Introducing a genealogy of OpenAI's business model

This paper shows the genealogy of accumulation at OpenAI. While each phase in the genealogy is noteworthy for content, scale, and dimensional phase shifts, the constant seems to be a business model of always accumulating more, more, more. It shows OpenAI as an accumulation organisation that is trained to perpetually shift towards new forms of resources from social life, or the non-capitalist environments on which capital accumulation depends, to increase its capital and find opportunities to put that to work in novel structures and practices that will allow further expansion. That’s its business model.

In some ways this accumulation has made OpenAI an infrastructure of and for the marketisation of socio-technical relations. The rapidity and breadth of growth of these relations seem novel but also are seemingly predictable using classic ideas of political economy.

Here we first highlight five distinct phases of accumulation in OpenAI’s short but narratively rich history. OpenAI shifted its business model again and again to make new markets; as of writing this means we’re seeing the Assetisation of everything from high order mathematics to simple moments of everyday family life. We then reflect on the theoretical context that can explain these phenomena.

On a methodological note, Narayan (2024) suggests studying organisations’ operating, business, and labour models or the structures and practices that coordinate organisational activity, manage workers, and generate profit.

OpenAI’s operating model has already been a riveting narrative for the tech press (See Hao 2025). Critical debates on its labour practices – tied to transforming human utterances to model value – are before the courts. Here, we are more interested in the business model of OpenAI, and its noticeable phase changes of accumulation.

To help explain the phase changes, with some cheek, we term OpenAI a “Generative Post-trained Transformer” organisation. This subverts the acronym Generative Pre-trained Transformer (GPT) to more accurately model how OpenAI generated a post-training process that increasingly “just accumulates” via transforming its own business goals, structures, and practices.

Or, maybe we can think about OpenAI really buying into reinforcement learning. OpenAI, as a learning “agent” seems to observe the current state, enact a policy, for which it is rewarded (or not) and that updates the environmental state. It then adjusts its policy/values.

Using the language of those you’re studying to try to explain what they are actually doing can limit your grasp of what is at stake and the conceptual toolkit you can bring to bear (see Cohn 1987on national security and nuclear war). But in this case, we think this language is quite useful analytically to think about how OpenAI’s business model based on accumulation and how we can discern 5 discrete phases of OpenAI’s accumulation. Each phase has social and technical components – various pressures that steer towards new necessities and goals, and show a political-economic history of how OpenAI came to be, and what it is accumulating from Society. We offer a genealogy of accumulation for OpenAI.

Phases of Accumulation

Phase 1: Social Capital (2015–2019)

“how is this?” Sam Altman wrote Elon Musk, when penning the mission, structure, and financing of Open AI on 8 December 2015. His words went live online by 11 December – less a phrase about low pay – for the world to meet OpenAI.

While the note mentioned a $1 Billion USD of investment (that they ‘expect to only spend a tiny fraction of … in the next few years’), the main focus was accumulating the goodwill of public capacity to form and norm AI for humanity. OpenAI was focused on creating a new public good, with public research(ers) that “free from financial obligations”.

This phase coincided with scientific forms that held ‘deep learning’ was starting to create interesting results for AI. For instance, OpenAI’s scientists took to heart that “Attention is all you need” (Vaswani et al. 2017), in a way that helped set a specific path for future AI model progress. By 2020 OpenAI was ready to explain to the world one of the consequences: accumulating goodwill of the top scientific minds was no longer enough: “Scaling laws for Neural Language Models” (Kaplan et al. 2020) signalled a phase shift in what OpenAI now needed to accumulate.

Phase 2: Accumulation of infrastructure and culture (c. 2019–2020)

In this phase, model size, dataset size, and amount of compute were directly aligned to improving performance of AI models – according to researchers from OpenAI. OpenAI had to shift its model of accumulation in two new dimensions of capital. Or as reflected on by OpenAI’s Chief Scientist in 2026, the science of scaling meant that they “sought out access to much more compute than we had originally planned, and increasingly oriented our research around a small number of very scalable directions. We believed that was the only way for us to be at the frontier of AI research, and influence the impacts of AGI” (Pachocki 2026).

First, to address crafting larger parameter model sizes and compute hardware infrastructure to train and run them, corporate capital was required. Chips. Servers. Energy. For always increasing exponential scale for continued progress, OpenAI needed backers of the biggest of tech firms, and settled on Microsoft. OpenAI started to accumulate capital in the traditional financial sense, but more specifically ‘compute’ in such a way that actually changed how the language of how compute was spoken about. The precision of floating-point operations per second shifted to the newly accurate forecasting needs: raw measures of electrical power deployed – compute measured in gigawatts. Gigascale infrastructure was needed for AI to be created, and ostensibly shared with humanity.

At the same time, more data was required. Humanity had to share itself with AI to create AI. Cultural data needed to be accumulated at scale for AI to exist. Quality of the data captured, their provenance or the legal and moral rights around the data mattered less than having their ingestion. Human-produced utterances and aesthetics became capital; and OpenAI structures and practices focussed on accumulating this as fast as it could.

So this phase was marked through accumulating corporate-compute and cultural-commons as new forms of capital. We note some of what we call cultural-commons fell under copyright, some didn’t – but it didn’t matter to the accumulation, which transformed whatever existing socio-economic relations existed or transformed cultural environments into capital for novel markets. This latter culture-capital – whether understood classically as OpenAI-produced assets for leveraging, or more critically as being unjustly extracted from the labour of others – was what enabled OpenAI’s models to produce value. This value, it was hoped, could later be exchanged in ways that might offer profit, adjacent to “open” AI for all.

We note the end of this phase with parallel research developments. As OpenAI thought about how to leverage the ostensible products that would come from accumulating compute and culture, the stochastic critique (Bender et al. 2021) of ever larger LLMs was forming: poorer data employed, Gigawatts of environmental cost, and lost opportunity for alternative directions in theory offered a very narrow (read white male English) way forward for AI. This second phase of accumulation – Scale Maxxing – of OpenAI required corporate partners, and thus hinted at future corporate plans, which defined the next turn in accumulation.

Phase 3: Products and Profit (deferred), 2020–2022

Phase three of OpenAI’s accumulation genealogy shows a new dimension from cultural and compute was added to its Accumulation practices: a timeline for a business model to profit. Here we understand business models as structures and practices meant to create profit (Narayan 2024), or if not profit, at least capital. In this third phase, commercial access to models via APIs and strategic licensing had the research organisation rethinking its own structures and practices to develop products.

Hirings reflect this not merely via shifts from “accumulating talent to accumulating head count” (Hao 2026, p267), but by qualifying the purpose of those hires. The Applied group in OpenAI formed in 2020, and had a handful of engineers by 2021, about 30 by 2022 and 130 by summer of 2022 (Morikawa). Likewise, Aliisa Rosenthal notes her June 2022 hiring at OpenAI was for a go-to-market strategy that saw her sales team grow from 2 people to hundreds (Rosenthal 2025). OpenAI continues to accumulate scientists researching AI, but the purpose to do so, and the structures and practices created in this phase started to prepare for profit oriented accumulation. The founders’ thinking on these shifts to openness is documented quite well (see reporting from Hao).

Phase 4: Labouring Users, c. 2022–2024

ChatGPT changed everything for OpenAI, again. The extremely rapid growth of public use of the ChatGPT interface had OpenAI accumulate a user base faster than any company in history. ChatGPT acquired 100,000,000 users 2 months after launch. Facebook took about five years to reach a similar number. More than the quantity, the qualitative importance of this accumulation was what those users were doing with ChatGPT and to OpenAI.

In the Web 2.0 era, of ostensibly social media, user generated content was circulated by platforms in ways that attracted attention. Attention-based metadata tied to users established forms of data relations (Couldry & Mejias 2019) that were monetised through pervasive surveillance of online life (Zuboff 2019).

Users of ChatGPT form a different sort of data relations. Rather than scrape existent data relations to ostensibly sell or steer behavioural traits, ChatGPT is accumulating generative uses of its own technologies: ChatGPT use accumulates the forming and norming of novel human-machine relational activities. From sexual companionship to self reflection, from tax advice to relationship advice, users are performing a new type of labour for OpenAI. These interactions help craft its understanding of how humans and machines can interact. Here, many relational interactions that used to define social environments are being shifted into the market; OpenAI is transforming social relations between people and machines into capital for future use. This might be thought of as a relational economy, instead of an attention oriented one. However, relations are not transactional. Both parties in a relationship give something up to build a third thing together, outside themselves. ChatGPT users build no such entity. They transact by learning from each other new applications of everyday practice.

User labour is in some ways a widely distributed post-generative training for OpenAI that rewards both users and the application through positive feedback loops. Sometimes this is explicit “do you prefer response A or B”, the ChatGPT interface asks.

But even without these A/B signals, the corpus of data relations ingested is unprecedented. Whether ‘agents’ can be thought of as things built by OpenAI and users together, which define a new way of thinking about relational economies remains to be seen. We note that there are differences between data processed to provide a service, data retained, and data used for training. And while more work to distinguish OpenAI’s practices here is warranted, in all cases, the labour of users is determinative of OpenAI’s product evolution – measured across spectrum from n=1 experience or macro scale of model development.

Regardless, the types of relations that become data-relations continue to expand and might signal a new phase in both quality and quantity of accumulation.

Phase 5: 2025+

The bounds of OpenAI’s accumulation continue to expand. For example, at time of writing, over 700 manuscripts that seem to account for years of future human work in the discipline of Math were uploaded by OpenAI. Or consider how the current agentic turn of AI product marketing – including OpenAI’s “dots” – requires total access not only to human metadata and content (eg. What you buy from Amazon, what your flight times are and when you message your mom) but also full disc access of your hardware.

OpenAI accumulates its agents on users’ machines, along with previously unprecedented access, information, and agency to perform tasks across them and the internet. It is accumulating ever new categories of social and technical forms of organisation. While the commodification path for the capacities of access to action continues to evolve, the point is that OpenAI’s business model continues to expand its accumulation and is generative of new forms of accumulation.

Taken together, each of these phases are ‘post trained’ in the sense that OpenAI’s business model of accumulation continues to shift its environment and actions. The public goodwill accumulated in phase one gathered researchers who decided to pursue a specific mission – and type of scaling to grow AI; that growth required accumulating immense compute and cultural ‘capital’; that accumulation necessitated corporate partnerships, and go-to-market strategies for products and profit; products accumulated users and their machine relationships; which is generating unprecedented accumulations of relational data of humans labouring to relate with machines.

Further theoretical context

We noted above our tendency to see things from a political and economic lens. Here we understand organisations have operational, labour, and business models that structure their practices to respectively coordinate organisational activity, manage labour utilisation, and generate profit or capital (Narayan 2024). We are specifically interested in the business model, so have to consider what ‘labour’ and ‘business’ actually mean, if anything, in an age of ubiquitous digital data and enticing their connections.

Luckily, people like Nick Couldry and Ulises Mejias (2019) have thought about the Costs of Connection a lot and came up with terms like data colonialism to describe how corporations continuously extract personal data from human life for profit. Simply put, our online connections and attention are surveilled and sold (Zuboff 2019). This means things that did not used to be assets for sale, like conversations with ‘friends’, or which band you listen to on Thursday, became economically valuable. There was a transformation of lived experience into sources of data-based economic value. And the value was ‘extracted’ from users and turned into the profits of platforms.

While these digital stories of extraction are useful, there are other precedents to describe what is happening with OpenAI as an accumulation organisation. In other words, digital media do not signal the first time the transformation of non-market things into market things was critiqued by political economists. Responding to what she felt were some shortcomings in Marx’s writing, in 1913 Rosa Luxemburg wrote a book called The Accumulation of Capital to describe and explain capitalist expansion into non-capitalist production.

For Luxemburg “capitalist production, by its very nature, cannot be restricted to such means of production as are produced by capitalist methods… [there] is unrestricted utilisation of all substances and facilities afforded by nature and soil….” Her turn of phrase did not mean to just single out our natural resources as the inert things around us. She included public, social, and personal connections themselves: “the accumulation of capital as an historical process, depends in every respect upon non-capitalist social strata and forms of social organisation.”. Which is where we get our definition of accumulation: increasing capital by means through expansion into non-capitalist environments, drawing on markets, resources (social or otherwise), and labour that progressively erode the conditions on which that accumulation depends.

Appropriating resources from social life, or the non-capitalist environments on which capital accumulation depends, is separate from capitalisation – or how that accumulation becomes means for further production. In OpenAI’s phase shifts, new forms of capital enabled new forms of production within its own business.

Reflecting on this, we can chart how 106 years of thinking riff off each other as new forms of accumulation organisations come into being.

This critical understanding of the political economy applied to OpenAI around accumulation offers an informed take on different phases of OpenAI’s practices and structures, and its relation to society. More informed, we think, than their own 2026 self reflection (see Altman and Pachocki 2026) that describes research, products, and safety as the three iterative phases that define OpenAI’s organisation. What that history leaves out is what AI ‘takes in’ or accumulates from society in its mission to create AI. It is important to pay attention to both.