Zuckerberg Denies Meta Compute Overcapacity: Market Demand Is Very High

Meta (META-US) was recently reported to be planning to rent out idle computing capacity, sparking market concerns about compute overcapacity and briefly dragging down global chip stocks. In response to these 'compute overcapacity' concerns, Meta CEO Mark Zuckerberg has publicly denied such claims in a recent media interview.

Zuckerberg stated that Meta needs as much computing power as possible. However, amid the current tight market environment for AI product operations and development, he is also considering whether renting out part of Meta's AI infrastructure could create higher value.

He said, 'External markets are offering very high prices for computing power, so in some cases, renting out compute or adopting similar partnership models may be more reasonable than reserving it entirely for internal use.'

Earlier this month, media reports indicated that Meta is planning to develop a cloud business, aiming to generate direct revenue through its data centers and other compute collaboration initiatives.

In response, Zuckerberg said, 'The potential to build a cloud business certainly exists if we choose to do so.'

However, he emphasized that this does not mean Meta has overbuilt or is experiencing compute overcapacity. 'I don't know of anyone in the industry who thinks they have excess compute,' he said, adding that Meta is currently making full use of all its computing resources.

According to reports, Meta is evaluating multiple cloud service models. One proposal resembles Amazon's Amazon Web Services (AWS) Bedrock platform, offering developers access to AI models deployed on Meta's infrastructure. Meta would operate the data centers and AI chips supporting the models and charge customers based on usage.

Last Thursday, Meta announced it is offering its latest AI model, Muse Spark 1.1, to developers via API—an additional example of a cloud business model.

Under this model, Meta runs Muse Spark 1.1 on its own AI infrastructure and charges developers based on the number of tokens used.

While Zuckerberg did not confirm whether Meta would allow competitors' AI models to be deployed on its servers and made accessible in the future, he said, 'I think this is absolutely something we could do and worth considering.'

Additionally, reports indicate Meta is planning to launch a 'bare metal compute' rental service, similar to those offered by AI cloud computing providers like CoreWeave, directly selling computing resources to enterprises.

Zuckerberg said, 'Even if, for some reason, we don't need all our compute in the future, or for other considerations, I believe there is still very strong market demand. We could absolutely sell compute through long-term contracts, just like AWS, Microsoft Azure, or Google Compute.'

Zuckerberg also mentioned he is very interested in recent moves by SpaceX, led by Elon Musk.

SpaceX is leasing computing capacity from its large Memphis data center to Anthropic and has also partnered with Google. Market estimates suggest this strategy could help SpaceX's AI business generate over $50 billion in revenue by 2028 and exceed $100 billion by 2030.

Zuckerberg said SpaceX's model of short-term, high-premium compute leasing is 'very interesting,' and Meta is currently receiving numerous similar partnership proposals, which the company will evaluate one by one to select the most suitable collaboration model.

Nonetheless, he emphasized, 'For now, this compute still has extensive internal use cases.'

On the other hand, a report last Thursday revealed an internal Meta memo indicating the company plans to begin mass production of its in-house AI chip starting in September 2024.

According to the plan, Meta aims to increase its total computing capacity to 14GW by 2027—approximately double its deployment scale in 2024.