Meta Platforms made two significant moves this week that are reverberating in Friday's premarket session: the public developer launch of Muse Spark 1.1, the company's most capable AI model, and confirmation via an internal memo reviewed by Reuters that Meta will begin manufacturing its proprietary data center chip, code-named Iris, in September. The combination is the clearest signal yet that Meta's strategy of developing vertical AI infrastructure is transitioning from announcement to execution.
Muse Spark 1.1 is a multimodal reasoning model designed for agentic and coding tasks, released through Meta's developer portal as a public preview with API access available via a waitlist. Meta AI chief Alexandr Wang described it as the company's "strongest model for agentic and coding work yet," highlighting a million-token context window, parallel subagents, and advanced computer-use capabilities. Developers signing up for API access receive $20 in free credits, after which usage-based pricing begins at rates Wang described as "very aggressive and attractive" relative to competing offerings from OpenAI and Anthropic. Meta claims Muse Spark 1.1 outperforms Google's Gemini 3.1 Pro across several coding and multimodal reasoning benchmarks.
On the infrastructure side, testing of the Iris chip — the fourth generation of Meta's in-house Meta Training and Inference Accelerators program — was completed in six weeks without major issues, a faster-than-expected certification timeline that clears the path for a September production start. The chip includes 51% more high-bandwidth memory than its predecessor and supports enhanced MX8 and MX4 data formats, while delivering a claimed 400% performance improvement versus the prior generation. Broadcom is partnering on design and TSMC is handling fabrication. If Iris ramps on schedule, it would reduce Meta's dependence on Nvidia and AMD GPUs and provide a meaningful cost-down lever against a backdrop of AI infrastructure spending expected to reach between $125 billion and $145 billion this year.
The dual announcements are directly relevant to Meta's 2026 financial picture. The company had one of the weakest performances among the Magnificent Seven in the first half, dropping 14.5% through June before rebounding in July. The Muse Spark and Iris catalysts pushed shares up roughly 8.3% over the four trading days through Thursday, and Friday premarket data shows the stock up another 1%, putting it on course for its best weekly gain since mid-April. ARK Invest's Cathie Wood added more than $20 million in Meta shares after the Thursday release, moving sentiment on Stocktwits to "extremely bullish" from "bullish."
Analysts are focused on two monetization pathways. The first is the Meta Model API itself, which positions Meta to generate external developer revenue rather than using models exclusively for internal social advertising. The second is the advertising read-through: stronger AI capabilities embedded in Instagram and WhatsApp can improve targeting precision, supporting ad revenue in a competitive market. Seeking Alpha contributors noted that analysts "believe Muse Spark 1.1 and related advances will boost Meta's chances of increasing monetization through new AI offerings, advertising share gains, and possible external licensing revenue."
Risks center on execution. Iris has historically faced engineering delays, and its real-world performance against Nvidia's latest hardware remains unvalidated at scale. The Manus AI acquisition is also reportedly being unwound, with Tencent in discussions to become the largest investor — a development that removes one piece of Meta's original agentic AI strategy. Nonetheless, the combination of a production-ready chip timeline and a competitive frontier model launch in the same week marks a measurably different posture for a company that has lagged its AI peers through most of 2026.