Why Google Will Win the AI Arms Race & OpenAI Will Not

So at the very bottom of things, ZML is a framework that runs any models on any hardware. We sit ultimately at the infrastructure layer. We enable anybody to run their model better, faster, more reliably, but on any compute whatsoever. Doesn’t really matter. It could be NVIDIA, it can be AMD, it could be TPU and whatnot. And we do all that without compromise. That’s the key point because if there’s a compromise, then it’s not really, you know, agnostic. Right?

Yes. You you actually can see it. It’s been happening for a while. Models now are not the right abstractions, at least. If you look at closed source model, they’re not really models. They’re more like back end. And there are a lot of tricks that you feel like you’re talking to one model, but ultimately, you’re talking to a constellation and assembly of back ends that produces, you know, a response. Probably the number one, you know, I would say obvious thing would be that if you ask a model to generate an image, then it will, you know, switch to a diffusion model, right, not an LLM.

And there’s many, many more tricks. The Turbo models and OpenAI do that. There’s a lot of tricks. So definitely, models in the sense of getting, you know, weights and running them is something that is ultimately going away because, you know, in favor of, like, full blown back ends. Right? You feel like you’re talking to a model, but ultimately, you’re talking to an API. The thing is that API will be running locally in your own, you know, cloud, you know, instances and so on.

No. Absolutely. You can get order like, probably an order of magnitude more efficiency depending on the hardware you run on. That is substantial. Not a lot of people have that problem at the moment. Things are getting built as we speak. But, you know, a simple example is if you, you know, switch from NVIDIA to AMD, on a seven TB model, you can get four times better efficiency in terms of spend. Right? So that is substantial. That is very much substantial. Now the problem is getting some AMD GPUs.

Right? I’m really

So there’s a few reason. The probably the most important one is the PyTorch CUDA, I would say Duo, and that’s very, very hard to break. These two are very much intertwined. Can you just explain to us what what PyTorch and CUDA are? Oh, yes. Absolutely. Yeah. Yeah. PyTorch is the ML framework that people use to build actually trained models. Right? You can do inference with it, but by far the most successful framework for training is is PyTorch, and PyTorch was very much built on top of CUDA, which is NVIDIA software.

Right? Let’s just say the strings of PyTorch make it ultimately very, very bound to CUDA. So of course, it runs on, you know, it runs on AMD, it runs on, you know, even Apple and so on. But there was always, you know, the tens of little details that not exactly run like, you know, you would expect, and there’s work involved. But then also there’s supply. So probably that’s the number one thing. The second thing is there’s a lot of GPUs on the market. Pretty much all of them are are NVIDIA.

The reason being that if you think, you know, in layers and you say, alright, I’m going to buy, let’s say, GPUs, and I’m going to sell them to folks to maybe not even do training. Right? Just do inference. Then most likely, if you look at it that way, you’ll end up buying NVIDIA because everybody will want to run on NVIDIA because nobody knows really how to do whatever, and they’ve trained on NVIDIA. So they’re like, I can just reuse my code and so on. So there’s like this self perpetuating circle of people just buy NVIDIA because they want to resell, and people just use NVIDIA because it’s there.

Right? But it’s by far not the not the most efficient platform. And arguably, even in terms of software, it’s not the best software platform. So that is probably two of the most I would I would wager the most important reasons. Can I ask before,

Because the chips are there. There’s a lot of things. But in in my opinion, there’s going to be a need for inference. Very hard to say whether it will be worth, you know, everybody’s money to do it on h 100. That is a a bubble that I think will blow sometime. I’m kind of afraid of that, be honest. Why do you think

Because it was built on the a 100, I would say, financial model, which was a generation zero, we do training. But when it’s last generation, we do inference, and it worked beautifully. Right? For a 100, then h 100 comes along, and inference is it’s worth five times the price, and it maybe runs twice in terms of performance. On inference, that is. On training, it’s a lot better. But on inference, it’s like maybe twice as fast. When it actually, when it came out, it rained at the same speed than the a 100.

So there’s a money gap that’s going to have to, you know, be bridged sometime. Right? And the the part that worries me is that I see, you know, amortization plans on on, like, you know, six, seven years, right, with the GPUs at the collateral. And I’m like, well, I’m not sure how it’s going to work because at least when they came out, they were were five times the price, and they’re just two times, you know, faster. Something has got to give.

Not much. Ultimately, the two things that could really much very much shake the industry, the chip industry, in my opinion, is our agents and reasoning.

I think this is where NVIDIA can be can be attacked. I mean, why agents and why reasoning? The the difference is for agents and reasoning, you need to wait until the end of the request to get whatever it is you came for. You don’t really care about the speed at which the text, you know, outputs, which is what you want in a chat. Right? You only care about how much time does it take between the beginning of my request and the end. And so that fundamentally changes the incentives from throughput bound to latency bound.

And so GPUs, let’s say you’re running a GPUs at, let’s say, 10,000 tokens per second, you very much like to do it, you know, a 100 times 100. Right? And they can do that. But they cannot do they cannot give you 10,000 tokens per second only on you, per stream, what we say. But in terms of, you know, agents are reasoning, this is exactly what you want because you don’t wanna wait, like, you know, fifty seconds for whatever thinking. Right? And agents, it’s it’s the same.

So these two, I think, are the shot that might make NVIDIA change its course with respect to chips. I mean, they’re not idiots. Right? How should agents change

Hard to say because NVIDIA is has a very, very vertical approach. They do more of more. Right? Like, if you look at Blackwell, it’s actually crazy the what they did for Blackwell. They assembled two chips, but the surface was so big that the chips started to to bend a bit, which further perpetuated the problem because it then didn’t make contact with the heat sink and so on. So they are very much in the power envelope. They push it to a thousand watts, it requires liquid cooling and so on.

So they are very much in a very vertical foot to the pedal in terms of GPU scaling. But the thing is GPUs are, you know, are a good trick for AI, but they’re not built for AI. It’s not a specialized chip. It is a specialization of a GPU, but it is not, you know, an AI, you know, chip.

the way it worked is that you can think of a screen as a matrix. And if you have to render, you know, pixels on a screen, there’s a lot of pixels, and everything has to happen in parallel, right, so that you don’t waste time. Turns out, you know, matrices are very are very important thing in AI. So there was this cool trick in which we essentially tricked the GPU into back that was, like, probably twenty years ago, we would trick the GPU into believing it was doing graphics rendering where actually we would making it do parallel work.

Right? It was called GP GPU at the time. Right? So it was always a cool trick, but it was not dedicated for this. The pioneers probably were, of course, Google with TPU, which are very much more advanced on the architectural level. But essentially the way they work, it kind of works for AI, but for LLM that starts to, you know, to crack because they’re so big and there’s a lot of memory transfers and so on. Actually, that’s why Grok achieves, not Grok, but Log Grok, Cerebras and all these folks, they achieve very high performance single stream is because the data is right in the chip.

They don’t have to get it from memory, which is slow, which GPU has to do. So there’s a lot of these things that ultimately make it a good trick, but not, I would say, dedicated solution per se. That said, though, the reason probably NVIDIA won, at least in the training space, is because of Mellanox, right? Not because of the raw compute. Because you need to run, you know, lots of these GPUs in parallel. So the interconnect between them is ultimately what matters. Right? So how fast can they exchange data?

Because remember, when you do a matrix multiplication, let’s say, you read the the matrix is read like hundreds of times during the multiplication. So there’s a lot of transfers going on. And so far, Mellanox with, you know, InfiniBand had the the best technology. So that’s why, you know, a lot of people and when you do training, by the way, it is the name of the game, the interconnect. When you do inference, not so much. You don’t care when you do inference.

I would divide it in two categories. Well, three categories. The GPUs you can buy or rent, the TPUs you can rent, and the TPUs you can buy. This is how the market is structured today. Right? Right now, if you are you wanna go dedicated, there’s at least in the cloud, there’s two options. It’s TPUs and Trainium. TPUs on Google, Trainium on Amazon. So these are, you know, available chips. You can rent them today. If you wanna buy GPUs or rent GPUs, you know, they’re GPUs. We we we know it all the time.

And there’s this new wave of of computing, which are dedicated, you know, chips you can actually buy. The tensed torrent, the etched, the viscera. So I think it will be a mix of, you know for instance, let’s say you are in Google Cloud, of course, you don’t wanna do NVIDIA. You get ripped off. Here’s the dirty secret, is that NVIDIA like, a TSMC sells you at 60% margin. NVIDIA sells you at, you know, 90% margin. And on top of that, there’s Amazon that takes, let’s say, a 30 margin.

So you are a very thin crust on a very big cake. It’s a bit of a losing game if you, you know, go all in on one provider. You want optionality. With

Absolutely. Yes. Yeah. Yeah. Yeah. Here’s the problem, though. Let’s say you are on Google Cloud and you run TPUs. Suddenly, you just remove that 90% chunk on you know, on the spend. The problem is is that for multiple software reasons, are, you know, which we are solving at ZML is that they’re not really, I would say, a commercial success. They are very much successful inside of Google, but not much outside of Google. Amazon, same, is pushing very, very hard for their, you know, training chips. So the future I see is that you use whatever, you know, your provider has because you don’t wanna pay, you know, 90% outrageous margin and try to make, you know, a profit out of that.

Okay. So when we move to

So these two obey fundamentally different, I would say, tectonic forces. So in training, more is better. You want more of everything, essentially. And the recipe for success is the speed of iteration. You change stuff, you see how it works, and you do it again. Hopefully, it converges. And it’s like, you know, changing the wheel of a of a moving car, so to speak. So that is training. On inference, this is a complete reverse. Less is better. You want less headaches. You don’t wanna be woken up at night because inference is production.

You could say that training is research and inference is production, and it’s fundamentally different. In terms of infra, probably the number one thing that is the number one difference between these two is the need for interconnect. So if you do, you know, production, you if you can avoid to have interconnect between, you know, let’s say a cluster of GPUs, of course, will not go, you will, you know, avoid that, right, if you can. And this is why models have the sizes they have, is so that people can run them without the need to connect multiple machines together.

It’s very constraining in terms of the environment. So that is probably the fundamental difference, the need for interconnect. And number two is, ultimately, do you really care about what your model is running on as long as it’s outputting whatever you want it to output?

Think of it like doing a painting and doing a million paintings. The tools you will use, the process you will do. If you do one painting, what you favor is the speed at which you can do a stroke and do some iteration. If you do a million, what you want is a process, a process that is reliable, that can deliver you efficiently a million paintings. So that is the same for for training versus inference. If you run around, you know, millions of instances of a model, you cannot, you know, hack your way to do that.

By the way, people do hack their way today, but this is probably the fundamental difference.

There’s a lot of duct tape. Here’s also probably one of the problem is that training on first principle is actually two passes, forward and backward. Right? It’s called forward pass and backward pass. Right? Inference is running only the forward pass. So that’s how things are today. There are people who are trying to specialize a bit because, you know, at some point, duct tape doesn’t really work out. And when you’re on big scales, that makes a problem. And it’s a problem that’s growing because a lot of people are coming on the market with the need for inference.

That wasn’t the case, you know, a year and a half ago or a year ago. OpenAI had this problem. Right? Maybe Anthropic had this problem. But it wasn’t a universal problem yet, and now it’s becoming a universal problem.

for instance, probably the number one thing, depending on how you deploy, but if you’re deploying inference, the number one thing that will get you is what’s called auto scaling. So as your, you know, systems get more and more loaded, you want to provision because, you know, these things are tremendously expensive. You want to provision them as you scale. Right? So you wanna say, I have a thousand GPUs, you know, twenty four hours, even, like, if there’s, like, nobody under production, I will pay for them, which is mind you what people are doing today.

This is crazy. So what you want to do is you want to, you know, provision compute as you grow your needs. Right? And you want to do it up and you want to do down. Probably the number one thing that, you know, gives you a lot of efficiency in terms of spend. Like, we’re talking, you know, multiples, like, you know, five, you know, sometimes 10 x, you know, improvement. The thing is, this is a problem, at least in that in, I would say, back regular back engineering, this is a problem everybody knows.

Right? Everybody is doing it because the the the savings are so huge. But on AI, nobody really had the problem. So now they’re coming up to it. So this is the one example.

That’s one example. Yeah. Another one is choosing the right compute. It’s like kind of, I would say, a vicious circle because provisioning compute is very hard. So if you lose compute, it’s very bad. You are essentially incentivized to overbuy in the case of, you know, Amazon or Google that would be buying reserved compute, which you’re not gonna use because if you buy it on demand, you will get tremendously ripped off. So that creates this, like, face scarcity of compute because that people buy preemptively because they rate shit ton of money, and they’re not using it.

Alright? So this is a major problem too.

It might well be. Yes. We are being spared a bit because Black Whale is late and others are getting canceled. And so h series, I would say, are still, you know, in the active. But, yes, absolutely. But, you know, what choice do you have? This is the thing.

So I might tell you that I think it already started. I’m getting cold emails for, you know, discounts, you know, from services I never heard about. And I started getting these emails probably around October, November. Some people are left with a lot of CapEx that they don’t know what to do with. You know, it’s a different thing to to build a cluster and run a training and do a training run than it is to build literally a cloud, you know, provider or hyperscaler or whatever you wanna call it.

So there are a lot of people who do their training runs on the regular providers, but then move to regular hyperscaler when they do production. So I very much worry there will be an oversupply of these chips. The problem is that, you know, remember, the chips are the collateral. So, you know, somewhere in The US or whatever, there’s going to be a data center with like a thousand GPUs that people may buy, you know, 30¢ on the dollar. You know? This is what might happen.

Technically speaking, he is right. But realistically speaking, I’m not sure I agree. The thing is these chips are on the market. They’re here. Altab, on Chrome and get one. That is something that I don’t take lightly. Availability, that is. Right? I think NVIDIA is here to stay, at least if not for the H100, you know, bubble bust, because these chips are going to be on the market, and people will buy them and do inference with them. Remains to see, you know, the the OpEx and the electricity, etc.

But the thing is, the only chips that are really, you know, frontier on that sense are probably TPUs and then the upcoming chips. But the thing is, they’re great chips, but they’re not on the market. Or like there are outrageous prices, like millions of dollars to run a model. So what chips are great and why aren’t they on the market? Let’s say for instance Cerebras. Incredible technology. Incredibly expensive. So how will the market value the premium of having single stream very high tokens per second? There is a value into that.

Right? As we saw with Mistral and Perplexity. But I think that was done at a loss. I don’t know. I don’t have the details, but I think it was done at a loss that Cerebras, you know, put it out. So today, there’s three actors on the market that can, you know, deliver this. I think this will be, I would I would say, the the pushing force for change in the inference landscape, agents, and reasoning. So that is, you know, very high tokens per second only for you.

What

So there’s this trick. Because here’s the thing. There’s no magic. This little trick is called SRAM. SRAM is memory on the chip directly. So that is very, very fast memory. But here’s the problem with SRAM, is that SRAM consumes, you know, surface on the chip, which makes it a bigger chip, which is very hard in terms of yield, right? Because the chances of like problems are higher and so on. So SRAM is, I would say, very, very, very fast memory, which gives you a lot of advantage when you do very, very high inference, but it’s terribly expensive.

And if you look at, for instance, Grok, they have on their generation this generation, they have two thirty megabytes of SRAM per chip. A 70B model is 140 gigabytes. So you do the math. Right? Cerebras has 44 gigabytes of SRAM into what they’re called their wafer scale engine, which is a chip the size of a wafer. I mean, most likely, it’s interconnected, but it’s huge. Right? And it’s it has to be water cooled. They have a copper, you know, I would say needles that touch the chip too.

It’s crazy stuff. Very, very impressive technology by mind you, but very, very expensive. So my bet is I think there will be, you know, chips on the market that do that at at much lower price. And there’s two companies I see going in that direction. One is called etched, and the other one is called vSor. That’s the two I see. Because if you can deliver this at the, I would say, the price that is comparable to GPUs,

It’s hard to say. I mean, need some SRAM, but if you can, you know, have a smaller process node, but if you can hook yourself with external memory, then yes, you you can do that a lot better. But the thing is, if you go, like, full blown Islam, then, you know, there’s no magic. You will have to pay the price. I’m

Pushed by reasoning. So reasoning not in the sense that you see on deep sea and whatever. Right? Reasoning and what’s called latent space reasoning. Latent space reasoning and agents will push the market towards different types of compute. Can

So the way models reason today is the reason it was in in tokens. So it’s as if if you think to yourself, you would, you know, say out loud what you’re thinking. So yes, it works, but it’s a bit, you know, inefficient. Right? And you lose information doing this. The latent space reasoning is this without going, I would say, to English or whatever. Right? So staying, you know, in what’s called the latent space, which is where all the information of an LLM, let’s say, an LLM an LLM lives.

Right? So this is very much how we how we, you know, work as humans, and we move toward what Yann LeCun calls energy based model in which we have different types of longer or shorter, I would say, thinking times, if you will. Right? So that fundamentally GPUs cannot deliver deliver this. Plain and simple. At scale. Why can’t GPUs deliver it? Because the access to external memory prevents it. So HBM is all the rage. Right? But HBM compared to SRAM is absolutely, you know, dark slow. So this is the problem you get.

So HBM is like the best we can do, but it’s still slow versus versus SRAM.

No. You you want HBM, to to be clear. No. SRAM, this this this will not deliver. It’s a dead end in terms of scaling SRAM means scaling the surface, mean you get, you know, depreciating problems. It explodes everywhere. Right? So you need some SRAM. Right? So, you know, we’ll have bigger amounts of SRAM in into chips, and, of course, bigger what’s called external memory into chips. The issue with HBM is that it’s still slow, and yes, maybe NVIDIA has a stronghold and they can prevent you from getting some.

So that would be like I call it the the Nutella situation, in which, you know, Nutella, they owns 80% of the hazelnuts market. Right? So yes, you can do a competitor, but who will you buy the nuts from, right? So there will be a need for HBM, there will be a need for SRAM, I would say better, more dedicated architecture will be able to deliver these things. And then there’s like the next frontier after that, which is called compute in memory. There’s two companies that are on that market.

One is called RAIN, RAIN dot AI. Sam Altman is one of the investor. There’s no surprise. The other one is called Fractile. So this is the next frontier. And the idea is that instead of like transferring the data between external memory and the CPU and do the compute there, you actually, you know, bring the CPU to the memory and you do everything. It’s crazy stuff, but it’s coming. How

Maybe not this year,

It means you get maybe not SRAM level performance, but you get very a lot faster performance in terms of compute. And if you translate that to LLMs, let’s say, you get much much higher tokens per second. In a single stream, which is exactly what you want when you go into reasoning. You want your model to maybe think, let’s say, for, like, half a second, and then boom. You don’t wanna wait fifty seconds and, you know, context switch to some other thing, which is the problem everybody has today, mind you.

So, yeah, I think inference will be pushed. The compute landscape will be pushed to change because of these two constraints.

In five years, I would say 95% inference, 5% training.

Depends on the supply. I think that there’s a shot that they don’t. Because here’s the thing, you know, even if we take, you know, same amount of, you know let’s imagine we get we get we have a new chip from Amazon. Right? That is the same amount oh, wait. We do. It’s called Trainium. You know, why would I pay 90% margin of NVIDIA if I can freely change to Trainium? My old production is run on runs on AWS anyways. Like, if you run on the cloud and you’re running on NVIDIA, you’re getting, you know, squeezed out of your money.

Right? So if you’re on production on dedicated chips, of course, you know so maybe through commoditization that, you know, hey, I’m on AWS. I can just click and boom. It runs on AWS’s chips. Who cares? Right? I just run my model like I did, you know, two minutes ago.

They are. They have a product called NIM. Sort of does that. The thing with NVIDIA is that they spend a lot of energy making you care about stuff you shouldn’t care about. And they were very successful. Like, who gives a shit about CUDA? I’m sorry, but I don’t wanna care about that. Right? I wanna do my stuff, and NVIDIA got me into saying, hey, you should care about this because there’s nothing else on the market. Well, that’s not true. But ultimately, this is the GPU I have in my machine, so, you know, off I go.

If tomorrow that changes, why would I pay 90% margin on my compute? That’s insane. This is why I believe it ultimately goes through the software. Because the software, like, if my entry this is my entry point to the ecosystem. So if the software abstracts away those idiosyncrasies, as they do on CPUs, right? Then the providers will compete on specs, and not on fake modes or circumstantial modes. Right? So this is where I think, you know, the market is going. And, of course, there’s there’s the availability availability problem.

There is, you know, if you, you know, piss off Jensen, you might need to kiss the ring, you know, to get back in line. Right? But, I mean, ultimately, this is I don’t see this as being sustainable. When we chatted

So all, I would say, chipmakers have a GTM problem. All of them. Whether, you know, it’s Google, whether it’s AMD, whether it’s Stenstor. The problem is is that there’s, I would say, probably two fundamental problems. The number one is if you maintaining multiple stacks today is very, very, very hard. So you don’t. So let’s say I buy, you know, AMD. I want to buy AMD. Right? That means I’m going to abandon NVIDIA. Oh, crap. You know, I have a six year amortization plan on that. Oh, man.

What do I do? So do I need to support both stacks? Unclear. Maybe until AMD tells me, hey, you know, you have, I don’t know, let’s say, a thousand NVIDIA GPUs, you’re about to buy a 100,000 of AMD. I mean, come on. Right? And I’m like, okay, that is, you know, makes it worth my while. Right? But that is ultimately the fundamental problem is that the steps are very high. Right? I need to have a lot of incentives to buy into that ecosystem. So I need to buy a lot of them.

So if you’re AMD, that is already a problem. But then Microsoft comes along and buys it all, makes, by the way, OpenAI, or at least on the inference side, puts OpenAI in the green because of the efficiency gains. Can I just try and understand, sir?

It’s actually both. So the buy in is very high. So to make it worth it, you have to buy a lot. And if you buy a lot, this is, you know, what we talk to all of them. They always have the same questions, and it’s completely understandable. They say, this is great, but who’s the customer? Because on the other side, let’s take Amazon, for instance, with Raynium. Apple just came and said, hey, we’re going to buy a 100,000 of them. So you want to buy, you know, 10,000, you feel like the big shot, right?

Yeah, but go back to the queue because there’s Apple before you, right? So they have to have very high commitments. You you cannot be incrementally better. It’s very hard. Right? And also very hard, I can give you I can give you one metric if you want. I know for a fact that being seven times better and whatever take whatever metric you want, whether it’s spend, whether it’s whatever, it’s not enough to get people to switch. People will choose nothing over something. So this is a very hard market to enter into because you cannot also compete of incremental gains.

It’s very hard. Right? So you have to convince with other people. Maybe you can go the Middle East route in which, you know, they sprinkle everything and they, you know, evaluate everything. That’s not, you know, very sustainable, I would say, strategy in the long term, at least in the midterm.

multiple. Absolutely. The right approach to me is making the buying zero. If the buying is zero, you don’t worry about this. You just buy whatever is best today. How do you do that by renting? Oh, because this is what we do. This is our promise. Our thesis is that if the buying is zero, you know, you completely unlock that value. Because you’re free

It means that you can freely switch, you know, compute to compute, like, freely. Right? You just say, hey. Now it’s AMD. Boom. It runs. You just say, oh, it’s 10 storey. Boom. It runs. Right? How do you do that then?

Oh, yeah. Yeah. Yeah. Not agreements, but, like, we we work with them to support their their chips. But the the thing is my at least as, you know, I would say a user myself of, you know, our tech is that if it’s free for me to switch or to choose whichever provider I want in terms of compute, right, AMD, NVIDIA, whatever, then I can take whatever is best today, and I can take whatever is best tomorrow, and I can run both. I can run three different platforms at the same time.

I don’t care. I only run, you know, what is good at the moment. And that unlocks to me a very cool thing, which is incremental, you know, improvement. If you are 30% better, I’ll switch to you.

is actually a great question. I think that if you are doing it bottom up, infra to applications, you will lose because nobody will care, as they don’t today. Right? If you look at TPUs, they’re available. They’re great. Nobody cares. Why does nobody care about TPUs? Sorry. Because the cost of buying. It’s always the same. Right? You have to spend six months of engineering to switch to TPUs. And mind you, TPUs do training. They are the only ones with Trainium now. But AMD can do training, but it’s also but in terms of maturity, the by far the most mature software and compute is TPUs, and then it’s NVIDIA.

Right? So the buy in is so high that people are like, we’ll see. Right? I’m not on Google Cloud. I have to, you know, sign up. Oh my god. Right? So these are tremendous chips. These are tremendous assets. Now in terms of the risk, I think if you want to do it, you have to do it top to bottom. You have to start with whatever it is you’re going to build and then permeate downwards into the infrastructure. Take for example, Microsoft with OpenAI. They just bought all of AMD supply and they run, you know, ChatGPT on it.

That’s it. And that puts them in the green. That’s actually what make them, you know, profitable on inference. So or at least, let’s say, not lose money. Right?

Because I can give you, like, actual numbers. If you run eight h 100, you can put two seventy b models on them because of the RAM. Right? That’s number one. Number two is if you go from one GPU to two, you don’t get twice the performance. Maybe you get 10% better performance. Yeah. That’s the dirty secret nobody talks about. I’m talking inference. Right? So you go from, let’s say, a 100 to a 110 by doubling the amount of GPUs. That is insane. So you’d rather have two by one than one by two.

Right? So with one machine of h eight h 100, you can run two seventy b’s model if you do, you know, four GPUs and four GPUs. Right? That’s number one. If you run on AMD, well, there’s enough memory inside the GPU to run one model per card. So you get, you know, eight GPUs, eight times the throughput. While on the other hand, you get eight GPUs, two to maybe two and a half times the throughput. So that is, you know, a Forex right there. Just, you know, by virtue of this of this.

So that is, you know, the the the compute part. But if you look at all of these things, there are tremendous amount of, you know, we talk to companies who have chips are coming with almost 300 gigabytes of memory on it. Right? So that is, you know, a model like one chip per model. This is the best thing you want if you run seven TBs. Right? So which is what I would say, not the state of the art, but this is the the regular stuff people will use for serving.

So if you look, you know, top to bottom, and you know what you’re going to build with them, then it’s a lot better to do the efficiency gains because four times is a big deal. Right? And mind you, these chips are 30% cheaper than NVIDIA’s. It’s like a no brainer. But if you go brought them up and say, I’m gonna rent them out, people will not rent them. Simple. So that’s why, you know, I think it’s it’s a good way to attack it from the software because ultimately, do you really care about that your MacBook, let’s say, is an m two or m three?

It’s like, oh, it’s the better one. Alright? And that’s it. Right? And imagine if you had to care about these things. That would be insane. When I listen to you

Yeah. I used to think the market was efficient. So probably I would go today at least I would go with NVIDIA still because the supply. But, you know, if we play our cards right, we ship our stuff. Hopefully, I will come back and tell you to buy AMD as much as you can or attend Storrent, you know, if they go public or whoever else. These chips are amazing, by the way.

Probably not a lot of people are accustomed to what it entails to run production. So that inference is production, and production is hard. Somebody has to wake up at night. And I used to be that guy, right? I don’t want to do it again. So production is hard. Thankfully, we have a lot of software nowadays to do that a lot better, but there’s not a lot of reuse because the AI field, at least, is not really accustomed to that yet. It’s changing, but, you know, the discussions I had a year ago and the discussions I had today are not the same.

They’re going to the right direction, but they’re not there exactly yet. So probably that would be the number one thing. That is only, you know, training code running only forward pass. Right? This is not what it is.

I mean, they’re still going after training. So there’s still this frontier. Probably it’s why also NVIDIA is the better buy right now. Because on the NVIDIA side, if you do training, it’s incremental. If you have bought a thousand NVIDIA GPUs and you buy 1,000 new NVIDIA GPUs, that gives you 2,000 GPUs. Right? But if you buy 1,001 AMD, that gives you twice 1,000. Right? It’s a bit different. So they’re still going after training, definitely, and they’re very pragmatic in doing so. But, I mean, they have the CapEx to spend.

They’re not making their money out of it, probably. The only one, by the way, that owns their compute are Google. There’s like this triangle of, I would say, of wind that I this is my mental model, mind you. You have the products, the data, and the compute. Who has all three? And you get everything flows from there. Product data compute. Who has all three? Google? Amazon? Amazon, they don’t have products. They have Amazon. Right? They have AWS, but they don’t have actual products. Google has, like, you know, Android, Google Docs, whatever they have, everything they can sprinkle everywhere.

This is the sleeping giant in my mind. If they’re not busy doing a reorg,

I mean, OpenAI is amazing, but it’s not their compute. It is Microsoft’s compute. And if you own your compute, you own your margin is essentially what you’re saying. Yeah. Microsoft even Microsoft, they bought when they weren’t running NVIDIA, they bought NVIDIA at, you know, some outrageous margin. I talked to a lot of people that build data centers, and I tell them, you know, they’re mind you, these people, like, buy tens of thousands of GPUs. And I asked them, hey. Do you get at least a discount or something?

And they’re like, no. The only thing we get is the supply. So, I mean, ultimately, if you don’t own your compute, you’re starting with, you know, something at your ankle. Definitely. And so this is why I like to think in this, like, this triangle, product data compute. And you can see where everybody sits and their weaknesses and their strengths.

There’s like a brute force approach to this. It is a very American approach, more and more and more. But the thing is, you look at, for instance, the the xAI cluster. It’s not a 100,000 GPUs. It is four times 25,000. You’re starting, you know, see some because InfiniBand and in the case of Rocky, which is, anyways, the technology they used to bridge their GPUs together, you have upper bound. Right? At at some point, you’re fighting physics. So you can push it’s like, you know, trying to get to the speed of light.

As you approach it, the the amount of energy you need is a lot higher and a lot higher, and it grows and grows. So there’s two, I would say, counter to that would be that number one is we still scale, but there’s a lot of waste and excess, you know, spending on the engineering side, which is the deep seek approach. Right? Very successful at that, mind you. They said, yeah, if we do this and this differently, then we get, you know, multiples sometimes. Right? So virtually, you increase your compute capacity because you’re more efficient.

And the other approach is Yann LeCun’s approach, which is this is not scaling, and at some point, you need we need to look the problem in the face and do something better. Right? So, of course, we push and push and push because there’s capital still, but I’m more of the of these two approaches. I think you can do more with less. At

I think until somebody does it, deep seek was a good wake up call. Right? Suddenly, efficiency is in. That’s number one. And number two is until there’s a new architecture that comes out and changes the game. So in the case of LLM’s, for instance, you have these what’s called non transformer models that changes fundamentally the compute requirements. So that might be a frontier that completely obsoletes the transformers. And if the transformers are the, I would say, the building block by which current model works. Right? So the way they work is that for each token or syllable, if you will, the model will look at everything behind it.

So you can see that as you add more text, you have more work to do. So there are these new architectures that do not require this, that might change, you know, these things and probably shift the amount of compute needed to do training or to do inference. And then there’s the new thing, is Yann’s thesis, which is the word model. As in LLM’s are at that end, what we need is something that understands the world fundamentally, and this is it’s JEPAS thesis, it’s called. I’m very bullish on this, but it’s it’s very frontier.

Why are you bullish on it? And why is it so frontier? Because it’s Yann LeCun. It’s hard to he’s no bullshit. Right? So he explained to me how it worked, I was blown away. But it makes a lot of sense. We are creeped out because the machine talks back to us. But it’s not a new thing. Right? It used to, you know, this is not new technology when it came out. Well, like, when it when it exploded, it wasn’t new technology. But suddenly it was talking back, and that freaked us out.

And we got crazy on it. Right? But language is one form of communication, but it is ultimately a very narrow window into the world. We use it to describe the world arguably with some loss. Right? And so the J PAL approach is, long story short, is that you have essentially two things you want to do, and you try and minimize the energy to do them. And from this understanding emerges, physics emerges and etcetera, because you’re trying to minimize the amount of energy to go from one state to the other.

And that actually makes sense. Like, if you try and, you know, pick this AirPod, you know, case, I’m not gonna go round trip around the block to get it. Right? I just get it. And in my brain, it’s wired to just do the the thing. If I go and, you know, talk to myself out loud, put the the hand down, move to the left and whatever, that feels very, you know, inefficient. So probably this will be something that that changes.

And in the case of LLMs, there’s good work also on what’s called diffusion based LLMs, which means like instead of thinking, you know, what’s called auto aggressively, that means you get a new token, you reinject and you redo, etcetera, they think more like what we do, which is in patches. Right? Imagine a paragraph of text and words appear until it’s done.

I think it’s fair game, to be honest. I will not shed a tear. It’s fair game. If you there were, like, some people who tried to ask I think it was I I remember if it was an OpenAI model. So a diffusion model image. Right? They asked it to generate an image from a Star Wars movie at whatever timestamp. And it came out with the Star Wars movie, you know, screenshot. Obviously, it was trained with it. I think it’s fair game because there’s no free lunch.

Right? It was trained with data. You had a good ride. Somebody was was sneaky and and took it, but you took it from the beginning too. So let’s just accept it’s fair game. And you and you

Absolutely. Absolutely. I, you know, take my my cup. I enjoy it very much that movie every single day. You

I’m a bit split on this. There’s a part of me that said that if you reinject data into the system, the system deteriorates. That feels a bit, I would say intuitive. But if you look at AlphaGo, for instance, the moment it’s, you know, ramped up in its skills is when they started generating games, or synthetic games. Right? So I’m a bit, you know, split, but there are some verticals that very much benefit from this code LLM’s, for instance. We can run code. Right? So this is the poolside thesis.

Because you don’t use the mod the AI model to generate output. You use the machine. You use the you you just run the code. Right? And you see what it makes, and you run all this code, and you create data out of it. Whereas if you’re on an LLM and you say to an LLM, alright, generate me 2,000,000,000,000 tokens of text, it will do it with its you know, so you may inject and stuff. So there’s a lot of tricks, but ultimately, my guts tell me that it feels wrong, right, because you reinject data that was there.

And so it will deteriorate. There’s, you know, there’s loss. So, yeah, I’m a bit bullish. I’m not sure exactly on what vertical. Code is one. We’ll see. Distillation is, in some sense, a bit like that. You create synthetic data from a bigger model into a smaller one. Probably the most, I would say, mind blowing thing about distillation is that sometimes the smaller models become better than the the than the bigger model through distillation. So Smaller

The one theory is that the smaller model is better at generating output that you would want it to generate, essentially. It’s not better in the general sense. It’s better at the task at which you were measuring it. This is what it learned to imitate.

Some sometimes it’s wasteful to run big models. A lot of times it’s actually wasteful to run big models. I think there’s going to be a lot of of smaller models for efficiency reasons, but there’s a there’s a but, which is you talk to people at DeepMind, and they don’t even fine tune anymore. Because they have such, you know, what’s called big context window, which is what the model, you know, the data, the model you inject, right, at runtime that nowadays they just dump data into it and just say, do whatever, you know, that data tells you to do instead of fine tuning as we used to do.

So if the efficiency gains we’re not there yet. Right? But if their efficiency gains, I would say, pass the threshold, we’ll just do it at runtime. We’ll just have a a great model that will just specialize at each request. But that’s not for tomorrow, I think.

It’s a very, very clever trick. What you do is you represent knowledge into what’s called the vector space or latent space. And what you do is through what’s called vector search. So imagine you have, let’s say, a three d space that represents all knowledge, all of everything. And let’s say a cat sits here, a dog sits close because it’s an animal, but it’s far from some other property and so on. So what you do is you run the user’s request through this same system, it’s called an embedding, And that will give you a vector and you will take whatever is closer to you, what’s called semantically close.

And then it’s actually very clever. You actually insert those pieces of text before the request. So it’s as if you would say, knowing the following, and you give the data, let’s say it’s law or whatever, please answer my request. And that’s it. So that’s a bit of a clever trick. It’s a bit dirty because, of course, you know, you are limited by the amount of data you can input. Right? So there’s this problem in which how do you chunk, you know, the data that you input?

A

versus

we’re It is. It is. Depends on how it works, but yes. Yes. Sometimes it is. But think of it as in it’s like a preamble to your question. Knowing the following, and the following is a is a tiny window into the content, please answer my question. And, of course, as you talk more and more, it will forget because that window is fixed. And

What pushes smaller models are efficiency, roughly, speed. You know, less is better. So if we can do with less, then less it is. Simple as this. Right? In terms of RAG, the key frontier is what we call attention level search, but this is something we’re working on. You have the exclusivity now, I’m putting it out there. It doesn’t push, I would say, model sizes. What really pushes model sizes are the efficiency rather than specializing. Meaning that if you can do the same performance with a smaller model that is fine tuned with RAG or whatever, then you’ll do it with a smaller because, again, less is better.

Oh, I love it. Constraint is the mother of innovation. Yes. You know, we can we can, you know, troll a bit about, you know, the Singapore, you know, gray market and all of these things. But ultimately, like, they had no choice. Here’s the thing. If you can buy more, why would you give a damn? Right? You can just buy more. So if you are pushed to efficiency, then you will deliver efficiency. These are very, very skilled people. This is the coolest thing to me about AI honestly is the geography doesn’t matter anymore.

You can just do things. You appear out of nowhere. Boom. You know, you’re on the map. And so I’m very, very glad that they did. I I found the the reaction very entertaining, to be honest. So, yeah, I mean, constraint is is a very good driver of efficiency.

I’m not sure who is a threat to OpenAI at the moment. Here’s why. You look at the numbers. I mean, we live in a bubble. We, you know, we follow every new episode, the whatever new model, whatever who said, who what, and so on. But, you know, I go to my mother, and I ask her, you know, do you know ChatGPT? She says, yes. And you know I don’t know. I don’t want to dunk on anybody, but do you want to know some other model? And she says, what it is?

What is it? Right? Even Gemini. Right? Like, Google. Right? So they have a strong brand. They have a strong product. But there’s a balance between the product and the models, honestly. So this is Gary from FluidStack actually, who told me that his mental model in terms of model providers, they’ll be like carmakers. There’s no winner take all. Everybody will have their own, because ultimately also human knowledge is everybody has everything. So we’re converging. But I like that analogy. Yes. Deep Seak made a very good you know, made waves, but it was it was, you know, waves that were amplified by the media and the narrative and and the drama.

Do

Today, maybe. Tomorrow, I’m not sure. They’re a bit late in terms of, you know, ASIC. They are like a 100 level, but they have probably, I would say, one of their unfair advantage is that it’s like, you know, when you do exercise in the water. Right? It’s like this. So this is their state. They are constrained, so they are bound to do better. They can just not buy their way into better compute. So I think it hinders their their success, but I think it’s short term to think that way.

No. I don’t care. I have zero I I this is something I it makes me wonder sometimes. I understand the narrative and so on, but I am absolutely not fearful. Let’s be successful first, and then we’ll talk about the politics. I have so far, you know, but again, I’m not Mistral, I’m not, you know, I’m not building gigawatt data centers and so on. So if you build gigawatt data centers, you run into these problems. But maybe you run into these problems. But the thing is, if you’re successful, everything flows from there from there.

They are very competent. I think it’s easy to spread FUD. There’s a lot of FUD going around, especially about regulation and everything. But here’s the here’s the thing. I look around me, and I don’t see, you know, what I read. Right? So I am hardly convinced about you know, everybody was saying that they they were dead and boom. They came out with their their release and it was insane. So what I know is that I hope they don’t have too much money. That’s for sure. You want to be clever.

Right?

My first impression was that I don’t buy it. I would say, you know, American style, right? You start with the claim and we’ll figure it out later. I don’t buy it and ultimately ultimately I’m not sure I care that much about it. Let’s imagine it’s true, right? Congratulations, amazing. But it is more of the same. It it is a vertical scaling. And as you know, my days are spent on efficiency. So I look at these things as being like, alright. This is a bigger you know, this is an American car of AI.

It’s big. It consumes a lot of gas. But ultimately, you know, it’s not a good car. Right? I think there has to be, you know, sufficient capital, but at some point, I’m not sure it is really a differentiator. That was prior to DeepSeek, then DeepSeek came. There was always, you know, my thesis, but, you know, you need money. You need infrastructure. You need but what is ultimately the the probably the two limiting factor today is talent and energy. That’s it. The rest, you know, yes, of course, you can buy 500,000,000,000 of GPUs.

I by the way, 90% margin. So if we work on that margin, we can, you know, shrink that number probably. So I’m not easily entertained by these numbers. I’ve seen how the sausage is made way too many times.

Oh, yeah. Latency. Reasoning. Definitely. This year. What does that mean in So the shift from throughput, so how speed my answers to how long it takes for my answer complete to to appear. That is probably one of the fundamental, like, this year. Right? Longer term, I’m very rooting for non transformer models that will change the compute also landscape. And, of course, you know, world models. Right? Yes. And or energy based models.

Probably the number one thing I would say is do not resell compute if you can. A lot of AI startups that are building on top of AI are trying to make a margin on top of a very big cake. And ultimately, what they sell is compute. If you look at the dollar of spend, you know, for $1 of spend, maybe 98% of it goes to somebody else’s margin. So if you do AI as much as you can try to verticalize on the product, but not on the compute.

If you are, you know, if your business model implies buying a lot of tokens, it’s a very hard circles to square to, you know, put that into $20, right, a month. So, you know, I I always say, like, please, you know, look at it from that angle. And if you can, try and avoid it. What’s the biggest challenge that Jensen Huang faces today? The highs are very high, but they don’t last forever. So probably it’s how to navigate the downslope. Black Whale is probably something that keeps them awake at night.

Because orders are getting canceled. Why are they getting canceled? They have a lot of problems with these chips. So a lot of people, you know, are canceling their orders. These chips are like on the frontier of scaling. And so, you know, they were supposed to come out last summer. But that heat dissipation and, you know, matter bending problem, used to be called the people who are very privy to silicon told me this is what we call a pretty big fucking problem. Right? Probably how to navigate the downslope?

Maybe you don’t know, but the supply of H100 was actually smoothed out over the year so that they decided so that they didn’t have like a big spike in deliveries and then a quarter or less, right? Which pissed a lot of people, mind you, who bought a lot of them. Some of them even haven’t received their order from last year. And they already see, like, the new chip, the b 200, and then the one after, you know, and they’re super pissed. There will be a downslope at some point.

The question is, you know, when, how, like, if there’s like the h one hundred bubble, of course, it will impact NVIDIA. But Black Whale is, I’m probably going to get a lot of flag for this, but, know, I’ve seen some very worrying numbers about it, and varying testimonies about people who operate these things. Right? So that ride will stop or at least, you know, slow down.