一口气了解英伟达，芯片新王凭什么是他？

On May 24, Nvidia released their first quarter earnings Many investors are calling this an unprecedented once in a lifetime release Nvidia relied on the dazzling data giving Wall Street a slap in the face The stock soared by 30% on the day Market cap reached trillion of dollars propelled Nvidia into becoming sixth-largest company in the world surpassing Tesla and approaching Amazon Who would have thought that a company selling graphics cards would become the biggest winner

in the AI ​​war in 2023 For the past few years Nvidia participated in almost all of global tech innovation cloud computing cryptocurrency Metaverse Artificial Intelligence Nvidia is main player in all these Majority of the AI model you’ve heard of are trained with Nvidia graphics cards.

Not only are they an industry leader but they monopolised global AI training industry by occupying 95% market share Even the quantity of owning Nvidia A100 graphics cards an indicator to measure a company’s computational power The founder, Jensen Huang said Sounds arrogant right But that’s the truth With his foresight from over 20 years and his unchanging outfit style becomes the Godfather of AI You must be wondering what’s so great about Nvidia How did they monopolise?

Why nobody can compete against them Today let Lin take you down a trip to the story of Nvidia We’ll also talk about the secret behind graphic card and chip industry Let’s begin with the rise of Jensen Huang and Nvidia In 1963 Jensen Huang was born in Tainan, Taiwan which means he is 60 years old this year At the age of 9, he moved to US after graduating from college he worked in two semiconductor companies

focusing on chip design one of them is AMD a company that fought with Jensen for half of their lifetime until now After finishing his Master at Stanford, Jensen turned 30 With two other tech guys who were also in their 30s with big ambition they planned to do something big They believed that 3D graphics processing has great potential in the future So in 1993 they established NVIDIA specialises in graphic processing chip Jensen Huang is the CEO.

he still is today With recommendation from his former boss he obtained $20 million capital investment from Sequoia Capital After all gaming industry needs real-time rendering when playing You can’t say there’s no 3D games but definitely can’t play with normal computer Many games are quite classic but the graphic is basically, well as long as you can see a figure there because this kind of 3D image processing is computationally intensive

It was difficult for the CPU at that time Normally they need a specific chip to process the graphics and this chip is graphic card In the early days, the graphics card was very simple it was a 3D accelerator card at most Now when you hear 3D accelerator It sounds like a small workshop business That's right actually at that time 3D games and 3D rendering was at seeding stage Graphics card companies like Nvidia are actually a lot

at least 50-60 companies There was no uniform standard for both hardware and software Whoever comes out with better research can publish their own standard.

Often times when you finally came up with a graphics card but it actually is not compatible with other people’s standard Just one word Chaotic One of the most famous company at the time is a company called 3dfx It was established in 1994, a year later than Nvidia At that time, a graphics card called Voodoo was all the rage Many popular games during that time relied on Voodoo graphic card It wasn’t going so well with NVIDIA

Although they obtained capitals and had very professional team but their NV1 was not successful NV2 was aborted By 1997 NVIDIA was hanging on by a thread 9 more months till they run out of money The company downsized its staff from 100 till there were about 30 people left Jensen Huang took a gamble Just when the company only had 6 months of operating capitals They released Riva128 graphic card carrying NV3 With its good price/performance ratio,

they finally occupy a place in the market and allow NVIDIA to survive Actually Jensen and his team are very strong in their R&D After they figure out the market direction they managed to enter fast lane quickly They reached a long-term strategic cooperation with TSMC At the same time they cooperate closely with Microsoft supported the Direct 3D display standard introduced by Microsoft Finally they rose in the sea of competition in graphics card industry With RivaTNT it helped

NVIDIA to become industry leader in graphics card industry In 1999, NVIDIA successfully listed on NASDAQ After listing NVIDIA had more money and in September 1999, they released the epoch-making GeForce256 which made them the leading force among their competitors I believe that gamers should be familiar with this GeForce series.

It has also become Nvidia’s the flagship line of consumer graphics card Jensen Huang named GeForce256 the world’s first GPU The first truly dedicated graphics card The claim is basically accepted by everyone Hence some people might generally claimed that Nvidia invented the graphics card.

As for why this dedicated graphics card is so powerful, we will talk about it later.

At that time, Microsoft happened to be working on Xbox With GeForce 256 powerful performance, NVIDIA managed to score $200 million worth of order After building image processing hardware for Xbox they then score another with Sony’s PS3 From 1999 to 2002, Nvidia’s revenue almost doubled every year to $2 billion becoming the only player in the market They started to acquire competitors in the same industry one of them we mentioned earlier the once popular 3dfx

Another major player in the market ATi was acquired by AMD And so in the early 2000s after a series of merger and acquisition happening in the market There are only two players left in the market Nvidia and AMD.

Until now, the dedicated graphics card market has been dominated by these two companies.

I don’t know if you have heard of the legendary N card and A card actually refer to the graphics cards of these two companies.

Gamers started to argue whether N card or A card is better It doesn’t matter which one is better after that there isn’t a third company exists in the market like B card of C card or X card It’s a two-horse race However NVIDIA is gradually eating away AMD’s market share from 60% in 2010 slowly expanded to 80% in 2022.

becoming global GPU hegemon The speed of development of GPU technology itself is also jaw-dropping The rapid development of gaming industry has supported Nvidia, At the same time, NVIDIA’s graphics card development has promoted the development of the gaming industry.

Look at the new games coming out every year so much improvement in image quality Even if you don’t understand game you can see the speed of progress In fact, the overall graphics card market is in a tripartite state Three biggest player – INTEL, NVIDIA, AMD Intel occupied 71% of market share Nvidia is 17% and AMD is 12%.

You must be wondering why is there Intel and their market share is way higher Didn’t you just say that Nvidia is the biggest player?

Actually this graphics card is not the same graphics card Because graphics card is divided into dedicated and integrated graphics card If you were to compare both then Intel indeed is the biggest player but they sell mostly integrated graphics card Integrated graphics card is placed together with CPU they share memory So Intel taking advantage of their position in CPU industry monopolise the integrated graphics card market share However the integrated graphics card

is quite weak, I won’t explain in details here Comparing with NVIDIA’s dedicated graphics card although they are both graphics card but they don’t belong in same market Just from dedicated graphics market POV NVIDIA occupied 80% of the market share Some of you might start to get bored now Alright we know NVIDIA designs graphics card and chips they are very good at it I’ve been talking a lot about 3D rendering and gaming

How does it have anything to do with AI Why are all these AI companies want to buy graphics card and it has to be NVIDIA’s graphics card Don’t worry, we’ll talk about graphics card characteristic in a computer, Central Processing Unit The purpose of its design is that it can do everything It’s sequential computing and it can carry out very complex logical reasoning However image processing doesn’t care much about sequential computing It’s more concerned with computational volume

For example, a 4k video has 10 million pixels Let’s say there’s 30 frames per second then each pixel and frame has to compute correspond colour based on shadow and action This requires non-stop, very fast and massive simple calculations is Graphics Processing Unit It’s especially designed to do this kind of computation The foundation of the chip design is to optimise parallel computing So for CPU is 64 or 128 core at best

while GPU could have thousands of core computing together at the same time See this video is giving a very good explanation CPU is like a very precise very strong gun firing one shot at a time The shots are fired in clear order but slow, GPU on the other hand is like having thousands of this gun firing at the same time Due to GPU special feature Jensen Huang started to think about how to maximise its potential

It’s definitely as simple as 3D image processing and rendering Can they carry out more of General Purpose Computing General Purpose Computing General Purpose Graphics Processing But it’s not simple to use GPU to do this kind of general purpose computing Because the purpose of its design is not for this so the programming is very difficult Not anyone can do this job Jensen Huang was thinking that if graphics card was to realise its greater potential

it needs to be programmable By chance he saw a project by a PhD student in Stanford Using C language programming to let GPU do some computing Jensen thought this idea is amazing and offered this fella a job at NVIDIA He appointed him with a very important job and let him lead the team to carry out R&D in making GPU programmable Finally in 2006 NVIDIA officially released CUDA making GPU programmable

In order to build this CUDA system, NVIDIA invested a large and unreasonable amount of capitals and human resources into it Originally those graphics card that to support CUDA Originally those graphics card that specializes in 3D graphics processing need many top engineers to make it programmable But now anyone can do it by buying an NVIDIA graphics card and use it with CUDA library Through CUDA NVIDIA expanded the boundary of graphics card

from gaming and 3D image processing to a whole realm of accelerated computing Like aerospace, biopharmaceuticals Weather forecasting, energy exploration and so on are actually using large amount of NVIDIA graphics card to carry out computation Many have tried to create software like CUDA to challenge Nvidia’s position.

But Nvidia has a monopoly on hardware itself.

They can try everything to merge their hardware, graphics card with the software, CUDA and make them work really well with each other Through hardware and software merger they formed a very strong moat Does this make you think of another company?

That’s Apple They all build something that’s had been bandied about in business world an Ecosystem Microsoft Adobe they all have their own strong ecosystem The amount of capital Jensen invested in CUDA might sound reasonable to you now It makes sense but if you look at short-term return It is very unreasonable.

Wall Street is quite displeased with this thing because although the computation performance of GPU is outstanding but their application is not much For a long period of time it can only focus on area that requires massive computing To put it bluntly, it's not profitable at all Who would’ve thought that AI could be so popular in 2023 right In fact, it is not AI that first makes the accelerated computing capabilities of graphics cards realise its commercial value.

It was a coincidence A trend that’s not related to AI at all Something that even someone like Jensen could not have expected The explosion of Bitcoin brought a huge demand for mining Mining in essence is mindless computing to encrypt and decrypt In order to do mining in faster way then you will need to use graphics card and you have to use NVIDIA’s graphics card I believe most of you only know about using graphics card to do

computation from the mining part The huge demand for mining is like a godsend for NVIDIA a big big gift This makes NVIDIA’s graphics card in constant short supply for years NVIDIA was very nice they designed a GPU specifically for mining Of course many would say Mining pollutes the environment, meaningless computation all sorts of problems Graphics card capability was seen by many and NVIDIA did make lots of money because of it According to analysts

Between 2018 to 2021 a period when bitcoin was very popular NVIDIA could earn up to $1 to $3 billion annually NVIDIA’s market cap even surpassed the giant Intel During the time when bitcoin was very popular their market cap even approached trillion dollars Although mining has made NVIDIA a lot of money but it is after all not its main business After the crash of crypto market NVIDIA’s stock plunged by 46% So far mining is only

a sideshow at best We all know what’s really been helping NVIDIA lately Artificial Intelligence We mentioned that graphics card massive parallel computing capability is very suitable for deep learning and machine learning AI has to keep learning up to billions of times So GPU parallel computing capability is where it fits in Jensen Huang showed that to train a large language model in comparison with CPU GPU server can complete it at 4% of the cost

and 1.2% of power.

Therefore, GPU CPU are not on the same level This is determined by its underlying structure.

Jensen Huang just wanted to subtly tell you that to train large language model only a fool would use CPU You have to use GPU If you use GPU, you’d better use CUDA.

Then you have to buy the graphics card from NVIDIA Actually a decade ago no one knows graphics card can be used in AI AI itself is more theoretical than practical The change occurred in 2012.

At that time, there’s a very famous computer competition called ImageNet.

Everyone was competing whose algorithm could better recognise the content of the image From No.2 to No.4

From No.2 to No.4

their error rate is about 26% to 29%.

A team called AlexNet made it to 16.4% Ten points ahead of the second place and won the comepetition they used neural networks to train their models with NVIDIA’s graphics card We’ve talked about this in ChatGPT episode the theory of neural network has been around for a long time but it had not been realised the problem was with computational power Thanks to NVIDIA’s graphics card the theory of neural network has been able to realise

It made a sensation in the academic world Jensen is also very serious about GPU’s application in AI They went all in after 2012 and let NVIDIA’s graphics card do accelerated calculations easily and conveniently Apart from investment focused on CUDA there are also optimisation on AI for graphics card including software support platform support and etc Later there is a consensus in the field of AI If you want to do AI then no doubt

you have to buy NVIDIA’s graphics card Google Amazon Microsoft Baidu all used NVIDIA’s graphics card to train the models The most famous one in this wave is NVIDIA’s A100 ChatGPT was trained with over 10,000 of graphics cards.

A100 has also become the standard for training large model For major AI companies if they couldn’t come out with significant result but still want to catch the wave What do they do They fight in owning graphics card They’ll tell you how many NVIDIA’s A100 graphics card they bought ranging from thousands to tens of thousands So this caused NVIDIA’s graphics card to be in short supply for a long time

Their price went up to ten of thousands of dollars Last year NVIDIA released upgraded version of A100, H100 It has four to six times the performance of the A100 So don’t think that NVIDIA is just in luck just because they soar the highest They spent a decade and a lot of money and effort to build their own wings Just waiting for the right wind Now that the wings are completed the wind has come

NVIDIA definitely will want to fly high This is A100 system board released by Jensen Huang 3 years ago a big guy over 20kg became the world’s largest GPU Looks intimidating This year he released another thing DGX GH200 supercomputer.

The size is in 1:1 ratio with the actual thing They use 240km of optic cable is kilometre that’s 240,000 metres Its weigh is equal to 4 adult elephants with internal memory of 144TB This GPU is connected using technology like NVLink, NVSwitch And these 4 elephants is actually a mega graphics card it is specifically used for AI computation It's expected to be ready by the end of the year

Google Cloud, Meta, Microsoft are the first to access it With such big first-mover advantage NVIDIA is not limited to making only graphics card and chip design In 2019, they spent $6.9 billion to acquire an Israeli chip company called Mellanox and came up with something called DPU This thing is very powerful too Jensen Huang said that this acquisition is the most successful strategic decision he has ever made NVIDIA has started to merge GPU, CPU and DPU together

and created a server with incredible computing power Supercomputer is slowly creeping towards CPU market and launched a variety of products The eighth supercomputer in the world DGX BasePOD Accelerated superchip GPU Grace Hopper computing platform BlueField-3 DPU and so on Good names right but you probably don’t understand In short, Nvidia covers everything from chips to supercomputers they’ve got it all covered What if you can’t afford it or don’t want to buy so many hardware?

It’s okay NVIDIA can lease it to you just get connected online and use NVIDIA’s server This is AI cloud business It provides services to end user as well as upstream companies They came up with a software called cuLitho to help TSMC, ASML these upstream chip makers Improve the performance of Inverse lithography technology by up to 40 times By the way, I have a Spanish friend who told me that the name cuLitho in Spanish sounds like

a booty One word Incredible See, NVIDIA expanding across hardware, software, services and etc becoming the biggest winner in this AI wave US sanctions towards China bans the sale of A100 and H100 to China.

The impact on Nvidia is actually quite big.

China occupies a quarter of NVIDIA market They don’t want the cake fly away just because of US government When a reporter interviewed Jensen Huang asking him how big is the impact of this for NVIDIA His answer is impermeable First we will definitely work closely with US government’s policy Express their position first At the same time we are trying within the rules to satisfy the demand of Chinese consumers He offends neither side

NVIDIA released A800 graphics card bypass a bunch of sensitive technology that has been sanctioned This is made especially just for China Alright we understand the entire background of Nvidia, it’s very clear now when you look at their earnings report Currently they divided their business into 4 segments Gaming, data centre automotive, professional graphics processing Previously NVIDIA’s ace business was gaming and data center is about accelerated computing lAI, Cloud service and etc are all in this segment

These two segments are the main ones In 2018, Gaming occupied half of their business Data center is a quarter by 2022, Data center occupies 56% of their business Gaming dropped to 33% The earning report released on 24th May their revenue dropped significantly due to sluggish global demand in gaming This is actually what Wall Street expected.

However data center segment is strong 18% growth from previous quarter Most importantly the revenue in second quarter blinded Wall Street Wall Street predicted that their revenue for Q2 would be $7.2 billion but NVIDIA came out and said Your estimation is incorrect Our revenue is $11 billion, 50% more For automotive segment it is a segment with great potential Not only have to do chip for cars also need to create car system Another big cake

But this is at initial stage It’s difficult to say what’s going to happen in future So we’ll not discuss it Before this NVIDIA released Omniverse to bet on Metaverse Although we haven’t seen much return yet but I’ve watched the promotional video It really is quite cool If Metaverse ushered in explosive period Then NVIDIA will again be on of the biggest winner It’s definite Anyway, the general picture is that gaming needs graphics card mining needs graphics card

AI computing needs graphics card Need graphics card look for NVIDIA NVIDIA stock price rose by over 1000 fold from its initial IPO becoming the sixth largest company in the world And it looks like it's just getting started From the stories earlier you would’ve believe Jensen is almost godlike Every step is so precise.

but he actually made a lot of mistakes along the way For the continuity of the story I ignored it In the early 2000s There was a series of graphics card failures almost defeated by ATI there were also insider trading committed by employees SEC did a thorough investigation on them There was also problem with over-marketing They even tried to enter mobile phone chip market but it was all a failure Looking at their stock price

Although it has risen so much as a whole but in 2002 it dropped by 90% 2008 dropped 80% with more than 50% of retracement This occurs every 3 to 5 years Everyone knows Jensen is great but he is no fortune-teller His obsession with leather jacket I don’t know what else to say Alright Let’s talk a little bit about NVIDIA’s stock price I’m worried that many would rush to buy NVIDIA’s stock after this video

I wouldn’t say you cannot buy but just don’t go buy it on a whim The estimated value for NVIDIA is too high from every perspectives Their Price to Earning ratio is over 200 Price to Sales ratio is 38 For Apple, Microsoft and Google their Price to Earning is less than 40 Tesla is only over 70 Compare them with those in same industry Their revenue is less than half of Intel’s

but market cap is 7 times more than Intel revenue is similar to AMD but market cap is 5 times more than AMD One word Expensive Of course they are expensive for a reason For a company in an industry like this moreover it’s a monopolistic company Its valuation is no longer one of the main criteria for judging stock price One of the main reason do you know what it is?

There are some professional institutions dare not to not invest in NVIDIA because AI is the biggest wave with biggest opportunity in the market And NVIDIA is the biggest player in this AI wave When these funds invested in NVIDIA even if the stock price fall investors wouldn’t complain much but if NVIDIA stock price keeps rising and they didn’t invest because the price is expensive then investors would really be pissed Cathie Wood liquidated NVIDIA stocks

in January through her ETF Lately NVIDIA stock price is rising she was reviled in the investment circle Choosing to not invest in this type of company would risk fund their reputation There’s a term in English that best describe this behaviour It is FoMO Fear of Missing Out They fear losing the potential rise more than buying at expensive price This is similar to Tesla two years ago EV has great future and Tesla was the only option

even if its valuation is astronomical compare to other automotive companies but funds still continued to buy The behaviour of FoMo in turn pushes the stock price of these companies to an even higher position You can say that for these type of companies they just have huge risk premium but just because of this you can’t say that their stock price is too high That's not necessarily true Because it they can sustain this momentum

and maintain this development trend Then their stock price will continue to rise There's a metaphor that many people often use The wave we mentioned is similar to gold rush It’s hard to bet where the gold is or who will find the gold but there is a sure-fire deal and that is you can sell shovels Nvidia is like selling a shovel at digital age at AI age This metaphor sounds right and quite reasonable

But whenever I hear this analogy it doesn’t seems right Those who understand economy would know that if you sell shovel during gold rush could you be making a fortune?

The first one would perhaps earn a little but if more people are selling it and it’s easy to make shovel then the marginal profit would quickly be eaten up so in most industries if you sell this so-called shovel the upstream production tool Due to its low entry barrier the competition in the industry will be very fierce so the profit margin will be low However for NVIDIA they can achieve a monopoly and a market value

of trillions by doing this so-called shovel Its similar to TSMC more basic than shovel, they make hammer a tool to make shovel And they can monopolise that Have you ever thought why?

We’ll get deep into the feature of chip industry Have you heard of Moore’s Law?

On an integrated circuit the number of transistors that can be accommodated double up every 18 months You can understand it as The speed of chip can be faster every 18 months Actually for the past few years The speed of CPU evolution is difficult to catch up with Moore’s law However Jensen found that the advancement of GPU dedicated graphics card is faster than prediction of Moore’s Law Performance increase more than triples every 2 years

This pattern even has its own term Huang's Law No matter which law it is This is a distinctive feature in chip industry The initial cost to invest is very high and requires a lot of talents and equipment The problem is that the iteration rate is too fast This is an industry that is constantly running the iteration rate of this industry is too fast From business perspective this pose a tricky problem

it’s difficult for company to build their own moat Moat is very important to a company For a traditional company if they spent a lot of money to build a factory, a railroad then they have large-scale advantage and cost advantage These advantages work as a moat This moat will protect the company for a long time The moat of internet company is stronger Once a network effect is established Wechat Tiktok Facebook users are your strong moat.

For chip industry there are Moore’s Law, Huang’s Law No matter how good of a chip you release today it’ll be outdated in two years You won’t know who’ll make it next year and take you down with his new technology In the 90s during the graphics card war A company established less than two years 3DFX quickly became the industry standard with their Voodoo graphics card 5 years later they became dreary and was acquired by NVIDIA

This is because their moat hasn’t finished building and was destroyed In this industry, it’s difficult to rely on a single product or a single technology to form a moat that lasts more than 2 years Of course you can slowly accumulate many patents to defend yourself but the truth is people will always find a way to get around those patents it doesn’t provide much protection You have to keep running and run faster than everyone

This running ability is your moat In chip industry the moat is R&D You have to develop a complete set of talents, facilities, organisation structure These form the moat for chip company This is why R&D is expensive First is the initial cost is very high requires a lot of talents most importantly you have to continuously keep the iteration and running speed high Not a lot of people can stand it Those tech giants are

actually very focused on technological innovation Amazon Google Microsoft their R&D investment is 10% to 15% of their revenue For NVIDIA they have to keep invest around 25% of their revenue for R&D The Huang’s Law is not a natural phenomenon It is the result of a rat race he created If we were to tell a story we can say NVIDIA has been monopolising graphics card market since 2006 end in just one sentence However for many times

they overturned their own framework and technology They pushed themselves to release new generation of chip every 6 months For example their latest ray tracing RTX technology completely overturned the program accumulated before They use deep learning method through 1 pixel they guess what the other 8 pixels surround it looks like to speed up image processing.

This intense running speed makes only those who run fastest with largest pool of talent and wealthiest companies can make money.

The rest could only run behind This is why we often see that after a war in chip industry it always ends in acquisition It's not just the difficulty of developing chips it’s mainly because it’s not economical Even those tech giants with plentiful talents won't set foot in the chip industry unless they are forced to But things are different now The AI field is one of the biggest possible battleground for the tech giants

In this battle, graphics card is too important after all it determines computing speed Non of them wish to be strangled by NVIDIA during AI era right Although currently they are NVIDIA’s biggest customer but at the same time they delve into R&D and chip making in full swing Google already developed a chip specifically designed for AI training called TPU According to them, it is more efficient than Nvidia's graphics card.

In 2017, Meta already used over 20,000 NVIDIA's graphics cards for AI training but now they are also all in in researching chip, calling it MTIA Actually from 2020 NVIDIA has been actively seeking Masayoshi Son’s Softbank to acquire their chip company ARM at the price of $40 billion But Qualcomm, Microsoft and Google are all strongly against it The deal fell through From here you can see that those tech giants

are actually very afraid of Nvidia.

Emerging from the early days graphics card industry is his ability Having laid out CUDA since 20 years ago is his vision Continuous monopoly in graphics card industry is his endurance Meeting the demand for massive computing power in mining is his luck The unchanged fashion style is his devotion When long accumulation period meets the AI wave NVIDIA naturally obtain first-mover advantage

Facing new vast potential market and competitors looking to get in on the action Can NVIDIA fight their way out again in this new wave of AI fight?

It’s hard to say