Zuck tries to justify AI splurge with talk of 'superintelligence' for all

Meta is plowing tens of billions of dollars into GPU bit barns the size of Manhattan Island, and yet The Social Network has struggled to upstage rivals like OpenAI or Anthropic.

So, CEO Mark Zuckerberg is moving the goalpost and refocusing his efforts on a nebulous new target: AI superintelligence.

"Over the last few months we have begun to see glimpses of our AI systems improving themselves," he wrote in a blog post on Wednesday. "The improvement is slow for now, but undeniable. Developing superintelligence is now in sight."

There isn't much consensus on what exactly constitutes AI superintelligence, and despite waxing poetic about how it will improve everything and change the way we create and discover, Zuck's post offers little clarity on the matter.

Regardless of what it actually is, Meta not only aims to build it, but instead wants to give everyone their own personal superintelligence to enrich their lives — after all, who needs friends when you can talk to your buddy Llama.

And unlike all the other superintelligences, Zuckerberg insists Meta's isn't trying to put you off a job and instead will empower users to pursue their individual aspirations.

Personal superintelligence that knows us deeply, understands our goals, and can help us achieve them will be by far the most useful

"Personal superintelligence that knows us deeply, understands our goals, and can help us achieve them will be by far the most useful," the creator of a social network to rate the attractiveness of his Harvard co-eds wrote.

The road to superintelligence is paved in bit barns

The hype-riddled blog post comes just hours before Meta is set to reveal its Q2 earnings results, and as Zuckerberg faces growing scrutiny and concerns among investors and analysts over the company's infrastructure spending and lavish pay packages for its newly formed "superintelligence" team.

As we reported in June, Meta offered one AI researcher an eight-figure payout to join Zuckerberg's team of AI elites. Other reports have said the offers are an order of magnitude or two higher.

Meta has also been pushing ahead with some truly massive infrastructure projects. This includes a 2.2 gigawatt AI supercluster in Richland Parish, Louisiana, which will be deployed in phases over the next five years or so.

Earlier this month, Zuckerberg announced plans to deploy several "multi-gigawatt" datacenters over the next few years, with the first, called Prometheus, coming online in 2026. The bit barn will reportedly pull more than a gigawatt of power.

By the end of the year, Meta aims to have in excess of 1.3 million Nvidia and AMD GPUs churning away, generating tokens and training models.

Of course, none of this is cheap. In 2025 alone, Zuckercorp plans to plow up to $72 billion into capex, primarily to build new bit barns — and that figure could grow even larger by market close as it reports Q2 earnings.

• Datacenter lobby blows a fuse over EU efficiency proposals
• Cisco donates Agntcy project to Linux Foundation in the hope it gets AI agents interacting elegantly
• Stacking up Huawei's rack-scale boogeyman against Nvidia's best
• China proves that open models are more effective than all the GPUs in the world

Impressive investments, underwhelming AI

Yet despite these investments, Meta has struggled to compete with more established players like OpenAI or Anthropic. Back in April, Meta revealed its much-hyped Llama 4 herd of large language models (LLMs). As you may recall, prior to the launch, Zuckerberg predicted that in 2025, Meta AI would serve "more than 1 billion people" and that Llama 4 would be "the leading state of the art model."

Unfortunately for Meta, the first of these models, codenamed Scout and Maverick, which were met with a less-than-stellar reception, made worse by allegations of a benchmark bait-and-switch.

The models were apparently so underwhelming that Meta has reportedly pulled the plug on its biggest and supposedly most capable Behemoth model, which was supposed to weigh in at 2 trillion parameters, and rival OpenAI and Google's biggest models.

During Wednesday's earnings call, analysts will undoubtedly be looking for signs that Zuckerberg's AI spending spree is justified and superintelligence isn't the Metaverse all over again.

Before ChatGPT kicked off the AI boom in late 2022, you may recall Zuckerberg was convinced virtual reality would take over the world. As of Q1, the company's Reality Labs team has burned some $60 billion trying to make the Metaverse a thing. ®