Code, Spark, Glimmer & Compute—Decoding Meta’s AI Strategy

Meta’s Artificial Intelligence (AI) strategy entered a new phase in the last few months. While market sentiment on its position in the AI market was strong, largely due to the initial launch of its Llama model family, it has declined since then, especially on the model/application layer of the AI stack as newer models significantly outperformed the Llama models. However, its recent moves spanning compute infrastructure, model development, and the application layer suggest a major overhaul in its strategy moving forward.

In July 2026, Bloomberg reported that Meta was potentially exploring plans to commercialize its internal AI infrastructure to operate like a neocloud business. Meta had established its Meta Compute business unit to be in charge of its AI capacity buildout, and it would be able to expand this group to offer AI compute to external parties. Meta has been aggressive in expanding its infrastructure, with its custom accelerators and Central Processing Units (CPUs), as well as partnerships with both AMD and NVIDIA to leverage their compute offerings. During its shareholder meeting in May, Chief Executive Officer (CEO) Mark Zuckerberg stated that Meta entering the cloud business was “definitely on the table.”

Further up the AI stack, Meta seems to have overhauled its model strategy. Its Meta Superintelligence Labs (MSL) team launched Muse Spark in April 2026—a multi-modal reasoning model. In July, they upgraded it with Muse Spark 1.1, which focuses more on agentic task capabilities, and released dedicated image and video models (Muse Image and Muse Video, respectively). Fast-forward to August, they released Muse Spark 1.2—a more coding-focused model—and introduced an open-weight model part of the Muse family—Muse Glimmer.

Alongside the release of Muse Spark 1.2, MSL introduced a beta terminal coding agent—Muse Code. It operates with a single agent loop alongside a set of async background agents that remain active throughout a session, rather than spawning for individual tasks.

Meta’s interest in expanding Meta Compute is likely driven by several factors, including SpaceX’s recent validation of the strategic value of monetizing large-scale AI capacity through third parties such as Google and Anthropic. Both deals, worth US$920 million and US$1.25 billion per month, respectively, suggest that frontier-scale compute can command substantial scarcity premiums, estimated at roughly 2X to 4X the market value of underlying capacity. For Meta, this strengthens the case for externalizing at least part of its internal AI infrastructure buildout, particularly as it looks to maximize returns on increasingly large data center and compute investments.

However, Meta’s compute strategy differs materially from SpaceX’s. Rather than relying on Graphics Processing Unit (GPU) aggregation, Meta is pursuing a more diversified infrastructure approach built around GPUs, custom silicon, and its own agentic CPU development efforts. If frontier labs were willing to lease capacity running on Meta-designed infrastructure, this would not only validate Meta Compute commercially but also strengthen market confidence in Meta’s broader infrastructure strategy. Longer term, this could create a pathway for Meta to commercialize parts of its silicon and infrastructure stack more directly, in a manner similar to how Google has Tensor Processing Unit (TPU) access to reinforce its own AI cloud positioning.

A second path for Meta Compute would be model access rather than raw capacity alone. Meta could offer model access to third parties in a similar fashion to traditional hyperscalers such as Microsoft, Amazon, and Google. This could be further strengthened through a deep strategic partnership with a frontier lab, with Meta hosting private instances of frontier models, similar to the Amazon-OpenAI arrangement on Amazon Web Services (AWS) Bedrock. Such a strategy would be more useful for strengthening Meta’s own model and application-layer offerings than for competing directly as a broad enterprise cloud provider, particularly given Meta’s more limited depth in cloud services and weaker positioning relative to hyperscalers. At the same time, integrating third-party frontier models into Meta’s own application layer would give users greater model optionality, which is increasingly important in a multi-model market where token consumption is distributed across multiple providers.

Meta’s model strategy also appears to be broadening. Muse Spark 1.2’s focus on coding workflows makes commercial sense, as software engineering remains one of the clearest enterprise AI use cases with measurable productivity benefits and more straightforward Return on Investment (ROI) justification. Simultaneously, Muse Spark 1.1 has been positioned more toward agentic task execution. Together, these launches suggest that MSL is building the Muse brand into a broader model portfolio, similar to how competing labs segment model families by modality, use case, and performance tier. At present, the Muse family appears to be segmented more clearly by modality and workflow than by distinct performance tiers, but the broader direction is one of portfolio expansion rather than reliance on a single flagship model.

Perhaps the most strategically significant move on the model and application layers is Muse Code. As Meta’s first coding agent, Muse Code represents a step into the application layer of the AI stack, where value increasingly comes from orchestration, workflow ownership, and user stickiness rather than model access alone. Starting with coding agents makes sense because it allows Meta to drive developer experimentation and gather feedback in a high-value, technically sophisticated environment. Over time, Meta could extend those learnings and capabilities into broader agent experiences across its platform portfolio, strengthening user stickiness across its consumer and advertising businesses.

It is important to assess Meta’s AI strategy holistically rather than as isolated product and infrastructure updates. ABI Research expects Meta to pursue a selective commercialization strategy for Meta Compute, likely prioritizing a deep frontier lab partnership to host private instances of its models alongside some bare-metal capacity leasing. This would be less about building a full stack AI cloud and more about prioritizing model partnerships to benefit Meta’s existing offerings while maximizing the ROI on its data center investments. Over time, an important strategic signal will be whether and how Meta pairs its infrastructure strategy with the application layer of its AI strategy, allowing different models and services to be routed according to user context, preferences, and task requirements.

For the competitive landscape, ABI Research expects bare-metal neoclouds to face the greatest effects if Meta formalizes external capacity offerings, as Meta is well regarded on its infrastructure strategy and would have a competitive edge relative to most neoclouds when combined with its custom silicon business. Providers that have moved further into managed services and high-level workflow support—such as Nebius, Vultr, and CoreWeave—will be relatively insulated by comparison.

Hyperscalers are unlikely to face near-term disruption from Meta’s entry, given their deep integration into enterprise Information Technology (IT) environments and the breadth of their cloud service portfolios. Inference platform vendors could explore capacity partnerships with Meta Compute, but Meta may prioritize higher-margin, strategically valuable agreements with frontier labs over more basic capacity agreements.

On the application layer, ABI Research is skeptical that Muse Code will emerge as a leading coding agent offering against more established products such as Cursor and Claude Code, particularly given Meta’s late entry into this market and its more limited presence in developer tooling. Instead, there is more strategic value for Muse Code to serve as an internal foundation for broader agent development. ABI Research recommends that Meta treat Muse Code as an early proving ground for harness capabilities for specific workflows within Meta’s consumer platforms and advertising business.