Everyone is developing chips that compete with NVIDIA’s. They are in the wrong race

Qualcomm advertisement on Monday that it is working on AI accelerator chips, which means there will be new competition for NVIDIA. The company that dominates the AI ​​hardware landscape is seeing a large group of competitors try to erode that position, but the problem for all of these companies is not the chips, but something else. A CUDA call.

what has happened. Qualcomm has announced the AI200 chip, which will begin selling in 2026, and the AI250, which will do so in 2027. Both will be able to work in rack-type systems that have liquid cooling. Qualcomm servers may have up to 72 chips based on the Hexagon NPUs of the company’s Snapdragon SoCs.

Inference yes, training no. The company has revealed that its chips focus on inference (the execution of AI models) and not training. Their rack-based systems will have lower operating costs than cloud system providers, Qualcomm says. Each rack consumes 160 kW, a figure comparable to the consumption of some racks based on NVIDIA GPUs. There are no details about the price of these chips, the cards or the racks that will integrate them, nor about how many NPUs can be offered in each rack. What we do know is that Qualcomm’s accelerator cards will support up to 768 GB of memory, more than what NVIDIA or AMD offer in their current models. according to CNBC.

Chips for third parties. The other important point is that Qualcomm will sell its AI chips and other components separately, allowing large AI companies to “customize” their own racks based on Qualcomm chips. It is an identical philosophy to the one they have adopted in the world of their mobile SoCs. Investors viewed the news with exceptional optimism, and Qualcomm shares rose 11% in Monday’s session.

NVIDIA dominates with an iron fist. In the AI ​​chip segment, the king is NVIDIA. The company is the absolute protagonist of this market and according to CNBC it maintains a 90% market share, which has allowed it to skyrocket its valuation to 4.5 trillion dollars. That dominance could now be threatened by the avalanche of chips that are arriving from various manufacturers.

All against NVIDIA. AMD has its excellent Instinct, Google has your TPUsAmazon their TrainiumMicrosoft their Maia and Huawei has your Ascend. All of them make really striking proposals for NVIDIA chips, and little by little these solutions are being integrated into more and more data centers. But the real problem is not in the hardware, but in the software.

The great challenge is to defeat CUDA. The de facto standard in the AI ​​industry that developers use It’s CUDAa platform that allows you to take full advantage of the capabilities of NVIDIA chips in the field of artificial intelligence. This hardware+software combination is much more mature than that of its competitors, who have the hardware part resolved (or are on the right track) but do not have a platform comparable to CUDA. AMD has ROCmwhich is especially interesting because it is Open Source, but at the moment its features still do not reach those of CUDA.

Reinvent the wheel? CUDA has been on the market for almost two decades, which means that the majority of academic research and pioneering models—such as ImageNet—were written for CUDA. It is not a language, it is a vast collection of libraries, optimized frameworks (like cuDNN), debugging tools and a huge community. Developing a competitor is basically like reinventing the wheel, and migrations are expensive and companies and startups will not have an easy time assuming it.

China is also in the fight. And of course, if there is another great protagonist in this race, it is China. The Asian giant, previously dependent on NVIDIA, is seeking to get rid of this manufacturer, and along with the development of advanced AI chips they are also trying to have its own AI software to surpass CUDA.

In Xataka | AI is the best thing happening to nuclear fusion. The construction of ITER is already accelerating