China proves the AI industry is overvalued in one single move
Deepseek AI is making Silicon Valley sleepless, and they're not the only ones.
If you like this article, please don’t hesitate to click like♡ and restack⟳ ! It’s a quick way to make sure more people will get to enjoy it too.
And don’t hesitate to subscribe to my newsletter for free to receive my new articles right in your inbox!
Incoming U.S. president Donald Trump announced the “Stargate project” on January 21, 2025, just one day into office. This project, estimated to cost 500 billion dollars (or half a trillion dollars), aims to secure U.S. leadership in their own self-declared “AI race,” betting on AI being the next cutting-edge evolution for the tech industry.
This project is now dead in the water just four days later.
If you didn’t hear the news, a Chinese business — previously unnoticed in the West — managed to develop a Large Language Model as efficient as all the current top-of-the-line consumer models (OpenAI’s GPT-4o, Anthropic’s Claude, or Meta’s Llama).
It’s no secret that AI is highly costly in terms of energy, and this is why Silicon Valley giants such as Microsoft have announced the (re)opening of nuclear power plants just to power their AI needs.
Basically, U.S. AI needs a lot of power: not just to train the model, which makes it able to generate legible responses, but also to run it. OpenAI’s GPT model for example can be freely accessed from https://chatgpt.com/, and you can chat with it right now. But, every query made has an energy cost — behind the scenes, thousands of graphics cards (powered by Nvidia) are processing these queries and calculating, mathematically, the response the AI should give. Multiply this by millions of users making millions of queries every day, and the cost becomes clear: AI is a resource hog that needs tons of electricity to run and water to cool the cards.
This has been a long-standing problem in the tech industry, and the ever-intensive Silicon Valley companies are coming up with more and more ludicrous plans to “answer” the valid criticism levied at their products.
But it’s clear that when they are able to secure a half a trillion dollar deal to continue in this direction, they won’t come out and say “actually AI is overvalued and inflated, it’s not worth all the taxpayer money you’re giving us.”
Perhaps this is why tech figures such as Sam Altman, the CEO of OpenAI, are increasingly seeing themselves as god-kings over the pleb. In such a structure where everything is handed to them (see the Stargate Project), there is no need to optimize a product. Sam is no stranger to ludicrous claims, and now even faces his own accusations of sexual abuse. For example, Altman is now looking at nuclear fusion— for the moment only studied in small-scale lab experiments, but one that he promises he can deliver commercially by 2028 — as a source of energy generation to fuel his company’s growing needs. He also believes that his company is capable of reaching Artificial General Intelligence (AGI), even though their model has been shown to answer programming queries wrong over half the time. More recently, Altman has even suggested the need for a new “social contract” — so that he could more freely develop his pet AI project without being constrained by such silly things as laws, wages and ethics.
Larry Ellison, co-founder and current CTO of tech giant Oracle, has described his vision for a system of interconnected cameras, all managed by an AI, that would put “Citizens on their best behavior” (while he himself is vacationing in his private island of Lanai, Hawaii, of which he owns 98% — certainly no AI cameras will be present there).
This isn’t just the Friday night ramblings from a group of friends after one too many beer. These are people with the full backing of the government, given money and talented employees to make their individual vision of society a reality for all of us.
In a capitalist system such as that of the United States, these are indeed our rulers. Under the justification that they provide jobs and innovation (despite that the push towards generalized AI is meant to destroy jobs rather than create them, with companies already replacing entire teams with AI), CEOs are given leeway and power that would not be afforded to the common person. They are edified to the rank of visionaries, sometimes even conflated to cult leaders. It’s no surprise that all of these tech CEOs were seen huddling together at Trump’s inauguration.
All of this trajectory — the economics incentive, the technocrat dystopia, the lobbying to push their product on everyone — was thankfully (and perhaps surprisingly) upset in less than 24 hours.
Enter Deepseek-v3.
DeepSeek AI was developed by a startup in Hangzhou, People’s Republic of China, at a measly cost of only 5.5 million dollars. You can test their latest model here: https://chat.deepseek.com/.
Benchmark tests are already showing Deepseek’s AI to be as performant as the currently available commercial models, as charted here: (performance in Y-axis, cost in X-axis).
According to the chart above, Deepseek is the only commercial AI in the optimal performance-to-price ratio at this time, and its generative language capabilities are as good if not better than the leading GPT-4o and Claude 3.5 models.
This has quickly sparked discussions on social media. Here is an anonymous engineer at Meta explaining just how upsetting DeepSeek is for Meta’s Llama model:
Deepseek’s true innovation is two-fold: firstly, it’s as good as the leading models but only cost 5.5 million dollars to develop — at a time where, as we’ve seen, Silicon Valley players need near-infinite amounts of money, or so they say, to develop more AI. For comparison, GPT-4o’s training costs alone (excluding every other cost) were estimated to be 63 million dollars.
Secondly, Deepseek’s other innovation is that its environmental cost is severely reduced. The graphics cards Deepseek requires to train the model and run its chat version are older Nvidia cards, which consume much less power.
The repercussions are huge. In just one move, Deepseek has positively proven AI to be an overinflated industry worth only a few million dollars. Its open-source model also fosters more innovation, as people are allowed to use it as is for their own projects (whether commercial or hobbyist) without having to train their own model. Deepseek can be run on your own PC, requiring not a farm of graphics card being cooled by a nonstop stream of water, but by your own graphics card that you might already use to play games, edit videos or browse the web.
Deepseek’s innovation was driven by constraints, which are a huge driver of creativity. Facing a tariff war and sanctions from the United States government (that started under the first Trump presidency but continued all the same under Biden), China has had to think differently about how it approaches modern-day problems. While Silicon Valley enterprises think that there will always be infinite money and infinite power to fuel their ambitions, Chinese companies instead operate creatively within rigid constraints. Deepseek’s AI was thus forced to not only optimize energy and chip usage. For example, their model uses only 676 billion parameters, which are similar to neurons, whereas GPT 4o boasts 1.8 trillion parameters. Yet, it still beats GPT 4o in terms of output quality.
But Deepseek is not the only Chinese AI model to exist under such constraints. Bytedance, the company behind Douyin (and its international version, TikTok) is also working on several AI models. According to benchmark tests, their just-released Doubao 1.5 model is so cheap to run that it wouldn’t even make sense to charge for API calls, making it completely free to users — whereas GPT 4o charges $10 per 1 million output tokens. Some users have already reported that Deepseek AI only cost them 50 cents for their API calls whereas GPT 4o might have charged them 100 dollars or more. With Doubao, this cost would be 0.
It’s not the first time China petrifies Western arrogance. Previously, China has brought down the cost of solar panels (a much-needed product in the modern age that should adorn every building) that make production as low as 15 cents per watt for the consumer, compared to 40 cents for US solar panels and 30 cents for European panels. The Western response was to sanction these products instead of welcoming them. Chinese production of solar panels accounts for 80% of the global production.
Likewise, China now leads the world in the electric car market — as much as this might upset Elon Musk at Tesla. From production to adoption, China was responsible for 58% of global EV sales in 2023. Chinese models are also affordable to consumers, with models that would be considered luxury in the West being priced at ~$30,000 retail — see for example the Xiaomi SU7.
The SU7 boasts all the amenities modern cars do (360° sensors, dashboard screen, A/C, electrical controls, etc.) along with other amenities — if gimmicky — such as a small fridge in the center console. You might think that at such a price, there must be a catch. But according to reviews, the SU7 drives exceedingly well. Meanwhile, $200,000+ Tesla cars often fall apart in the driver’s hands right after purchase. Yet, while Tesla is being offered packages to open a factory in Germany (which included 2 billion euros in state funding, 1.3 billion euros in grants and 800 million euros in tax incentives over 15 years), Chinese companies like BYD, Xiaomi or Huawei — all of which are producing cheap EVs — are coming across hurdles every step of the way. Importing the SU7 to Germany, for comparison, would add between 25 and 50% of the vehicle’s cost just in tariffs which have to be paid out of pocket.
European and American consumers are asking for these innovative and cheap products, but at every step of the way they are being denied from us. But how come China is able to produce cutting-edge cars and solar panels on the cheap and still make a profit?
These differences are explainable: the West, as the hegemon, has ossified and has no reason to innovate. Its wealth is built on the exploitation of the Global South with coups and massacres that keep labor available and cheap. Our system in the West is completely different from the way China understands business and production, to such an extent that we simply don’t understand how they are able to keep costs at scale and still make a profit, because we are not concerned with these issues. Our economic system in the West exists to give the likes of Musk, Bezos and Zuckerberg more money and access to labor so they can become richer and influence law.
Instead of competing with China or learning from their model, our governments prefer to sanction them and keep their innovations away from us, the common people. They invent excuses such as “[Chinese] state subsidies” to make claims of unfair competition, while ignoring that our governments subsidize tech as well in the West — and heavily so. Alternatively, accusations of “data harvesting” are made, despite the fact that data harvesting in the West for the benefit of intelligence agencies has been thoroughly documented for over a decade ever since Snowden blew the whistle (and between you and me, I’d rather a foreign country own my data than my own. What are they gonna do with it? Advertise BYD electric cars?)
It’s difficult to concisely underline just how much Deepseek has upset the entire industry and the trajectory it was starting to go on. But with their help, we might yet avoid the corporate tech overlord dystopia.
Update 28 January: Predictably, NVidia’s stock price crashed when the markets opened on Monday (they always close on the weekend, and Deepseek was released on Friday). Prices fell from $142 per share down to $118 by closing time on Monday. The company’s market capitalization fell by over half a trillion dollars on that day, marking the biggest loss of any traded company in modern history. Market capitalization is calculated as the price of shares times the number of shares in circulation.
I did almost no editing on this one as I really wanted to get the ideas out on paper while I still had them in mind. Hope it's legible.
Excellent piece thank you so much for the information!