First Paid Model Launches, In-House Chips Mass-Produced in September: Meta Eases Wall Street Capex Anxiety With Two-Pronged Approach

Meta's newly released Muse Spark 1.1 is its first paid AI model, which developers can access via the brand-new cloud-hosted platform Meta Model API, signaling Meta's official entry into enterprise-grade programming services.

Unlike the first-generation version in April, which was only open to specific partners, the new version has entered public beta, and developers can register for the waitlist to gain access. Currently, the API is limited to Meta's own platform and is not yet open to third-party marketplaces such as OpenRouter.

Alexandr Wang, head of Meta's AI business, said that Muse Spark 1.1 is the company's 'most powerful model to date for agentic tasks and programming work,' outperforming competitors in integrating with third-party programming tools and automated interaction tasks.

The model is competitively priced, with new users receiving a $20 initial free credit, and subsequent charges set at $1.25 per million input tokens and $4.25 per million output tokens, significantly lower than comparable offerings from OpenAI and Anthropic.

This shift marks a major pivot in Meta's AI strategy—transitioning from its previous focus on the open-source community's Llama series of models to a commercialization model of selling access to proprietary AI models.

However, Meta has not completely abandoned its open-source route. Wang revealed that the Superintelligence Lab is developing an open-source variant of Muse Spark, though no specific release date has been announced.

In addition, Meta released the image generation model Muse Image (internally codenamed 'Mango') this week to attract creators and advertisers to its AI products, further enriching its commercialization matrix.

In addition to progress on the software front, Meta has also achieved a milestone breakthrough in the hardware sector.

According to an internal memo, the in-house AI chip code-named "Iris" will begin mass production in September this year. This chip is part of Meta's in-house MTIA (Meta Training and Inference Accelerator) fourth-generation product roadmap. It completed testing in just six weeks with no major issues found, bringing a positive inflection point to this in-house chip project, which had faced sluggish progress over more than five years.

The Iris chip was co-designed by Meta and Broadcom ( AVGO) and manufactured by TSMC. Fully customized for Meta's own business needs, the chip is positioned not to replace GPUs from Nvidia and AMD, but to complement them in accelerating AI training and inference tasks. Meta hopes to reduce its reliance on external suppliers through in-house chips while optimizing its data center cost structure.

Deutsche Bank analyst Benjamin Black believes that by combining the use of Iris chips and Nvidia chips, Meta is expected to reduce data center costs by as much as 35% by 2027, particularly in inference and core recommendation workloads.

To support the development of its AI business, Meta has also announced an ambitious computing power expansion plan, deploying 7 gigawatts of AI computing infrastructure in 2026 and doubling it to 14 gigawatts in 2027.

To this end, Meta expects its AI infrastructure investment to reach up to $145 billion this year, accounting for a significant portion of the global tech industry's total AI investment.

To ensure the smooth progress of its computing power expansion, Meta has signed long-term supply agreements with suppliers such as Samsung Electronics, SanDisk, and Sumitomo Electric, securing the supply of memory chips, flash memory products, and optical fiber equipment.

Meta's AI strategy adjustments have triggered mixed market reactions.

On the one hand, investors welcomed its AI commercialization progress, believing that paid models and cloud service products can provide additional incremental revenue beyond core advertising revenue, offering a clear path to returns on AI investments.

BNP Paribas analyst Nick Jones noted that these developments "demonstrate a clear and direct path to high returns on Meta's AI investments."

On the other hand, the market remains concerned about Meta's massive capital expenditure plans. On Thursday morning, news that Meta plans to double its computing capacity to 14 gigawatts sent its stock price down by more than 3% at one point, as investors worried that aggressive infrastructure spending would squeeze the company's profitability.

However, Deutsche Bank analysts believe that the cost optimization and high-margin incremental revenue opportunities brought by in-house chips could be more substantial than market expectations, and the additional costs might not be as high as feared.

For a long time, Meta's AI strategy has faced criticism, with investors believing the company spent billions of dollars on top researchers and data center construction without generating sufficient returns.

The commercialization of Muse Spark 1.1 and the mass production of the Iris chip are seen as key moves for Meta to prove the value of its AI investments to the market. Meta CEO Mark Zuckerberg is facing pressure from Wall Street to demonstrate returns on AI spending, and these two developments are undoubtedly positive signals.

This content was translated using AI and reviewed for clarity. It is for informational purposes only.