How many GPUs does Nvidia need to sell?
Back in the 1970s, the game developers for Space Invaders realised they could get much more performance if they separated their computing into two sections. They could have one chip responsible for performing all the calculations, like player scores, how many ships there currently are in play, and the speed at which they descend.
And then there would be a second, separate chip, and all it would do is control the colors on the screen.
Eventually, this separate chip came to be called a Graphics Processing Unit, or a GPU. Over the years, rendering graphics became more complicated as screens became higher resolution, computer processing times became faster, and new technologies emerged.
GPUs evolved to fit these needs, and the primary thing that GPUs got really good at, was arithmetic operations on very very small or very very large numbers. What GPUs excel at (compared to Central Processing Units, or CPUs), is arithmetic operations on matrixes of very large or small numbers.
I don’t want to get too technical here because most of you have probably stopped reading already. But multiplying matrixes of very large or very small numbers turns out to be very important when rendering text or images on a screen very quickly.
So in the early 2000s, you had computer nerds buying GPUs to play Call of Duty at the best speed and highest possible resolution, and the companies that built these chips, AMD, Intel and Nvidia, made pretty good money.
And then, in 2008, the Bitcoin whitepaper was released. People read it and thought “yes, a new currency that isn’t accepted anywhere with no money laundering or terrorism protections and you lose everything if you forget your password, that sounds great, where do I sign up?”
Again, I don’t want to get too technical, but performing division on very very large numbers is how you mine bitcoin (among other cryptocurrencies). And dividing very very large numbers is something that GPUs were great at!
So AMD, Intel and Nvidia sold pickaxes to the gold miners and made a tonne of money.
Then in November 2022 OpenAI released a sentence-finisher called “ChatGPT”. People used it and thought “yes, a machine learning model that uses statistics to come up with the average next word in a sentence, I would love all my writing to only ever be average, where do I sign up?”1
Almost all machine learning is just multiplying together matrixes of very small numbers. Do you happen to know any kind of computer chips that might be good at that?
AMD, Intel and Nvidia are absolutely printing money at this point, and government subsidies are fanning the flames. These companies are rushing to move their manufacturing plants away from Taiwan and South Korea, in case a war breaks out. TSMC is building a huge plant in Arizona, and Intel is putting $10 billion into a factory in Kiryat Gat in Israel. Hopefully a war doesn’t break out!2
The star of the show has been Nvidia, which was founded in the 1990s in Sunnyvale, because Sunnyvale is so boring that thinking about multiplying matrixes is exciting. Since ChatGPT was released, Nvidia’s stock has done this:
Nvidia is now the largest company by market cap in the world, Jensen Huang, one of the founders and the CEO has done an absolute victory lap, speaking on podcasts and conferences and even signing the chest of a fan. The company is now worth $2.9 trillion (!!!!) dollars and investing in Nvidia represents a bet on AI in general.
But is it actually worth that much? If investors today think Nvidia is worth almost three trillion dollars, how many GPUs do they need to sell to justify that valuation?
If we take the value of a company to be the sum of its future cash flows, plus some terminal value, discounted to the net present value, we can estimate this.
You can see my whole working here.
But to put it simply, Nvidia needs to grow at 20% per year for the next eight years, and then 10% per year for each year after that in perpetuity, in order to justify the valuation.
Now, a very simple model of what Nvidia does is they sell GPUs for money. They break down their business into four different sources of revenue: Data Center (which is what we are mainly interested in), Gaming, Professional Visualization and Automotive (ie chips for self driving cars). The breakdown of their business is as follows:
Data Center: 87% of sales
Gaming: 10% of sales
Professional Visualization: 1% of sales
Automotive: 1% of sales
So let’s focus on the Data Center, which is where all that AI growth goodness is going to come from.
According to TechInsights, Nvidia sold 3.76 million GPUs in 2023, which gave them an operating income of $28 billion dollars, roughly. So with some back of the napkin maths, we know that one GPU would get them about $7400 of profit per GPU sold to a data center.
So how many of these bad boys would Nvidia need to sell each year to justify their valuation?
And then growing by 10% per year each year after that.
So there you have it. Can Nvidia sell five million GPUs this year? And can they keep pace with that growth? This isn’t accounting for inflation, or capital expenditure. This is just sales. Can it be done?
Don’t come after me, I am literally building an AI company.
The Intel factory in Kiryat Gat is 13 miles, or 20 kilometres from Gaza. Less than a half marathon!