Developments in artificial intelligence (AI) are increasing the demand for computing power, which in turn intensifies competition among chip manufacturers. Recent events in this field are further highlighted by a new agreement signed between Nvidia and Meta.
According to Redaksiya, the multi-year agreement between Nvidia and Meta involves the social media giant purchasing billions of dollars worth of Nvidia chips to power its large-scale infrastructure projects. This agreement will also include Nvidia's processors (CPUs). Meta had previously estimated it would acquire 350,000 H100 chips from Nvidia by the end of 2024 and possess a total of 1.3 million GPUs by the end of 2025.
As part of the new agreement, Nvidia stated it would build hyperscale data centers optimized for both training and inference to support the company's long-term AI infrastructure roadmap. This includes a large-scale deployment of Nvidia's CPUs and millions of Nvidia Blackwell and Rubin GPUs. Meta became the first tech giant to widely adopt Nvidia's Grace CPU as a standalone chip. Nvidia also emphasizes its offering of technology that combines various chips.
This move indicates Nvidia's recognition of the importance of running an increasing number of AI workloads on CPUs. Analysts note that the acceleration in CPU usage is related to AI training and inference support. However, CPUs are still only one component of the most advanced AI hardware systems. The number of GPUs Meta is acquiring from Nvidia still significantly outweighs CPUs. Meta plans to substantially increase its spending on AI infrastructure this year, allocating between $115 billion and $135 billion for this purpose.
Nvidia has long stated that its hardware can be used for inference needs in addition to cutting-edge AI training. In December, Nvidia announced it had spent $20 billion to license technology from chip startup Groq and attract Groq's top experts, including CEO Jonathan Ross, to Nvidia. This acquisition was Nvidia's largest to date.
Nvidia's deal with Meta comes at a time when prominent AI labs and multi-trillion dollar software companies are seeking to diversify their sources of computing power. OpenAI, Anthropic, Meta, XAI, and many others have relied on Nvidia hardware to train and deploy generative AI models over the past few years. Currently, in many cases, they are building or customizing their own chips, which encourages Nvidia to offer more services. Microsoft uses a mix of Nvidia GPUs and custom-designed chips for its AI cloud services. Google also uses Nvidia chips for its cloud services but primarily relies on its in-house Tensor Processing Units (TPUs). Anthropic uses Nvidia GPUs, Google's TPUs, and chips from Amazon (one of its main shareholders) for its Claude AI models. OpenAI is working with Broadcom to build its own AI chip hardware and networking systems. OpenAI has also signed a deal with AMD, under which it plans to purchase up to 6 gigawatts of chips from AMD over the next few years. OpenAI also plans to add "750MV ultra-low latency AI compute" to its platforms using technology from Cerebras.