Okay; let's say that by launching new chip factories in the US,
Posted: Tue Jan 21, 2025 9:21 am
Accordingly, to support the hardware for such a staggering volume of calculations, it will be necessary to manufacture at least tens of times more chips suitable for AI calculations than are made today. Which is extremely difficult from a purely production point of view: practically all graphic accelerators - both for PCs and for servers - are currently manufactured for the entire world by a single Taiwanese company TSMC, commissioned by Nvidia, AMD and Intel, and the potential for increasing its production capabilities is by no means limitless. This refers to limitations in the availability of natural resources (the production of chips is an extremely energy- and water-intensive enterprise), and human resources (training qualified engineers and even workers for the semiconductor industry is an especially expensive and time-consuming business), and, of course, purely financial ones.
South Korea, Vietnam and mexico whatsapp resource other countries, TSMC, Intel, Samsung Electronics and other global chip suppliers will be able to provide enough hardware by 2030 to meet customer demand for AI computing. However, these calculations still need to be done, which also requires energy. Analysts point out that a standard server rack today consumes from 5 to 15 kW, while a rack densely packed with powerful video cards for solving AI problems already consumes under 100 kW, and in some cases even closer to 150 kW.
Over the next three to five years, the power consumption of one such rack threatens to grow to 250, if not 300 kW, even despite the fact that future graphics processors should clearly be more energy-efficient than current ones. There is no logical contradiction here: generative AI models scale perfectly, and the more complex they are, the more (with appropriate adequate training, of course) the results they produce correspond to customer requests. Therefore, along with the increase in the capabilities of video cards, the hardware appetites of AI models will also grow, which will lead to a surge in the power consumption of a single server rack. And how many such racks does a cloud AI provider need to satisfy all customer requests - hundreds, thousands, tens of thousands?
Collision with reality
According to McKinsey, by 2030, the world should have 3 to 9 new chip-making factories capable of processing up to 15 million silicon wafers per year using the technological standards of "7 nm" and less. And this is only for logic chips (graphics processors, neural processors): the production of memory chips will require another 13 to 21 million wafers - and, accordingly, 5 to 18 factories processing them. And this is not even mentioning the NAND chips for flash drives, which will store the generative models themselves and the results of their processing of user requests, as well as other necessary components for assembling the corresponding servers (service chips for motherboards, all kinds of controllers, power elements, power supplies, fans, etc.)
In principle, chip makers have a certain reserve: “thanks” to the restrictions of the US Department of Commerce regarding microprocessor production in China, similar factories in Taiwan, South Korea and other countries gravitating towards the US have alreadyare experiencing a clear underutilization of production capacities, which reaches 40% for microcircuits manufactured according to mature technological standards. In the fight for IT sovereignty, mainland China localizes the production of chips as much as possible, starting, of course, with service ones, manufactured according to standards that are formally outdated today - thereby freeing up quite significant capacities for chipmakers outside its borders.
South Korea, Vietnam and mexico whatsapp resource other countries, TSMC, Intel, Samsung Electronics and other global chip suppliers will be able to provide enough hardware by 2030 to meet customer demand for AI computing. However, these calculations still need to be done, which also requires energy. Analysts point out that a standard server rack today consumes from 5 to 15 kW, while a rack densely packed with powerful video cards for solving AI problems already consumes under 100 kW, and in some cases even closer to 150 kW.
Over the next three to five years, the power consumption of one such rack threatens to grow to 250, if not 300 kW, even despite the fact that future graphics processors should clearly be more energy-efficient than current ones. There is no logical contradiction here: generative AI models scale perfectly, and the more complex they are, the more (with appropriate adequate training, of course) the results they produce correspond to customer requests. Therefore, along with the increase in the capabilities of video cards, the hardware appetites of AI models will also grow, which will lead to a surge in the power consumption of a single server rack. And how many such racks does a cloud AI provider need to satisfy all customer requests - hundreds, thousands, tens of thousands?
Collision with reality
According to McKinsey, by 2030, the world should have 3 to 9 new chip-making factories capable of processing up to 15 million silicon wafers per year using the technological standards of "7 nm" and less. And this is only for logic chips (graphics processors, neural processors): the production of memory chips will require another 13 to 21 million wafers - and, accordingly, 5 to 18 factories processing them. And this is not even mentioning the NAND chips for flash drives, which will store the generative models themselves and the results of their processing of user requests, as well as other necessary components for assembling the corresponding servers (service chips for motherboards, all kinds of controllers, power elements, power supplies, fans, etc.)
In principle, chip makers have a certain reserve: “thanks” to the restrictions of the US Department of Commerce regarding microprocessor production in China, similar factories in Taiwan, South Korea and other countries gravitating towards the US have alreadyare experiencing a clear underutilization of production capacities, which reaches 40% for microcircuits manufactured according to mature technological standards. In the fight for IT sovereignty, mainland China localizes the production of chips as much as possible, starting, of course, with service ones, manufactured according to standards that are formally outdated today - thereby freeing up quite significant capacities for chipmakers outside its borders.