Meta's AI thesis is unique among hyperscalers: rather than selling AI as a service, Meta uses AI to make its advertising engine more effective. Every dollar spent on AI infrastructure flows back through better ad targeting, higher ROAS for advertisers, and ultimately higher ad prices Meta can charge. This creates a direct, measurable feedback loop between AI investment and revenue.
Is the AI investment self-reinforcing or a treadmill?
Bulls argue Meta's AI spend creates a virtuous cycle: better AI improves ads, which generates revenue to fund more AI. Bears argue diminishing returns are inevitable — the jump from 20% to 22% ROAS improvement costs exponentially more than the jump from 0% to 20%. At $125B annual capex, Meta needs each marginal GPU to generate substantial incremental ad revenue.
Will Meta's $115-135B FY2026 capex deliver proportional improvement in ad revenue, or are we seeing diminishing returns?
Llama is the most downloaded open-source AI model family, but its investment thesis rests on strategic value (AI supply chain independence) rather than direct monetization. The Llama API announced at LlamaCon 2025 is the first attempt to generate direct revenue, but the open-source model fundamentally limits Meta's ability to charge for what it gives away for free.
AI ad optimization is the clearest path from Meta's massive AI infrastructure investment to revenue. Unlike other hyperscalers that sell AI as a service, Meta's AI directly improves its core advertising product. Every percentage point of ROAS improvement translates to higher ad prices and more advertising dollars flowing to Meta's platforms.
The quality-over-volume shift
The most telling metric in Meta's recent results is the decoupling of impression growth from revenue growth. In Q4 2024, impressions grew 6% while price per ad grew 14%. This means Meta's AI engine is making each ad impression more valuable — charging more for better-targeted ads rather than simply serving more ads to more users.
Meta's competitive position in AI is paradoxical: it dominates social advertising with 67% market share and has the most widely downloaded open-source AI model (Llama), yet faces intensifying competition on every front. The question is whether Meta's unique asset — 3.35B daily users feeding data into AI models that improve ads that generate revenue to fund more AI — creates an unassailable flywheel, or whether AI itself commoditizes and erodes Meta's advantages.