If You Missed Nvidia, Then How About Xiao-I

This is not being entirely ridiculous - it's diving deeper than normal

So, yes, we're all aware that artificial intelligence, or AI, is the big new thing in computing. ChatGPT is going to put online writers like me out of business, and isn't that a relief for the rest of society?

Except, well, we do need to think about who is actually going to gain here. One company that obviously has in this immediate timespan is Nvidia Corporation (NVDA). Their recent results, stock price boom and forecast sales all were based on that selling shovels in a gold rush idea. Sure, Large Language Models (LLMs) are going to take over the world, but whose chips will power the LLMs?

And fair point, too. They're clearly going to make out like bandits at this point in the cycle. Just like they did in crypto and for the same reason, too. Which requires a little bit of thinking about the chip cycle.

The chip cycle

So, the first thing that happens is that people use a standard processor to do something. Pretty obvious, really. But anything that is math heavy doesn't work so well on a standard processor (look, I'm so old I actually recall the day someone came around and put an 8087 in my IBM XT). For math-heavy operations, we want what is called a GPU - graphics processing unit. The name comes from the fact that graphics is math heavy, lots of people want lovely graphics for game playing and so on. So, math-heavy chips have largely been developed for graphics uses - GPUs. Which is where Nvidia comes from, of course.

So, along comes another math-heavy application, crypto, and the rest of us couldn't buy a high end graphics - especially Nvidia - card for love nor money. They were all locked up in the mines along with the pit ponies.

We've actually all seen this happen these past few years, we all know this is true.

But we all also know that there's one more stage - ASICS, or Application-specific integrated circuits. We put the very specific math process we want to run into the chip itself. These chips are terrible at doing anything other than their specific and only task - but they're great at that.

It didn't take long for it to become impossible to gain money from mining crypto on standard processors, only GPUs would do. And what, 18 months maybe before GPUs only worked if you had really, really, cheap energy, ASICs were what you had to have to be serious?

AI and the chip cycle

It's not just that we could posit that the same cycle is going to happen with AI, it's that we know, absolutely, that it is. We can already see the second stage of it - that's exactly what is driving the Nvidia surge right at the moment, those GPUs. But the next stage is already starting. Not that I think Brainchip Holdings is going to win here, but we're already seeing people making ASICs to handle AIs. If we're not so keen on a specific example company, then some background theoretical stuff on using ASICs to drive AI.

A very simple and absolutely bound to be correct prediction is that give it 18 months to two years and we will be seeing ASICs for AI. The use of GPUs will decline.

Ah, but which AI?

Now with crypto it's not so difficult. Because we're trying to embed the particular equation to be solved into the silicon. Clearly, Bitcoin (BTC-USD) is where it's at, although there are specific ASICs for a few other coins out there. So, design the ASIC to be good at the bitcoin hash (I may get some of these buzzwords not quite right, but the meaning should still be clear enough) and we're good to go. New generations of these chips come out often enough and so on.

But AIs - or if we want, LLMs - aren't so simple. Not in the sense of the coding into silicon, that's not what I mean. But which AI should be coded into silicon? At present we've a pretty large number of them. We're going to have many more in the near future, too. Bard, ChatGPT - and so on and on. Which one gets put into the ASIC?

There's one more specific point here. Each and every AI - or LLM - is only good up to the point of the end of its training data. That's just a limitation we're all going to have to put up with. But that does also mean that we're all used to the idea of a cut-off point - which is fine for encapsulating into silicon.

OK, but which AI goes into the ASIC?

There's another way of thinking of this, and it's that the data doesn't go into silicon, only the subroutines. Which is fine - but they'll still differ over LLMs. So will ASICs. Which brings us back to the same problem. Which models make it into silicon? Because that's what is really going to determine which will be the dominant model.

So China

China is different, that we know. One of the ways it's different is in information spread. This is to be entirely agnostic about that, it's just an observation. One thing this does mean is that the varied Western LLMs - AIs - aren't going to work in China. Not just the language issue. Anything trained on Western data is going to have all of those Western assumptions in it. It's simply not possible to use an AI trained on Western data and meet the rules for use inside China. And if that's not true right now, then the authorities are going to get there real soon now.

AIs that work in China are going to have to be trained on data making the assumptions that China requires. That bifurcates the market. Which is where we start to think about Xiao-I Corporation (AIXI). They've already thought this through. From their recent PR missive about their recent launch:

Xiao-I's LLM can directly integrate with business systems, empowering enterprise-level software. It can also generate a controllable output, be rapidly deployed to a user system, and achieve systemic delivery to realize "controllable, customizable, and deliverable" goals.

OK, fairly boilerplate, but this:

The Model is "controllable" in terms of ideology, laws and regulations, computing power algorithms, cultural values, ethics, and morality. It ensures controllable data security at the national level and controllable content output at the enterprise level.

See what I mean?

Today, with the release of GPT, the global open-source community is working to reduce the cost of the model. This has led to many open-source achievements, making it easier for participants to access the model. However, the true success of a large model lies in building a new commercial ecosystem. Merely possessing the model or showcasing demos is meaningless. Xiao-I's approach emphasizes the importance of controllability, customizability, and deliverability. Without these, the model 39's entry into the commercial market holds no value in China, and also why Chat GPT got banned in China so soon. These principles separate the few winners from the majority in this competitive landscape. Just like Prime Number Capital analyst Hao said "XIAO-I 39's specialization in Mandarin NLP gives it a competitive advantage in serving Mandarin-speaking users. Furthermore, the strict regulatory environment for foreign tech companies operating in China creates barriers to entry for competitors, providing XIAO-I with a favorable market position. Not only is Xiao-I positioning itself ahead of the AI trend, but the company's financial metrics are pretty nice - 48% top line growth last year.

My view

This is less an insistence upon the one specific investment and more an insistence on these are the things that need to be thought about in the structure of the market. The Chinese experience is going to be different from that outside that country. The particulars of the marketplace and how LLMs are constructed makes that simply true.

The other is that there's going to come a time when AIs get coded down into ASICs. This is going to depend upon who has the dominant model at the time that that encoding starts. BTW, do note that ASICs are easier than truly leading edge processing chips. The effort is put in at the design stage, not at the fab stage. So it's not necessary to have leading edge 7 or 3 nm fabs to make ASICs work.

The investor view

Everything is always hazy at this early stage of a technological revolution. Pretty much by definition consumers are going to make out like bandits from this AI one. But which of the producers? There are good arguments that Xiao-I will do well within the Chinese market because of the social characteristics of that market. Further than that is difficult to go at this stage.

How is the AI market going to work out?