Nvidia share price rebounded 8% Tuesday after plunging 17% on Monday on the DeepSeek news. It then dropped 4% Wednesday and rebounded 1% Thursday.

YTL Power shares dropped 10% last Friday and another 10% on Monday after its bonus warrants proposal last Thursday and the DeepSeek news over the weekend. It dropped another 9% Monday to a low of RM2.93 before rebounding slightly to close at RM3.11.

US responses to DeepSeek News

Donald Trump says news of Chinese AI system DeepSeek should be seen as a positive. The US President said it should just spur American tech firms to do better. The release of DeepSeek AI from a Chinese company should be a wake-up call for our industries that we need to be laser-focused on competing to win, because we have the greatest scientists in the world.

Even Chinese leadership told me that. Trump was speaking in Miami on Monday after news about DeepSeek sparked a global stock sell-off.

BBC Reporting on US responses on 29 Jan 2025

Chinese AI firm DeepSeek says it is facing "large-scale malicious attacks", which are affecting its services.

Chinese state media is citing cyber security experts who say the attacks are originating from US-based IP addresses, which the BBC is unable to verify.

This comes hours after White House press secretary Karoline Leavitt said that the US National Security Council (NSC) is "looking into" DeepSeek.

The little-known Chinese firm's reportedly cheap yet powerful AI model surprised Silicon Valley, which has splurged billions on AI infrastructure - just last week Trump announced an AI plan worth half a trillion dollars involving top US firms.

DeepSeek did not immediately respond to a request for comment from BBC News.

"Due to large-scale malicious attacks on DeepSeek's services, registration may be busy," a banner on the company's website said on Wednesday.

This is the second such attack DeepSeek has reported this week - but Yuyuan Tantian, a social media channel under China's state broadcaster CCTV, claims the firm has faced "several" cyber attacks in recent weeks, which have increased in "intensity".

DeepSeek shot to fame only last week as AI geeks lauded its latest AI model and people began downloading its chatbot on app stores. Its rise caused a slump in US tech stocks, many of which have since recovered some ground.

But America's AI industry was shaken by the apparent breakthrough, especially because of the prevailing view that the US was far ahead in the race. A slew of trade restrictions banning China's access to high-end chips was believed to have cemented this.

Although China has boosted investment in advanced tech to diversify its economy, DeepSeek is not one of the big Chinese firms that have been developing AI models to rival US-made ChatGPT.

Experts say the US still has an advantage - it is home to some of the biggest chip-makers - and that it's unclear yet exactly how DeepSeek built its model and how far it can go.

But the White House has raised national security concerns amid reports that the US navy has banned its staff from using DeepSeek's apps.

According to CNBC, the US navy has sent an email to its staff warning them not to use the DeepSeek app due to “potential security and ethical concerns associated with the model’s origin and usage”.

"I spoke with [the National Security Council] this morning, they are looking into what [natinal security implications] may be," said Ms Leavitt.

Speaking on Fox News, the recently appointed "White House AI and crypto czar", David Sacks, suggested that DeepSeek may have used the models developed by top US firm OpenAI to get better.

This process - which involves one AI model learning from another - is called knowledge distillation.

"There's substantial evidence that what DeepSeek did here is they distilled the knowledge out of OpenAI's models," Mr Sacks said. "I think one of the things you're going to see over the next few months is our leading AI companies taking steps to try and prevent distillation... That would definitely slow down some of these copycat models."

OpenAI echoed this in a later statement that said Chinese and other companies are "constantly trying to distill the models of leading US AI companies."

"As the leading builder of AI, we engage in countermeasures to protect our [intellectual property]... and believe as we go forward that it is critically important that we are working closely with the U.S. government to best protect the most capable models".

At his confirmation hearing on Thursday, Trump’s nominee for Commerce Secretary, Howard Lutnick, also shared concerns about theft and raised the prospect of further US action to protect US AI companies.

“What this showed is that our export controls, not backed by tariffs, are like a whack-a-mole model,” Lutnicck says.

As DeepSeek rattled markets this week, US President Donald Trump described it as "a wake-up call" for the US tech industry, while suggesting that it could ultimately prove to be " a positive" sign.

"If you could do it cheaper, if you could do it [for] less [and] get to the same end result. I think that's a good thing for us," he told reporters on board Air Force One.

He also said he was not concerned about the breakthrough, adding the US will remain a dominant player in the field.

Naomi Haefner, assistant professor of technology management at the University of St. Gallen in Switzerland, said the question of distillation could throw the notion that DeepSeek created its product for a fraction of the cost into doubt.

“It is unclear whether DeepSeek really trained its models from scratch,” she said.

“OpenAI have stated that they believe DeepSeek may have misappropriated large amounts of data from them.

“If this is the case, then the claims about training the model very cheaply are deceptive. Until someone replicates the training approach we won’t know for sure whether such cost-efficient training is really possible.”

Skepticism on DeepSeek’s low cost claims

Dylan Patel of chip consultancy SemiAnalysis has estimated that DeepSeek and its sister company, the hedge fund High-Flyer, has access to tens of thousands of Nvidia GPUs, which were used to train R1’s predecessors.

“DeepSeek has spent well over $500mn on GPUs over the history of the company,” Patel said. “While their training run was very efficient, it required significant experimentation and testing to work.”

Rival company insiders and investors have expressed skepticism about the low costs cited by DeepSeek in developing its models. In December, the company said its V3 model, which its app’s chatbot runs on, cost $5.6mn to train.

However, this figure was only for the final training run, not the complete cycle, and excluded “the cost associated with prior research and …experiments on architectures, algorithms, or data,” it added.

Pat Gelsinger, recently forced out as chief executive of Intel, was among those buying his former rival Nvidia’s stock on Monday. “The market reaction is wrong: lowering the cost of AI will expand the market,” he said in a LinkedIn post. “DeepSeek is an incredible piece of engineering that will usher in greater adoption of AI.”

Data Security

Wiz Research has identified a publicly accessible ClickHouse data base belonging to DeepSeek, which allows full control over database operations, including the ability to assess internal data. The exposure includes over a million lines of log streams containing chat history, secret keys, backend details, and other highly sensitive information. The Wiz Research team immediately and responsibly disclosed the issue to DeepSeek, which promptly secured the exposure.

As DeepSeek made waves in the AI space, the Wiz Research team set out to assess its external security posture and identify any potential vulnerabilities.

Within minutes, the team found a publicly accessible ClickHouse database linked to DeepSeek, completely open and unauthenticated, exposing sensitive data. It was hosted at oauth2callback.deepseek.com:900 O and deve.deepseek.com:900.

This database contained a significant volume of chat history, backend data and sensitive information, including log streams, API Secrets, and operational details.

More critically, the exposure allowed for full database control and potential privilege escalation within the DeepSeek environment, without any authentication or defense mechanism to the outside world.

Extracts from Bloomberg TV responses on 28 Jan 2025

Does this worry you? Well, I think there's so many headlines out there, there's so much news flow, so let's take a quick step back and think about what's happened over the last month, because this news that's really getting all the headlines today and got all the headlines over the weekend with DeepSeek really started, well, I guess a year ago when they released their model called V2, which introduced a lot of these potential breakthroughs. Those breakthroughs came to fruition in a model they released a month ago. So a lot of this has been out there.

Last week, they released their latest reasoning model that coincided with the inauguration. There's been a lot of headlines since then. So what my take on DeepSeek is, in particular, is I think a lot of the advancements that they've made are very real, and it's something to pay very close attention to if you're an NVIDIA shareholder.

I want to read to you a statement that NVIDIA emailed to our reporters, Ian King here at Bloomberg, basically saying that it is an excellent AI advancement, and they say the work illustrates how new models can be created, but they kind of dismiss some of these concerns as somehow this is a negative for NVIDIA itself. Well, I mean, I think there's clear negatives for NVIDIA. There's also potential longer term positive, I’d say too.

So on the negative side, what's clear is DeepSeek has created a way to, I'd say, do more with less. What they've done, the way I think about it in simple terms, is take a professional certification exam like I took the CFA exam. There's a lot of ways to study for that exam.

One way is you read all 5,000 pages of the textbook. On the other extreme, you just do practise questions. What DeepSeek did is when they trained this reasoning model, they kind of took the practise question approach.

They took some shortcuts but created something that's definitely good enough and capable enough that was done very cheaply. I love that analogy. By the way, I did the practise route, too, and I failed level two twice, and then eventually I just read the book and I passed.

I was like, but then you passed. So it's okay. And it was interesting to see also NVIDIA talk about how DeepSeek was export control compliant. It was like, please don't restrict us anymore when it comes to exports.

Do you buy the dip? So again, I think with NVIDIA, you don't trade the stock. You own it for the long term.

And I think two points I'd make on sort of the positive side for NVIDIA, what could this do for the industry more broadly? If the cost of AI inferencing, the cost of AI training comes down, it could make the applications more broad-based. It could unlock new pools of demand, new potential customers. So this could be good for innovation in the AI space broadly, which would be good for NVIDIA in the long term.

The other point I'd make is, you know, I think this is good for big tech as well, separate from NVIDIA. I think lower cost of inferencing, very good for a company like Meta, very good for a company like Google, very good for a company like Amazon, companies that have AI embedded in their P&Ls. And then back to NVIDIA, the other point I want to make is this happens from time to time in tech investing.

Big tech stocks get left for dead and they come back. This happened with Microsoft many years ago. This happened with Apple after Steve Jobs passed away.

This has happened many times with Meta, many times with Google. Amazon was never going to turn a profit. Netflix was never going to turn a profit.

These existential questions tend to come up from time to time. That's what's happening today with NVIDIA. I think, you know, there are plenty of near-term questions that these companies are going to have to address during earnings this week, but I still am a holder of NVIDIA long-term.

By the dip is not really how I generally trade NVIDIA, but I'm still owning NVIDIA for the long-term. At the end of the day, though, if we get into a world where we can reduce the cost and you don't actually, you can't charge us a lot for AI services, like the everyday consumer, who is that bad for? Well, I think, so I'd say, first of all, it's probably bad near-term on the training side. I do think it's not, it's seeming to be not as important to invest massive amounts of dollars in training to get the best output.

I think DeepSeek is showing us that there's ways to optimise your AI spending in a way that hasn't been done yet. So I think that's, it's probably bad for them. For someone like OpenAI and some of the model companies, so Google Gemini would be another one there.

I think this is showing that there is, those models commoditise pretty quickly. I think it could be bad for these model companies. We'll see.

Again, it's very early. This is evolving very quickly. That's where I'd start as kind of the losers.

Then the derivatives for some of that stuff. So some of the industrials, some of the power companies, we'll see how this unfolds. I think near term, probably fewer massive training data centres constructed, but long-term, could there be more inferencing data centres constructed?

Could there be more AI run at the edge? Could there be more AI infrastructure built in that way? We'll see.

I am curious as to, I mean, you mentioned some of the earnings we're going to get. We're going to start to hear from some of the hyperscalers and some of the, basically the buyers, the ones that have been investing a lot of money in this. Do you think we're going to hear any sort of material change in the amount of money that they've committed to these projects? Well, Romain, I mean, that's one of the really interesting things about the last week.

So this R1 model from DxE was launched on Monday. And then the next day you had this Stargate announcement with Oracle and OpenAI, and we can debate how real that is, but a $500 billion AI investment programme. And then on Friday, Meta pre-announced, unexpectedly, their 2025 CapEx plans, $60 to $65 billion, about $10 to $15 billion above the street.

So clearly, Meta is going all in on this. And I think their LLAMA models are the closest deep-seek comp, by the way, instead of open-source models like deep-seek. So to answer your question, will these companies backtrack on the CapEx plans? I don't think so near-term.

I think the question is more next year, 2027, is the investment cycle finally starting to crest? We'll see there.

Extracts from Notes on Schwab Network on 28 Jan 2025

NVDA, A.I. Potential's Massive Upgrade with DeepSeek

Fast market. I'm Diane King Hall at the New York Stock Exchange. I'm alongside Kevin Hanks.

He's over at the SIBO. Time now for our cash tag segment. For that, we want to welcome in Andy Swan, co-founder of Likefolio.

Today, we're taking a look at NVIDIA, Andy. All right, that's stock down 16 percent. It is being absolutely rocked on the heels of this news about Deepsea, this China venture developing this open source, large language model for a fraction of the cost.

It's when I looked at the Apple App Store, and Kevin pointed this out earlier, it is at the top. It's still at the top of the downloaded apps list. Now, ChatGPT, OpenAI's ChatGPT is number two today.

But what do you make of all this, Andy? Well, I think it's a mistake to just accept the claim that this was done at a fraction of the cost. When I first heard the news, the first thing I did was go into DeepSeek, which is a phenomenal AI tool, and asked it to have a conversation with me about Tiananmen Square, and it would not do that. This is a company that is owned and controlled in some way by the Chinese Communist Party, not exactly known for being forthright with its disclosures.

I think we can discount that factor a little bit. The thing that is encouraging to me about this is that what DeepSeek did do is it created a better mousetrap, and it created a better way of thinking, of training, and it definitely created a cheaper way of doing so, especially in terms of the training data necessary to achieve high outcomes and the way that the model can teach itself through reinforced learning. These are big breakthroughs.

To me, the story here is that AI overall's potential just got a massive upgrade, that the ceiling rose in terms of what AI can accomplish. I think we will see a lot of this over the coming years, because this is the sector that is getting the heaviest investment in probably the history of technology, massive companies chasing everything they can get, putting everything they can put into it. I think the big takeaway for me here is that, first of all, it is not just a large company game anymore.

If you accept what DeepSeek said, you do not have to have $50 billion to participate in this market. I think that will open the AI up to a lot of smaller companies. It will broaden NVIDIA, AMD, Intel's customer base quite a bit.

It will bring a lot of new players into this arena. There was a feeling out there before that you needed $50 billion to compete, and you needed to be one of these four to six companies with those kinds of resources to even compete. Now, that is a little bit shattered.

I think you will start to see new entrants into it. I do not see anyone from Microsoft, Oracle, Google, Amazon walking into the offices today and saying, wow, did you see what DeepSeek did? I think we should spend less on AI. I just do not buy that narrative.

Andy, there are so many questions that are still out there for this company. They may have, in one release, done it cheaper. When I keep reading, there is no guarantee that the next one is going to be as inexpensive.

Then, Andy, in your using common sense, as we like to do on this show, how is an American company ever going to trust their AI to a Chinese company? In what world, with everything we have learned for the last 24, 36 months about Chinese companies and restrictions, and how like Alibaba and all these companies, who is going to trust a Chinese AI company? I just do not see it happening, no matter what they can do. Yes, could it possibly affect prices and put pressure on them to come down lower? Yes. Could it help improve the American companies doing it? Yes.

But I do not see American companies switching to this. Do you? No, absolutely not. I do not think that is a real threat.

In fact, I think that as of this moment, DeepSeek has just restricted users to Chinese mainland phones here in the last hour or two. This is something that is going to get clamped down on. Of course, the big story here is that they supposedly did this for $5.5 million.

I do not believe that at all. But again, what I do believe and what they documented really well is that there is a better way to train these models. There is a better way to get them to teach themselves.

That is an extraordinarily bullish signal for the AI sector. That tells us that the limits that we thought were there just a few months ago have already been broken. Again, I think this will happen over and over again.

This is the first time it has really happened where you see a step change in capability, which can be really scary. It can reprice some models and some forecasts, certainly. But this sector will continue to see step change technological advancements over and over again.

Wall Street is going to have to get used to it and going to have to get used to the idea that their models are going to be broken within three to six months every single time. The real question is, are these companies really going to decrease how much they're willing to spend having the best AI because someone in China proved that they could do it better with less? I think the answer is clearly no. It's so funny, Andy.

I got chills when you made that mention about trying it out and what you looked up with regard to Tiananmen Square. I just did a quick search on ChatGPT to see what it pulled up. It pulled up a variety of information, including the historical information.

It shows that open AI still has that leadership, at least in terms of informing with regard to diverse information. But I do want to ask you in the context of devaluations, because there are some analysts who have argued that devaluations have been stretched within mega cap tech, especially they talk about NVIDIA being one of those. What do your channel checks or what does your data show you with regard to what CapEx looks like for NVIDIA and the like and how it compares across the group that you're watching? Yeah, it's exponential growth for NVIDIA.

You can see it in the reported expenses of the major companies, how they're forecasting out their CapEx for the next three, four, 10 quarters. I think it's very clear that NVIDIA has by far the best chip. And that's where the money is going to continue to flow, in my opinion.

Latest results

Microsoft and Meta Platforms have in combined spent a total of $37.4 billion in the quarter ended December 2024, according to the companies’ earnings reports Wednesday afternoon.

My Initial Assessments

Initial assessments of the potential impacts of DeepSeek to YTL Power’s AI data centre business have given rise to several points for take-away:

1. The low costs of development cited by DeepSeek is being challenged by numerous industry experts. The total costs may well exceed $500mn, almost a 100 times larger than what was claimed by DeepSeek.

2. AI experts commented that low end chips may be used for AI inferencing, but not for AI training work which definitely require high end AI chips.

3. US President Donald Trump is urging US tech firms to catch up quickly. He was not concerned about the breakthrough, adding the US will remain a dominant player in the field.

4. The US National Security Council (NSC) is "looking into" DeepSeek.

5. The White House has raised national security concerns amid reports that the US navy has banned its staff from using DeepSeek's app.

6. The White House thinks that DeepSeek have distilled the knowledge out of OpenAI's models as said by David Sacks who thinks that over the next few months the US leading AI companies are taking steps to try and prevent distillation... That would definitely slow down some of these copycat models.

7. DeepSeek shows that there are ways to optimise AI models at lower costs, which may commoditise AI models quickly. That is bad for AI model developers like OpenAI and Google Gemini, as reported by Bloomberg TV.

8. If the cost of AI inferencing, the cost of AI training comes down, it could make the applications more broad-based. It could unlock new pools of demand, new potential customers. So this could be good for innovation in the AI space broadly, which would be good for NVIDIA in the long term.

9. In near term, probably fewer massive training data centres will be constructed, but long-term, there could be more inferencing data centres constructed, some experts said to Bloomberg TV.

10. Experts have the view that US hyperscalers are unlikely to materially change their capex plans (eg. $500bn investments for Stargate, Meta’s announced capex plan of $60-65bn for 2025, etc.) after the emergence of DeepSeek, as they will probably need to catch up quickly as Trump said.

11. AI overall's potential just got a massive upgrade, that the ceiling rose in terms of what AI can accomplish. The big takeaway is that, first of all, it is not just a large company game anymore. An AI player does not have to have $50 billion to participate in this market, and hence it will open the AI up to a lot of smaller companies. It will broaden NVIDIA, AMD, Intel's customer base quite a bit. It will bring a lot of new players into this arena.

12. American companies are unlikely to trust a Chinese AI company and switch to DeepSeek models, as opioned by experts interviewed by Schwab Network.

13. DeepSeek will not have any impact on the AI data centre development by YTL Power, at least in the near term or on the initial phases already secured for Nvidia and the US hyperscaler, as DeepSeek’s development costs are likely to be high as Point 1. suggests, US government may just ban American companies to use it on national security concerns as Point 4 & 5 suggest, and American companies will unlikely switch to a Chinese AI model as Point 12 suggests.

14. The investments by US hyperscalers in data centres in Malaysia will likely go ahead with no material change as Point 3 & 10 suggest. Nvidia and the US hyperscaler have already signed agreements with YTLP and committed to their investments in the new AI data centres in Kulai, and these plans are unlikely to change, especially the physical DC buildings are already substantially completed and the required chips are already on the way for delivery to Malaysia later this quarter.

15. The emergence of DeepSeek may be good for the long term development of AI space as Point 8 & 11 suggest. This argues well for the potentials of YTL Power securing more AI players, big or small, into its new data centres in Kulai in the longer run. Its green DC Park in Kulai still have capacity for 300MW more.

That’s my initial assessment based on information available and news flows from the US over the past few days.

Net short positions on YTL Power have increased to 42.2 million shares at close Tuesday. Traders beware, but for long-term investors, the share price drop has presented an excellent opportunity to accumulate more at bargain prices.

I see that besides the potential earnings dilution of about 10% from the bonus warrants proposal, nothing has since changed to YTL & YTL Power’s fundamentals. The DeepSeek episode will just pass in no time, as Donald Trump sees no threat to US dominance in the AI space. Hence, I see good chances for the share price of YTL Power to rise back to RM4.00 level pretty soon, especially once the short selling activities subside in coming days.