Data centers have long ceased to be just warehouses of servers - now these are factories that literally sell digital "tokens", and it is around the acceleration of their production Nvidia built a presentation of its new platform Vera Rubin.
Last week, the company invited journalists to a closed briefing and showed a mini-date center in Silicon Valley, one of four such Nvidia laboratories. They demonstrated that the assembly of the Vera Rubin NVL72 computing unit now takes one minute instead of 90 minutes, which were required to build the previous generation GB200. This acceleration was achieved by automating the installation of equipment.
The Vera Rubin platform combines six types of chips: Vera processor, Rubin graphics accelerator, NVLink 6 switch, ConnectX-9 network adapter, BlueField-4 data processor and Spectrum-6 Ethernet switch. According to Nvidia, in the CoreWeave tests on the DeepSeek-R1 model, the new system produces ten times more tokens per watt of power than the previous generation of the GB200 NVL72.
The company’s bet is made on the fact that for the work of autonomous AI agents, not the number of computational cores is more important, but their speed and ability to hold a long context with a minimum latency. According to Nvidia, the Vera processor speeds up the operation of agents by half and reduces delays by 40 percent thanks to the memory of the LPDDR5X.
The economy of this approach is simple: the more tokens the data center produces within the framework of fixed energy consumption, the higher its income. But the growth in the number of equipment manufacturers and competition among AI companies simultaneously reduce the market price of the tokens themselves, so the industry expects that the fall in price is offset by the growth of demand.
The Nvidia spokesperson cited the calculation in which the use of AI in software development tripled the productivity of programmers around the world, estimating the effect of nine trillion dollars. However, this is rather an illustrative assessment than a confirmed fact, and against the background of reports that some companies limit the limits on the use of AI-tokens, the real effect for all organizations remains in question.
Simultaneously with the presentation, the data centers remain a point of social tension: just the day after the briefing, Dutch activists threw a data center under construction, designed for Microsoft infrastructure, water balls with a chemical mixture in protest against the climatic and political consequences of the development of such facilities.
Last week, the company invited journalists to a closed briefing and showed a mini-date center in Silicon Valley, one of four such Nvidia laboratories. They demonstrated that the assembly of the Vera Rubin NVL72 computing unit now takes one minute instead of 90 minutes, which were required to build the previous generation GB200. This acceleration was achieved by automating the installation of equipment.
The Vera Rubin platform combines six types of chips: Vera processor, Rubin graphics accelerator, NVLink 6 switch, ConnectX-9 network adapter, BlueField-4 data processor and Spectrum-6 Ethernet switch. According to Nvidia, in the CoreWeave tests on the DeepSeek-R1 model, the new system produces ten times more tokens per watt of power than the previous generation of the GB200 NVL72.
The company’s bet is made on the fact that for the work of autonomous AI agents, not the number of computational cores is more important, but their speed and ability to hold a long context with a minimum latency. According to Nvidia, the Vera processor speeds up the operation of agents by half and reduces delays by 40 percent thanks to the memory of the LPDDR5X.
The economy of this approach is simple: the more tokens the data center produces within the framework of fixed energy consumption, the higher its income. But the growth in the number of equipment manufacturers and competition among AI companies simultaneously reduce the market price of the tokens themselves, so the industry expects that the fall in price is offset by the growth of demand.
The Nvidia spokesperson cited the calculation in which the use of AI in software development tripled the productivity of programmers around the world, estimating the effect of nine trillion dollars. However, this is rather an illustrative assessment than a confirmed fact, and against the background of reports that some companies limit the limits on the use of AI-tokens, the real effect for all organizations remains in question.
Simultaneously with the presentation, the data centers remain a point of social tension: just the day after the briefing, Dutch activists threw a data center under construction, designed for Microsoft infrastructure, water balls with a chemical mixture in protest against the climatic and political consequences of the development of such facilities.
