Lesser Gods: Labor in the AI Labyrinth, Part 3

What follows is the second of three parts of an extensive draft on the AI supply chain, based on my own firsthand experience as a worker assembling models within the “AI factory.” The first part was posted late last year. In that post, I explain the origin of the project and how it quickly grew out of hand into a short book. That first post provided a brief introduction attempting to disenchant AI technology and then a longer history looking at the evolution of the industry, with a focus on labor deployment and the close relationship between economic crisis and gig work in data annotation and content moderation. The second part then dug into the actual AI assembly process and its associated supply chain structure.

This third and final part, examines competition within the industry and argues that the sector offers a window through which to explore intra-elite conflicts within the US. In particular, it understands the AI industry as a key foothold being used by Silicon Valley interests attempting to enter into the military-industrial nexus, detailing the increasing use of machine learning systems in military and carceral applications, including targeting systems and spy databases recording information on migrants, activists, and the population at large. The piece finishes with a very rough draft of my conclusion, which speculates on some possibilities for labor organizing among AI workers.

As usual, please keep in mind that this is an early, rough draft. So it shouldn’t be quoted as a finished source and, if you notice any major factual errors, please feel free to reach out here on Substack to note them. In addition, information on the industry has been difficult to grasp. If you or anyone you know works at any of the companies mentioned and would be willing to do an anonymous interview, please reach out! Comments and other contributions will be particularly helpful as I rework these three parts into a more cohesive book project.

The dispossession and dismemberment of labor is a secular trend which, viewed at great distance, appears to advance in a gradual fashion. Viewed more closely, however, its progress is punctuated: specific crises create the conditions in which the newest forms of labor deployment and workplace discipline can be experimented with and then rolled out at expanding scales. Each crisis inaugurates a new front in the class war, projecting the power of capital into new realms of life. But they are also chaotic events, triggering incendiary reprisals by the proletariat while also intensifying the fraternal conflict between fractions of capital.[1] Thus, even while crises often bolster the power of capital in general, they also threaten specific capitals with annihilation. By placing existing industries under strain, triggering waves of bankruptcies, and enabling the emergence of new firms wielding new technologies, the crisis effectively destabilizes the existing corporate order, pitting legacy sectors against ascendant ones and, in some cases, even determining the fate of national fractions of capital.

Sometimes, this disruption is spectacular and catastrophic. The 1991 crisis in Japan, for example, saw the entire range of Japanese Fortune 500 conglomerates systematically dethroned. At the end of 1989, the world’s five largest publicly-traded companies (measured by market capitalization) were NTT (Nippon Telegraph and Telephone), the Industrial Bank of Japan, Sumitomo Bank, Fuji Bank, and Dai-Ichi Kangyo Bank, and Toyota. After the crisis, Japanese firms fell out of the ranking entirely, displaced by American transnationals. By late 1992, the top five were ExxonMobil, Altria, Walmart, General Electric, and AT&T.[2] But the same sort of restructuring can occur within national fractions of capital as well, where the process can be more gradual. By the mid-2000s, that same crop of US and European transnationals had fattened into hyper-financial holding companies even more dependent on rising asset and commodities prices than the Japanese firms they once replaced. In late 2005 and early 2006, the top five was composed of ExxonMobil, General Electric (by that point operating as a diversified financial conglomerate), Microsoft, Citigroup and BP, with AT&T, Bank of America, and Shell trailing close behind. After the Great Recession, only ExxonMobil and Microsoft remained and, once the commodities bubble deflated, ExxonMobil fell from the rankings entirely.

Over the course of the 2010s, a new cadre of companies then arose in the space opened by the collapse: Apple ascended into the top five in 2010 and held first place from 2012 to 2018, Google (now Alphabet) joined in 2013, Amazon and Facebook (now Meta) in 2017, and Microsoft, present throughout the decade, briefly ascended to first place in 2018 before being displaced by the record-breaking IPO of Saudi Aramco. Within the financial sphere, JPMorgan-Chase grew into the lynchpin of the financial system, joined by massive investment fund conglomerates like Berkshire Hathaway and ascendant asset-management firms like Blackrock and Vanguard. Meanwhile, subcontractors and suppliers for lead firms like Apple were pulled upward alongside, often assisted by substantial infusions of capital from investment funds and asset management firms. In this way, Taiwanese contract manufacturers like Foxconn and TSMC were elevated into the upper ranks of the corporate order, as were affiliated firms like Nvidia (Taiwan) and Avago (Singapore, later to acquire and take the name of US firm Broadcom), and even competitors like Samsung (South Korea) and Tencent (China). These intricate supply chains and complex webs of mutual investment then serve to gestate new competitors who can stage sudden challenges, particularly in the context of new crises. Thus, the Silicon Valley financial infrastructure built by the ascendant tech giants incubated OpenAI, Tesla, SpaceX, while the AI boom saw formerly subsidiary companies like TSMC, Nvidia, and Broadcom elevated to leading positions within the corporate order, alongside resurgent lead firms like Oracle.

Within this field of fraternal warfare, contract manufacturers play contradictory roles. On the one hand, their independence is a structural necessity: Foxconn and TSMC have been able to refine and radically cheapen production precisely because of the scale and diversity of their contracts, which derive from the entire range of lead firms. Were they to be directly acquired and vertically integrated into these companies, competing lead firms would no longer contract with them, their revenues would shrink drastically, and other contract manufacturers would capture their former market share. On the other hand, their rapid growth and relative monopoly over key segments of the production process eventually makes such firms a potential threat, capable of undermining the integrity of the very companies they supply. TSMC overtaking its customer Intel is a case in point and has served as one of the key arguments for the partial nationalization of Intel. Lenovo acquiring segments of and eventually growing larger than its main customer IBM followed the same pattern. Similar threats are posed by second-order competitors able to retain closer links to this outsourced production base, as when Broadcom was acquired by Singaporean firm Avago in 2015 or when Chinese smartphone brand Xiaomi, initially specializing in cheap iPhone knockoffs, rose to challenge lead firms like Apple over the course of the 2010s.

As established lead firms find themselves unsettled by the AI boom, they’ve scrambled to acquire as much of the market as possible. At the same time, they’ve had to depend on subcontractors for all the cheapened inputs that make their final products functional. Scale AI again provides a good illustration of the contradictory position occupied by contract manufacturers. When start-ups fat with venture capital and established tech giants backed by the largest financial conglomerates all raced to catch up to OpenAI, they poured massive amounts of money into contracts with AI training firms to assemble and refine their own proprietary models. Through early deals with nearly all the major players in the industry (including work on GPT-3.5, the model behind the first iteration of ChatGPT) and major investments from Silicon Valley venture capital interests like Peter Thiel’s Founders Fund, Scale AI soon grew into the sector’s largest contract manufacturer, securing its position through rapid build-outs up and down the supply chain which allowed it to market itself as a one stop shop for model assembly and testing. Meanwhile, it began using its experience to win lucrative government contracts and design its own purpose-built models. Throughout, it worked closely with Silicon Valley financial firms such as Accel to secure further rounds of investment, including from its own major customers such as Amazon and Meta.

In fact, these circular investment structures soon came to define the financial architecture of the AI boom. By the mid-2020s, the top 10 stocks accounted “for roughly 32% of S&P 500 earnings and over 41% of total market capitalization – levels not seen since at least 1980…”[3] Six of these ten were directly driving the AI boom: Microsoft, Amazon, Alphabet, Meta, Nvidia, and Broadcom. Two were AI-adjacent manufacturers: Apple and Tesla. And the other two were the major wall street interests bankrolling the boom: JP Morgan and Berkshire Hathaway. Related companies such as Oracle are often added to the list as well. Meanwhile, leading AI firms like OpenAI and Anthropic were not publicly traded, leaving them off the listings, though they lay at the center of the investment boom and have speculative valuations now matching or surpassing the valuation of all but these top firms. In part, these speculative valuations have been made possible by continual infusions of capital from the lead firms themselves. In 2025, for example, Anthropic partnered with Google, Nvidia, and Microsoft in investment deals that also gave the AI firm access to expanded computing capacity. Meanwhile, by that same year, OpenAI had entered into “deals with Nvidia, AMD and Oracle Corp. that altogether could easily top $1 trillion” despite the fact that the firm “doesn’t expect to be cash-flow positive until near the end of the decade.”[4] In addition, the company had earlier started a partnership with Microsoft, again exchanging investment for processing power, which then evolved into an ongoing Microsoft stake (roughly 27% in OpenAI’s for-profit wing as of 2025) that included a revenue-sharing deal.[5]

The core of the financing circuit is found in the triangular relationship between equipment makers, model designers, and the firms that control major cloud computing platforms and their underlying data center infrastructure. As described by economic journalist Grace Blakeley, as of October of 2025, “Three companies have become particularly embedded in these financing loops: Nvidia, Oracle, and OpenAI. Nvidia makes the chips that power AI models. Oracle provides the cloud infrastructure. And OpenAI builds the models.”[6] Nor is this simply a pattern of passive co-investment via linked stock holdings. These financing structures are instead purpose-built to guarantee demand for core products, thereby inflating their price: “OpenAI will use Nvidia’s investment to buy Nvidia’s own chips to power its new ‘AI factories’: data centres designed to train and deploy its models at scale,” all while OpenAI enters into similar deals with Oracle, spending “$300bn to purchase data centre capacity” from the firm, “which is now rushing to build new data centres packed with (you guessed it) Nvidia chips.”[7] Similar deals have been signed with other chipmakers like AMD and new circuits of co-financing have begun to arise in other linked sectors like power generation and military systems.

The core of the financing circuit is found in the triangular relationship between equipment makers, model designers, and the firms that control major cloud computing platforms and their underlying data center infrastructure.

These circular financing structures closely resemble those of the dot-com bubble of the late 1990s. Much like AI-adjacent firms like Nvidia today, early network equipment manufacturer Cisco saw 3,800% growth in its stock price from 1995 to mid-2000, using this influx of capital to finance the very internet service providers buying from it. In some cases, equipment vendors even took substantial equity stakes in their customers, allowing them to further inflate the cost of their own goods. Meanwhile, these telecoms were taking on massive amounts of credit to build out the physical infrastructure of the modern internet. When the bubble began to burst and these companies collapsed, so did the stock price of vendors like Cisco – despite the fact that its revenues remained stable and the company continued to grow throughout the early 2000s.[8] Today, many warn that companies like Nvidia and Oracle have placed themselves in a similar position.[9] The density of looped capital flows between AI-adjacent firms has thereby come to serve as a barometer for the building bubble. However, while firms like Oracle have suffered severe setbacks in the stock market after failing to hit their revenue goals following an initial spike in valuations, Nvidia’s price-to-revenue ratio is currently healthier than that of Cisco in the 1990s and debt levels are lower across the board. Meanwhile, many of the lead firms are far more massive than their dot-com predecessors and more closely linked to both the financial system and the macroeconomic management infrastructure of the state itself, allowing far more resources to be poured into the nascent industry than was possible in the dot-com era.

Throughout, the goal of these lead firms has been to desperately buy their way back into the forefront of the industry. Faced with leaner, more agile competitors, these established tech juggernauts have pursued a multifaceted strategy, building their own models, profiting from their control over necessary infrastructure and existing platforms (especially in cloud computing, operating systems, and data centers), planting strategic investments across the sector, leading an aggressive series of acquisitions, and poaching top talent. When they were then outcompeted yet again by DeepSeek in 2025, these strategies grew even more aggressive. Lead firms stressed the geopolitical dimensions of their competitive struggle, invoking anti-Chinese sentiment to enlist the aid of the US government, which promised a series of support measures. Partnerships between both established and ascendant lead firms in the industry also accelerated as the federal government loosened regulations to encourage mega-projects like the “Stargate” initiative, involving OpenAI, Oracle, and Softbank. Meanwhile, Meta, Google, and Microsoft partnered with companies like Scale AI and Palantir to build out computing infrastructure for state agencies.

For the contract manufacturers training AI models, intensifying competition has been both a blessing and a curse. On the one hand, demand for model training has renewed and, after an initial overexpansion and retrenchment in late 2024 and early 2025, a new wave of hiring had begun by the end of the year. On the other hand, success has also made the strongest sub-contractors into targets for one another and for the lead firms above. Following its conquest of the entire middle range of the supply chain, Scale AI was effectively acquired by Meta in June of 2025 for $14.8 billion in exchange for a 49% stake. The deal is what is known as an “acquihire,” which avoids the messy legal ramifications of an outright acquisition but nonetheless allows the larger firm to strip the smaller one of its talent and patents. And this is exactly what happened, as Wang shifted from his CEO position at Scale AI to a top position within Meta’s new “superintelligence division.” Similarly, Meta was able to mine its former subcontractor for valuable IP related to the model assembly workflow and poach other top engineers, all while keeping Scale AI intact as a nominally independent firm capable of continuing to subcontract for Meta’s competitors. Shortly after the acquisition, Scale AI mobilized an undisclosed amount of cash to settle its outstanding labor lawsuits in the state of California, almost certainly drawing from funds and legal support supplied by its new de facto owner. In retribution against the workers who filed the lawsuit, Scale AI has stopped hiring California-based contractors entirely.[10]

As might be expected, in the wake of the deal, all of Scale AI’s major customers aside from Meta quickly cancelled their contracts, starting with Google. By violating the neutrality of the industry’s largest contract manufacturer, Meta had essentially salted the earth. Thus, less than a month after the investment, Scale AI laid off 14% of its full-time staff (200 of its 1400 legal employees) and 500 additional high-end contractors, including its entire “generalist” AI team in Dallas.[11] In the leadup to the acquihire investment, contracts had already been put on hold and work on the company’s platforms grew sparse. In those same months, I watched as project after project paused and pay plummeted. Often, I’d do an unpaid training only to discover after passing that the project had no tasks and all its workers were EQ. On the forums, everyone complained of the same conditions. Meanwhile, many of the remaining higher-rung employees not poached by Meta jumped ship, as VPs, chiefs of staff, and researchers bled from the company in the following months. Many were quickly hired into competing firms. Similarly, as customers cancelled their contracts, they then flocked to major competitors like Surge AI, as well as smaller firms like Mercor, Invisible Technologies, Appen, Labelbox, Prolific (not to be confused with the semiconductor contract manufacturer of the same name), and Micro1. According to the CEO of annotation firm Turing (not to be confused with model designer Turing AI), customer inquiries had increased tenfold in the months after the deal. Lead firms also began to spread their contracts over a variety of vendors so as to avoid overreliance on a single company – mirroring a practice that had already become common in a manufacturing sector stricken by trade wars and rising tariffs.[12]

As a result, gig workers also jumped ship. After work at Outlier dried up, I myself switched to simple annotation tasks on Labelbox platform Alignerr. Nor was I alone. According to Rowan Stone, CEO of competitor Sapien (not to be confused with model designer Sapien AI), some 40,000 workers flooded onto its platform in the first few days after the deal closed, most of them from India and the Philippines, where Scale AI subsidiary Remotasks once employed hundreds of thousands of contractors. In an interview shortly afterwards, Stone joked that the company’s “servers are currently melting.”[13] In response, firms scrambled to build out sufficient infrastructure to absorb the tens of thousands of AI workers seeking new jobs and a flood of contracts from lead firms seeking new vendors. Capital poured in at a similar pace, as venture capitalists hoped to catch the next unicorn. Mercor raised $350 million in its Series C funding round in October of 2025, bringing its valuation to $10 billion.[14] Surge AI, which had never raised money from outside investors, entered into talks to raise $1 billion at a valuation of $30 billion.[15] Some smaller firms like Invisible Technologies (valued at $500 million in 2025) took a different approach, taking out loans from major financial conglomerates to buy back shares from its earlier venture capital investors in what was, effectively, a debt-backed management buy-out.[16]

The feeding frenzy also saw AI training firms gradually sort into various specialties. For example, Sapien focuses on data gathering via audio recording and structured surveys as well as detailed annotation of 3D and 4D data like LiDAR and other remote sensing content, while Mercor has come to specialize in elite-level training across a variety of academic disciplines, recruiting “International Math Olympiad medalists, Rhodes Scholars, and Ph.D. students,” and managing some 30,000 contractors by October of 2025.[17] In the midst of the US shakeup, Appen has increasingly begun to focus on winning contracts from Chinese AI start-ups.[18] Meanwhile, firms like Labelbox and Surge AI began marketing themselves as readymade replacements for Scale AI, offering flexible, full-service data pipelines spanning annotation, RLHF and supervised learning for a wide variety of models. Even before the Meta deal, Surge AI had been the largest training firm by revenue. By 2025, it reportedly employed roughly 1 million annotators via its Data Annotation platform.[19]

The systematic disempowerment of workers within the AI supply chain also enables the literal weaponization of these AI systems. In 2018, a small group of high-level software engineers at Google engaged in an effective work stoppage, refusing to contribute to a project designed to help the company win military contracts. This then drew attention to other military projects, including involvement in the Pentagon’s “Project Maven,” in which Google AI technology would be used to analyze drone surveillance footage (annotated by Scale AI workers, in one of the firm’s first major contracts).[20] Thousands of employees signed onto a letter demanding that the company end its relationship with the Department of Defense, with a dozen or so resigning in protest.[21] In response, the company scrapped the controversial project being protested by the engineers, pulled out of Project Maven, and released a set of guiding principles promising that it would not build AI-powered weapons. However, without any form of independent organization among workers, even these relatively skilled engineers had no way to exert power against the company over time. Nor did they have the ability to organize against the wave of layoffs that would begin a few years later, when the company cut roughly 6% of its global workforce.[22]