He argued a dozen years ago Marc Andreessen, Silicon Valley investor-guru, that software was “eating the world.” He meant that, with a now mature technological infrastructure, it would be the companies of intangible innovation, data, codes and programs, that would change the face of the economy. Reading it again now, at the dawn of the era ofArtificial intelligence, that famous thesis rings only half true. The magic of generative AI is still a question of software, of the large data with which it is trained and of the algorithms that make it “think”. Yet, like the human brain, the synthetic one also requires sophisticated hardware. Silicon neurons, microchips, for which the AI industry now has an insatiable hunger. And whose production depends on two companies so dominant as to be de facto monopolists.
The masters of the chips
The first of these chip masters is Nvidia, an American multinational that designs the most used super-processors to train large AI models. Everyone, from Big Tech giants to governments, is trying to grab as much of it as possible, and on the wave of record sales and profits Nvidia has taken off beyond two trillion dollars in capitalisation, the third in the world behind Microsoft and Apple. The second champion – “the most important company you have never heard of,” someone said – is the Dutch Asml, which is a step above in the supply chain because it builds the machinery necessary to “print” the smallest and most advanced chips, including those from Nvidia. In recent times it too has seen orders, turnover and value increase, reaching almost 400 billion dollars, third in Europe behind the Danish Big Pharma Novo Nordisk and LVMH. While waiting to understand who between Microsoft, Google & Co will win the gold rush of AI as software, AI as hardware already has two winners.
That microchip makes Nvidia everyone. The productive Eldorado promised by AI drags global stock markets
by Filippo Santelli
Hunt for Nvidia
And best wishes to those who think of challenging them, given that decades of innovations and winning intuitions are embedded in their jewel-products. Take Nvidia: the founder and CEO Jen-Hsun “Jensen” Huang, an American of Taiwanese origin, an engineer with a leather jacket, was among the first to believe in the AI revolution. At least since, at the beginning of the last decade, some researchers began to use Nvidia GPUs, processors created to animate the graphics of video games, in supercomputing experiments, given that they were able to manage a multiplicity of of operations. Huang realized it was the future, directing efforts there. Product development, whose crown jewel is the H100 processor. And cooperation with programmers, developing a software platform – CUDA – with which they can best distribute computing resources.
Result: Nvidia is the standard today and controls between 70 and 90% of the AI chip market, little squares worth 40 thousand dollars each. The queue at the door is such that Huang has had to ensure that he will process orders without favouritism. Nvidia is a company “fabless”: it has no factories, it designs the chips and then contracts production to so-called “foundries”, such as the Taiwanese Tsmc. This guarantees very high margins and the possibility of quickly scaling production. A perfect recipe that will be put to the test by fierce competitors in the coming years. Silicon rivals, such as AMD and Intel, are investing to catch up. But even large customers such as Microsoft/OpenAI, Google and Amazon – owners of large computing centers and some of the most powerful AI models – try to develop “custom-made” chips in-house, which are more powerful and economical. For its part, Nvidia has created a new division to design these “third party” microprocessors and avoid being bypassed. But his numbers show that, for now, leadership is not in question.
The most complex machine
The advanced chip “foundries” capable of churning out the smallest and most powerful processors, the sub-ten nanometer nodes needed for AI, can be counted on the fingers of one hand: the Taiwanese Tsmc, Samsung, soon Intel. But only one company, the Dutch Asml, produces the machinery needed to print them all. They are the deep ultraviolet (DUV) scanners and especially the new generation a extreme ultraviolet (EUV)«the most complex machines that man has ever built», price 200 million euros.
Like that of Nvidia, the history of Asml also comes from afar, born in 1984 as a spinoff of Philips. But its activity, which makes it a sort of large techno-industrial consortium, could not be more different. Part of the intellectual property, particularly on lasers, arose from US government research programs, has been developed over the years with the contribution of customers such as Intel, Samsung and TSMC, and remains in an independent American subsidiary. Each lithographic apparatus then includes tailor-made parts made by dozens of hyper-specialized manufacturers who operate “in symbiosis” with Asml, such as the ultra-thin mirrors of the German Zeiss. And in the headquarters of Veldhoven the very delicate final assembly takes place.
Catania at the chip frontier, a pilot plant for electric cars and satellites is on the way
by Filippo Santelli
This flow has rigidities. It’s difficult to accelerate: the company delivered a total of 421 machines last year. And it is tied to a limited number of clients, exposed to their investment cycles. But at a time when investments are being made in advanced chips and the order book is full, this same complexity makes Asml unrivaled. Its machinery is non-replicable and indispensable, at least until the way to create more powerful chips is to make them smaller, the famous Moore’s law.
This uniqueness is also why ASML has come under the spotlight of the United States as it seeks to contain China’s technological rise. In recent months, Washington has progressively restricted Nvidia’s ability to sell the most advanced chips in the People’s Republic, forcing it to create “weakened” versions. But above all he convinced – or forced – the Dutch government to stop exports of Asml machinery to China, the most advanced EUVs and then also the previous generation DUVs. A terrible blow to Beijing’s dreams of hi-tech autarky, because an advanced chip can be designed, as Huawei did, but producing it industrially is much more complex. Geopolitics also reveals how important hardware is at the dawn of this AI era.
#Nvidia #share #price #rally #history #quadrupled #capitalization #months
– 2024-03-15 06:45:31
