The Chinese AI company Deepseek is shaking up the business model of the emerging industry.
Tomasz Konicz, 06.04.2025, Originally published on konicz.info on 02/16/2025
The shockwaves that the Chinese AI model DeepSeek sent through the American high-tech industry also produced ironic, downright comical moments. ChatGPT developer OpenAI, which is backed by Microsoft, accused the Chinese startup of data theft and espionage. The business model of the American AI pioneer was built on “stealing data from the entire internet” and now they are “crying because DeepSeek may have trained on the outputs from ChatGPT,” as tech critic Ed Zitron said to PC Gamer.[1] The team led by AI guru Sam Altman is now being given a taste of their own medicine, Zitron ranted. OpenAI had designed a “plagiarism machine” and was now complaining that its plagiarisms were being used to generate new plagiarism machines.
Link: https://exitinenglish.com/2025/03/27/the-ai-revolution-devours-its-children/
Knowledge distillation is what the industry calls this process, in which a lot of money and resources can be saved by using the output of a large language model specifically to train a smaller, cheaper model. It is no coincidence that OpenAI is complaining loudly about the low-cost Chinese competition, which allegedly completed its model for just under six million dollars – ironically, the pioneer of the AI industry, which likes to aggressively propagate its potential to bring about total economic rationalization, simply seems to be losing its business model. The proprietary, closed AI systems were actually intended – due to their gigantic training costs –to be monopolized and sold by the tech giants of Silicon Valley, since machine learning has so far been able to devour billions of dollars. OpenAI would, however, become obsolete in its current form once the innovations of the Chinese language model, which is largely open source, are generalized.
DeepSeek triggered a disruptive shock in which proprietary software is beaten by the open source principle, which enables far faster, global collaboration and innovation (only the latecomer Meta has also pursued an open source approach with its large language model Llama –precisely because Facebook & Co. are not dependent on revenue from the AI business).[2] The dreamed-of software profits of the AI industry giants would thus be largely shattered, because soon every medium-sized company will be delighting its customers with similarly annoying AI tools, as Microsoft has done with its already much-hated Copilot – essentially the Clippy[3] of the AI age – at a cost of billions of dollars.[4]
An analogy from the market for operating systems can illustrate the disruption that is now unfolding: The AI industry wanted to pursue a model like the one Microsoft has practiced with its Windows operating system since the 1990s, in which the software itself is the monopolized product. With DeepSeek, the software becomes free and/or cheaper, while now the services and customizations, the “service” so to speak, must be monetized – similar to what Red Hat does with its Enterprise Linux. This is a realistic business principle, but this potential volume of the AI market is much smaller, even before its widespread realization.
However, the hardware manufacturers whose computing capacities made the AI boom possible also saw their share prices plummet after the DeepSeek shock. The graphics card manufacturer Nvidia not only discovered a goldmine with its computing cards adapted for AI processes, but also largely monopolized it to increase its share price almost tenfold within two years. After the release of DeepSeek, however, its share price slumped by 20 percent. The entire AI boom, which is effectively only keeping the U.S. financial market in a speculative boom (the EU is already largely decoupled), is in danger of running out of steam. What if the hopes for a new accumulation regime, new markets and job-generating economic sectors burst as abruptly as they did during the deflation of the dot-com bubble at the turn of the millennium? One of the most important pillars of the U.S. economy, which is actually only able to maintain its exceptional position thanks to the U.S. dollar, has been clearly cracked by a stock exchange massacre of around one trillion dollars in February.[5]
DeepSeek is not only undermining the U.S. financial market boom, the AI tool also poses a geopolitical and military challenge to Washington’s dominance, which can now only be maintained thanks to the power of the U.S. military machine. This is why the White House – aside from Trump’s empty phrases about the innovation-promoting effect of competition – immediately moved to minimize the app’s reach and simply ban its use in government agencies.
The timing of the publication of DeepSeek was probably also intended to humiliate the terawatt gigantomania of Trump and his techno-oligarchs, who a few days earlier announced Stargate, a 500 billion dollar AI investment program that now looks simply ridiculous.[6] The signal that Chinese state capitalism is sending out is clear: Chinese efficiency beats the American brute force approach. China has also demonstrated the ineffectiveness of American sanctions on high-tech products, which were intended to prevent the development of competitive Chinese AI against the backdrop of the hegemonic struggle between Washington and Beijing – precisely because of the frightening potential of military applications of AI systems.
On the contrary, DeepSeek claims to have made a virtue out of necessity, with several innovations in the training phases of AI leading to the use of Nvidia chips being limited to older 2048 H800 models (DeepSeek has not confirmed the alleged knowledge distillation that scandalizes OpenAI).[7] In the meantime, however, a study by the IT think tank SemiAnalysis has cast massive doubt on precisely these cost advantages of the Chinese competition.[8] According to the study, the Chinese hedge fund High-Flyer, which financed DeepSeek, has computing farms of around 60,000 Nvidia cards, and the expenses for the highly qualified personnel and the development of the new training methods are not included in the DeepSeek creators’ cost calculation. So the true expenses of the hedge fund High-Flyer in the “People’s Republic” are said to amount to one billion dollars.
Even if large parts of this western cost counter-calculation were to correspond to reality, its implicit logic is wrong. DeepSeek is open source, its development costs play no role in its further use, the process innovations that went into its development are not kept under wraps, they have become common property – and they inevitably lower the price of the AI-based services that America’s IT industry wanted to monopolize. The AI pie is melting away. And these innovations are real, they are not just a cheap copy, as the MIT Technology Review acknowledged – U.S. competitors are now working hard to copy these innovations, which are being promoted by Washington’s sanctions.[9] New compression methods such as multi-head latent attention have reduced memory consumption and minimized bottlenecks resulting from inadequate memory bandwidth.[10]
Another key innovative step that DeepSeek has taken is the extensive automation of the multi-stage training phase of the automation machines. According to the Financial Times (FT), DeepSeek’s “major innovation” is to minimize the use of human labor in the correct “labeling” of data.[11] This technique, which is used in the final training phase and is known within the industry as “reinforcement learning from human feedback” (RLHF), is expensive and time-consuming, according to the FT, as it requires a “small army of human data labelers.”[12] The day laborers of the AI age, most of whom are paid less than two dollars an hour and recruited in peripheral regions such as Latin America or Africa, spend their working day repeatedly tagging digital data with appropriate labels for the AI – not unlike the captchas of traffic lights, bicycles or dogs that used to be requested when passwords were entered.
And these slum jobs, which number in the hundreds of thousands and whose exploitation in the context of the RLHF is bringing the high-tech industry of the 21st century back to life in the 18th century, will soon become obsolete. According to the FT, DeepSeek was able to automate reinforcement learning through digital reward mechanisms that are activated when the AI system gives the right answers. As soon as this process is repeated often enough, the large language model begins to “spontaneously solve problems without human supervision” once a tipping point is passed. An “aha moment” occurred when DeepSeek began to evaluate questions again and adapt its computing time to the different questions, according to the financial journal, echoing reports by Chinese AI researchers. To replicate this, it no longer needs AI day laborers, but only “its very strong, pre-trained model” and a very good infrastructure to carry out “this reinforcement learning process at a large scale.”
AI is also devouring its miserable children. However, the wage earners on the periphery of the late capitalist world system, who are now at risk of losing even their precarious slum jobs, will soon be followed into obsolescence by millions of employees in the core. Although AI will radically transform the societies of the core in a similar way to the internet and the first phase of digitalization, this will not bring about a long-term economic boom in the sense of a new accumulation regime that would valorize masses of labor power in the production process of capital.
The opposite will be the case. The desubstantialization of capital and the displacement of wage labor from the production of goods and the service sector will continue. This is why fears of a slump in demand for AI chips are unfounded; at least Nvidia will continue to enjoy healthy demand. Wherever “experienced people still press the same buttons in a race” (FAZ), the market-mediated rationalization pressure will prevail.[13] The price reductions for training units for large language models will only lead to an accelerated adaptation of this technology in the valorization process of capital, which has already been able to maintain its zombie existence for decades thanks to the production of credit-generated demand and fictitious capital on the world financial markets. The last echo of this global bubble economy of the declining neoliberal era is the AI bubble in the U.S.
I finance my work mainly through donations. If you like my texts, you are welcome to contribute – either via Patreon or Substack.
[1] https://www.pcgamer.com/gaming-industry/the-brass-balls-on-these-guys-openai-complains-that-deepseek-has-been-using-its-data-you-know-the-copyrighted-data-its-been-scraping-from-everywhere/
[2] https://www.nytimes.com/2025/01/29/technology/meta-deepseek-ai-open-source.html?searchResultPosition=6
[3] https://9to5mac.com/2017/04/26/clippy-microsoft-office-mac/
[4] https://www.zdnet.com/home-and-office/work-life/the-microsoft-365-copilot-launch-was-a-total-disaster/
[5] https://www.dqindia.com/news/deepseek-sparks-1-trillion-tech-stock-meltdown-8662575
[6] https://apnews.com/article/trump-ai-openai-oracle-softbank-son-altman-ellison-be261f8a8ee07a0623d4170397348c41
[7] https://www.dw.com/de/deepseek-ki-aktie-b%C3%B6rse-nvidia-v3/a-71434687
[8] https://winfuture.de/news,148575.html
[9] https://www.technologyreview.com/2025/01/24/1110526/china-deepseek-top-ai-despite-sanctions/
[10] https://towardsai.net/p/artificial-intelligence/a-visual-walkthrough-of-deepseeks-multi-head-latent-attention-mla-%EF%B8%8F
[11] https://www.ft.com/content/ea803121-196f-4c61-ab70-93b38043836e
[12] https://www.cbsnews.com/news/labelers-training-ai-say-theyre-overworked-underpaid-and-exploited-60-minutes-transcript/
[13] https://www.faz.net/aktuell/feuilleton/medien/bullshit-oekonomie-deepseek-und-die-maerchen-der-ki-branche-110266144.html
Originally published on konicz.info on 02/16/2025