Zuckerberg says Meta is 'very close' to substantially stronger AI models

Meta CEO Mark Zuckerberg said the company is 'very close' to having substantially stronger AI models, though he did not name the models or provide a release schedule. Meta expects 2026 capital expenditures of $130 billion to $145 billion, including finance-lease principal payments, and spent $31.08 billion in the second quarter alone. Zuckerberg's comments came in an interview about his essay 'The Future is for Everyone,' outlining Meta's strategy of broad distribution and open models.

"We are very close to having substantially stronger models," Mark Zuckerberg @finkd https://x.com/finkd/with replies told Alex Heath in an interview published by sources.news https://sources.news/p/zuckerberg-pro-ai-manifesto . The Meta founder did not identify the models, give a release schedule or provide independent evaluations. But his claim puts a near-term marker on Meta's effort to close the model gap with OpenAI, Anthropic and Google after committing extraordinary sums to AI infrastructure. Meta expects 2026 capital expenditures of $130 billion to $145 billion https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/default.aspx , including finance-lease principal payments. The company spent $31.08 billion during the second quarter alone, while free cash flow fell to $784 million as its infrastructure buildout accelerated. Zuckerberg, who founded Facebook in 2004 after studying computer science at Harvard https://investor.atmeta.com/leadership-and-governance/person-details/default.aspx , made the prediction while discussing a 6,500-word essay he had spent the past month or two writing. He said he wanted Meta's values understood before the stronger systems arrive. That timing makes the interview more than a forecast about model performance. Zuckerberg is preparing customers, policymakers and investors for a release strategy built around broad distribution, open models and limited delays, even as more capable systems raise concerns about cyber activity and other risks. Meta has already resumed releasing some open models. The company published the weights for Muse Glimmer https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model , a 30-billion-parameter agentic model designed to run locally within a 24 GB or 32 GB memory envelope. Meta says the model exceeds similarly sized Gemma and Qwen systems on several selected evaluations, though those comparisons remain the company's own tests. Muse Glimmer followed Muse Spark 1.1 https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/ , released July 9 with a public preview of Meta Model API. On August 5, Meta introduced Muse Spark 1.2 and the Muse Code terminal agent https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2 , calling the update a step toward frontier performance while acknowledging that larger models remained in development. The sequence lends some support to Zuckerberg's assertion that a capability jump is approaching. It does not establish how large that jump will be. Until Meta names the systems and allows outsiders to test them, the market is being asked to accept Zuckerberg's direction of travel rather than independently verified performance. The release strategy behind the prediction In "The Future is for Everyone" https://www.meta.com/thefutureisforeveryone/ , Zuckerberg argues that concentrating powerful AI systems inside a few laboratories, governments or corporations is more dangerous than distributing their capabilities. Meta plans to offer free versions to billions of users and paid access through an auction-based system for computing capacity, he wrote. That position also serves Meta's immediate competitive requirements. The company needs more computing infrastructure, freedom to train smaller models from the outputs of stronger ones and room to ship products without lengthy external reviews. Zuckerberg endorsed model distillation, opposed policies that could delay American model releases and called for faster construction of data centers and energy capacity. His discussion of recursive self-improvement /article/lilian-weng-returns-openai-recursive-self-improvement further raises the stakes of Meta's spending. Zuckerberg assumes that once models can improve their own training and efficiency, every frontier laboratory will need to devote substantial computing power to that process or fall behind. His proposed response is a broad expansion of capacity that lets laboratories pursue self-improvement while reserving most AI output for human-directed work. The argument supplies a public-interest case for Meta's construction program. Zuckerberg presents additional data centers as necessary to distribute intelligence, preserve US leadership and prevent a small number of institutions from controlling advanced models. The same infrastructure gives Meta the capacity it needs to compete with better-performing systems from rivals. Zuckerberg paired the essay with a $1 billion Future Is for Everyone Fund for US communities where Meta owns and operates data centers, according to Axios https://www.axios.com/2026/08/10/zuckerberg-ai-manifesto-meta . The money is intended for teachers, first responders, and energy and water infrastructure. Meta is also promoting America's Workforce Academy, a skilled-trades program offering free training and guaranteed jobs with Meta partners near data center projects. A test for Meta's board Zuckerberg also said Meta's independent directors will receive authority to approve safety criteria for model releases and review whether individual releases satisfy those standards. The arrangement creates a formal checkpoint, but Zuckerberg remains Meta's controlling shareholder and serves as chairman and CEO. Meta's 2026 proxy statement https://www.sec.gov/Archives/edgar/data/1326801/000162828026025532/meta-20260416.htm says the board's standing committees consist of independent directors and can establish special committees when needed. Zuckerberg's voting control still determines the broader corporate structure in which those directors operate. The stronger models he described in the interview will provide the first practical test. The directors' influence will depend on whether they can delay a release when Meta faces competitive pressure and has already committed billions of dollars to training and infrastructure. Critics interviewed by The Associated Press https://apnews.com/article/meta-ai-mark-zuckerberg-artificial-intelligence-df8a4e7d7825470d09e8090367457c2c argued that broad access does not resolve the risks posed by systems capable of autonomous cyber activity. Digital-rights advocates also warned that Meta benefits if its models become the foundation for widely distributed AI, even when their weights are open. For founders and developers, Zuckerberg's most consequential statement was not his defense of that strategy but his warning that stronger systems are close. Their capabilities, release terms and safety review will determine whether Meta's massive infrastructure investment has produced a genuine competitive leap or another company-defined benchmark gain.