Original video in Chinese.
Key Takeaway
- GenAI is only the vanguard: generative AI let the world recognize the value of GPUs, and Nvidia made its first pot of gold from it, but its essence is “language statistical probability,” so it will eventually become a red ocean and barriers to entry will collapse.
- General scientific computing is the main force: Nvidia’s real ambition is to move $100 trillion of real-world industry onto GPUs—drug discovery, materials discovery, automotive wind tunnels, oil exploration, and more—using digital twins + physical simulation to replace expensive experiments, with far higher premiums than selling cloud compute.
- The moat decides the winner: competitors like TPU sacrifice generality for AI training, and once they return to “solving physical equations,” they become scrap iron; Nvidia uniquely owns CUDA + decades of physics algorithm libraries + Omniverse, and once industry adopts it, it can’t leave. In the long run, this is a market bigger and deeper than GenAI.
If I could only hold one stock, I’d definitely choose Nvidia, and I’d hold it for ten years. Because general scientific computing is a market bigger than generative AI, and a bigger opportunity.
The “digital twins” and “AI for Science” that Huang has been talking about these past two years are exactly what general scientific computing is about. Its essence is “simulating the physical world.”
In this field, Nvidia has a dominant position. “General scientific computing” is currently almost synonymous with “accelerated computing using Nvidia GPUs.”
More importantly, Nvidia’s competitors cannot catch up. Because their products are all born for AI, tailor-made for model training and inference. They have cut away too much generality. That instead locks them into the AI market.
Once the computing paradigm shifts from “language statistical probability” back to “solving physical equations,” TPUs become scrap iron.
So, in the short term, say within three years, AI will definitely be the market hotspot. But if you stretch the time horizon, general scientific computing is a much larger and deeper market than generative AI.
If you’re in investing, you may not need to allocate to it right now, but you can’t not know about it. Otherwise, how are you any different from the uncles and aunties speculating in A-shares?
Hello everyone, welcome back to my channel. Modestly speaking, I’m one of the few creators in China who can clearly explain the Why and How of AI. What I offer is far more valuable than tutorials. Remember to hit follow. If you want to connect with me, come join our newtype community. This community has been running for 600 days, and more than 1,900 friends have paid to join.
If you’re a domestic user, you can join through Knowledge Planet (Zhishi Xingqiu). If you’re an overseas user, you can join through Substack. My first course, daily newsletter, and exclusive videos are all available in the community.
Back to today’s topic: Nvidia’s general-purpose accelerated computing.
In the video before last, I introduced TPUs and talked about why TPUs cannot replace GPUs.
To train and run large language models, ASICs represented by TPUs cut out high-precision computing units and focus on FP16, FP8, and even INT4. Because generating text doesn’t require high precision.
This cut does make TPUs very powerful in AI. But that comes at the cost of generality.
What Nvidia has always insisted on is general accelerated computing. Their making money from AI was purely accidental.
CUDA was released in 2006 to 2007. Deep learning didn’t take off until 2012. There was a full six-year gap in between.
During those six years, Huang was under enormous pressure. He painfully tried to sell CUDA to physicists, chemists, oil exploration experts, and financial analysts. At the time, many people on Wall Street and in Silicon Valley thought he was crazy, making graphics cards so complicated that costs skyrocketed, only to serve these very niche scientists.
In 2012, the reason Alex chose Nvidia GPUs to train neural networks wasn’t because Nvidia had done anything especially for AI. It was purely because this was the only affordable parallel computing hardware on the market that could be programmed in C.
So GPUs were not born for AI; they were built to serve general scientific computing. By design, they preserve two key characteristics, two characteristics that today’s ASICs cannot surpass.
First, extremely high programmability.
Specialized ASICs are often optimized for a certain class of algorithms. Once the algorithm changes, the chip is obsolete. CUDA is general-purpose. If the algorithm changes, you just rewrite the code and it runs. That’s also why AI papers are almost always reproduced first on Nvidia chips.
Second, an obsession with precision.
As I just said, TPUs are optimized for AI and cut out the high-precision parts. But if you’re doing scientific research, you need high precision, such as FP64. Because scientific computing is extremely sensitive to error. This is something TPUs can’t do, but GPUs can.
The strength of ASICs/TPUs is as “matrix multiplication accelerators.” But scientific computing is extremely complex. For example, protein folding is not just matrix operations; it also involves complex geometric search and graph theory algorithms. These are exactly Nvidia’s strengths. GPUs and CUDA are highly suitable for them, and Nvidia has accumulated that advantage over many years. Scientists simply cannot switch camps, because the migration cost is too high—roughly equivalent to rewriting the entire software infrastructure of a discipline.
What I’ve said so far is logic. Let me give you data too. The website top500.org lists the supercomputer rankings. You can download the table yourself and filter out the Nvidia entries, and you’ll see how strong they are in this field.
Everything above explains why, in general scientific computing, only GPUs are viable, while TPUs simply don’t work. So the question is: how big is the market for general scientific computing? Will it be bigger than generative AI?
I’ll analyze this from three dimensions.
First, the nature of the market.
The essence of the generative AI market is the “bit economy.” Its ceiling is human attention. No matter how many videos and articles you generate, humans can’t consume them all, so the commercial value will see diminishing marginal returns. So for now, generative AI is a digital economy market worth roughly several trillion dollars.
The essence of the general scientific computing market is the “atom economy.” Its ceiling is the efficiency of the physical world. Human demand for cheaper drugs, more efficient batteries, and cleaner energy is almost infinite. So in aggregate, this is a real economy market worth nearly $100 trillion.
Second, the revenue model.
The awkward situation generative AI faces today is that many companies buy GPUs to do AI, but are still desperately searching for a business model.
But scientific computing is a necessity, because it directly taps into the R&D budgets of global industry.
For example, the world’s top ten pharma companies spend more than $100 billion on R&D each year combined. If Nvidia’s simulations can cut the cost of developing a new drug from $2 billion to $1 billion, or double the success rate, pharmaceutical companies will absolutely hand over large budgets to Nvidia without hesitation.
Also, automotive wind tunnel testing and oil exploration drilling can each cost millions of dollars in physical expenses per attempt. If GPU simulation can replace even half of that, this is a market shift worth hundreds of billions.
In this field, Nvidia is not selling “compute”; it is selling a “digital laboratory.” That pricing power is far greater than selling to cloud computing companies.
Third, the moat.
The generative AI market will eventually become a red ocean, because inference devices are getting more numerous and the barriers to entry will get lower and lower. But the general scientific computing market won’t, because it requires high precision and decades of accumulated physics algorithm libraries. Right now, only Nvidia can do this.
Once industry gets used to using Omniverse and CUDA for R&D, they’ll never be able to leave, because it involves underlying data assets.
To sum up, you can think of the relationship between these two markets like this:
Generative AI is the “vanguard.”
It made the whole world realize how powerful GPUs are, and it earned Nvidia its first pot of gold.
General scientific computing is the “main force.”
Huang wants to move the world’s $100 trillion real economy onto Nvidia’s chips to run. That is his true ambition.
The end of compute is physical simulation, not just language generation. This is a bigger opportunity than generative AI.
OK, that’s it for this episode. If you want to understand AI, become a super-individual, and find like-minded people, come join our newtype community. See you next time!