How hardware is (still) eating the world

Stay informed with free updates

It was quite the Christmas present. In December, the Dutch technology company ASML started shipping 250 crates to Oregon to install a €350mn machine for the US chipmaker Intel. The Twinscan Exe: 5000, to give the machine its full name, is probably the most complex piece of equipment ever built. Weighing as much as two Airbus A320 jets, it took a decade to develop and will require 250 engineers to make it operational next year. 

Its purpose? To “print” tiny 8-nanometer lines on a silicon wafer, compared with the 13-nanometers in earlier models. That sounds like a microscopic difference and indeed it is, but it has giant implications. By deploying the latest iteration of its extreme ultraviolet lithography technology, ASML can enable 2.9 times more transistors to be packed on to a chip, significantly improving computing power, memory and energy efficiency. 

Many of those chips will be used to meet the near-insatiable demands of the tech companies developing the latest artificial intelligence services. That makes ASML an intriguing prism through which to view the evolution of the new tech economy.

More than a decade ago, the venture capital investor Marc Andreessen famously declared that software was eating the world. But the hardware needed to power that software is still hungry, and growing hungrier. Some investors are now calculating that hardware may be a surer bet than software when it comes to exploiting the AI revolution. This is a classic “picks and shovels” play during the AI gold rush. 

It is telling that ASML’s stock market value is now 1.85 times that of Europe’s biggest software company, SAP. Similarly, the market worth of the US chipmaker Nvidia, which sells the graphics processing units that power the latest AI models, recently overtook Alphabet and Amazon.

And Sam Altman, chief executive of OpenAI, who helped trigger the generative AI frenzy after launching ChatGPT, has talked about the need to invest as much as $7tn to produce the chips, energy and data centres to run the future tech economy. If Altman is not hallucinating, that would require a staggering amount of new kit. According to one estimate, semiconductor companies have spent just over $1tn on chip manufacturing equipment since the birth of the industry.

Now, it may well be, as my colleague June Yoon has argued, that investors are getting way ahead of themselves in their enthusiasm for AI-related hardware companies. There is excessive hype about the impact of the technology. There is overcapacity in several segments of the notoriously cyclical semiconductor industry. There is geopolitical risk associated with China, one of the world’s biggest chip markets, which is being squeezed by US export restrictions. A salutary market correction is probably heading our way.

But there are two reasons to believe the longer-term demand for the leading-edge hardware companies’ products will remain strong. First, it is mind-bendingly difficult, and costly, to do what Nvidia and ASML do. Last year, ASML spent €4bn on research and development, aiming to extend the exponential increase in computing power, known as Moore’s Law, for the next two decades. 

In an interview with the FT last year, Peter Wennink, ASML’s outgoing chief executive, talked about the “exponential increase in complexity” in chip design now needed to keep Moore’s Law alive. That complexity creates enormous barriers to entry in the industry. ASML’s “near-monopoly” position, in the words of one analyst, enables the company to command margins of more than 50 per cent.

Moreover, Altman is probably roughly right about the direction of travel, even if — like everyone else — he will be precisely wrong about the speed of the journey. “We are in an arms race to develop intelligence on a scale we have never imagined before,” says Brett Simpson, partner at Arete Research.

For the moment, we are still in the early, research phase of AI as the tech companies tune their models, says Simpson. But we will soon enter the deployment phase when pretty much every company and government department will seek to adopt AI. “There will be a decade-long investment cycle. We are going to see enormous innovation,” he says. “We have not really started the deployment phase yet and that is when the big gun will go off for investors.”

As ever, the challenge is to distinguish between a routine market cycle and a secular business shift. Whatever the short-term market oscillations, it would be unwise to bet against that shift.

[email protected]

link

Leave a Reply

Your email address will not be published. Required fields are marked *