A reading of the AI Now Institute’s 2025 report in light of entrepreneurial realities

by Pascale Caron — for the EntrepreneurIA project

Artificial intelligence is not merely a technical revolution. It is first and foremost a reconfiguration of power. The “Artificial Power” report published in June 2025 by the AI Now Institute — directed by Amba Kak, Sarah Myers West and Kate Brennan — offers a lucid, often alarming diagnosis of the AI industry’s current trajectory. Beyond technical prowess and the hopes carried by LLMs and other autonomous agents, it is an economic, political and social model unfolding before our eyes. And it deserves to be questioned.

In a landscape dominated by a few sprawling firms — Microsoft, Google, Meta, Amazon, xAI, OpenAI, Anthropic — the report paints an unflinching portrait of a sector that claims to disrupt the world while consolidating monopolistic positions. It is less about innovating than capturing: capturing infrastructure, data, imaginaries… and public policy.

AI and platform capitalism: old reflexes in new clothing

The generative AI saga has brought back into the spotlight an old promise: that of machine intelligence serving humanity. But for the researchers at the AI Now Institute, this promise is now being instrumentalized by large tech platforms to expand their power. AI, as industrialized in 2025, does not aim for emancipation. It seeks profitability, often without the slightest proof of its real social utility.

OpenAI, valued at over $300 billion, has nonetheless lost five billion in a year. Anthropic, backed by Amazon and Google, shows a deficit of $5.6 billion. And yet, investors are rushing in. Why? Because the discourse around AGI (Artificial General Intelligence) — that superhuman intelligence always imminent but never demonstrated — functions as a self-fulfilling prophecy. It legitimizes an unprecedented race for technological equipment and concentration of resources.

The report emphasizes: the AGI myth allows critical questions about the utility, limits and risks of current AI to be evaded. As long as we debate the machine that “will surpass humanity,” we avoid talking about the tools already in place that evaluate job applications, modulate social benefits or influence judicial decisions — often without transparency or recourse.

An unbalanced ecosystem, where technology acts on citizens more than it serves them

Contrary to the dominant discourse on co-construction and democratization of uses, the report insists on a far more asymmetric reality: AI is most often used on individuals, not by them. In educational, health and justice systems, AI does not reinforce human capacities, but monitors, filters and automates them. The result is a loss of control, opacity of decisions, a dilution of responsibilities.

The report notably cites uses in immigration services, social rating platforms or work automation tools: errors are frequent, structural biases persistent, avenues for contestation virtually nonexistent. “Why would society accept such a compromise?” the text asks. The answer lies in the illusion of efficiency and the inability — or unwillingness — of institutions to regulate the sector.

The great imposture of algorithmic productivity

One of the report’s densest chapters deconstructs the productivity discourse. According to the AI Now Institute, the announced gains benefit neither workers, nor users, nor even the real economy. They reinforce the position of those giving orders, accelerate job precarity, and mask a reduction in service quality under the guise of innovation.

Among developers, teachers, caregivers and administrative staff, people are now asked to “supervise” AI tools supposedly helping them — without having been able to choose, configure or question them. Result: more mental load, more standardization, less freedom. Work becomes an algorithmic protocol to follow, sometimes against common sense, often against ethics.

An economic model based on structural instability

The report is unambiguous: the AI economy is to date an economy of debt and betting. LLMs are expensive to train, very expensive to host, and no truly profitable application has yet emerged at scale. The model relies on capturing public funds, government contracts (notably military and police), and speculation on future value.

But the authors warn: we are no longer in the 2010s. The economic context is tense, markets saturated, tolerance for tech bubbles is diminishing. When this bubble bursts — and it will burst — who will pay the price? Citizens, through subsidies. Communities, through infrastructure. States, through their strategic dependence on private companies.

An extractive dynamic, not only digital, but energetic and political

AI consumes. And it consumes a lot. In 2025, AI-related data centers represent nearly 5% of U.S. energy consumption. And this figure could double within five years. The illusion that AI could “solve” the climate crisis actually masks an ecological disaster in progress. Because the obsession with scale — training ever larger models — implies an environmental cost that is hard to justify.

Worse still: this massive energy demand is used to justify new environmental exemptions, investments in fossil fuels and the construction of data centers at the expense of local populations. A social crisis is compounded by an instrumentalized ecological crisis.

Between national sovereignty and global deregulation: the paradox of industrial policies

Faced with China’s rise to power, the United States has activated an urgent strategic discourse. AI becomes a “national priority,” and tech companies become “essential sovereignty partners.” This translates into regulatory exemptions, public aid, military contracts. The report speaks of an AI Arms Race 2.0, where geopolitical competition justifies internal deregulation.

But this digital nationalism poorly masks a structural reality: AI startups are almost all dependent on Google, AWS or Microsoft clouds. In other words, there is no sovereign AI without infrastructure rupture. And this applies to Europe as well as the United States.

What alternatives? Towards an AI anchored in the commons, social justice and collective autonomy

Fortunately, the report is not limited to criticism. It proposes a genuine roadmap for reversing the trajectory. It begins with a change in narrative: making AI a power issue, not a technical progress issue. Moving beyond euphoria to question governance, production conditions, social impacts.

The main lever identified is that of labor. It is employees, collectives and unions that today have the greatest capacity to refuse certain uses, to impose deployment rules, or even to demand alternative models. Provided they are supported.

Next, ambitious regulation, based not on “trust” but on “reasonable mistrust,” is essential: prohibition of the most toxic uses, life cycle system oversight, independent certification, citizen control.

Finally, the report calls for rethinking innovation as a common good, not as a rent. This implies opening public alternatives, supporting open source models, investing in shared infrastructure, and revaluing slow, frugal, ethical research.

For an artificial intelligence serving an intelligent society

What “Artificial Power” reveals is a major dissonance between AI’s potential and its current use. It is not the technology that needs to be questioned, but its power architecture. Far from being inevitable, this trajectory can be redirected — provided we name responsibilities, recognize power relations, and restore power to field actors.

For the entrepreneurs interviewed as part of the EntrepreneurIA project, this report constitutes a precious compass. It invites us to think about AI differently: not as an automatism to follow, but as a field of choices to structure. A technology is never neutral. It is what we make of it.

Reference

Brennan, K., Kak, A., & Myers West, S. (2025). Artificial Power: AI Now 2025 Landscape Report. AI Now Institute.