Meta has placed what may be the largest order yet for NVIDIA H100 GPUs — the most in-demand AI processors on Earth. A record-breaking 350,000 GPUs are allocated for Meta AI data centers, at a total cost in the billions of dollars. It could take up to a year to fulfill, but when 2025 rolls around Meta will be in a front-running position for the AI race. CEO Mark Zuckerberg’s goal of open-source AGI is now within reach.

Is Meta building the world’s biggest AI data center?

Meta has 350,000 NVIDIA H100 GPUs inbound, probably forming the largest AI data center in the world. To put this into perspective, each unit can cost around $35,000 – $40,000 if bought individually on the secondary market. The total cost to Meta is estimated to be roughly $10.5 billion. This is a ‘reasonable’ wholesale cost, roughly $3 billion less than if bought on the secondary consumer market.

For the low price of someone’s annual salary, you’ll get an 80GB GPU for PCIe5.0 capable of 26 teraFLOPS (FP64). Pulling the most impressive number, you’ll see a wildly different figure of 3,026 teraFLOPs. This figure is “with sparsity”, and specifically refers to FP8 tensor core performance of the standard PCIe model. The higher bandwidth H100 SXM model increases this figure to an almost unheard-of 3,958 teraFLOPS.

Some services will allow you to ‘rent out’ a H100 or H200, at an optimistic cost of $2/hour+. Given the extreme corporate demand for the processor, this is the only way that smaller businesses will be able to get their hands on one. Enterprise data centers like those of Meta AI will benefit from the NVIDIA AI Enterprise initiative. The social media giant is an NVIDIA HGX H100 Partner with NVIDIA-certified systems.

Essential AI Tools

Meta is creating open-source AGI ‘for all’

Speaking via the Threads platform, CEO Mark Zuckerberg states that “it’s become clearer that the next generation of services requires building full general intelligence”. Here, the Facebook founder is referring to artificial general intelligence (AGI). This is a level of computer intelligence equal to human intelligence. Of course, computers are already much faster at processing data than we humans and can retrieve a far greater variety of information than any person could possibly remember.

This processing speed and information retrieval are not what defines AGI. ‘Full general intelligence’ is something that we have not yet achieved. It is the simulation of intelligence, thinking in ‘the same way’ as a human, with all its inferences and biases. An AGI will be proactive in telling us information that we don’t even know that we don’t know — something that traditional search engines cannot do.

It goes without saying that most businesses cannot afford to compete with a $10.5 billion hardware investment. This may even put them ahead of their leading competition — OpenAI and Microsoft — at least in terms of H100 acceleration.

When will LLaMA 3 be released?

The open-source AI most likely to emerge first from this world-leading hardware advantage could be LLaMA 3. The next generation of Meta AI models is expected to become available at some point this year, but there is no official release date for LLaMA 3 just yet. One thing’s for sure — 350,000 AI accelerators will bring this forward substantially.

About the Author

Steve Hook

Steve is an AI Content Writer for PC Guide, writing about all things artificial intelligence. He currently leads the AI reviews on the website.