The Department of Homeland Security AI inventory documents ICE’s use of facial recognition. An ACLU investigation traces Palantir’s role in ICE’s detention and deportation infrastructure. A Government Accountability Office review describes the Pentagon’s Maven program using machine learning to identify potential targets from geospatial imagery. The AI Now Institute’s 2025 Landscape Report documents AI’s energy and water demands, worker displacement and exploitation, workplace surveillance, deepening inequality, and risks to critical infrastructure and national security. Debate over hypothetical loss of control can obscure the harms already imposed on people by systems functioning as designed.

Pacing the Frontier does little to challenge the structures producing these harms. Its limits are especially clear on labor. A genuine slowdown could reduce the pace of worker displacement, resource extraction, and demand for exploitative or psychologically harmful data work, but it would not by itself change the ownership structures, working conditions, or power relations governing AI production. The letter does not include these present harms in its case for restraint. It defines danger primarily through frontier capabilities, geopolitical competition, and possible future catastrophe, not through low wages, algorithmic management, psychological injury, labor displacement, or the transnational contracting chains already sustaining AI development.

That narrow conception of danger also shapes the safety framework envisioned by Pacing the Frontier. Evaluating and stress-testing models, reviewing content, and monitoring systems after deployment would still require human labor. Given the tech industry’s entrenched labor practices, this framework would almost certainly continue to rely on workers who toil in the shadows. The letter calls for none of the protections or decision-making power these workers would need: labor standards, transparency about who performs the work and under what conditions, collective bargaining rights, protections against workplace surveillance, or a meaningful role for data workers in determining what their labor is used to produce.

The concentration of authority in laboratories and government experts can fortify the companies already at the top. Frontier regulation may reduce genuine dangers, but expensive testing, security, and reporting requirements can exclude universities, public-interest researchers, and smaller developers while the largest firms absorb the costs. Because governments depend on corporate expertise, oversight can become a wall shielding Big Tech from public accountability and leaving the labor regime on which it depends untouched.

On labor, then, the letters are less opposed than they appear. Open Weights would diffuse models whose dependence on human labor remains concealed; Pacing the Frontier would subject the most powerful models to expert oversight without democratizing the labor relations or restructuring the supply chains behind them.

The Ideology Behind the Two-Tier System

Both letters treat AI’s advance as inevitable, leaving the public to adapt and government to manage its deployment. That technological determinism merges with Silicon Valley solutionism: questions about ownership, labor, surveillance, military power, and the right to refuse AI shrink into matters of design and oversight. Within that shared premise, Open Weights gives institutions greater freedom to deploy AI, while Pacing the Frontier places corporate laboratories, experts, and officials at the center of determining whether and when AI development poses an unacceptable danger. The public remains a population to manage rather than a political force empowered to decide whether and how the technology should be developed and used.

The division between the letters is not ideological alone. It also corresponds, although imperfectly, to different positions and material interests within the AI economy. The largely industry-based coalition behind Open Weights spans chipmakers, cloud providers, model developers, software companies, security and military contractors, and venture capital firms. Pacing the Frontier emerges from a different institutional position. It was signed in individual capacities by employees, researchers, and executives of frontier AI companies, including some of the industry’s most influential figures, rather than formally endorsed by the companies themselves. Their concerns cannot simply be reduced to their employers’ economic interests. Yet their place within the industry matters: they work inside the small group of elite laboratories where computing power, investment, proprietary knowledge, and technical authority are most heavily concentrated.

The letters, however, should not be treated as expressions of completely separate blocs of capital. Google, Meta, OpenAI, and other companies appear in the Open Weights coalition, while employees and executives of the same firms signed Pacing the Frontier. A single corporation can benefit from broad diffusion across the AI market and concentrated control at the frontier: it can encourage widespread adoption, sell the infrastructure and services that adoption requires, and retain control over its most advanced systems. What appears to be a conflict among different sections of AI capital can therefore also reflect competing strategies within the same corporations and investment networks. This is primarily a dispute within AI’s corporate and technical elite over the terms of expansion, not a challenge to corporate ownership or concentrated control.

Open Weights echoes themes found in Marc Andreessen’s Techno-Optimist Manifesto, effective accelerationism, and techno-libertarianism. These currents treat technological expansion as a social good, private markets as engines of progress, and regulation as an obstacle. Pacing the Frontier departs from techno-libertarianism’s hostility to government restraint, but it does not overturn the distribution of power that techno-libertarianism legitimates. Frontier systems remain privately owned, development remains organized through corporate competition, and technical and corporate elites retain a privileged role in defining danger. The letter therefore represents less a rejection of techno-libertarianism than a technocratic effort to contain the dangers created by unrestrained corporate competition while leaving its underlying structures of ownership and control intact. At the same time, its emphasis on possible future catastrophe and expert-led restraint more closely resembles longtermism, which can assign greater moral importance to hypothetical future populations than to people suffering now, and singularitarianism, which anticipates that superhuman AI will transform society beyond recognition. These currents overlap with transhumanism, which envisions using technology to enhance selected human capacities and ultimately create supposedly superior “posthuman” forms of life.

Despite their differences, these ideological currents converge around the pursuit of ever-more-powerful AI. Accelerationists seek more computing power, faster deployment, and fewer restraints, convinced that technological growth will produce abundance and solve social problems. Longtermist safety frameworks may urge greater control over the pace of development, but they often leave that underlying project and its private corporate ownership and direction unquestioned. In the framework developed by Timnit Gebru and Émile P. Torres, these visions share a tendency to centralize power, elevate technical elites as humanity’s guardians, and invoke “safety” or universal benefit while pushing out of view the exploitation, discrimination, and environmental destruction already affecting dispossessed peoples.

They trace transhumanism, singularitarianism, longtermism, and related ideologies to Anglo-American eugenics. In their account, transhumanism recast improving the “human stock” as creating enhanced or “posthuman” beings. The language shifted from compulsory breeding to innovation and personal choice, but the sorting logic endured: ideas about race, intelligence, productivity, able-bodiedness, and capacity for enhancement continue to shape whose lives and futures are valued. Eugenics has not disappeared. It has been updated for an age in which powerful elites imagine redesigning humanity through technologies that reproduce existing hierarchies.

Techno-optimism turns this hierarchy into a story of liberation, presenting corporate technologies as neutral instruments of progress while concealing who owns them and who bears their costs. At its authoritarian edge, it overlaps with neo-reactionary or “Dark Enlightenment” thought, which rejects egalitarian democracy and imagines society governed like a corporation by CEOs or technocratic elites. Such an order need not announce itself as a dictatorship. It can emerge through privately owned systems that monitor, classify, restrict, and punish people without meaningful consent or appeal.

Silicon Valley’s supposed hostility to the federal government is selective. The industry resists the state as regulator while embracing it as investor, customer, military partner, police force, and border authority. It does not necessarily seek a smaller state, but one that protects a techno-capitalist oligarchy, suppresses resistance, and uses privately developed systems to govern.

That bargain has entered the machinery of the Trump administration. Reuters reported that technology executives and companies gave more than $300 million to support Trump’s 2024 campaign and an allied political committee. Elon Musk spent more than a quarter-billion dollars, while David Sacks and figures from Marc Andreessen’s network entered or advised the administration. Whether driven by ideology, profit, or access, this alliance advances domestic authoritarian priorities (including ideological policing, mass surveillance, and deportation) and an imperial project of military dominance and worldwide dependence on American technology. The state gains private infrastructure and expertise. The companies receive deregulation, subsidies, contracts, and support for global expansion.

The administration’s America’s AI Action Plan turns that bargain into policy through accelerated adoption, fewer regulatory barriers, rapid data center construction, and global export of American chips, models, software, and standards. A related executive order on “Preventing Woke AI” directs agencies to buy models conforming to the administration’s definitions of truth and neutrality while casting systemic racism, intersectionality, and what it calls “transgenderism” as ideological threats. The state is not removing politics from AI. It is using public purchasing power to present its own politics as neutral truth.

None of this means every signer or developer endorses the Trump administration or shares the ideological commitments described here. The point is structural: both letters operate within a political order built around private infrastructure, concentrated wealth, U.S. geopolitical dominance, and interstate competition for AI-enabled military and technological superiority. The race (especially between the United States and China) to develop AI for weapons, intelligence, surveillance, cyberwarfare, logistics, and military planning turns continued expansion into a national-security imperative and makes restraint appear to be strategic surrender. Open weights accelerate diffusion, frontier-safety governance can legitimize control by insiders, and public institutions supply the money, data, infrastructure, legitimacy, and users. AI spreads throughout society, but ownership, power, and wealth remain concentrated.