Nvidia has a very strong and not quite unexpected ally against Depseek: TSMC

Deepseek has convulsed the industry of the artificial intelligence (AI). The irruption of this free and open source Chinese model has questioned the need to use in training and inference processes Chips for very powerful and expensive. Like those designed by Nvidia, which leads with an unappealable roundness the hardware market for Ia. Even so, just ten days ago Deepseek caused Its market value fell abruptly.

Since then the hardware used by this Chinese company has generated a lot of distrust. The Deepseek responsible argue that the infrastructure they have used to train their agglutin model 2,048 chips H800 of Nvidia. And that this process with 671,000 million parameters has cost 5.6 million dollars. However, some analysts defend that these figures do not reflect reality.

The very juicy report prepared by SEMIANALYSIS He maintains that, in reality, the infrastructure used by Deepseek to train his AI model approximately 50,000 NVIDIA GPU with Hopper MicroAritecture. According to Dylan Patel, AJ Kourabi, Doug O’Laughlin and Reyk Knuhttsen, at least 10,000 of these chips are GPU H100 of Nvidia, and at least another 10,000 are GPU H800. The remaining chips, according to these analysts, are the cuts cut H20.

The TSMC cowos packaging is a very strong support for Nvidia

As we have just seen, at the current situation it is reasonable to have doubts about the hardware that Deepseek has used in the training of its model (in inference seems to be using the GPUs Ascend 910c of Huawei). And also about the real cost of this process, which could be much higher than the officially announced by this Chinese company. In any case, Depseek has poured uncertainty about the AI ​​market, and this is the reason why so many US technology companies have lost value.

Whatever Nvidia has just received a very strong support from TSMC, which is the semiconductor manufacturer that produces its GPU. And is that, according to Digitimes Asiathis Taiwanese company has decided launch an expansion plan five years long for integrated circuit manufacturing capacity using its advanced cowos packaging technology (Chip-on-Wafer-on-Substrate). According to Beth Kindigof the I/O Fund consultant, this technology will monopolize between 50 and 60% of the market in 2025 compared to 15% it supported during 2024.

The high demand for GPUs for AI with Blackwell MicroAritectura de Nvidia is largely responsible for the implementation of this plan. The company led by Jensen Huang can respond better to the needs of its customers and will see how its competitiveness is increased in a phase in which Depseek and other Chinese companies represent a challenge. In March 2024 TSMC officially announced which was building two cowos packaging plants in the town of Chiayi, housed in southern Taiwan.

However, this is not all. He also shuffled the option to put a plant more specialized in this advanced packaging technology in Japan, presumably on the island of Kyushu, in which This company is building at the moment two semiconductor production plants of avant -garde. In any case, there is something else. And it is that Chiayi plants will be trained to work, in addition to the packaging cowos, with advanced Info and Soic technologies (System on Integrated Chips).

It is evident that TSMC wants to cover your back well and look to the future to prevent its production capacity from being threatened by a bottleneck. An interesting note: currently the Cowos packaging is being used with the AMD Instinct Mi250 chips and with the A100, H100, H200, B100 and B200 NVIDIA GPUs, as well as in its derivatives. The review used in these last two chips, the B100 and B200, is known as Cowos-L. In 2025 TSMC will be able to process no less than 60,000 wafers per month using its advanced packaging technology.

Image | TSMC

More information | Digitimes Asia | Yahoo! Finance

In Xataka | Samsung is preparing to give TSMC a bars where it hurts most: the manufacture of the chips for ia