How trillion-dollar chipmaker Nvidia is powering the AI ​​Goldrush by John Naughton

iIt’s not often that Wall Street analysts’ jaws drop to the floor, but this is what happened late last month: Nvidia, the company that makes computer chips, released sales numbers that lit up the collective psyche of the Street. Last quarter it generated $13.5bn in revenue, which was at least $2bn more than the above-mentioned Financial Genius estimates. Suddenly, the company’s stock price surged in May, turning it into a trillion-dollar company.

Well, to a point, anyway. But since 1998 – when it released the revolutionary Riva TNT video and graphics accelerator chip – how did the company become worth a trillion dollars almost overnight to gamers? The answer, oddly enough, can be found in the folklore that emerged during the California gold rush of the mid-19th century, when it became clear that some prospectors made fortunes in gold, but the suppliers who sold them picks and shovels prospered. great

We’re in yet another gold rush – this time focused on artificial intelligence (AI) – and Nvidia’s A100 and H100 graphical processing units (GPUs) are the pick and the spade. Immediately, everyone wants them – not just tech companies, but also petro-states like Saudi Arabia and the United Arab Emirates. So the demand greatly exceeds the supply. And just to make the squeeze really great, Nvidia nimbly prebooked rare (4-nanometer) manufacturing capacity at a Taiwanese semiconductor manufacturing company, the world’s only chip-fabrication outfit that can make them, when demand was low during Covid-19. 19 Epidemic. So, at the very least, if you want to get into the AI ​​business, you need an Nvidia GPU.

What is special about GPU? Well, here comes the video gaming connection. In gaming, graphics images are made up of polygons (mostly small triangles) – rather than images produced by digital cameras made up of rectangular pixels. The more triangles you have, the higher the resolution of the resulting image. For gaming, polygons are defined as the coordinates of their vertices, so each object becomes a large matrix of numbers. But most objects in video games are dynamic, not static: they move and change shape, and for each change, the matrix has to be recalculated. Underpinning video games, therefore, is an alarming amount of continuous computation.

And for the game to be realistic, this calculation has to be done very quickly. Which basically means that traditional central processing units – which do work sequentially, one step at a time – are not up to the job. What makes GPUs special is their ability to perform thousands or millions of mathematical operations Parallel – Therefore, when you play Grand Theft Auto VThe goodies and baddies move fast and smoothly around a convincingly rendered fictional version of Los Angeles in real time.

In the 00s and especially after 2017, when Google introduced the “Transformer” model on which most generative AI is now based, AI researchers realized that they needed the parallel processing capabilities offered by GPUs. At that point it was clear that Nvidia was the outfit that had the start over everyone else. And since then the company has smartly capitalized on that advantage and strengthened its lead by building a software ecosystem around its hardware that is catnip for AI developers.

So is Nvidia going to be the next Apple, or at least the next Intel? For the next few years, its dominance looks pretty secure, in part because more of its revenue is coming from cloud-computing companies eager to equip their datacenters with parallel-processing kits that will meet their expected needs, not just traditional servers. AI gold rush. They are good customers who pay on time and will need at least two years to reconfigure their cloud infrastructure.

But nothing lasts forever. After all, it wasn’t long ago that Intel’s dominance of the semiconductor industry was complete. And now it is a shadow of its former self. Curiously, when Nvidia crossed the trillion-dollar mark, it was not Intel that was on everyone’s mind, but Cisco, a well-known maker of networking and telecom equipment that was once in the right place at the right time. The first Internet boom began in the mid-1990s. Between 1997 and 2000, its revenue tripled as demand for routers and other networking equipment increased. Then came the bust, and by 2001 Cisco’s share price (and consequently market valuation) had fallen by 70%.

Could this happen to Nvidia? The key question, says Ben Thompson, the smartest tech guru around, is: What will be the end market for AI when the frenzy dies down? No one knows the answer. Either way, Nvidia’s picks and shovels will make some people a lot of money.

What I was reading

the definite article
Consciousness is a great mystery. It’s Definition Isn’t is an interesting post by Eric Hoel on his Intrinsic Perspective blog.

Intelligence test
Benedict Evans addresses this still unresolved issue in a typically laconic and thoughtful essay on his website, Generative AI and Intellectual Property.

Expected results
A wonderful lecture on how a misreading of Adam Smith contributed to the death of despair. The Boston Review By Nobel Economics Prize Winner Angus Deaton.

Leave a Comment