Did you know Meta’s CEO Will Bring Company’s AI Vision for 2025 to Life With Mega $65 Billion Investment

 

Mark Zuckerberg has shared with the world how he feels 2025 is going to be the year for AI. This is why he made the tech giant’s plans for a major investment worth $65 billion public.

The news is a major rise from that witnessed in last year’s budget which stood at $38 billion. As per Zuckerberg, a huge chunk of the investment will go to designing more data centers for the company. He added how it was crucial to produce more data centers to give Meta the computing it needs to design more AI-related products.

The initiative as per Meta is not only linked to building more infrastructure but also related to more innovation and ensuring tech leadership in America. Through such a strategy, Meta hopes to acquire more than 1.3M GPUs by this year’s end.

Meanwhile, in other news, another top player from China has shared more about how it’s rolling out more cost-effective models dubbed DeepSeek-V3 and DeepSeek-R1. They ended up outperforming so many other top models from OpenAI and Meta. For this reason, it’s getting a lot of necessary attention for its reduced cost and more effective capabilities.

Deep-Seek underwent training using just 2048 GPUs with a cost hitting only $6 million. This is a fraction of what other top models need. In contrast, tech giant Meta needs a staggering $60 million and 30.8 million GPU hours for its Llama models. Let’s not forget how these models are open source which means they’re very much like Llama so anyone has the chance to run them through hardware.

The API pricing for DeepSeek is also much less than other leading AI rivals like OpenAI. For instance, DeepSeek only charges $0.14 for one million tokens of input when compared to that seen for OpenAI which charged $7.5. Now the question remains, what is it that makes DeepSeek so cheap in design?

The model makes use of the MoE framework which gives it the chance to activate a single portion of its parameters while processing. This gives the company more efficiency and less computational requirements when compared to other classic models such as Llama. The latter might not be able to make use of such techniques as efficiently.

Other than that, models like o1 from OpenAI make use of supervised fine-tuning. As per DeepSeek, it also uses reinforcement learning which can give rise to advanced reasoning capabilities on its own. Hence, it’s going to be interesting to see how top AI firms compete to give rise to better and more effective models this year.