DeepSeek Disruption

Dear friends and well-wishers,

I recorded an interview with Ausbiz earlier this week.

DeepSeek got all the press, but Alibaba released a model that outperformed it shortly after - only it wasn’t open-weight, so it’s unlikely to get major success outside of China. We’re also expecting imminent updates to OpenAI and Llama’s latest models, so it will be interesting to see what they’ve achieved from a much larger spend.

Personally, I don’t think the US megatech founders see a successful Chinese startup and decide to fold and roll-back their plans to win in AI… but we shall see. META and ASML reported this week, but more of that in another note.

Best regards

Michael

Andrew Geogeghan: As we have heard all day, Chinese startup DeepSeek launched a free AI system that says it uses less data at a fraction of the cost of incumbent services. And by Monday, the assistant had overtaken US rival, ChatGPT in downloads from Apple's App Store.

What have you made of it? What do we know of DeepSeek and the potential impact it has?

Michael: Oh, it definitely has a big impact. I mean, there's been some question over the quality of foundational models as a business. And this really brings to the fore the kind of challenges that whole business model has. DeepSeek has actually released the weights of the model, which means you don't have to use their servers in China, you can take the weights, you can take the model and host it yourself. And it's also extremely good.

So that really brings into question why you would pay a firm like OpenAI $300 Aussie dollars a month when there's an equivalent model that has been released for free. And it's not just deep seek. This will happen in the foundational model sector again and again. Every time there's a development, every time a really good model releases its weights, they'll almost invalidate all the spending that has been done to date everywhere else. And, you know, the headline grabbing figure was that they spent six million USD on the training run. Now obviously it was much more than that.

That was just assuming they're renting some GPUs at a set rate, doesn't take into account salaries. There's also questions over whether they had access to Nvidia's latest hardware. As a Chinese company they can't exactly admit to it, but it's possible. But there's no doubt that it was done for a fraction of the cost and it's certainly being delivered at a fraction of a cost, it’s free to use. So again, that really throws the whole business model of these foundational model companies into question.

Andrew Geogeghan: What does this do for valuations and in particular talk about capex, the huge spend that's been undertaken at the moment. And I mean a case in point I guess, we had that announcement last week with Stargate backed by the Trump administration. Where does that leave that now?

Michael Frazis: Yeah, I mean that seems almost dead in the water, 500 billion, but it's not like it was fully funded. You know Masayoshi Son came in from Softbank. He has a track record of kind of top-ticking markets as well. It looks like DeepSeek used existing foundational models to train their own. So they used existing models to basically rate their answers to generate more training data to improve their own models.

So it makes it hard. I mean, how can you invest a large amount of money if that model is just going to immediately be used to train another model at a fraction of the cost? And can then be delivered for a fraction of the cost?

It's probably too much to say the Chinese have taken the lead, but they're certainly up there now with this model. Will the hyperscalers and the big US tech companies scale back capex now because of that? Because of the risks? I'm not sure. You know, it might also spur some competition.

Andrew Geogeghan: But are investors are going to balk at that?

Michael Frazis: Well it has certainly brought the whole context into question and I think there's a difference between the companies that will benefit from this. There's actually quite a few winners. For example, Apple was up that day when most big tech and certainly all semiconductors were down.

They didn't invest heavily in foundational models but this one's so good and so efficient that it's likely to be able to run on their hardware, which is an advantage. AWS is probably another winner because you've got a much cheaper model that's more widely available, that really opens up more use cases, that's going to still need more data, more compute capacity, and lot of that will end up on AWS servers.

The firms that are most challenged are companies like OpenAI, which were intending to charge for their model as a business.

Andrew Geogeghan: You take a look at the impact locally, not surprisingly data centres have been hit. So too uranium miners, obviously which are powering the nuclear reactors, which are powering the data centres, what's your view on those particular stocks then?

Michael Frazis: I think there's two sides. There's a market side where that was all tied up in one trade and it became a very consensus trade. There's some of the top performers, particularly uranium miners and energy providers.

In the first leg Nvidia and semiconductors rallied, and then in the second all the derivative plays, including those the energy plays. So part of what we saw was just an unwind.

There's also a whole new set of business models which are basically serving up excess capacity for big tech companies, the so-called Neo Clouds. They're basically setting up data center capacity to fulfil the excess demand.

But the problem is they have no real competitive advantage, they're just a commodity supplier.

So if there is a downturn, if CapEx does come off the boil, then it's the neoclouds that will be hit the most. And that includes data center players in Australia. It includes companies like CoreWeave in the United States. They're far more at risk than the big tech companies. The big tech companies will just end up using their own capacity.

It's almost a surprise that the demand was so great that they had to go to these external parties to deliver excess supply. If there's a downturn, those excess supply providers will be the first to go.

Andrew Geogeghan: So Michael, are you reassessing your investment strategy with exposure to some of these potential stocks then? Given, and I note, your outperformance in your index in your fund last year, where you had that significant exposure to perhaps AI-related tech stocks?

Michael Frazis: Yeah, look, semiconductors are extremely cyclical. So we're managing this risk with models. So we're kind of right on that borderline where we'll exit the whole position or keep it on. It's one of those situations. And I really think because semiconductors are so cyclical, you have to have a serious approach to risk to manage those downturns.

Because remember, it takes a long time to scale up this capacity and bring it back down. These purchase commitments can be done a year in advance. That supply can't just be switched off.

If there is over supply, you might start seeing some startups not get funding. They'll then have to sell their hardware on the market. You'll see secondhand prices for Nvidia chips collapse. This has happened two or three times in the last five years. And you'll see that in the secondhand prices, both for rentals to rent that capacity and also for the chips themselves. And so I'm sure it's quite likely we do enter one of those downturns in the next year or so.

Andrew Geogeghan: So you're not tempted to buy Nvidia after the pullback.

Michael Frazis: I don't think this is one where you just pile head in. I don't think it's clear. And I think it really depends what a few people decide.

There's only a handful of people making these decisions, right? It's Satya Nadella, Sundar Pichai. There's probably seven or eight people on the planet who'll decide whether to scale back their capex or to push forward.

Now they're all at the moment quite ideologically in favor of mass investment. They all think we're on the cusp of generalised AI. And the fact that a Chinese company has delivered such gains and such good model so efficiently, that doesn't invalidate the idea that you can apply more compute to what they’ve learned and get an even better response.

So success from here in the short term will probably depend on those CEOs and what they say and what they do and what they decide to on CapEx.

Andrew Geogeghan: Now Michael, aside from tech and AI, you've also had a focus in your fund on Australian biotech. Among them, Clarity Pharmaceuticals. We just interviewed them earlier this hour. What are you seeing in that space that's attracting you at the moment?

Michael Frazis: Look, Radiopharmaceuticals is really interesting. It also went through one of those cycles of mass investor attention.

It went from a space where nobody knew anything about it to a space where probably every Australian investor has looked at Telix and Clarity. And there's no doubt that Telix has had huge commercial success. That's an enormous company now based on revenues.

It does feel like there was a takeover wave last year that perhaps has now come off the boil. And it's a very competitive space in prostate cancer. There's multiple companies targeting that late stage cancer with different approaches.

But this is an interesting year for Clarity. We're about to get some good data on the latest cohort on their treatment. They're going head to head with Telix in diagnostics.

That should be recruited and we may see data from that this year. So it's an interesting one, but that company's also sold off quite significantly now.

Andrew Geogeghan: So you're mainly focused on Australian healthcare and biotech stocks?

Michael Frazis: Yeah, think it's surprising how much is going on at the moment. You've got EBR, which is on the cusp of FDA approval and will start rolling out their device. Orthocell has been on a tear. They're also expecting US approval and a rollout in the United States. They're already selling around the world. There's another company that's not widely known called Anteris, which just relisted on the NASDAQ and raised 88 million US. That's definitely one to watch as well. It was, it actually fell pretty significantly last year, but it was really that funding overhang after they announced they were going to relist or dual-list in the United States. There was always going be a huge capital raise as part of that. And obviously once you announce a big capital raise six months in advance, there's very little reason to own the shares. Often that actually becomes a good short. But once that capital raise occurs and the overhang disappears, then it's a different picture completely.

Andrew Geogeghan: And Michael, how are you feeling about the Australian market at the moment given its underperformance relative to Wall Street last year? What's your weighting at the moment, if you like?

Michael Frazis: We're pretty significantly invested in Australian healthcare, so probably 30 to 40 percent in Australia. In terms of the market, it's hard to tell. It does seem like the currency will continue to weaken. It does seem like the current government is doing most of the hiring. The public sector is doing significantly better than the private sector. I think that's pretty clear from the numbers. We'll just have to see.