Meta AI Updates: November 4, 2026

1. Meta Seeks $20 Billion to $25 Billion in a Dollar and Euro Bond Sale to Fund AI Spending

Meta went to the bond market on November 3 looking to raise $20 billion to $25 billion, about two-thirds in U.S. dollars and one-third in euros, according to people familiar with the matter cited by CNBC, with final pricing expected that afternoon; Meta declined to comment, and Bloomberg first reported the plan. The sale comes days after S&P Global Ratings revised Meta’s outlook from AA- stable to AA- negative on October 30, citing “uncertainty” about its capital spending and negative free cash flow. In its third-quarter report Meta guided to full-year capex of $135 billion to $145 billion, raising the low end by $5 billion, and posted negative free cash flow of $2.12 billion for the quarter, as Alphabet and Amazon also did. The new debt adds to a $25 billion bond sale in April and the $95 billion equity plan announced the previous week, which starts with $35 billion in underwritten public offerings followed by a $60 billion at-the-market program. Meta shares rose 2.7% on November 3 after falling in 12 of the previous 14 sessions and ending October with their worst month in four years. Source

2. Meta’s F3 Lightning Releases Each Ads ML Feature Independently and Ships Them About 3.6x Faster at No Extra Compute

Meta described F3 Lightning, a layer on top of its Facebook Feature Framework (F3), which authors, compiles and serves the features behind its ads ranking models, where a single request can need more than 20,000 features. F3 compiles every feature into one merged DAG so shared reads, filters and aggregations run once per request, but that graph was also the release unit, so one invalid definition, or the expensive model-related tests required for a feature a live model already reads, held back every other feature. F3 Lightning compiles each feature into its own self-contained configuration that is validated, deployed and rolled back on its own, pulled from a release queue prioritized by impact and risk. Early per-feature prototypes used substantially more CPU because work was no longer shared, so the runtime restores sharing at service startup in four steps: compiler identifiers derived from each operation, its parameters and its inputs, so identical computations match across features released weeks apart; a single topologically sorted execution plan that keeps one copy of each shared node and flattens it into a list, where a failed computation marks only its downstream nodes as not ready; one memory buffer per request with a separate readiness vector, which closed a CPU gap that remained after deduplication because scattered per-feature containers caused cache misses; and dense integer slots that turn every lookup into a direct array index. Meta reports that comparable extraction features now reach production about 3.6 times faster than on F3, with no increase in infrastructure compute cost. Source