AI Replacement Theory

AI generated content is everywhere: flooding our social media feeds, responding to us when we call support hotlines, and in our jobs as employers insist we use AI to “improve our productivity.” As socialists, we strive for a world where labor is reduced to its minimum and workers are free to spend most of their time living. AI CEOs purport to offer such a future world, with Anthropic CEO Dario Amodei suggesting future generations may only need to work 15 to 20 hours a week, focusing only on the most important and fulfilling tasks while leaving the rest to AI. 

However, under capitalism, employers use AI to increase their profits, not to make workers’ lives better. Workers replaced by AI aren’t freed from toil: they are cut loose into a growing unemployed population to fend for themselves. Whatever rhetoric capitalist leaders once employed about easing workloads and boosting workers’ conditions cannot conceal their personal satisfaction with this outcome. Oracle CEO and AI investor Larry Ellison succinctly unmasks tech evangelists priorities when he brags that AI will ‘eliminate human labor and human error,’ all in an endless drive towards ‘lowering production costs.’ 

If these marvelous innovations are as efficient as AI supporters claim, their financial benefits have yet to reveal themselves: most AI companies have yet to turn a profit. In fact, the only firms making actual money off this gold rush are the shovel sellers: the construction companies building AI datacenters, and Nvidia, which manufactures the chips powering AI. In reality, our economy is currently being propped up by a massive AI bubble, fueled by circular investment and creative accounting that under-reports the real cost of serving AI. AI hype is reaching levels high enough that, even shoe companies are trying to get in on the action and seeing their stock jump 600% overnight.

All this points to an economy based on vibes and imagination, far removed from any underlying value creation. However, as socialists, we can look beyond the technocratic mysticism of AI CEOs by studying the actual process of how AI is created and sold. Training and running AI models requires massive amounts of computation, forcing AI companies to acquire or build thousands and thousands of servers. To run models, they need electrical power that does not exist yet and could take years or decades to build. They turn to sucking up any existing capacity — employing fossil fuel generators to make up the difference — negating any advances in renewable energy that we might have been making. AI faces a crisis as it hits these physical limits. Marx’s theories of value and crisis can help us demystify the underlying dynamics at play and understand the true fragility of the US economy. 

AI Intensifies Work

Already workers have begun to see AI’s promise of easier, more fulfilling work as the lie that it is. Instead, those using AI are finding their days longer and more draining. Who could have predicted this? 

Marx, in the first volume of Capital, explained this dynamic over 150 years ago, long before computers existed. He describes the process where technology, which could be used to reduce workers’ labor, actually increases it under capitalism: 

Machines are transformed in capital’s hands, becoming objective means that are systematically employed to squeeze more labor out of the worker in the same period of labor. This happens in two ways: by increasing the speed of the machines and by having a single worker supervise more machinery—in other words, by enlarging his field of labor. Better machines are needed in order to put greater pressure on workers, but at the same time, such advances naturally go with the intensification of labor, since the limits imposed on the workday force the capitalist to economize as rigorously as he can.

This prediction maps nearly word-for-word to a Harvard Business Review study on AI’s effect on work:

“AI introduced a new rhythm in which workers managed several active threads at once: manually writing code while AI generated an alternative version, running multiple agents in parallel, or reviving long-deferred tasks because AI could “handle them” in the background. […]

“Over time, this rhythm raised expectations for speed—not necessarily through explicit demands, but through what became visible and normalized in everyday work. Many workers noted that they were doing more at once—and feeling more pressure—than before they used AI, even though the time savings from automation had ostensibly been meant to reduce such pressure.”

AI, instead of reducing hours, increases them and intensifies the pace of work employers expect. Not only that, employers cite AI as they fire more workers. Under capitalism, AI will always be used to lay off workers and to intensify the labor of those that remain. Profits, or rather the relentless search for them, are the driving force behind AI under capitalism. Only when we have a system driven by satisfying human needs could AI ever potentially be used to free us from the need to work.

Marxist Value

Employers also face challenges with AI, but to understand their difficulty in locating the technology’s profitability, we first need to understand where capitalist profits, and the value underlying them, even comes from. 

In the capitalist system, things aren’t produced for their direct use to fulfill a human need (what Marx calls use-value). Instead, they are produced to be sold on the market. For Marx, such products are created primarily for the purposes of exchange; the value at which a product can be exchanged is its ‘exchange-value’. 

When things are created solely for exchange on the market, they become commodities. In the process, they lose their individual qualities – when one is dealing with 500 tons of corn, the characteristics of each ear of corn are averaged out. On the market, individual types of commodities are not directly comparable: for example, how does one equate a ladder to a coat quantitatively in order to compare their value? 

Instead, Marx points out something that commodities have in common: that they were made by some person having to spend a certain number of hours working in order to create them. To Marx, their value is determined by the time a person has had to work to create them. This is known as the labor theory of value (LTV), first theorized by English Economist Adam Smith, Marx expands on that theory – adding on the concept of socially necessary labor time, clarifying that a commodity’s value is determined by the socially necessary labor time that went into creating it. 

“Socially necessary” is an important distinction because it means that creating something that is not useful to at least someone does not create value. Additionally, because individual commodities lose their individual characteristics on the market, their value is not determined by how much work went into any commodity in particular, but by the prevailing average amount of labor time required to create that type of commodity. For example if Steve makes a ladder in his backyard with basic tools and Ladder Corp makes one thousand ladders an hour with specialized machinery, the value is determined by the average number of hours needed to create an average ladder. If someone is really bad at making ladders, any time spent beyond what the average amount of time takes to build a ladder is no longer socially necessary, and does not contribute any value. 

Another consequence of this is that when a new technique or technology comes out that makes building a ladder take less labor time than before, once this technology becomes widely adopted the value of ladders will go down. Ladders not made with the new technique will be spending labor time that is no longer socially necessary. 

Despite this, companies still introduce technology to reduce their labor costs, because they are in competition with one another. The first company to innovate on a process gets to make higher-than-average profit rates by pocketing the savings from reduced labor costs, because the average labor time needed, and therefore the price, has not yet fallen. When their innovation becomes ubiquitous, the value falls, erasing any previous advantage the company had. And then the cycle begins anew with more investment.

Echoes of Dot-com

Capitalists are investing in AI with huge sums of money, and finding capital to finance AI has not been difficult the last few years. CEOs and investors are stoking the hype, promising huge profits once the infrastructure is built. This has led to an explosion of investment in AI infrastructure that has parallels with the dot-com bubble. 

In the late 1990s into 2000 the internet was new and exciting (as opposed to the never-ending slop factory it is today) and dot-com companies were going public for huge sums of money. The common sense at the time was that everyone and everything would soon be connected to the information superhighway. Anticipating growing demand for faster speeds and more bandwidth, internet infrastructure was seen as a guaranteed investment. Companies like Intel and Enron dumped billions of dollars into fiber and internet routers that would power this future. During the inflation of this bubble, infrastructure investment took up the largest portion of US GDP growth, amounting to 20% of growth. Companies built thousands and thousands of miles of internet fiber on the promise of huge returns.

But when profits didn’t come, when capital investment dwarfed value producing labor — when demand didn’t match with the supplies being built — it came crashing down. 

In 2025, AI amounted to 17% of GDP growth and companies are making bigger and bigger commitments for data center capacity. Last November OpenAI projected that they will make $1 trillion in infrastructure investments over the next 8 years. Perhaps hinting at growing strain in the bubble, the company has since walked back some of its estimates to $600 billion by 2030. 

Depreciating Chips

Even more troubling for AI companies is that the massive expenses they incur to build their infrastructure are likely drastically underestimated. The problem arises in some clever accounting used to calculate the value and lifetime of their physical infrastructure. Companies will account for costs for large infrastructure expenses (for example, AI data centers) over the years of that infrastructure’s lifetime, instead of reporting one huge expense the first year, and no expense in later years. This math works so long as their estimate of its useful (i.e., profitable) life is correct.

Meta, Google, and Microsoft all account for their AI chip infrastructure to depreciate over a period of 5-6 years. However, there is evidence to suggest that this is a dramatic over-estimate. AI workloads typically run GPU chips at very high temperatures with no downtime, causing their useful life to degrade significantly (some estimate that the useful life halves for each 10 degree increase).

Even if the chips do continue to work, they can still depreciate due to advances in technology. Nvidia releases new generations of chips every year, with each release improving power and computational efficiency, making older chips obsolete. Nvidia representatives downplay this by saying that older chips aren’t useless, they can still process AI requests after new chips are introduced. However, Dr. Michael Burry, an investor of some notoriety for predicting and shorting the 2008 Housing Crash, counters this argument: 

“The accounting standard – GAAP – refers to how long an asset will be economically productive and justify its marginal cost, not necessarily how long it will last as a physically functioning widget … An older chip will be a residual value unit that is an energy hog and extremely costly, not to mention difficult to run outside a same-vintage data center. “

The issue is not whether or not old chips can continue to run, but whether they can run profitably. As we saw above with Marx’s theory of value, firms are incentivized to introduce more efficient technology because it will give them temporarily higher-than-average profits. When Nvidia releases a new chip that uses less power, the first company to adopt that chip will receive higher profits. This puts companies in a double bind: AI companies are in competition with one another to build larger, faster models, competing for investment capital and users, so they are forced to adopt newer technologies, even before their old chips are paid off, or face being passed by their rivals, older chips then become unprofitable, and so late adopters must buy the new chips too. Every company is incentivized to act in a way that makes their own previous investments worthless.

Not only do the chips themselves lose their value, but changes in power and chip cooling infrastructure between chip generations mean that datacenters themselves may depreciate much faster than the 15 year schedule that the big companies use in their accounting. Microsoft CEO Satya Nadella describes this problem: 

“We just saw the GB200 [chip models], the GB300s are coming. By the time I get to Vera Rubin, Vera Rubin Ultra, guess what, the data center is going to look very different because the power per rack, power per row, is going to be so different. The cooling requirements are going to be so different. That means I don’t want to build out a whole number of gigawatts that are only for a one-generation, one family.” 

The subtext here is that data-centers are not necessarily useful when the next generation of GPUs comes along, meaning that if the GPUs used in a datacenter become obsolete due to competition with newer chips, the building themselves may also become obsolete. With Nvidia releasing chips at an accelerated rate, those datacenters will also need to be rebuilt or replaced. All this points to significant over-valuation and over-statement of profits of AI companies, with Michael Burry estimating “every one of these hyperscalers will overstate earnings by double digits, and each one will have tens of billions of overstated assets vulnerable to write down.” A correction of this overvaluation could come in the next few years, pushing the bubble to its breaking point. 

The AI Rat King

AI companies continue to seek investment in order to prop themselves up and prevent the bubble collapsing, as they find fewer options for cash infusion, they have pursued more risky circular deals. Their own suppliers and clients now invest in them heavily, often with one or more parties taking on debt to do so. Their cross investments in one another not only pump up their valuations, but also mean that when one firm falls, others will soon follow. 

OpenAI, now exceeding a valuation of $800 billion dollars, is one of the biggest culprits. Nvidia has invested $100 billion in OpenAI, which they in turn use to purchase chips from Nvidia. Microsoft has had a long term partnership with OpenAI, investing over $135 billion since 2016 in exchange for Azure compute commitments for $250 billion as well as intellectual property rights. OpenAI came to a similar arrangement with Amazon for $50 billion in February, which they will spend on Amazon Web Services cloud compute (possibly violating a cloud exclusivity deal they had with Microsoft). They also partner with many smaller venture capital firms and start-ups, with those companies investing in OpenAI and receiving AI services or licensing in return.

Anthropic, the other major “foundation” model company, has similar deals with Google, Microsoft, and AWS, fueling their expansion with investments from companies that are also their customers. They will continue to pump their numbers chasing profitability, but they will not find it. 

The Crisis of Value

Companies don’t invest in new technology because it is cool, they do so to lower their labor costs in order to compete on the market. But when the cost of labor is less than the cost needed to automate work (i.e. when labor is cheap, or when technology is expensive), then they will hire the cheaper labor instead of introducing technology. 

At the moment individuals and companies are only using AI because they are not bearing the actual costs. AI is being used to replace workers, but the real (unsubsidized) costs are so high that it is not actually cheaper than hiring the workers it replaces. To give a sense of the level of subsidisation, in some circumstances users are allowed to spend 10 to 20 times more tokens than what their subscription actually pays for. So far, AI companies have found new investors to keep this unsustainable arrangement going, but that cannot go on forever. At some point they will have to turn a profit, meaning that companies will have to reduce their services or raise their prices to actually cover their cost to serve. OpenAI and Claude have both already degraded their services, frustrating customers. Much of the demand for AI is propped up by subsidized prices and has begun collapsing as those prices normalize to their real cost.

Even assuming companies can get a handle on costs and make AI energy efficient, the more fundamental problem facing AI companies is that value in the market comes from human labor. As more human labor is replaced by AI, capitalism will run up against Marx’s law of the tendency of the rate of profit to fall.

In Capital volume III, Marx theorized that as workers are gradually replaced with technology — replacing human labor (which is where value comes from) with machines (which don’t produce value) — the rate of profits will tend to go down. As the ratio gets larger more capital investment is needed to produce the same or less value and as this trend continues the rate of profit will approach zero. Individual firms are incentivized to invest in technology to temporarily increase their profits, but the effect on the system in the long run is to reduce profits. Marx called this the tendency of the rate of profit to fall (TRPF), and AI may be the most extreme end point of this process.

Marx’s theory of the TRPF is one of his most controversial positions, with many economists purporting to have disproven his theory, and others coming to defend it. The tendency of the rate of profit to fall is called a tendency because there are countervailing factors: destruction of capital, whether through war or bankruptcy, can restore profitability by reducing the ratio of capital to labor. Increasing productivity or intensity of work can also work to negate the TRPF. Defenders of the TRPF will argue that those factors can only temporarily restore profits but the overall trend is still for profit to fall, citing empirical findings to back up these claims. Regardless, we may soon find out the truth, as capitalists move forward with an experiment in what happens when you try to eliminate workers entirely with AI & machinery.

Even though AI may not be financially viable in the long run, or even in the medium term, it will still cause a lot of harm to workers on its way down. Workers are already being displaced, notionally because companies use AI to replace them. We can expect this trend to continue as the recession gets worse. So long as we are under a capitalist system, a world without work — rather than being a utopia — is a dystopia. 

The ruling class will happily discard the working class as a surplus population, a permanent underclass with no work. While they build AI data centers that use massive amounts of power, water, and raw materials, they are preparing for a world of greater scarcity from the climate change caused by their endless pursuit of profits on a finite world. As the planet becomes more hostile to life, they are banking on a world where they don’t need us to survive. But they are in for a rude awakening.

Socialism or Barbarism?

As Trump and his administration lay bare the realities of the US empire, more people are questioning US’s imperialist hegemony and our ties to Israel. The US and Israeli military are partnering with US companies such as Palantir to employ AI to identify and bomb people in Iran and Palestine, and to surveil us at home. AI will strain these systemic contradictions and exacerbate the general crisis of world profitability explained by Marx’s TRPF. But this process does not guarantee the end of capitalism, and even if it did, it would not automatically bring socialism: the option “or barbarism” still exists. The only means of staving off calamity and correcting this remorseless course of capital is through class struggle.

Many workers now understand what AI will really mean for our lives. We are seeing the media flooded with slop content to keep us occupied and AI used to justify firing us. When bosses force us to use AI to increase our productivity, they are also compelling us to train the machines that they ultimately will use to replace us. As socialists, we must harness anti-AI sentiment and anger by engaging in campaigns that connect this emerging consciousness with the ultimate need to end capitalism. We can do this by building campaigns to disrupt AI’s expansion and connecting them to struggles for workers’ rights.

One strategic weak point the left should exploit is the fragility of AI datacenter financing. We can make building AI datacenters more costly, using direct action and law to cut off their access to our energy grids and water sources. By disrupting AI companies’ ability to fulfill their contract requirements, they could be pushed into defaulting on their debts. 

Communities are already putting up organic resistance to datacenter construction on their own, with the left participating in those struggles. Socialists should help to coordinate those efforts and connect them to wider struggles to dismantle capitalism. Demands to ban datacenter construction could also be linked to demands to nationalize data centers and AI models to use for the public good, recognizing AI as being only possible through the collective knowledge produced by all of humanity. Socialists can articulate and organize for a vision of a world where AI is used for the specialized tasks it is suited for, and leave the art and self expression to humanity.

AI will also produce new opportunities for labor organizing as previously privileged white-collar jobs face declining prospects. Tech workers in particular have been historically difficult to organize, favoring individual negotiation over collective action, but with tech being hit heavily by layoffs, workers may be more open to organizing. Tech workers’ individualism has made them susceptible to allowing their labor to be used to build the systems used to discipline – or even kill –  other workers around the world. Some already understand this, but without knowing any effective ways to fight back against the actions of their companies, they have either learned helplessness, or resorted to ineffective individual acts of disobedience, usually resulting in termination. Only with collective action can tech workers have control over what their labor is used for, and we must help them become conscious of this fact.

On the other hand, tech workers’ proximity to systems of control could be useful if they could be brought into the larger workers movement. Tech workers operate datacenters that power the majority of the world’s computation, used to power logistics systems, commerce, surveillance, weapons systems, and generally powering global capitalism. Strategic control of tech is vital for workers’ to take power from the ruling class. We must use this opening created by AI to organize those disaffected workers and develop new tactics for working within those industries.

More generally, as AI intensifies the existing capitalist polycrisis, we must use every opportunity that the crisis presents to agitate within the working class for the systemic rupture needed to make way for a socialist society. We must build robust organizations now, before the AI crisis breaks. The structures and alliances of trust and community we create now will be necessary to remain united under worse and more chaotic conditions.  Only by destroying capitalism can workers escape from being “transformed into the thinking and speaking appendage of a specialized machine,” and instead use machinery and AI to reduce human labor and provide a thriving life for all.