Omdia’s research allows us to estimate the quantities in which Nvidia’s major customers purchased computing accelerators this year. Since only accelerators of the Hopper generation are currently being mass-produced, we will talk about them, without taking into account future Blackwells. Microsoft turned out to be the largest buyer of Hopper accelerators.
Image source: NVIDIA
According to a source cited by the Financial Times, Microsoft purchased 485,000 Nvidia Hopper generation accelerators this year. The closest US company that has managed to get closer to Microsoft is Meta✴ Platforms, but its 224,000 Hopper accelerators are more than twice as fast as Microsoft’s result. The latter increased purchases of Nvidia Hopper accelerators more than three times compared to last year.
It is noteworthy that the second and third places in the world ranking of Nvidia’s largest clients were the Chinese ByteDance and Tencent, which, even under American sanctions, managed to order approximately 230,000 Hopper generation computing accelerators, including those modified to take into account H20 export restrictions.
The exact price of Hopper chips is unknown, but it is generally accepted that one Nvidia H100 accelerator costs about $30 thousand. Thus, Microsoft could spend at least $14.5 billion on Nvidia accelerators themselves alone. This is more than the annual budget of many countries; for example, the budget of Belarus for 2024 was about $13.2 billion.
Microsoft, with such scale of purchases, can create advanced computing infrastructure not only for itself, but also for its customers. OpenAI, which receives full support from Microsoft, used the Microsoft Azure cloud infrastructure to train its advanced o1 model.
In fifth place among the largest buyers of Nvidia computing accelerators were Elon Musk’s Tesla and xAI, they purchased about 200,000 Hopper accelerators. In sixth place was Amazon with 196,000 accelerators, and in seventh place was Google with 169,000 Nvidia accelerators. It is worth recalling that Amazon, Google and Meta✴ are also developing their own computational accelerators, including specialized processors. Google and Meta✴ each installed approximately 1.5 million of these processors in their own systems this year. Amazon is slightly behind them with 1.3 million of its own chips. From this point of view, Microsoft is more dependent on Nvidia, since the company has installed no more than 200,000 of its own processors from the Maia family this year. AMD products are not so popular against this background, but Meta✴ this year purchased 173,000 Instinct MI300 accelerators, and Microsoft purchased about 96,000 units.
In total, technology sector companies spent a combined $229 billion on building their server infrastructure this year, with about 60% of all investments distributed among the ten largest market players. Nvidia received about 43% of the companies’ funds aimed at developing their computing server infrastructure. Most likely, this value will be the peak for it, since competition in the market for computing accelerators does not stand still.