There are few areas of the global economy where fever dreams of crisis, competition and war are so tightly interwoven as they are in the production of what is misleadingly called Artificial Intelligence (AI). The phalanx of smirking tech bros behind Donald Trump during his inauguration, the global “chip wars” over semiconductor manufacturing, Labour’s policy burble about AI as fuel for reviving Britain’s flagging productivity rates and the operations of Israel’s automated assassination factory in Gaza—all point in this direction.
This article offers a brief sketch of the AI industry from a perspective that is grounded in an analysis of changes in global manufacturing, particularly in the hi-tech digital economy, which laid the foundations for the recent explosion of investment in AI production. In contrast to accounts emphasising the novelty of AI products, I underline the incremental and cumulative processes of industrial development. These have led to the current explosion of investment and production in the industry.
The following story should be familiar to anyone with even a passing knowledge of Karl Marx’s work on the development of global systems of industrial machinery in the 19th century. This process saw the mobilisation of vast quantities of raw materials, energy, water and labour on a planetary scale, dragooning millions of workers into factories but also opening up new frontiers of struggle and organisation. The machines of this era, like AI systems today, systematically appropriated workers’ creative faculties, their physical dexterity and mental skills, appearing as alien and superhuman powers. As Marx put it,
The individual machine has been replaced by a mechanical monster whose body fills an entire factory building and whose demonic power, obscured at first by the measured almost solemn movements of its gigantic parts is now on display in the wild, whirling feverish dance of its countless working organs.1
The fact that the AI industry’s rise has been accompanied by a powerful ideology of machine worship is not surprising considering the sheer scale of human labour, energy and materials required to produce these technologies. However, it is the rising intensity of competition between capitals and states permeating all stages of the production process that is the most striking feature of the AI industry. This competition is playing out in a context of the prolonged polycrisis, increasing the pressure towards the fusion of industrial and military means and ends.
Understanding the inference industry
AI is naturally a loaded term, and currently it is hard to get behind the hype and the claim that these technologies are some kind of alien intelligence rather than products of human labour.2 This article focuses on a particular subdomain within AI research that has recently made the leap from academia to industry: the production of very large Machine Learning (ML) models. ML, a branch of computer science, uses statistical inference for the purposes of classification and prediction. Developments in ML systems, used for classifying images and texts over the last decade, have laid the basis for the emergence of a sector of industry now producing the large ML models lying behind services such as OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, Alibaba’s Qwen and DeepSeek’s R1.
More important, and likely to have a much greater impact than the current chatbot ballyhoo, are the principles underpinning the production process of these Large Language Models (LLM). Across the industry, corporations and states have drawn the conclusion that recent significant leaps in model performance are the result of a dramatic scaling up of the network infrastructures, energy, water, computational processing power and training data required to produce them. This should not be equated with simple linear models of growth in any of these inputs or reduced to a single “magic bullet” solution (such as cramming more transistors per nanometre on a semiconductor). There are complex and often chaotic interactions between technical, social and political factors shaping production. However, at an aggregate level, the trends demonstrate clearly that this is a technological race where large-scale and long-term investment matters.
Below, I trace how the rise of this new industrial process has transformed existing branches of hi-tech industry (for example, data centre operations), exponentially increased the consumption of electricity, water and computer hardware, and has created entire new systems of exploitation in training data annotation factories—while mobilising millions of workers in mining, logistics and transport. Similar patterns of rapacious extraction of minerals and natural resources combined with intense concentration and centralisation of capital investment emerge across all stages of the production process. The data centre factories where ML models are manufactured are characterised by extremely high ratios of fixed capital (property, plant and equipment) to labour. They stand at the end of a chain of industries of similar character: fossil fuels, petrochemicals, plastics and semiconductors, all of which are prerequisites for the production of ML models.
The word “model” here refers to a set of probabilities and the mathematical functions required to generate them. In LLMs, this is a statistical description of the relationships between the constituent elements of a collection of texts that can then be used to predict the next word in a sequence. The actual version of the process in current LLMs computes these probabilities not based on words but on parts of words known as tokens in natural language processing. These probabilities can in turn be used to generate new text, in effect creating a combination of letters and spaces statistically similar to the arrangement of tokens in the training data.3 If the collections used as training data are large and varied enough, these statistical renderings of the internal structures of the texts can be used for a much wider range of tasks. The model is now a statistical guide to relationships between billions of tokens, and the outcomes of its predictions take on the appearance of artificially generated language.4
Collections of images can be modelled in similar ways. In this case, the computational tasks are slightly different, but the general method is the same: the assembly of a collection of statistics and the mathematical functions producing them describe the totality of relationships between pixels across the whole collection. As with texts, these statistical guides allow users to find things they might be looking for in the collection (by ordering images according to their statistical similarity, or “nearest neighbour”). Moreover, with a few more operations, they can generate new images by assembling pixels in arrangements that are statistically similar to those in the model. These arrangements of pixels are very likely to appear identical to images of actual objects, people or places to many.5 However, deploying ML models in this way is just an accelerated version of what Alan Turing called an “imitation game”, using techniques in statistical inference to speed up the process of finding out what group of pixels or combination of letters a human being is most likely to label as “a picture of a cat” or “a quote from Shakespeare”.6
Extremely large models, which have been trained on enormous, varied datasets, are often called foundation models, because they can be used for a wide variety of tasks. The LLMs and Text-to-Image models highlighted above are some of the most familiar ones. However, new foundation models are being manufactured all the time, trained on different kinds of data, including geospatial data, such as satellite images and data from the sensors in wearable activity trackers.7 Most of the recent advances in model development have been achieved through the massification of both computation and training data in the production process. This massification is based on improvements in semiconductor technology—which have made it easier and cheaper to process larger amounts of data—and the creation of more efficient architectures for the computational neural networks producing the models.
The production process for very large ML models is time consuming. A typical large-scale model might require 1,000 graphics processing unit (GPU) hours in the training phase, when the neural network generates statistics based on the dataset its predictions rely on. This may be followed by a further 500 GPU hours for a process called fine-tuning: refining the model to produce outputs tailored to a particular set of tasks. Further time and computational resources are required for the inference stage. Here, the model responds to questions, generates text or produces classifications.8
The scaling up of production inputs is important because it has set in motion a global competition to produce new foundation models, including in the form of sovereign AI (foundation models produced with specific datasets reflecting national languages, cultural heritage and laws).9 Crucially, the large-scale production of foundation models has demonstrated a generalisable set of techniques for turning huge collections of unstructured, “messy” data into a programming language directing the actions of machines. As I will discuss in more detail, the foundation model approach has emerged out of a fit between the needs and capacities of digital platform corporations (including Google, whose engineers wrote the 2017 research paper outlining the Transformer neural network architecture that spurred the latest intensification of competition).10 In the future, other ways of producing AI technologies will certainly emerge, some of which will at least temporarily privilege factors other than the scale of past and future investments. For the moment, however, the impact of this massification of production inputs plays a crucial role in the intensifying economic competition and accelerating its transformation into military competition. The scaling-up of AI production through the fusion of inter-state with inter-capitalist competition is an important dimension largely missing from accounts analysing the rise of ML as an instance of automating the social division of labour.
The creation of large-scale general-purpose AI products is the result of decades-worth of prior investments in digital communications and distributed computing infrastructure, consumer electronics manufacturing and research in fields such as natural language processing and computer vision.
Changes to the way in which large private and state capitals have been using electricity grids, data networks and data centres have laid the foundations for industrial production of ML models, in turn becoming a factor accelerating further transformation of these infrastructures to better serve the needs of capital accumulation across the model production and deployment process. Since 2016, the exponentially rising electricity demand from data centres triggered by the demands of ML model production provides evidence that success in this competition is largely conditioned by the capacity for infrastructural investment. This will inevitably have the effect of intensifying the processes of concentration and centralisation of capital, a feature already visible in patterns of electricity demand for ML model production.
Recent surveys by the government of the United States and the International Energy Agency show how ML model production has led to a step change in energy demands from data centres, first in the US itself and then in competitor countries, particularly China.11 With 45 percent, the US still commands the lion’s share of data centre electricity demand, followed by China with 25 percent and the European Union (EU) with 15 percent.12 Between 2010 and 2017, electricity consumption by data centres in the US was largely stable at 60 Terawatt hours (TWh) with minimal annual growth. This reflects the primary role of data centres as locations for warehousing data, performing relatively low-intensity data processing and hosting connections in data networks. However, between 2017 and 2023, US data centre electricity consumption grew to 176 TWh, equivalent to 4.4 percent of total US national electricity consumption, a figure forecasted to grow to 12 percent by 2028.13 This dramatic growth in electricity usage is driven by the rush to develop larger and larger ML models, which requires intensive computation across several phases in the production process. China too has experienced both a similar leap in electricity demand by data centres and a significant shift in these data centres towards becoming industrial-scale engines of computation rather than acting as relatively passive infrastructure. Chinese government figures put demand at 77 TWh in 2022, 150-200 TWh in 2025 and an estimated 400 TWh by 2030. The US Energy Information Administration estimated a more conservative baseline of 100 TWh in 2024 but still projected a doubling by 2027.14 The EU, which accounts for the third largest share of global data centre electricity demand after the US and China, is also experiencing a boom in both data centre construction and projected increased electricity demand. Recently, Goldman Sachs estimated that the current wave of data centre construction could boost European electricity demand by 10-15 percent over the next decade.15
As researchers from the Berkeley Lab note, although data centres currently amount for a relatively small amount of electricity consumption, the challenge of meeting the power requirements of the industrial production of ML models is compounded by increasing demands for electricity from other sources. These include electric vehicle adoption, onshoring of manufacturing, hydrogen utilisation, and the electrification of industry and buildings.16 The challenges in all three locations—the US, China and the EU—are compounded by severe problems of clustering of energy intensive data centres: nearly half of data centre capacity in the US is in five regions.17
The scaling up of electricity consumption driven by energy demands of ML model production raises overall energy demand. Last year, Ireland’s data centre clusters, for example, consumed more energy than the entirety of its urban households combined.18 It also intensifies competitive pressures for the addition of new energy sources, threatening to slow down or derail moves towards decarbonisation of energy production.19 The campuses, where the new breed of supercomputers used to produce large ML models are based, are incredibly energy hungry: Colossus (owned by Elon Musk’s xAI company) consumes as much energy as a city of 250,000 people.20 In September 2024, Microsoft announced that it secured a 20-year deal to use 100 percent of energy produced in the re-opened Three Mile Island nuclear power plant to power its ML model production processes.21 The AI investment gold rush is also leading to the re-opening of coal fired power stations in the US, with the full support of Trump, claiming that “good, clean coal” is essential to maintain the pace of production.22
Very similar dynamics are at work in the race to find the vast amounts of water required for cooling data centres. Campaigners in the Spanish state are organising under the banner “Your Cloud Dries Up My River!” (Tu Nube Seca Mi Río) and calling for a moratorium on data centre construction as large companies such as Amazon guzzle millions of litres of water each year. The US data centre industry has clusters in locations suffering extreme drought, such as Mesa, Arizona.23 Higher rates of water consumption in data centres are driven specifically by ML model production. Despite tech companies’ claims to be introducing “water zero” facilities, the reality is that water-based cooling is promoted as far more effective than air cooling, especially for ML workloads.24
Chip wars: brute-forcing the laws of physics
The shift towards ML model production on an industrial scale has also intensified the demand for the most advanced semiconductor computer chips, capping decades of development of highly capital intensive industrial processes. In 1971, when Intel released the first major commercial example of a microprocessor, the 4004, this semiconductor contained 2,300 transistors around 10 µm in size. Within ten years, transistors had fallen to 1 µm and manufacturers were packing up to 100,000 transistors onto each chip. Exponential growth continued, reaching 1 million transistors per chip in the 1990s, rising to 10 million by the early 2000s and 100 million a decade later.25 Today, chip manufacturers commonly use a 5 nanometre (nm) process that packs 100 million transistors per square millimetre, although 3 nm is fast becoming an industry standard. The Taiwan Semiconductor Manufacturing Company (TSMC) is promising to start production of a 2 nm process, claiming it will lead to 10-15 percent performance improvements, 25-30 percent power consumption reduction and a 15 percent increase in transistor density compared to its previously most advanced semiconductors.26
TSMC’s announcement that it is moving towards mass production at 2 nm is significant for a number of reasons. It confirms TSMC’s dominance of an enormously specialised area of manufacturing. The Taiwanese company had 64 percent of the global market share in semiconductor manufacturing in 2024 and a near monopoly over the production of the most advanced semiconductors.27 TSMC’s chips are essential inputs for products made by Apple, Nvidia, Qualcomm and many other major corporations. TSMC’s ability to continue innovating at or near the physical limits of nanoscale manufacturing enhances not only the company’s geopolitical significance but also that of the Taiwanese ruling class, which has carefully positioned its chip foundries as a kind of “silicon shield” for Taiwan itself.28 This is illustrated by the choice of location for the roll-out of the first 2 nm chips in mass production at TSMC’s new microchip manufacturing plants in the Taiwanese cities of Hsinchu and Kaohsiung (due to start in late 2025).29 However, by the end of the decade, this will be followed by mass production at 2 nm and 1.6 nm at TSMC’s new factory in Arizona, partly funded by the US government through a $6.6 billion (£5.2 billion) grant under the Creating Helpful Incentives to Produce Semiconductors (CHIPS) and Science Act.30 In recent years, the US has pursued aggressive policies aimed at denying China access to the most advanced semiconductors, including export controls. The CHIPS and Science act itself promotes the “reshoring” of semiconductor manufacturing to the US.
Unsurprisingly, pushing the boundaries of physics in this way comes at extraordinary costs in resources and energy. The fabrication of chips consumes electricity and water at a scale similar to the patterns observed in the data centres discussed above. In 2023, Greenpeace Asia found that by 2030, semiconductor manufacturing from Samsung was on track to exceed 32 million tonnes of CO2 annually, more than Denmark’s total emissions in 2021. TSMC’s energy use was predicted to surge by 267 percent, with the manufacturer consuming as much energy as 5.8 million people, mostly generated using fossil fuels.31 Water consumption, especially ultrapure water required for the chip manufacturing process, has also soared with the expansion of advanced chips, optimised for ML operations.32
There are persistent patterns of concentration and centralisation of capital across the industry, with monopolies or near-monopolies emerging in key production processes often based on closely-guarded technical knowhow. Two examples critical to chip manufacture include the highly concentrated chip design industry and the manufacture of lithography machines. US-based tech giant Nvidia designs and markets the chips TSMC manufactures in its foundries and, in turn, represents a dominant force in the global ML chip market. Nvidia’s share of the market is estimated at 90 percent, and the company successfully lobbied Trump to rescind a rule that would have restricted its exports to China, Nvidia’s second largest market after the US.33 Extreme Ultraviolet (EUV) lithography is another critical process in semiconductor production. The only manufacturer of EUV machinery is the Dutch firm ASML, although Canon has developed a “low-cost” alternative using an alternative nanoprint lithography method. ASML’s machinery reportedly costs $370 million (£310 million), Canon’s version around a tenth of the price.34
Analysts at Epoch AI, a US-based non-profit research centre, argue in a recently-published paper that enormous supercomputer clusters have played a fundamental role in producing the current crop ofLLMs, such as OpenAI’s GPT series, Meta’s Llama and Google’s Gemini 2.0. The size and requirements of these facilities has grown exponentially during that period. The Colossus supercomputer runs on 200,000 advanced AI-optimised semiconductor chips. In 2019, supercomputers with more than 10,000 chips were rare. Although hardware and power needs doubled annually across the supercomputer cohort tracked by Epoch AI, computational performance doubled every nine months, driven largely by the acquisition of more and better semiconductors.35
Training data, exploiting humans
Another aspect of ML model production subject to the twin processes of massification and industrialisation is the training data required to create the model. The first generation of LLMs often used text collected from online sources without attribution or consent. WebText, the corpus that OpenAI used initially to train the early versions of its GPT series of large language models, was compiled by relying on text scraped from the internet in 2017, using ratings provided by users of Reddit as a guide. A 2021 study of the composition of the English Cleaned Common Crawl Corpus, a huge training dataset used in the production of several important LLMs, found that the most common source was the patents website hosted by Google, followed by Wikipedia, then the New York Times, Los Angeles Times and Guardian.36 Mixed among these relatively mainstream curated and edited sources is plenty of fan fiction (including 32 million words stolen from Archive of Our Own) and user comments on social media platforms, including plenty of far-right, misogynistic and racist content.37
Training data image collections reach from ImageNet, which provided the foundations for the development of image classification models in the late 2000s, to the five-billion-item LAION dataset used to train the StableDiffusion image generation tool released in 2023. They have also often been sourced from crawls of the open internet and, as a result, contain a wide range of images, from digitised museum and cultural heritage collections, to holiday snaps, artistic works and pornography.
However, the kinds of processing that turns such images and texts into the standardised collections later fed into machine learning pipelines includes the application of consistent labelling schemas, often derived from internationally recognised “benchmarking” systems (responsible for standardising images and removing some of the content considered “not safe for work”).38 Today, this is often achieved through poorly paid human labour in dangerous, factory-style conditions in the Global South. Content moderation models used by social media platforms such as Facebook are trained using the platform’s own data, with workers verifying and labelling all kinds of deeply traumatising content at unthinkable speeds, as Sonia Kgomo testified in February 2025:
For two years, I spent up to 10 hours a day staring at child abuse, human mutilation, racist attacks and the darkest parts of the internet so you did not have to… You could not stop if you saw something traumatic. You could not stop for your mental health. You could not stop to go the bathroom. You just could not stop. We were told the client, in our case Facebook, required us to keep going.39
This human labour works in concert with machine systems. The automation of essential elements in the pre-processing for ingestion into ML has been a key driver of the leap in scale that signifies the maturation of training data production as an industry in its own right.
A look at how the production of training data has changed over the past 25 years illustrates this neatly. Labelled image data collections used to be created through craft-like processes by highly skilled researchers. In 1998, Michael Lyons and collaborators created the JAFFE dataset of just 213 images of Japanese people’s faces, which were then labelled with facial expressions by 60 annotators. The JAFFE dataset played a foundational role in the development of computer vision models claiming to automate facial expression recognition, and it has been used in a variety of other models including image/document clustering and the evaluation of clustering algorithms.40 Lyons and his colleagues oversaw the whole process of creating this highly specialised research tool, including recruiting and interacting in person with the volunteers whose faces were photographed to produce the dataset.41
Within a few years, however, such painstaking methods were being abandoned by researchers in favour of much bigger collections of images scraped from online sources, among them the photo-sharing platform Flickr, founded in 2004. ImageNet, the first mass-scale collection of images to be used for machine learning model production, was created by a team of researchers led by Fei-Fei Li a couple of years later. It featured huge numbers of Flickr users’ photos, taking advantage of the Creative Commons licensing system, allowing reuse for “non-commercial” purposes, including academic research. ImageNet’s production also relied for the first time on waged workers at scale, whose commoditised labour power was obtained by the research team through Amazon Mechanical Turk (MTurk), a digital piecework platform.42 The combination of the “semiartistic”, skilled labour of academic researchers (who no doubt put in many hours of “voluntary” labour beyond their contracted hours) with a putting-out system of piecework executed out at home by anonymous MTurkers was not unique to ImageNet. It was a general feature of this phase in the development of the ML production process.
Some forms of data labelling relied on the work of “volunteers” or leveraged big digital platforms’ ability to turn data labelling into a kind of micro-transaction in order to access content and services through CAPTCHA tests. However, the general trajectory is often towards more traditional models of exploitation. One reason is that quality assurance is a problem for capitalists whose primary products are training datasets, just as much as it is for the makers of industrial chemicals or ginned cotton. MTurkers, for example, have been found using LLMs to speed up boring work.43
One obvious solution to this problem is to recruit and train workers and pay them a regular wage rather than simply by the click. This is part of the reason why data annotation as an industry serving US-based companies has shifted location from the US to countries where labour with relevant skills is cheaper—India, Kenya and Uganda. Another driver is the need for scale and speed in production. Alexandr Wang’s Scale AI data annotation business has essentially created an inhouse version of MTurk, a clickwork platform called RemoTasks, which sources labour from across the world. This version is overseen by directly-employed engineering and product development teams, part of whose role is to ensure that training datasets meet various quality assurance standards such as the requirements of ISO certification.44 Scale’s business model has been spectacularly successful to date, and the company lists an array of clients, including large sections of the US military. However, like any other industrial production process, it is vulnerable to disruptions in the supply of skilled labour through competition with other employers, geopolitical factors and the gyrations of the financial markets. These examples show why theories viewing everyone in our digitally mediated and networked world as a worker in a giant social factory are mistaken. Although the history of the training data business is certainly one of systematic plunder of human creative effort, the industry producing ML models as commodities at scale could not emerge without creating direct forms of exploitation—and, with it, workers’ resistance (illustrated by the emergence of collective action and organising by data labellers and other tech workers).45
Training data, like other inputs for the ML model production process, is increasingly the subject of fierce competition between corporations and even between states. A study published in July 2024 highlights growing restrictions aimed at stopping AI companies crawling parts of the open Web for training data.46 Rising protests by artists, writers and actors against the appropriation of their work by training data crawlers and legal action over copyright infringement are likely to be behind some of these restrictions.47 Major AI companies, Google and OpenAI, have responded by lobbying for weakened restrictions on training data use. In Britain, the Labour government is desperate to please them, promising to allow AI companies to seize anything they want online, unless rightsholders explicitly opt out.48 Labour is also keen to open up the country’s high-quality public datasets to AI companies, for instance, the US military firm Palantir, which now has access to large amounts of NHS data thanks to a multi-million contract. Palantir has forged a cosy relationship with Labour ministers, and Keir Starmer toured the company’s offices on a recent visit to the US.49
Disruption through model design
In January 2025, the Chinese company DeepSeek released an LLM, claiming better performance than several of the major Western companies’ LLMs. The market panic around this illustrates the importance of research and development to ML model production—and reveals yet another arena of fierce competition between state-capital clusters. By training a high-performing LLM on a cluster of GPUs, much smaller than those of its competitors and relying on far fewer specialised AI-optimised semiconductors, DeepSeek’s announcement wiped $593 billion (£438 million) off the market capitalisation of US chip designer Nvidia.50
DeepSeek’s researchers outlined their method in January 2025. They applied a technique known as chain-of-thought reasoning to the design of their model, which used open-source LLMs Qwen (built by Chinese company Alibaba) and Llama (built by Meta) as a base.51 Like much of the research literature in this field, the paper is saturated with anthropomorphic language and unquestioned assumptions about the “intelligence” and capacity for “reasoning” in the model, but the main points are relatively straightforward to understand. Chain-of-thought model architectures prompt the LLM to show its working by spelling out intermediate steps in the process of producing an output (for example the answer to a question), and uses a common technique called reinforcement learning (RL) to incentivise production of the answers the designers believe are correct. Based on sets of model tasks and solutions developed by other machine learning researchers, DeepSeek’s R1 model matched or beat the scores of one of OpenAI’s LLMs, at massively reduced cost.
The scale of efficiency gains and their potential to reshape the production of ML models should not be underestimated. The Transformer model architecture, developed by a research team from Google in 2017,, laid the basis for many of the current generation of large ML models and therefore for the massive scaling up of training data and leaps in performance at text processing tasks.52 In contrast to previous model architectures that processed data sequentially, Transformers allowed much more of the training process to be carried out in parallel. This made it possible to create models of much larger datasets than previously. The change in model architecture had dramatic real-world effects, including supercharging the race to build ever larger LLMs and hence powering the expansion of supercomputing outlined above.
Turning back to the state
On 22 January 2025, two days after DeepSeek published its R1 models and code, Alexandr Wang, the CEO of Scale AI, released a slightly panicked sounding letter to Trump, laying out ways that the US government could “win the AI war”.53 Wang chided the US administration, in stark contrast to China, for misallocating spending by concentrating investment on algorithms, rather than on the compute segment of the pillars of AI production (compute, data and models):
Not only are we spending less, we’re not investing well. To win, our government must adapt by changing our investment strategy to more closely reflect industry. Then we must not only match, but exceed China’s aggressive funding for AI focused on fielding and implementing AI solutions.54
Wang’s appeal for a more interventionist US state reflects a desire among a segment of AI companies, particularly those clustered around military industries, for more than an increased scale of investment. After all, many major tech companies providing networked infrastructure, as is the case with Amazon Web Services (AWS) and Google, have already committed billions of dollars in capital expenditure.
In February 2025, Amazon announced that it planned to spend around $100 billion (£74 billion) this year, a fifth of the estimated total spent on property, plant and equipment (PPE) since 2008.55 This illustrates the impact of growing demand for AI services despite an already steep upward curve in PPE spending over the last decade and a half. These spending trends are mirrored among China’s big technology companies, with Alibaba forecasting $52 billion (£38 billion) spending in 2025, ByteDance announcing spending plans of $20 billion (£14 billion) and Tencent forecasting investment in the “low teens” of billion US dollars.56
As Wang’s letter indicates, publicly declared investment figures by Chinese companies do not capture the full scale of the Chinese state’s commitment to AI development. US AI corporations and government officials are spooked by DeepSeek’s breakthrough, in part because the Chinese company achieved significant efficiency gains despite US export controls designed precisely to hobble Chinese AI research by denying access to several critical technologies. In fact, it is entirely plausible that US efforts to choke off access to advanced chips have stimulated rather than repressed research and innovation inside China’s AI production industry. There are plenty of examples from US history where geopolitical crisis has created a relatively closed “hothouse” system entailing conditions for a leap in technological innovation. A good example is the production of synthetic rubber in the US after the Second World War cut off access to natural rubber. Massive government intervention in collaboration with the petrochemical industry accelerated the development of an alternative that in turn formed a critical component of the petrol-powered expansion of US capitalism after the war.57
Confronting the military-industrial-god-complex
The new breed of defence tech AI capitalists—Wang or Palantir’s Alex Karp and Shyam Sankar—are fixated on precisely these kind of examples from the 1940s and 1950s. They fuse together a heady mixture of nostalgia for a golden era of American manufacturing, technological research and military industrialism with appeals to patriotism and the need to defend “the West”.58 This perspective has powerful supporters in the Trump administration, including vice president JD Vance. At the “American Dynamism” investment summit in March 2025, Vance laid out a vision with AI at the heart of an “American industrial renaissance”. Globalisation, Vance explained, failed due to the offshoring of jobs to places where labour costs are lower, helping competitor states on the road towards overtaking the US. He continued:
I think we got it wrong. It turns out that the geographies that do the manufacturing get awfully good at the designing of things. There are network effects, as you all well understand. The firms that design products work with firms that manufacture. They share intellectual property. They share best practices. And they even sometimes share critical employees. Now, we assumed that other nations would always trail us in the value chain, but it turns out that as they got better at the low end of the value chain, they also started catching up on the higher end. We were squeezed from both ends.59
Vance spelled out an alternative to the assembled crowd: returning manufacturing processes to the US, enhancing productivity on the assembly line with AI, powering the whole lot with abundant fossil fuel energy thanks to “our friends from the United Arab Emirates” and enforcing a militarised border regime in order to keep out migrant labour (another use for AI technologies supplied by Palantir and others). Vance’s closing words show that he is aware that war plays a role in this deadly race for industrial dominance. “Whether it’s the war of the future, the jobs of the future, the economic prosperity of the future”, he said, “we believe that we must build it right here in the United States of America.”60
There were plenty in the audience at American Dynamism making similar arguments. Drone manufacturer Adam Bry followed on from Vance’s keynote in a later panel, sketching a vision of how mass production of drones could help “move the needle” on the reindustrialisation of the country. He noted that the Ukrainian military is burning through 1,000,000 drones a year on the battlefield. Following from this, Bry argued:
This is the point about reindustrialisation where military purchasing power makes a massive difference. The consumer civilian quadcopter markets are measured in the scale of single digit billions. That’s comparable, I would argue, to what the military should be spending in this space.61
Drone warfare is one area where the military applications of AI research are being feverishly debated, particularly in relation to autonomous and swarming behaviour. Two decades ago, the US military deployed drones as the lowest level in a technological “stack” of imperialist air power against adversaries who were fighting with very different kinds of weapons. These days, drones are cheap to purchase, relatively straightforward to make, and play a wide range of roles on the battlefield.62 Commercial drones rely on GPS and communications links that are easy to disrupt through radio jamming. One of the frontiers of innovation in drone manufacturing is ML computer vision that would allow drones to navigate in conditions of electronic warfare. Swarming behaviours among drones is another area of current military AI research, for example at the Massachusetts Institute of Technology, where the Israeli army commissioned researchers to work on the topic in 2022-3.63
The killing fields of Gaza have played an important role in normalising levels of wanton destruction of human life rarely seen since the Second World War, and AI technologies are right at the heart of it. From Israel’s mass surveillance systems to the ML models creating kill lists and drones taking out wounded children in refugee camps, AI is everywhere in the unfolding genocide.64 One of the persistent lies the prophets of this new age of technological militarism like to tell is that what they are enabling is “precision” strikes. These, so the narrative goes, are better enabled through AI, apparently saving civilian lives. Once again, the eerie emptiness of areas at the epicentre of Israel’s bombing campaign in Gaza tells a different story, perhaps closer to the blasted wasteland at the centre of Hiroshima (Japan) after the nuclear bombing in August 1945 or the fire-scorched remnants of Dresden (Germany) following Allied bombing raids in February that year.
The new would-be masters of AI war are at least honest about comparisons with past weapons of mass destruction. As Bry put it: “On an individual level, deploying a nuclear weapon is like a miserable, terrible, terrible thing, but the only world worse than one where you and your adversary have nuclear weapons is one where only your adversary does [have nuclear weapons]”.65
Of course, neither the defence tech bros nor Vance represent the entire interests of the US capitalist class, not even the interests of the entire spectrum of major US technology companies. Trump’s tariffs came under massive pressure not just from the international bond markets, but also through intense lobbying from a wide section of US capitalists, including major tech firms, such as Nvidia, Amazon and Apple. Nvidia in particular seems to have secured a major climbdown by the administration over the AI Diffusion Rule, which was due to take effect on 15 May 2015 and would have required special government approval for exports of AI chips, thus hitting Nvidia’s profits from its international markets, including China.66 No matter that this was achieved through flattery rather than confrontation with Trump. Clearly, even in this critical area, the US state is not quite yet the all-powerful entity imagined in Karp’s fantasies and still has to negotiate with its own major capitalists.
There is a much more fundamental set of problems in the assumptions underlying the race to adopt AI technologies—one that Marx laid out in his analysis of large-scale machinery. The mechanisation of any productive process has the potential to benefit the capitalists who use this technique first by allowing them to grab a larger share of the surplus value they have wrung out of their workforce. Marx noted that can take place in a variety of ways, including increasing the intensity of work, lengthening the working day and expanding the workforce through the exploitation of new layers of the working class. Yet, the impact on the system as a whole is catastrophic since the proportion of living labour power, the source of all value, shrinks as living workers toil surrounded by machines, the products of the dead labour of their predecessors. This rising organic composition of capital over time is the primary factor in the long-term tendency of the rate of profit to fall.67 Put bluntly, no amount of robots will solve capitalism’s long-term crisis. On the contrary, building them will only intensify the trajectory of system-wide decline. This does not mean there is no way to recover temporarily. Cyclical slumps in production at times of economic crisis can destroy capital and restore the balance between dead and living labour across the system.
Wars are another mechanism for clearing out capital choking up the system, allowing a reset. One of the dangers of the current period is that we are seeing how the US ruling class and its counterpart in China, along with the majority of their allies and great power rivals, are ramping up the operations of a deadly mechanism of inter-state industrial competition directly connected to preparation for “the war of the future”. Focusing the means of production on creating the means of destruction, as we know from the past, creates intense competition across the construction of energy networks, through the seizure of minerals and water supplies, the expansion of industrial assembly lines and the mobilisation of countless armies of workers. This is ultimately why a clear analysis of AI technologies as the products of human labour is essential—there is no more dangerous illusion in today’s world than the one whispering that we can do nothing to stop the spiral into war. Resistance is always fertile, and uncovering the hidden labour powering AI can reveal new terrain where we can mobilise more effectively against our real enemies.
Anne Alexander is the author of Revolution is the Choice of the People: Crisis and Revolt in the Middle East and North Africa (Bookmarks, 2022). She is a founder member of MENA Solidarity Network, the co-editor of Middle East Solidarity and a member of the University and College Union.
Notes
1 Marx, 2024, p351.
2 The collections of statistics and mathematical functions in ML models get labelled as AI for many reasons, but an important one is that it is usually easier to get investors excited if you can convince them that these are really robot brains. Matteo Pasquinelli’s work on the social history of AI provides a necessary counterblast to such ideas and is an important contribution to Marxist theory in this area. There are limitations to Pasquinelli’s approach, however, particularly in relation to the lack of engagement with the question of how competition between many masters is fundamental to capital accumulation and a lack of attention to the role of states in this process—Pasquinelli, 2023.
3 If my training data consists of the works of children’s author Dr Suess, then the phrase “the cat…” is likely to be completed by the words “in the hat”. However, if my training data is the works of William Shakespeare, then “the cat…” is not closely associated with “hat” but with other words (dog and mouse for example). The distance between cat, hat, dog and mouse in the two sets of texts can be calculated and expressed as statistics, which, moreover, also turn out to be a useful way of differentiating between the two sets of texts based on their internal structure as datasets, rather than by the names of their authors. In order to discover the properties that make these texts distinctive, it is vital to look at each as a totality, as doing so allows us to see both what they have in common and what sets them apart. Because the texts are broken down into parts that are smaller than words during the process of “tokenisation”, neologisms (entirely new words) can be generated.
4 What comes out of a language model is just a text generated through a statistical process but the chatbot form which gives the impression of dialogue with another being creates a powerful illusion of artificial language.
5 Although not always—image classifiers sometimes fail in ways which are not visible to the human eye, see Karpathy, 2015.
6 Turing, 1950.
7 Narayanswamy and others, 2024.
8 BytePlus Editorial Team, 2025.
9 Lee, 2024. This particular source is a puff piece by Nvidia’s marketing department, but there is no doubt that competition to produce foundation models with ‘national’ characteristics is real. This mirrors a similar process for organisations which can purchase models tuned to their internal norms and rules from companies such as Open AI.
10 Vaswani and others, 2017.
11 Shehabi and others, 2024; Spencer and Singh, 2025.
12 Spencer and Singh, 2025.
13 Shehabi and others, 2024, p6.
14 Ye, 2025.
15 Goldman Sachs, 2025.
16 Shehabi and others, 2024.
17 Spencer and Singh, 2025.
18 Ambrose, 2024.
19 Beyond Fossil Fuels, 2025.
20 Pilz, 2025.
21 Sebastian, 2024.
22 Robbinson, 2024; Kimball, 2025.
23 Barratt and others, 2025.
24 Digital Realty, no date.
25 McKenzie, 2023.
26 Shilov, 2025. Apparently the 2 nm definition of this process is essentially now a marketing label rather than meaningful in material terms. Nevertheless, if TMSC’s claims to have continued to scale SRAM (the key type of memory circuit for semiconductors) are borne out in practice, this will be a major event.
27 Bowman, 2024, p1.
28 Miller, 2023.
29 Clarke, 2025.
30 Shilov, 2024.
31 Greenpeace, 2025.
32 Wang and others, 2023.
33 Weatherbed, 2025.
34 Trueman, 2024.
35 Pilz, 2025.
36 Dodge and others, 2021.
37 Codega, 2023.
38 Although this is done inconsistently if at all, researchers frequently find examples of racist, sexist and abusive imagery including images of child sex abuse in image training datasets.
39 Kgomo, 2025.
40 Lyons, no date.
41 Lyons, 2021.
42 Deng and others, 2009.
43 Veselovsky and others, 2023.
44 There is no need to take Scale’s claims at face value about the robustness or ethics of their production process, but simply to note that this is a fundamentally different process to the way MTurk worked 20 years ago.
45 Kgomo, 2025; Anwar, 2025.
46 Longpre and others, 2024.
47 Creamer, 2025.
48 Wiggers, 2025; Maher, 2025.
49 Quinn 2024; Amin and Geoghegan 2025.
50 Carew and others, 2025.
51 DeepSeek-AI and others, 2025.
52 Vaswani and others, 2017.
53 Wang, 2025.
54 Wang, 2025.
55 Morgan, 2025.
56 Reuters, 2025a; South China Morning Post 2025; Reuters 2025b.
57 Hanieh, 2024.
58 Karp and Zamiska, 2025.
59 Vance, 2025.
60 Vance, 2025.
61 A16Z, no date.
62 See HTS use of drones in December 2024 when capturing Damascus (Alexander, 2025).
63 MIT Coalition for Palestine 2024, 38–9.
64 Amnesty International, 2023; Iraqi, 2024; Al Jazeera, 2025.
65 A16Z, no date.
66 Weatherbed, 2025.
67 Roberts, 2022.
References