Meta Doesn’t Need to Build Frontier Models to Win AI Race

Meta once bet its future on the metaverse and virtual worlds, but the strategy failed spectacularly. Now, the company believes AI can become the foundation of an entirely new business spanning consumer assistants, enterprise software, AI infrastructure, and developer services.

Meta’s second-quarter results offered a glimpse of the company’s AI ambitions. CEO Mark Zuckerberg described a future where AI moves beyond supporting Facebook and Instagram to become a business in its own right.

“We are now at a point where our investments in AI are accelerating every major part of our core business,” Zuckerberg said during the earnings call. “They’re improving the experience for people using our apps, driving better performance for advertisers, and helping our teams build new experiences and ship faster.”

Zuckerberg believes that soon billions of people will interact with AI agents that work continuously on their behalf. 

However, AI agents remain limited to technical users willing to tolerate complexity and instituting human-in-the-loop checkpoints. Consumer adoption, he said, requires products that simply work without setup or configuration.

That’s where Meta can reliably tap into its largest distribution platform—WhatsApp. The platform hit a record 30 million messages per second during the FIFA World Cup final. That’s what gives it a structural advantage—and a low-hanging fruit—to position AI agents as primary interaction interfaces, especially with WhatsApp Business, which has over 200 million active accounts.

These agents can handle end-to-end tasks—booking appointments, processing payments, negotiating orders, and resolving complex customer service issues—without the user ever leaving the chat interface.

Coupled with its enormous existing user base across the family of apps—3.6 billion daily active users, up 3% year-on-year, with Instagram alone accounting for 2 billion DAUs—Meta has ample opportunities to embed and scale AI agents directly into feeds.

“Meta starts with a distribution advantage that few companies can replicate. Facebook, Instagram, WhatsApp, Messenger, Meta AI, and its wearable partnerships give it repeated opportunities to embed the same model engine into high-frequency consumer and creator workflows. Each surface can drive adoption, improve product learning, and create another path to monetisation,” Chirag Dekate, VP Analyst at Gartner, tells AIM.

LLM-Based Recommendation

Meta is processing every public Feed and Reels post on Instagram through an LLM before generating recommendations. These models analyse content across topic, tone, and context before feeding that information into ranking systems.

It is also deploying its Muse family of models to classify videos, summarise content, and improve moderation systems

“LLMs add a first-principles understanding of what the content is about and why it is compelling, as well as a deeper understanding of what people are interested in and what their goals are when they're using our apps,” said Zuckerberg.

Susan Li, Meta’s CFO, said generative AI is reshaping the advertising stack. She said the company introduced what it calls the Meta Generative Recommender, replacing traditional ad scoring systems.

“Rather than scoring every possible ad individually, we are now using LLMs to reason about ad content and user preferences together, and predict the best ad for each person,” Li said. Early pilots improved app-event conversions on Instagram by 1%, while newer ranking models generated an 8.3% increase in ad clicks and a 15.7% increase in conversions on Facebook.

Earlier this year, it also introduced Ads MCP Server, allowing developers to connect AI agents directly to Meta’s ad capabilities.

Enterprise AI Could Be the Next Growth Engine

While consumer AI has received most of the attention, Meta believes businesses may become one of its largest AI customers. The company has rolled out Meta Business Agents globally across WhatsApp and Messenger, with more than 1 million businesses onboarded. Instagram support is now being added.

Li said that Meta Business Agent Platform gives enterprises the infrastructure to build, customise and deploy AI agents on WhatsApp with enterprise-grade controls and governance. 

Beyond business agents, Meta is commercialising its AI model Muse Spark through APIs. The company recently launched the Muse Spark 1.2 model, adding tools for software engineering workflows and long-running development tasks to its AI coding portfolio. 

The company also introduced Muse Code, currently available in beta for macOS and Linux. It can plan code changes, write code, validate results, and coordinate multiple persistent subagents to work across large code repositories. 

The company is returning to its open-weights roots well. Meta is releasing the weights for Muse Glimmer, a 30-billion-parameter dense model that can run on a laptop or a single consumer GPU, while preparing to open source a version of Muse Spark 1.2 in the coming weeks.

Dekate believes that Meta should now prioritise capabilities that can be shared across its ecosystem, including model quality, inference efficiency, memory, identity, and safety, instead of building isolated experiences. 

“If the same infrastructure, model and agent platform strengthens every surface, breadth becomes an advantage. If each product develops its own fragmented stack, execution costs rise quickly,” he explains.

Similarly, Varun Mishra, Principal Analyst at Counterpoint Research, says Meta’s AI strategy should be viewed as an integrated effort rather than a collection of separate initiatives. 

“It’s about two things: owning multiple consumer touchpoints, and owning the entire infrastructure-to-intelligence-to-hardware stack,” Mishra notes. He adds that the strategy could deliver significant long-term gains if Meta’s smart glasses also become a mainstream consumer device. 

“On frontier capabilities, OpenAI, Google, Anthropic, and xAI might hold an edge. However, Meta’s biggest advantage is its huge distribution and user stickiness, which others are still trying to build.”

Building Infrastructure Before Demand Peaks

To support its AI ambitions, Meta is also building enormous computing capacity and infrastructure.

The company reported capital expenditure of $31.1 billion during the latest quarter, largely on servers and data centres. It also raised the lower end of its 2026 capex forecast, now expecting to spend $130–145 billion. 

Meta also recently announced a venture with BlackRock to develop a 1 GW data centre in the US while continuing to invest aggressively in servers, networking, and AI infrastructure. 

Li said Meta’s infrastructure strategy is built on the belief that the AI industry remains underbuilt.“The industry has underbuilt historically for the wave of AI adoption, making existing capacity, including our own, extremely valuable,” she said.

Meta is therefore planning infrastructure not only for its own products but also for external demand. Alongside APIs and enterprise agents, Zuckerberg revealed the company has already received offers to purchase compute capacity at significant premiums. At the same time, he argued that selling intelligence built on top of that infrastructure would ultimately generate higher returns than selling raw compute.

Mishra says Meta’s decision to commercialise its AI compute capacity could help offset the massive capital expenditure required to build AI infrastructure as the demand for AI compute continues to outstrip supply. However, the gains may be temporary.

“Supply-demand balance is expected to start restoring in 2028, as new capacity comes online, post which the prices of compute could start normalising,” he notes.

In April, the company had broken ground on 10 data centres over the previous 24 months. Its biggest projects include Prometheus, a 1 GW AI cluster expected to come online in 2026, and Hyperion, whose capacity was raised to 5 GW in July 2026 with an investment of more than $50 billion.

Separately, according to a SemiAnalysis report, Meta had contracted more than 5 GW of capacity in the first six months of 2026 across cloud providers and colocation facilities, on top of its own data centre construction. It is also reportedly in final talks with Anthropic for access to private instances of Claude, as part of a potential $10 billion compute deal that could kick off Meta’s strategy of monetising its infrastructure. 

Dekate says investors would need to look beyond Meta’s traditional operating metrics if the company moves deeper into commercialising its AI infrastructure. Measuring contracted backlog, external utilisation, revenue per deployed megawatt, gross margin, power costs, and customer concentration would become increasingly important.

“Running infrastructure for Meta’s own workloads differs materially from operating a reliable, secure, multi-tenant service for outside customers,” Dekate says. He adds that turning compute into a cloud business would require billing systems, service-level commitments, enterprise support, workload portability, security controls, and a developer ecosystem.

“A sustained valuation shift therefore depends on customer demand, differentiated economics, and disciplined capacity allocation,” he says.

Betting on Glasses 

For now, glasses remain Meta’s strongest consumer bet. It is now hoping that glasses will be the new interface for interacting with personal AI, especially when embedded with Muse Spark.

However, that will bring new challenges around privacy and social norms around always-on cameras. “A camera, microphone and assistant embedded in an everyday object require new expectations around notice, consent, context, and acceptable behaviour,” Dekate says. 

That said, Meta has already introduced visible capture indicators, safeguards that disable the camera when those indicators are obstructed, but these mechanisms will need to evolve as smart glasses shift towards edge-computed visual tasks.