Meta just turned AI from a cost center into an engagement engine—and proved the compute investment thesis while every other hyperscaler is still figuring out what to build.
The Summary
- Instagram screentime jumped double digits year-over-year in Q2 2026, driven by AI-powered feed and Reels recommendations that draw on deeper user history
- Meta's bet on hoarding compute is paying off—Zuckerberg said it would be "foolish" to give too much away while the company uses it to grow core business
- Every public Reel and post now feeds an LLM for topic and tone analysis, creating what Zuckerberg calls "a key building block toward greater personalization"
- Meta's cash flow plunged 91% from last year as capex spending hits home, but engagement metrics suggest the spending is working
The Signal
While Google's cash flow went negative for the first time in company history this quarter, Meta is showing what all that compute spending actually buys. Instagram sessions increased 15 basis points, with particular strength in reshares and time spent. Those are the metrics that matter when you're selling ads against attention.
The technical shift is specific. Instagram Reels now combine faster inference with a new architecture that pulls from deeper user history to improve predictions. Not "AI makes things better." Actual architectural changes that draw on more data to match content to users. The result is measurable: double-digit growth in time spent.
"Our recommendations are also becoming more personalized, surfacing more fresh content while giving people more direct control over what they see."
Here's the part that matters for the agent economy: Meta is feeding every public Reel and post to an LLM and analyzing them for topics and tone. This isn't human labeling. This is automated content understanding at platform scale. The LLM becomes the content comprehension layer, and that layer makes better recommendations possible.
Meanwhile, Zuckerberg is threading a needle on compute allocation. A "significant portion" goes to training models, powering agents, and growing core business. But Meta expects to "grow a large business serving large customers" too. Translation: they'll sell compute, but only after they've proven it works for their own products. That's different from hyperscalers who are selling compute capacity they're still figuring out how to use themselves.
The financial hit is real. Meta's cash flow dropped 91% year-over-year. Google and Meta both raised capex forecasts for the year, and Google signaled 2027 will be even bigger. But Meta has something the others don't yet: proof that the spending improves the product in ways users respond to. You can see it in the screentime numbers.
The Implication
Meta just showed the template for how to monetize AI infrastructure investment. Use it internally first, prove it works in production, then consider selling excess capacity. Compare that to companies raising capex without clear product wins to show for it.
For anyone building AI-powered products, the lesson is architectural. Faster inference plus deeper user history plus automated content analysis equals measurable engagement gains. That stack is now proven at Instagram scale. Expect every feed-based product to follow.