Zuckerberg came out to deny "excess computing power": No one would think there is too much computing power, and the economic value of selling computing power is higher
Zuckerberg denies that Meta has excess computing power. On July 9, local time, Meta released the latest version 1.1 of its large model MuseSpark, and for the first time started charging for its model API (application programming interface). Meta CEO Mark Zuckerberg said in an interview with foreign media that the company will adopt a "radical" pricing strategy to provide one of the most affordable options on the market.
Zuckerberg denies that Meta has excess computing power. On July 9, local time, Meta released the latest version 1.1 of its large model Muse Spark, and started charging for its model API (application programming interface) for the first time. Meta's CEO Mark Zuckerberg said in an interview with foreign media that the company will adopt a "radical" pricing strategy to provide one of the most affordable options on the market. At the same time, Zuckerberg responded to the discussion about excess computing power caused by Meta's recent decision to sell computing power: "Right now, we still have a lot of internal demand for computing power." Zuckerberg said that the company needs as much computing power as possible, but it may be more valuable to rent out some of Meta's AI infrastructure: "Sometimes, the price quoted by others to use these computing power is very high. In this case, it may make more sense to rent computing power or consider such transactions than to stay within the company and use it for ourselves." Zuckerberg noted that the potential for the company to launch a cloud computing business "is certainly always there if we want to do it." He emphasized that this does not mean that Meta’s computing power has been overbuilt or that it has excess idle computing power: “In the entire industry, I have not seen anyone feel that they have too much computing power.” Zuckerberg added that Meta is fully committed to using all the computing power in its inventory. He called launching a cloud business a "fallback" for the company: "Even if for some reason we don't need all the computing power ourselves, there is a huge demand in the market to sell this computing power in the long term like AWS, Azure or Google Cloud." Previously, according to foreign media reports, Meta is formulating a cloud infrastructure business plan to sell AI computing power and model usage rights to external customers. Meta not only plans to sell access to various AI models hosted on the company’s existing infrastructure, but is also considering directly selling the use of raw computing power. Some analysts pointed out that Meta’s logic is similar to that of SpaceX subletting the computing power of its data center Colossus to Anthropic and Google. Zuckerberg wouldn't confirm whether Meta is considering selling access to rival models it hosts, but said: "I think it's something we can do, and I think it's worth considering." Zuckerberg also said that the computing power trading method adopted by SpaceX is very interesting: "It is essentially reaching these short-term transactions that can obtain high premiums. We will also receive a variety of similar offers, and we will evaluate them to see what is reasonable." This pricing is higher than OpenAI’s entry-level model GPT-5 mini and Anthropic’s low-cost Claude Haiku 4.5, but lower than the latter’s high-end model Claude Sonnet 4.6. Zuckerberg said Muse Spark 1.1 has "the most advanced or very close to the most advanced" agent reasoning and tool usage capabilities, and has also been greatly improved in coding. In addition, Zuckerberg said he was satisfied with the current progress of the company's AI department MSL (Super Intelligence Laboratory): "Our overall performance exceeded our own expectations." He admitted that Meta still lags behind some larger AI laboratories, including Anthropic and OpenAI, but the company is about to launch another new model code-named "Watermelon" that can help Meta "touch the forefront of intelligence." Just this week, Meta launched an AI image generator called Muse Image. However, because the model automatically pulls public photos on Instagram account profiles, privacy controversy arose as soon as it was launched. In the past year, Wall Street's attitude towards Meta has been full of contradictions: on the one hand, it is the remarkable technological evolution in large models, and on the other hand, it is the astonishing "speed of burning money." Currently, Meta's 2026 capital expenditure guidance has reached a staggering range of $125 billion to $145 billion.
On the 9th, Meta (Nasdaq: META) stock price rose 4.70% to close at US$631.48 per share, with a total market value of US$1.64 trillion. Since the beginning of this year, the company's stock price has fallen nearly 3%. (
On the 9th, Meta (Nasdaq: META) stock price rose 4.70% to close at US$631.48 per share, with a total market value of US$1.64 trillion. Since the beginning of this year, the company's stock price has fallen nearly 3%. (