Openai wants to hit the table in the race to dominate artificial intelligence. At the beginning of September we knew that “a mysterious client” made an investment in Broadcom of 10,000 million dollars for the design of AI chips. Through sources close to the company, we learned that it was OpenAI, with the intention of form a strategic alliance with both Broadcom and TSMC for the manufacture of own chips. Called ‘Xpus’, they have a clear objective: stop depending on Nvidia, something that Many great technology want.
Have independence in AI is paid expensive. For Openai, depending completely on Nvidia to feed chatgpt and its future models is a capital risk, especially for having to deal with astronomical costs, and being limited to the technological evolution and stock of a company like NVIDIA. Creating your own chips allows OpenAI to optimize hardware specifically for your language models, reduce operating expenses and free yourself from the bonds of a single supplier. A strategy on the Apple line when He said goodbye to Intel to start making their own chips in their Mac. Although stop depending on Nvidia today is complicated.
The golden trio behind the XPUS. Broadcom Lead the design Of these specialized processors, contributing their experience in custom chips for technological giants. The company controls approximately 70% of the Custom IA processors market and has already collaborated with Google in His tpus For years. TSMC, the world’s largest semiconductor manufacturer, will be responsible for producing chips using their advanced 3 nanometers technology. Mass production would predictably start in 2026with the first shipments arriving that same year.
What are exactly the Xpus. These chips are not adapted gpus for AI, which is what is broadly marketing nvidia, but architectures designed from scratch With a single purpose: accelerate language models. While the NVIDIA GPUSs were born to rendering graphics and then adapted to artificial intelligence, OpenAi XPUS are specifically conceived for training and inference tasks of AI. They incorporate systemic matrices, high bandwidth (HBM) and integrated network capabilities, similar to the most advanced NVIDIA processors but optimized for the specific needs of OpenAI.
The rebellion against Nvidia has several fronts. Openai is not alone in his efforts to break the hegemony of Nvidia. Google has been developing its TPUS (Tensioner Processing Units), which already go for its seventh generation. Amazon created its chips Trainium and Inferentia For AWS. Microsoft designed Maia For Azure. Goal collaborates with AMD In alternative solutions. Even from China, Huawei threatens face face In the next three years. It is clear that all large technological ones want to have greater control of their products, and for that they must go down to the chip supply chain to start rethinking the scheme.
Nvidia’s real pit. Despite how the panorama is, Nvidia preserves An ace in the manga That will not be easy to overcome: CUDA. This development platform has become the industry standard for more than a decade. Virtually all the researchers and developers of the CUDA program, which generates a very powerful network effect. Changing to a different architecture not only implies acquiring new chips, but rewriting software, removing equipment and, in many cases, starting from scratch. This technology is the one that keeps Nvidia in a privileged position even before the avalanche of competitors.
A dual strategy. The paradoxical of the case is that Openai is not abandoning Nvidia. Parallel to this own chips project, the company maintains An agreement of 100,000 million dollars With the Santa Clara giant for its Stargate project, which plans to build huge data centers until 2028. On the one hand, the calculation power supply with NVIDIA is ensured, while on the other they develop their long -term alternative. The idea is that XPUS are initially used internally, especially for inference (apply already trained models), while NVIDIA GPUS will continue to be necessary for training the most demanding models.
The board is removed. Broadcom is clearly reinforced from this movement. The announcement of your agreement with Openai He shot his actions In the stock market and consolidate its position as a preferential partner for companies that seek custom chips. TSMC also wins: each new wave of specialized chips reinforces its indispensable role as a world reference manufacturer. Nvidia, on the other hand, saw his price slightly fall after knowing the news, although he continues to maintain A dominant position Very difficult to collapse. In addition, AMD could benefit if smaller companies, unable to develop their own chips, look for more accessible alternatives than NVIDIA.
Cover image | NVIDIA, VECTEEZY, Village Global
GIPHY App Key not set. Please check settings