Unlock the Editor’s Digest for free
Roula Khalaf, editor of the FT, selects her favorite stories in this weekly newsletter.
Nvidia CEO Jensen Huang has long claimed that the world’s data centers will have to be completely renovated to meet the demands of generative AI. He says it will cost $1 trillion over the next four to five years, essentially doubling the amount already sunk into the digital infrastructure, to train and use the new AI models.
Nvidia itself has been the most obvious beneficiary of this. The stock market’s dizzying rise has made it the most valuable company in the world.
But a flurry of earnings announcements and AI-related deals over the past two weeks have also provided encouraging evidence that the boom sparked by the launch of OpenAI’s generative AI chatbot ChatGPT is spreading. Of course, there’s no telling whether the spending spree will be sustainable or big enough to justify the massive rise in tech stocks, but it has at least brought some comfort to the bulls.
For example, shares of chipmaker Broadcom are up more than 20 percent since the last AI-driven sales surge. Another such increase would join the rare group of tech companies valued at more than $1 trillion, more than six times what they were half a decade ago.
Much of the increase comes from demand for AI accelerators, the chips that Broadcom custom designs for customers like Google to speed up their AI calculations. But the increasing growth also points to the more important role that high-speed networks have come to play in data centers.
The enormous amount of information needed to train and run AI models requires much faster connections between individual processors, as well as between the different machines running in data centers. Broadcom CEO Hock Tan estimated that networking will account for 40 percent of his company’s AI chip sales by the end of this year.
Meanwhile, shares in software maker Oracle, a company late to the cloud, rose 17 percent on news of a deal to train OpenAI’s large language models on its cloud infrastructure. The deal will see OpenAI’s partner Microsoft’s Azure cloud service brought to a massive new Oracle data center – a relationship between two of the tech industry’s oldest enemies that would once have been unthinkable.
Elsewhere, even Hewlett Packard Enterprise, which seemed to be missing out on the growing demand for AI servers that seemed to be pushing rivals Supermicro and Dell to the next level, finally got a break on Wall Street. Shares rose 24 percent after the earnings report, as investors reassessed their position in the AI boom.
Now that news like this has fueled hopes that the AI boom will spread to more vendors, several things are becoming clear. One is that the impact appears to be widespread and includes many different parts of the ‘tech stack’: the hierarchy of components, from chips to software, needed to run today’s complex IT systems.
Nvidia is still positioned to be the biggest winner by far. The majority of sales do not come from individual chips, but from entire servers, often connected in complete racks. Squeezing out the best performance comes from adapting every element of these systems to work together, using Nvidia’s proprietary technologies in areas like networking
Nvidia’s biggest customers are desperate to reduce their dependence on the company and are pushing for new standards in everything from networking to AI software, which could create more competitors. But those initiatives will take time.
The largest tech companies are also expanding their direct involvement in more parts of the infrastructure that AI requires. For example, a key part of Apple’s AI announcement last week involved the news that it is designing its own servers, reportedly based on internal chip designs. Apple has already taken control of most of the key components of its mobile phones: a similar move is likely in the data center, as the demands of AI force the company to bring more of the processing of its customers’ data back to its own facilities.
One result of these types of moves is that suppliers’ business models have had to adapt, leading companies like Broadcom to play a supporting role as customers gain more control. The so-called ‘hyperscalers’ – the largest cloud companies – will also account for a larger share of total demand, leading to dependence on a smaller number of large customers. This will increase the vulnerability of suppliers in the event of a recession. But for now, Wall Street is fixated on how many technology boats will be lifted by the rising wave of generative AI.
richard.waters@ft.com