Full Breakdown
Meta Accelerates AI Cloud Push with New Models, Paid API and Compute-Rental Plans
7/11/2026, 2:00:25 AM
Core Development: Monetizing Muse Spark 1.1 and Exploring an AI-Compute Cloud
In early July 2026 Meta Platforms announced three linked initiatives that reshape its AI strategy. On July 9 the company released Muse Image, a generative-image model, and on July 10 unveiled Muse Spark 1.1, an upgraded foundation model with stronger agentic reasoning and coding capabilities. For the first time Meta is charging developers for access through a Meta Model API that “will be very aggressive and attractive” in pricing. Simultaneously, CEO Mark Zuckerberg confirmed that Meta is evaluating offers to rent out excess AI compute and is developing a dedicated cloud unit, dubbed Meta Compute, that could sell both raw GPU capacity and hosted AI models to external customers. Production of the in-house Iris AI chip is slated to begin in September 2026, supporting a target of 14 gigawatts of compute power by the following year.
Background & Context
Meta’s AI spending has surged since 2024, with the company shifting from an early “open-source” posture to a closed-model approach that can be monetized. After a disappointing Llama 4 launch in 2025, Meta hired Alexandr Wang to lead the newly created Meta Superintelligence Labs. The lab delivered Muse Spark and Muse Image, and now the firm is positioning its AI stack as a revenue engine that can complement its core advertising business.
Data & Statistics
- Share price reaction: Meta stock rose 5.96 %–6.04 % on July 10 and has logged a 15 % weekly gain, erasing its YTD losses and moving above a 2 % YTD gain.
- Capital-expenditure guidance: The company’s April 2026 earnings call set a range of $125 billion–$145 billion for 2026 capex. BNP Paribas later projected a possible raise to $135 billion–$155 billion, while Wolfe Research estimates total 2026 spending could approach $200 billion.
- Compute capacity: Meta aims to deliver 14 GW of AI compute by the end of 2027, with 1 GW already online.
- Pricing: The Muse Spark 1.1 API is priced at roughly 25 % of the rates charged by OpenAI and Anthropic.
- Financial fundamentals (Q1 2026): Revenue $56.3 billion, net income $26.8 billion, operating cash flow $32.2 billion, free cash flow $13.2 billion after $19 billion of capex.
Official Statements & Responses
Zuckerberg emphasized the scarcity of compute resources, noting, “I don’t know anyone in the industry who feels like they have too much compute.” He added that high external offers “make sense” for renting capacity, and that Meta’s “pricing from some of the other labs is very extreme.” Analyst Justin Post (Bank of America) observed that Meta “may have engineered significant cost savings to get capacity cost per MW well below our and Street expectations.” BNP Paribas senior analyst Nick Jomes wrote that, despite “near-to-medium-term elevated capex,” Meta is “well positioned to generate ample revenue” from AI-model fees, subscription growth and a potential cloud offering.
Criticism & Opposition
Several analysts caution that the AI-compute business entails substantially higher capex and could require a capital raise. Wolfe Research highlighted a possible $200 billion spend versus the Street’s $160 billion expectation, flagging tighter free-cash-flow and financing risk. Critics also note that Meta’s models, while improved, still trail OpenAI and Anthropic in many benchmarks, and that entering a market dominated by Amazon Web Services, Microsoft Azure and Google Cloud presents a formidable competitive hurdle. Regulatory scrutiny in Europe over digital and copyright rules adds an additional layer of uncertainty.
Conflicting Reports & Gaps
- Capex outlook: Reported ranges vary from $125-$145 billion (Meta), $135-$155 billion (BNP Paribas) to an estimated $200 billion (Wolfe).
- Model performance: Bloomberg claims Muse Spark 1.1 outperforms Google’s Gemini in several agentic and coding tests, while other sources acknowledge that Meta still lags the leading OpenAI and Anthropic systems.
- Timeline for cloud products: Zuckerberg confirmed interest in renting compute and selling model APIs but provided no concrete launch dates or pricing structures beyond the API’s relative cost.
Verbatim Quotes
- “The offers that you get for using the compute are so high that it may make sense, in some cases, to rent out or consider those kind of deals instead of your own internal uses,” — Mark Zuckerberg, CEO, Bloomberg interview
- “Since this is not an open source model, this is I think the first time that we’re doing a real serious API,” — Mark Zuckerberg, CEO, Bloomberg interview
- “The pricing from some of the other labs is very extreme,” — Mark Zuckerberg, CEO, Bloomberg interview
- “We’re generally doing better than we expected,” — Mark Zuckerberg, CEO, Bloomberg interview
- “Meta may have engineered significant cost savings to get capacity cost per MW well below our and Street expectations,” — Justin Post, analyst, Bank of America
- “While we expect near-to medium term elevated capex, we believe Meta is well positioned to generate ample revenue to support its spending, driven by monetization of its own AI initiatives, advertising share gains, incremental subscription revenue, an optionality of cloud offering, and fees for external use of its AI models,” — Nick Jomes, senior analyst, BNP Paribas Equity Research
What’s Next
Meta is expected to update its 2026 capex guidance in the upcoming second-quarter earnings release, potentially reflecting the impact of the new AI-cloud initiatives. Production of the Iris chip begins in September 2026, and the company has signaled that the Meta Compute service could launch later in 2026, though exact pricing and product details remain undisclosed. Investors will watch for concrete revenue figures from the Muse Spark 1.1 API and any signed compute-rental agreements, especially in comparison with the cloud-service models of AWS, Azure and Google Cloud.
