Brian Nowak (Morgan Stanley) asked which new opportunities (consumer/business agents, API, compute rental) will scale first to show quantifiable ROIC, and for early 2027 capex philosophy.
Mark Zuckerberg said a substantial share of compute trains leading models while the rest spans core-business improvement, new consumer products, the API, business agents, developer tools, and selling compute directly (many premium offers), expecting meaningful growth across all; Susan Li declined a 2027 capex figure, reiterating a focus on maximizing 2026-2027 capacity with 2028+ flexibility since near-term capacity is more valuable.
Eric Sheridan (Goldman Sachs) asked how much of the enterprise opportunity extends the current ads business versus requires new go-to-market, and the philosophy on sources of capital.
Zuckerberg said business agents naturally extend the ads/marketing relationship (paid on results via an auction) while coding/productivity tools for other enterprises are a 'somewhat different muscle' Meta is building; Li said strong operating cash flow is supplemented by a greater mix of cost-efficient long-duration debt and partnerships like BlackRock to fund long-horizon AI infrastructure.
Mark Shmulik (Bernstein) asked whether consumer AI adoption can close the utility gap and if a breakthrough is near.
Zuckerberg said coding agents are the first real agentic market (technical, closed-loop users) but consumer personal agents will be a massive market, with billions likely to have a 24/7 personal agent within five years; the key is delivering a consumer product that 'just works' at billions-scale, which plays to Meta's strengths, with more to ship soon.
Doug Anmuth (JPMorgan) asked about the recommendations roadmap and how far along Meta is, and why Meta both sells and buys compute.
Li cited further headroom into 2027 via more personalized/relevant recommendations, LLM content understanding (every public IG Reels/Feed post processed), richer training data, and agentic ranking approaches; Zuckerberg said there is nowhere near enough compute for demand, so Meta builds out capacity for high-margin internal intelligence uses while opportunistically monetizing compute given the lead time before data centers come online.
Justin Post (Bank of America) asked how Meta Superintelligence Labs is performing a year in and what durable advantages it is building.
Zuckerberg said he is happy with the trajectory (impressive early-scaling-ladder models, larger models in progress), emphasizing the data/knowledge flywheel alongside intelligence, Meta's ability to scale products to billions, its advertiser/SMB base for business agents, and building a low-drama, well-managed research culture as durable advantages.
Ross Sandler (Barclays) asked about competing at both the low-cost and high-performance ends (Muse Spark 1.1 near the Pareto frontier) and the return to open source.
Zuckerberg explained the scaling-ladder process, wanting both efficient models to serve billions of prompts cheaply and more advanced models for hard problems; on open source, Meta will do a mix of open and closed and expects to release open-source models again soon, having kept MSL 'uninhibited' to build the most intelligent (harder-to-open-source) models first.
Ken Gawrelski (Wells Fargo) asked whether proliferating open-weight models reduce the need for Meta's own frontier models, and whether the 2026-2027 capacity focus is a demand or supply comment.
Zuckerberg said open-weight models are not yet as strong as frontier models and relying on competitors is risky, so full-stack model sovereignty is essential to Meta's differentiated, use-case-specific advantage (while open source still matters for the ecosystem and does not undercut the API opportunity); Li said the 2026-2027 focus reflects being demand-constrained (including the core business) plus supply-chain uncertainty, with 2028 planning centered on flexibility (land/power now, chip decisions later).