Believe it or not, Meta's AI investments made a meaningful difference to their advertising business in the second quarter. It's not like these models have everyone, including social networks, they cultivate billions of dollars into data centers a year.
It will take several more years more before infrastructure and AI visionary Zuckerp ramp-prolonged spending and AI visionaries began to pay off in Wednesday's second quarter revenue call.
“I don't think our genai work will be a meaningful factor in our revenue this year or next year,” she said.
For now, the Power Meta recommended system is paying bills, and improvements to the more common machine learning model.
The widespread rollout of the company's new AI-powered recommended models has led to an increase in ad conversions by about 5% across Instagram and a 3% across Facebook.
As their names suggest, these models are designed to increase engagement across the company's various platforms by connecting users to friends, posts and other content. And in the last quarter, Zuck boasted improvements to these systems, increasing the time spent on Facebook and Instagram's doom crawls by 5-6%.
More importantly, these recommended models are how Meta makes money by sorting dozens or hundreds of thousands of potential ads to find the ones that are most likely to be relevant to individual users. In fact, this process is so important that Meta has developed a custom accelerator to run as quickly as possible.
Going forward, Li and Zuckerberg say Genai will ultimately play a bigger role in its advertising business. Facebook parents already use Genai to allow users to create ad collateral.
“Now that nearly 2 million advertisers use video generation capabilities, image animation and video extensions, we can see powerful results from text generation tools,” Li said.
Li also revealed that Meta has begun to incorporate large-scale language models (LLMs) into the recommended system used in Meta's X-competent threads. “LLMS is currently driving a meaningful share of ranking-related time spent on threads,” she said.
Meta is also messing around with LLMS and other Genai tools in other areas as well. Zuck previously teased an AI engineer designed to speed up the development of his Llama model flock.
In a call Wednesday, he said the team had started using Llama 4 to build an autonomous AI agent and “improve Facebook algorithms to improve quality and engagement.”
Speaking of open weight models, Zuck said work is ongoing on Llama 4.1, 4.2 and future models, but has not stopped providing a specific roadmap.
All of this “is happening at a low volume right now, so I'm not sure the outcome is a major contributor to revenue this quarter, but I think the trajectory of something like this is very optimistic,” added Zuckerberg.
For now, Meta is focusing on building the infrastructure needed to accelerate the development of new generation AI models by investing in infrastructure and talent.
The company is currently building a series of AI computing clusters, including a gigawatt scale facility called the Prometheus Set, which will be online in 2026, and another Gigawatt scale facility called the Hyperion, which Zuck boasts, which will eventually scale to a capacity of 5 gigawatts.
Meta also lavishly spends 8 or more salaries to build AI Superintelligence teams to build new systems that not only surpass human intelligence but also exceed human intelligence.
So it's good that Meta's old school ML is still paying the bill. Because Li has made it clear that these strategic adoption-related compensation packages will become the second-largest driver of 2026.
“We really believe it's time to really invest in the future of AI, as we think we'll be opening new opportunities for us, in addition to strengthening our core business,” Li said.
For at least now, Zack's AI spending and super intelligence pipe dreams still won't put meta in poor homes. In the second quarter, Meta's profits rose 36% year-on-year to $18.3 billion with revenue of $47.5 billion. ®
