Mark Zuckerberg, CEO of Meta, revealed plans to spend $60-65 billion on AI resources in 2025 for his social media giant-corporation – and indicated his intention to continue spending for years to come.
This announcement comes only days after OpenAI, a rival model builder, made a similar announcement. In a meeting with US President TrumpStargate, a $500 billion AI infrastructure project, was revealed. It is a joint venture between SoftBank, Oracle and MGX.
Stargate quickly attracted attention. Elon Musk and the US Department of Government Efficiency’s (DOGE) Elon Musk expressed open skepticism about whether OpenAI CEO Sam Altman, and his friends, had the money to make this happen. The OpenAI boss refuted those claims. The OpenAI chief rubbished those claims. History Musk and Altman’s animosity is nothing new, but seeing Musk, a high-profile member of the Trump administration, publicly trash Stargate (a Trump-backed project) like a jealous teen has given us an insight into the stable geniuses of the country.
Microsoft, Google, Amazon, and others have all pledged to spend at least $200 billion on building AI infrastructure in the next year to integrate assistants, generative modeling, and more.
Zuck is clearly interested in this drama. In a He claimed in a Facebook post posted early Friday that his Llama 4 models would become “leading state of the art model.”
Mark. The highly capable and readily available DeepSeek from China, the rise of superlabs like Anthropic or OpenAI, or Google’s persistence with Gemini? Maybe.
Meta, like its competitors, has no problem burning money to get ahead in the artificial intelligence race. According to Zuckerberg’s statement, the Social Network will spend $60-65 billion on CAPEX in order to deploy roughly a gigawatt worth of new computing capacity. This will include more than 1.3 millions GPUs for training and serving models.
Meta told El Reg that it will be a “mix of commercially available GPUs and our in-house silicon,” or MTIA in 2023. Meta told Regit’ll be a”mix of commercially available GPUs and our in-house silicon,”aka MTIA. Meta has used large quantities of Nvidia Hopper and AMD Instinct MI300 series GPUs in the past to train and service its models. It has also developed its silicon to power recommender models but has not announced custom parts for LLMs or diffusion model.
Alongside training and running new model like Llama 4 some of these GPUs are going to power an AI engineer in the future, Zuckerberg insists. “will start contributing increasing amounts of code to our R&D efforts.”
No matter how many GPUs Meta has, it won’t be able to use them if the company doesn’t have the space for them. Meta began building a datacenter with a capacity of two gigawatts and more in Richland Parish, Louisiana, last month to power the accelerators. The facility is so big that Zuckerberg “it would cover a significant part of Manhattan,” boasted and shared an image with the facility superimposed over the island.
- OpenAI wants $500B for AI infrastructure, with the help of friends
- Trump reverses Biden’s AI safety order and other policies
- Biden rails in his farewell address against the tech industry complex and disinformation that is destroying democracy
- Musk shatters $500B Stargate AI plans, Altman responds
A $10 billion facility announcedearlier this month will cover 4 The facility will be built in stages, with construction continuing until 2030.
Unlike Meta, which is searching for a nuclear power provider to run the GPUs, this site will be powered by combined cycle combustion turbine plants that have a total energy production capacity of 2,262 Megawatts.
It seems Zuckerberg will continue his AI shopping spree, whether investors like it or not. “We have the capital to continue investing in the years ahead,” wrote he. “This is a massive effort, and over the coming years it will drive our core products and business, unlock historic innovation, and extend American technology leadership.” (r)
