What DeepSeek really Means for Tech Stocks

Volatile times ahead: Why the AI winners of the past may not necessarily be the AI winners of tomorrow.

‘A trillion dollars wiped out in a few hours’, ‘DeepSeek shakes markets’ - these were the headlines in the business media, as if something really bad had happened on the technology stock markets at the beginning of this week.

In reality, however, only Nvidia, Broadcom , Arista Networks and a few other providers of data centre infrastructure and suppliers such as energy stocks lost up to 20% of their lofty stock market valuation within one trading session on Monday. Most other tech stocks were just as unaffected by the DeepSeek quake as my own sample tech portfolio.

If your portfolio suffered major losses because of DeepSeek last week, then you are most likely over-invested in stocks that are directly or indirectly dependent on the further course of the AI revolution.

The AI Bubble

Once again, it is time to remember that share prices on the stock market often do not correspond to the actual value of a company. The share prices of the above-mentioned IT infrastructure providers are driven by fantasy and the supposed business potential arising from AI's insatiable appetite for resources.

The share prices of these companies have risen far too quickly in the last two years with the triumph of ChatGPT and co. Most recently, the share price gains have been driven primarily by euphoric sentiment. I have warned several times (and admittedly much too early) of an AI bubble and pointed out three ‘proof points’ for an AI bubble. I would particularly encourage all Nvidia shareholders to read the article on Sequoia's still unresolved $600 billion issue.

Such warnings, arguments and information have so far been ignored by Wall Street. Until now, it was considered a foregone conclusion that companies would have to invest more and more in AI infrastructure in order to remain competitive in the AI age.

Most recently, American AI gigantomania culminated in OpenAI CEO Sam Altman, together with Softbank CEO Masayoshi Son and 80-year-old Oracle founder Larry Ellison, meeting with the new President Donald Trump at the White House to present the ‘Stargate’ project, which aims to invest $500 billion in American AI data centres over the next four years.

What does all this have to do with DeepSeek?

DeepSeek is a Chinese AI company with only 200 employees. The Chinese AI researchers claim that with far fewer computing resources they have been able to train a generative AI that is at least on a par with the latest American models from OpenAI, Anthropic and the like.

Now this could be seen as skilful propaganda from China. And it is certainly no coincidence that DeepSeek went public a few days after the announcement of the $500 billion Stargate programme, which is intended to cement US AI supremacy.

It is probably true that DeepSeek used ChatGPT illegally to improve its own AI. AI experts call this procedure ‘distillation’ and although this is a violation of OpenAI's copyrights, it is common practice in AI labs. I find it hypocritical that OpenAI is now making a political scandal out of this, because after all they disregarded the authors' copyrights for years when training ChatGPT.

To prove that DeepSeek is not a China fake but a genuine innovation, the Chinese researchers have published their latest V3 model as open source. This means that anyone can view the source code of this LLM (Large Language Model), modify it and also use it commercially. V3 is a gift from China to the world, which is now being analysed in AI research centres around the globe.

What can already be said is that DeepSeek has found some extremely clever and creative optimisations for the training and querying (inference) of its AI model. It is estimated that the training could be done 45 times more efficiently than the leading US players such as OpenAI and Anthropic. As a result, the cost of training DeepSeek-V3 (comparable to GPT-4o) was reportedly just over $5 million. Previously, it was assumed that a three-digit million sum would be required to train such an LLM.

This all sounds plausible, as DeepSeek fully discloses the new ideas and algorithms they have used to achieve these incredible advances.

If you are interested in these things in detail, then I recommend the following blog post by Jeffrey Emanuel. You'll need a quiet hour and a little technical understanding to read it. But I promise you, it's worth investing the time and brainpower.

From everything I've read and heard in the last week, I assume that DeepSeek really is heralding a new era for AI. The AI revolution as a whole should receive a further boost and, overall, generative AI should be far less restricted than I and other sceptics have previously assumed. In future, it should be possible to get a grip on the problematic hallucinations of AI with the new inference algorithms and produce much more reliable solutions.

What does DeepSeek mean for tech stocks?

SaaS Stocks are Experiencing a Revival

The big winners will be the application service providers, i.e. the providers of cloud-based application software, as can be found here in the Nasdaq Emerging Cloud Index.

These SaaS companies are dependent on integrating AI functionalities in the form of AI co-pilots, AI agents or similar into their applications in order to remain relevant in the AI age. Until now, it had to be assumed that the operation of these AI functionalities would result in high operating costs, which would put pressure on the usually high gross margins of these companies.

If new technologies such as DeepSeek's now ensure that AI can be trained and operated at a fraction of the previously planned costs, then this is of course very good news for providers of application software.

The financial market has now also recognised this. The aforementioned Nasdaq Cloud Index is up 12% over the past two weeks, far outperforming the big tech-dominated Nasdaq 100 Index.

Development and composition of the Nasdaq Cloud Index

It is also interesting to note that this cloud index is still 35% below its peak reached at the end of 2021.

Infrastructure Cloud Providers Need to Rethink

Until now, it was assumed that application service providers would have to pay more and more money to the major cloud providers (above all AWS, Microsoft and Google) to run their AI workloads there.

Thanks to the DeepSeek research results, cloud providers will presumably require far fewer resources (e.g. energy, GPUs, high-performance network components, etc.) than previously planned in order to operate their customers' AI services. On the one hand, this is likely to lead to lower revenues. On the other hand, applications that previously required too many resources to implement will also become economically viable. This could lead to additional demand from application providers.

Overall, the impact of DeepSeek on the turnover and profitability of the major cloud providers is difficult to predict. The only certainty is that the cards will be reshuffled.

Cloud providers are probably currently frantically revising their business plans, as DeepSeek has changed important framework conditions. Even though Microsoft and Meta have said that they are sticking to their investment plans for the further expansion of data centres, at least in the short term.

The Nvidia Monopoly is Beginning to Falter

Nvidia NVDA 0.00%↑ is currently the de facto monopolist for high-speed chips (GPUs), which are required for training and querying AI models. This monopoly position enables the company to charge prices for its chips that are ten times the manufacturing costs. Nvidia achieves an estimated gross margin of 90% with these products, which is quite unique in the history of the semiconductor industry.

I assume that these margins will be a thing of the past in the foreseeable future. After all, according to the latest findings, which customer would still commit to a billion-euro investment today at the moon prices charged by Nvidia if significantly fewer resources will be sufficient for training and inferencing AI models in the future? Will the competition from AMD be good enough in the future? Or will the in-house chips that Amazon, Google and Microsoft already have manufactured today become a real alternative to Nvidia under the new technological conditions?

It will take some time before the fog clears and the effects of the DeepSeek innovations really become clear. But the uncertainty that is now inevitably arising among Nvidia's biggest customers could be enough to cause delays in ordering Nvidia chips. Or at least to a certain margin pressure.

The new situation will probably not yet be reflected in Nvidia's business figures for the next few quarters, as Nvidia's turnover for 2025 is already largely on the books. Investors and analysts will therefore be particularly interested in the company's longer-term outlook.

As a reminder, most of Nvidia's revenue is NOT recurring, but Nvidia is still at home in the cyclical semiconductor market. Nvidia's revenues have to be recouped year after year. Therefore, Nvidia's revenue should actually be valued much lower by investors than the recurring revenue of a SaaS company. At present, the opposite is the case: Nvidia shares are valued at 27 times sales.

Analysts still expect Nvidia to achieve revenue growth of over 50% and a net margin of over 50% in the coming 2026 financial year. I am curious to see to what extent the very optimistic Nvidia share price targets of analysts will have to be revised in the next few quarters.

In addition to these very bullish voices from sell-side analysts, I strongly recommend that every Nvidia shareholder read this Nvidia Short Case written by an experienced software engineer and AI expert. In my opinion, the crash of Nvidia stock is only a matter of time. I'm really surprised that no analyst has yet been brave enough to give Nvidia shares a sell rating.

OpenAI under Pressure

However, those investors who have invested in OpenAI over the counter should be even more concerned than Nvidia shareholders. I was highly critical of the mega-financing round of OpenAI a few months ago and now assume that the company could face existential difficulties in the medium term.

This is because OpenAI consumes large amounts of capital and relies on revenues continuing to grow as rapidly as they have in the recent past, in order to keep raising new financing rounds at ever higher valuations. Now DeepSeek is coming along and offering, among other things, an API that allows customers to integrate comparable AI models into their applications at a fraction of the cost of OpenAI.

I wouldn't want to be in Sam Altman's shoes. Now he is likely to suffer for having alienated many of his highly successful AI researchers. After all, the big lead that OpenAI had over its competitors just 1-2 years ago is probably gone now.

Stargate a Flop?

Larry Ellison, Masayoshi Son and Donald Trump, in particular, now look rather ridiculous. Just last week, they stood together with Sam Altman in the White House and claimed that they wanted to invest a total of $500 billion in AI data centres in the USA over the next four years.

However, the financing is by no means secured, and even the $100 billion planned for Stargate in 2025 is far from a done deal. Masayoshi Son, who is obsessed with the idea of creating a superintelligence by the end of his life, is said to have promised Altman $40 billion for Stargate. He apparently wants to finance half of it with loans and deposit the valuable Arm shares of Softbank as collateral.

In addition to the equity capital, which is to come not only from Softbank and OpenAI but also from Oracle and MGX (Abu Dhabi fund), Stargate, like other data centre operators, is to be financed to a considerable extent with debt. According to The Information, the company wants to deposit its own Nvidia chips as collateral.

How absurd is that? Everyone should know by now that the currently so coveted Nvidia chips are not exactly likely to retain their value, if only because of the rapid pace of innovation with ever new GPU generations.

I can't really take the reports about Stargate seriously. But maybe the Stargate creators will get lucky and it turns out that they won't need anywhere near $500 billion over the next few years thanks to DeepSeek.

Conclusion

DeepSeek is not a Chinese fake. It represents real innovation and new competition for the American AI industry.

I have never been as optimistic as I am today about the possibilities of AI. We really do seem to be on the threshold of a new AI era. However, this will probably also be characterised by more diversity and competition at all possible points in the value chain.

The AI winners of the past may not necessarily be the AI winners of tomorrow.

In any case, we have to prepare ourselves for extremely volatile times with our tech stocks. If you want to follow the further development together with me:

Disclaimer: The author and/or related persons or companies own a short position in Nvidia. This post represents an expression of opinion and of course not investment advice.