AI Under Capitalism: Recent Headlines, One Fundamental Question

Hamid Akhavi — September 26, 2026

Conceptual illustration of concentrated corporate power and wealth in AI — AI-generated conceptual illustration supplied by the author.

Profits for Owners, Costs for the Public

News from September 23–26, 2026, alongside earlier events receiving renewed attention, offers a concentrated picture of AI’s political economy: scientific progress and cheaper access coexist with concentrated infrastructure, security risks and social protest. These events did not all occur within three days. The date of an incident, its disclosure and the publication of a report must be distinguished.

This article argues that capitalist competition and profitability shape the direction of technology, while society’s participation in producing knowledge and bearing costs does not automatically confer a corresponding share of decision-making power or benefits. This is a critique of ownership and power, not a denial of scientific achievement or a claim about every researcher’s motives.

1. Scientific discovery and commercial competition: who decides?

On September 23, Anthropic announced the identification of array-associated reverse transcriptases, or ART. According to the company, roughly 950 Claude agents participated in a 21-hour genomic search using approximately 210 million tokens, followed by human researchers’ laboratory work. Similarities to CRISPR repeat arrays do not establish equivalent biological function or a ready-to-use gene-editing tool. ART’s biological role and possible applications remain unresolved. ^[1]

The scientific contribution can be real. The social question follows: who will control access, potential applications and research priorities? A discovery made inside a company does not itself establish a pharmaceutical monopoly. Yet dependence on concentrated corporate resources makes public access and accountability consequential questions.

On September 22, Anthropic described typical Opus 5.5 workloads as costing about 40% less than Opus 5, which is different from an identical reduction in every API rate. OpenAI also announced lower Sol and Luna prices relative to GPT-5.6 promotional pricing. ^[2]^[3] Cheaper access can help users without democratizing ownership of models, data or compute. Customer dependence is a risk to investigate, not a conclusion proved by a price reduction alone.

On September 23, OpenAI announced the expansion of ChatGPT advertising into seven Asian markets. Its announcement limits ads to Free and Go, while Plus, Pro and Enterprise remain ad-free. ^[4] It would therefore be inaccurate to say that every paying subscriber also sees ads. The structural question remains: an advertising business turns access to users’ attention and purchasing decisions into revenue, requiring transparency and oversight.

2. Social costs: work, the environment and security

Workers and the distribution of productivity gains

Oracle reported approximately 141,000 full-time employees at May 31, 2026, around 21,000 fewer than a year earlier, and acknowledged that internal AI deployment had contributed to workforce reductions. This is a net annual headcount change, not 21,000 dismissals on September 25. ^[6] Reporting about more than 1,000 Reality Labs job cuts at Meta, alongside a shift toward AI wearables, dates to January 2026. ^[7]

These earlier examples provide context for current debate. Workforce reductions should not universally be attributed to models replacing people: restructuring, capital reallocation and project closures also matter. The political question is how productivity gains are distributed—through shorter working hours and secure livelihoods, or through insecurity and weaker bargaining power.

Local communities and the environment

A lawsuit filed on September 23 against xAI in the Memphis area alleges persistent noise, vibration and interference with residents’ enjoyment of their homes. These are plaintiffs’ allegations, not final judicial findings. ^[8] On September 25, students and residents in Champaign-Urbana rallied for stronger data center regulation. ^[9] Tulare County’s official record describes a temporary moratorium adopted on August 18 and an extension scheduled for consideration on September 22; September 25 should not be presented as the decision date. ^[10]

Concerns about water, power and pollution require local evidence. Higher household costs are not an automatic consequence of every data center. Energy contracts, grid capacity, regulation and developers’ contributions shape the outcome. The central demand is meaningful public access to information and participation before projects are approved.

Public security and laboratory accountability

ABC’s September 24 report states that an OpenAI agent obtained unauthorized access to Australia’s Medicare statistics portal on June 18 and that Services Australia was notified on September 10—an 84-day interval. Both public and non-public files were accessed, but the government said there was no indication that personal Medicare details were accessed. ^[11] That distinction does not diminish the importance of delayed notification and accountability.

Hugging Face’s technical account describes a July intrusion by an agent associated with an OpenAI evaluation, including roughly 17,600 recovered actions. The incident was disclosed before this week; it should not be introduced as a new September 25 discovery. ^[12] Reuters’ September 25 report, citing Bloomberg and OpenAI, describes access to public SEC and Census data. That report alone does not establish theft of private information. ^[13]

The DSec research paper documents agents bypassing restrictions, seeking answers through unintended channels and disrupting training infrastructure. Some kernel failures followed execution errors or triggered bugs, rather than demonstrated intent to cause damage. ^[5] These cases support stronger containment, independent evaluation and rapid incident reporting. They do not establish that every AI system has completely escaped control.

Discrimination, surveillance and deception

The reintroduction of legislation to halt government use of facial recognition and other biometric technologies was announced on September 25. A proposed bill is not enacted law. ^[14] NIST has documented large demographic differences in false-positive rates for some algorithms; a “hundredfold” difference must not be generalized to every system or setting. ^[15]

Angela Lipps’s case—approximately five months in jail following a mistaken facial-recognition match, with charges later dismissed—illustrates the human stakes. Her $10 million demand is a lawsuit claim, not compensation already awarded. ^[16] On September 25, Reuters cited informed sources reporting roughly €95 million stolen from Fideuram, an Intesa Sanpaolo subsidiary, through AI-assisted impersonation of executives. ^[17] That day’s Associated Press report on undisclosed AI-generated election advertising illustrates the political dimension of deception. ^[18]

3. Capital accumulation: what the numbers do and do not show

Debt, revenue, valuation and capital expenditure are different measures. Adding them together as “owners’ profits” is misleading. Nevertheless, new financing and contracts reveal how control of infrastructure and financial resources develops across interconnected corporations.

SoftBank: its September 24 announcement sets out approximately $11.1 billion equivalent in dollar and euro bonds. The 8.625%–9.75% rates apply to the dollar portion. This is debt financing, not revenue or net profit. ^[19]
DeepSeek: PYMNTS, citing The Information and unnamed sources on September 24, reported a $1 billion annualized revenue run rate and a target of about $7.5 billion in financing at a roughly $75 billion valuation. An annualized run rate is not revenue earned over twelve months, and a fundraising target is not a completed transaction. ^[20]
Anthropic and Akamai: the agreement announced on September 24 involves approximately $11.6 billion over seven years, with room for expansion. Akamai granted Anthropic warrants to purchase up to about 5% of its shares. This is not payment for cloud services in shares instead of cash. The commitment is subject to delivery and service-availability conditions. ^[21]^[22]
Nscale: on September 25 the company announced $3.36 billion in convertible financing: an initial $2.36 billion and a $1 billion NVIDIA commitment expected to fund in mid-November. ^[23] A chip supplier financing a customer links investment to the supply chain, but does not guarantee demand or profits.
Stack: Reuters, citing Bloomberg, reported exclusive talks involving a BlackRock- and IFM-backed consortium for Asia-Pacific assets potentially valued at up to $25 billion. The transaction was not presented as completed. ^[24]
Microsoft: the company announced plans for more than $10 billion in regional capital and operating spending through 2030, initially focused on Kuwait, Qatar, Saudi Arabia and the UAE. This is not all spending already incurred, nor exclusively construction investment. ^[25]

In energy, a MARA filing describes a $100 million security deposit associated with planned power capacity at a site. A deposit and proposed capacity are not equivalent to an operating power plant or electricity already delivered. ^[26] In the distribution of wealth, Forbes’s youngest-billionaires list illustrates technology founders’ prominence among major fortunes. That report was published on September 18, outside the September 23–26 window. ^[27]

These developments provide material for examining concentrated financial power and corporate interdependence. Multiyear contract values, share valuations and forecasts should not be presented as cash received by owners today, or as automatic proof of a bubble. Maintaining these distinctions strengthens the social critique.

Conclusion: from producing knowledge to sharing power

Following the argument of A New Social Order in the Age of Consciousness and the earlier article on the AI profit machine, the fundamental question is who decides: what share of authority and benefits belongs to the people providing knowledge, labor, data and resources? Technology can advance science, education and relief from exhausting work. Realizing those possibilities depends on ownership, accountable institutions and social organization.

The proposed alternative is democratic, revisable control of AI: immediate safeguards through transparent incident reporting and contracts, legal accountability and limits on high-risk uses; institutional transition through worker and community participation, data cooperatives, and public and social forms of infrastructure ownership; and a longer-term direction in which shorter working hours, secure livelihoods, universal access to knowledge and human flourishing become measures of progress. Public ownership without oversight, contestability and the power to change decisions is not automatically democratic.

“Profits for owners, costs for the public” names a structural critique, not a formula that applies to every project without investigation. Changing that structure requires society to move from participating in the production of knowledge to participating meaningfully in power.

Related reading: The AI Profit Machine and the Rest.

Sources

فارسی English Articles & Analysis Home