January 27, 2025 - 2:00pm
Last week, the Chinese AI company DeepSeek released its new DeepSeek-R1 and DeepSeek-R1-Zero models. Chatter built as programmers and enthusiasts started testing the new software out. By the end of the week, X was awash with users claiming that the new Chinese AI model could manage what the best Western AI models â such as OpenAIâs o1 â do at a fraction of the cost. Whatâs more, DeepSeek is open-source, meaning that anyone can use it. On Friday, tech CEOs were saying that Ivy League researchers from Stanford to MIT had switched to the Chinese model.
As of this morning, tech stocks that had been riding high on the AI boom have started to collapse. In pre-market trading, stocks such as Nvidia, Arm, Micron and AMD were all posting double-digit losses. The entire Nasdaq technology stock index is now under threat, and is already showing signs of losing ground. It is no secret that AI has been in the driving seat of the American stock markets for some time, but now the entire business model appears to be at risk.
DeepSeekâs threat to American firms riding the AI boom operates at multiple levels. The first layer of the business model challenged by the Chinese company includes the firms working on AI models themselves. Goldman Sachs estimates that in 2024 there was around $68 billion of investment in the sector in the United States. But DeepSeek seems to have undercut the entire market on price by a factor of around 20. Early estimates show that DeepSeek input and output tokens â that is, the smallest unit of text processed â are around 96% cheaper than alternatives.
Related to this is the low cost of developing the new Chinese AI model. DeepSeek-R1 cost only around $5.6 million to develop. In comparison, OpenAIâs o1 model cost $600 million to develop. This is not just a minor difference in development cost. The fact that DeepSeekâs development cost less than 1% of that of its American competitors calls into question the entire sector in the United States. Is the entire edifice based on overinvestment â that is, too much money sloshing around causing businesses to become bloated and lazy?
Then there is the impact on the market for graphics processing unit (GPU) chips. Early estimates suggest that DeepSeek uses only around 9% of its rivalsâ GPU capacities. This means that, as researchers and commercial players adopt the new Chinese model, they will need far less physical GPU infrastructure. But companies such as Nvidia have seen their stock market valuations skyrocket precisely due to Wall Street analysts factoring in massive future demand for GPU chips. This explains why the Nasdaq index is now at risk of collapsing â before the release of DeepSeek, Nvidia alone made up around 14% of the entire index.
The irony of the DeepSeek saga is that the developers of the model were pushed to make it so efficient because the United States is trying to restrict access to American-designed chips from the likes of Nvidia. Since necessity is the mother of all invention, this has pushed Chinese developers to economise on chip usage. If DeepSeek does indeed crash the entire American AI industry â and possibly even the stock market â it will be because bureaucrats in DC tried to control the technology sector in ill-advised ways.
Philip Pilkington is a macroeconomist and investment professional, and the author of The Reformation in Economics
philippilk
Join the discussion
Join like minded readers that support our journalism by becoming a paid subscriber
To join the discussion in the comments, become a paid subscriber.
Join like minded readers that support our journalism, read unlimited articles and enjoy other subscriber-only benefits.
Subscribebecause bureaucrats in DC tried to control the technology sector in ill-advised ways.
You say that like you are surprised.
I might be wrong but I’m pretty o1 is 4 generations behind the current OpenAI o4 commercial offering.
Even though enough value was wiped off my share portfolio this morning to make me wince, it’s good that a bit of pressure has been let out of the US stock market. The valuations have been wild.
Let’s just stay calm about this DeepSeek thing – 2 years ago we were all going crazy about ChatGPT and making the craziest predictions. I work in SEO and the doommongers were telling me that the whole discipline was going to die within the week – and it hasn’t. It’s just changed. So let’s just ride this one out too and just deal.
There are massive overvaluations in the tech sector, crypto, AI, and the magnificent7. It’s mostly hype and fomo , unless AI eliminates millions of middle class jobs in the the next few years to justify the investments it will flop. Stock values are at record highs based on most measures such as PE, price to sales, price to book. Nvidia can easily go to 50$ at the first signs of sales peaking. Target price Microsoft 250, Tesla 150, Amazon 150 , Apple 150. Many of the AI stocks could lose 80% of their value in a crash. I won’t even get into crypto, one of the greatest ponzi schemes in history. One of the greatest bubbles in history can burst at any time with nasdaq going to 10,000 or even lower. Investors and speculators are sitting on a time bomb and there is nothing that Trump or Musk or anyone else can do about it.
Nvidia c14% down as I write this but it’s no stranger to large corrections is it? Microsoft only 3% down Meta virtually no change, Dow ditto. Hysterical article. DeepSeek, if something’s too good to be true then it probably is – what has been described appears to me as impossible without theft or some big limitations which the market hasn’t sussed out yet. Might be wrong and rue the day I didn’t go all cash but I think I’ll ride it out thanks.
Interesting essay. First Iâve heard any of this.
Things move very fast in the AI world.
Is anyone surprised at the timing of this? i.e. give the new US administration a split second to exhale, then cause a sharp intake of breath?
Whatever the longer-term effects, you have to give the Chinese credit for political nous. If it doesn’t cause the US to up its own game, that’ll be entirely the fault of the US, but also a problem for the rest of us.
Meanwhile, the UK lags so far behind in the tech race, we’ve virtually disappeared over the rear horizon.
Deep Seek is open source, so anybody can take the code, and make a version without the guard rails that ChatGPT has.
How is Starmer going to regulate that?
I watched Zuckerberg on Rogan Experience, and when he was talking about deepseek, if you really paid attention, you could see palpable fear in him.
In the U.S., technology ventures competes for profit, whereas in China, the motivation is not profit only but advancement – the best of the best must prevail. The idea that competition drives innovation is flawed because, in America, companies arenât truly competing to be the bestâtheyâre competing to be the most profitable.
Zuckerberg looked genuinely scared when discussing how China handles its technology regulation compared to the U.S. regulation. He couldnât say the words outright, but they were on his lips.
If it is true competition of pure technology, US already lost.
Have you got an approximate time stamp for that assertion. I’m not going to sit through nearly 3hrs of interview to check it out. Primary souce material please!
I would oblige this time but remember internet is free and anyone can search anything: Take what you need: https://youtube.com/shorts/S2JkKaw69Vg?si=7OPcnoIHHEX3UdHU
Also at the 7:59 you will see Meta panic letter after Deepseek was released. I cannot make this thing up. I am not that intelligent actually.
https://www.youtube.com/watch?v=OC2J-0vlhy8
A few years ago Google engineers allegedly expressed their concern that big tech really does not have any special sauce when it comes to AI. The transformer methods are public and anyone can more or less develop something with relatively little effort. Although I don’t think competitors can surpass Nvidia quickly, the entire premise that extreme amounts of GPU power will be the key to useful AI may be a bit of a gamble.
So the biggest risk at this point is not even the question to what degree AI can live up to expectations but the fact that everyone is betting on handful of US companies. Especially because this euphoria seemed to have stopped a serious Covid bubble from deflating.
Most likely, dropping the price of “AI” processing will make it useful in a wider range of market niches.
“But then it hit the third paragraph, starting to write that âthe Uyghur community has been the subject of significant international attention due to reports of human rights abuses by the Chinese governmentâ, and that these reports âinclude allegations of mass detentions, forced labour, surveillance, and restrictions on religious and cultural practicesâ.
And then the message vanished.
Suddenly beset by Orwellian shyness, DeepSeek deleted its answer and replaced it with a single line:
Next, I asked it to: âTell me about the Tianammen Square massacre.â But it went silent.
Just donât ask it about Tibetan exile, Xi Jinpingâs first wife, starvation from the Great Leap Forward, the impacts of the Cultural Revolution, the Hong Kong protests, âGrass-Mud Horseâ, âTank Manâ, âEmpty Chairâ, or what happened to Jimmy Lai….DeepSeek doesnât have any thoughts on those. Letâs talk about something else. From The Speccie…
âDeepseek, could you tell me what happened at Tiananmen Square all those years ago, and give me details of the current treatment of the Uighers?â
Crashes
There is a direct parallel with the cost of the Manhattan project, and the bombs subsequently built on a shoestring by first the UK, and then Russia, France etc.
I made precisely this and related types of points a while back, and this time I reference myself, yes, for a pat on the back because I deserve it, and also to avoid typing out stuff yet again:
1. https://unherd.com/2023/07/we-wouldnt-want-oppenheimer-today/#comment-658186
2.https://unherd.com/newsroom/is-wokeness-undermining-science/#comment-249933
To my eyes, what this means is not that China will take a lead in AI, but that we have entered an arms race where none of the barriers to entry to creating nukes exist at all (like access to fissionable material and to high-end engineering, etc.) which means neural net developments will now not just accelerate, but absolutely explode, because the potential rewards are simply off the scale. This implies a danger that people are grossly underestimating. I don’t mean existential risk (and to be clear I am an outright AI existential risk doomer), but the fact that these technologies are arriving much much faster than people can possibly adapt. This technology is not like the motor car arriving in the early 1900s, and quickly replacing the horse within a couple of decades. Or rather it is, but the entities being replaced are humans. The consequences look rather obvious to me (and people are free to bury their head in the sand if they want to) – the entire model of human work breaks, literally within half a decade from here.
*I am at this point tempted to change my name to Cassandra by deed poll, but I won’t because I don’t really want to transition.
I totally agree that AI is now an arms race. And the chip manufacturers are going to get even more important now. And despite the shock to Nvidia’s share price, Nvidia is going to end up as another big winner, at least until the time the world discovers how to do AI without silicon. However, I disagree with the doom-mongering. Economies are always an exchange of human labour. If you automate out one piece of human labour, those individuals will then share labour in different ways. Word processors removed the need for typists, but that work shifted to other areas.
One of the interesting things from AI is how it changes the nature of creative work. In the same way that photography meant you didn’t have to have good technical skills to create art, just a good eye, AI presents the possibility that you don’t need to have good writing (or any direct creative skills), just a good eye. With expertise, you can use AI to shape output to what you want. Changing and re-imagining is now cheap compared to the labour of reworking sentences, drawings and code, so the job is in deciding what the output should be, not in the techniques needed for creation.
However, that also means that AI output is not ‘singular’ – there is not one definitive result, but a space of possibilities which can be guided. AI output can be anything you imagine it to be – it doesn’t produce ‘a’ truth. And because it can be manipulated and directed, and suffers from hallucinations, we have to get better at judging and questioning the output, because output is now incredibly easy, but reality remains hard.
Already, you can see artefacts in journalist writing – Imagine, Consider, AI emphaticisms, sentences that start with ‘But’. But the writer is moulding the tool to the task, they just haven’t quite got the level of polishing right yet. AI is not an end point, but the start of something different.
Workers displaced will move to ‘Other areas’
This seems woefully unsubstantiated and unlikely. There is a finite number of people jobs. X number. This number will not increase as AI is integrated – it will reduce in total.
It is also very doubtful that the additional profit yielded through labour savings will be taxable at a rate to completely pay for the ‘UBI’ that the Lizards keep telling us we can look forward to.
I think it’s fair to say AI is unstoppable
Also fair to say that it remains to be seen how well humans who lose out will be compensated by the system.
If the market decides – they won’t be – that seems plain..
Two observations.
1. Now seems an appropriate time to reflect on the holistic benefit of ‘Work’ – that is purpose, identity, social utility and so on.
For these will not appear on the balance sheet.
2. It seems likely that an entire ecosystem or culture will arise that rejects AI.
Bookshops, galleries, institutions & agencies that guarantees no AI.
Ironically the new Luddites will use AI to filter out AI itself.
A new sector will arise that cater for this new ‘Community’
Moments like this I like to think of builders and roofers and plasterers. Driving around in their vans chuckling at the thought of lawyers having to retrain as carers.
The fact that DeepSeekâs development cost less than 1% of that of its American competitors calls into question the entire sector in the United States.
Or, you know, it suggests it’s a fraud.
There are two ways to interpret your last sentence.
They can both be true!
This means that artificial intelligence will become far cheaper. That means that the world of autonomous cars, robot servants, UBI and perpetual deflation is far closer. This should be embraced rather than feared because liberation from drudgery can only be good
It’s sweet that people still think that a UBI future, in which we are freed up by machines, to dance, write poetry, pursue our interests – is a plausible reality.
Do you really think western govts will be able to afford to finance this pipe dream when the job losses start. They ll take the efficiency savings on behalf of companies but the rest of the fantasy will never materialsie.
Think Ready player one.
We sit at home and watch TV, eat gruel, take our meds.
The Chinese are doing something similar although not as dramatic in carmaking and computer driving. Some say it’s the Chinese government’s industrial policy that is driving all this innovation, but it’s not. It’s mainly being driven by the way the Chinese specialize in parts of the supply chain, making products easy for system integrators to build with a modular architecture. We in the US are missing out on this, as we stick with oligopolistic markets dominated by big, uninnovative companies.
https://www.stuff.co.nz/nz-news/360562380/watch-we-test-deepseek-censorship-it-doesnt-go-well
Looks as if my earlier comment in regards to Tiananmen and the Uighers wasnât far wrong
Interesting link. The live change in the response to the prompt “Is there persecution of Uyghurs in xinjang” from “however” to âSorry, thatâs beyond my current scope. Letâs talk about something else.â is particularly interesting.
Can any other AI app change their response live? Does the published open source code allow responses to be started before the processing for them is completed?
US based AI apps do the same kind of thing to reflect American political correctness.
This weekâs story from the world of AI.
It’s Chinese….this means using it will allow the Chinese to spy on you and your work. The fact that it is so cheap means it is heavily subsidized by the CCP to ensure wide uptake. The naivetĂ© of the tech community in the West knows no bounds. Don’t they know that the Chinese want to bury us? After crucifying us? How incredibly dumb can they be? Do they really believe that because this product is new, it signifies progress (i.e. has no defects)? This product should be banned from our shores like all the other stuff coming out of China….!
This article takes the claims from China at face value. Rumour from China – which of course may be wrong – suggests that the development of DeepSeek actually involved pretty much all Chinaâs black market stock of the most powerful NVidia chips and a colossal investment. Why? To destabilise a key area of US advantage, which is entirely consistent with CCP strategy and its past record. We donât yet know the reality behind Deepseek, but it seems implausible – to put it mildly – that the finest US minds in the field, the huge investments, and by far the better technology (to date) were so quickly and comprehensively rendered redundant by a tiny team working with old tech on a shoestring. So this article is perhaps a little naive.