The $65 Billion Question

>

When Zuckerberg announced $65 billion in AI capex for 2026, Wall Street held its breath. That's more than the GDP of Croatia, double Meta's 2023 total, and roughly equal to Microsoft and Google's combined AI spend in 2024. The scale is unprecedented — Meta is spending nearly 50% of revenue on infrastructure, versus Microsoft at 32% and Google at 24%. The question: will this print money or incinerate it? $META has rallied 80% in the past year on the thesis that this time is different from the metaverse disaster. The data suggests the bulls have a point.

▶ Meta AI
Meta AI capex spending analysis

Llama 4: Open Source as Moat

Why give away a frontier model for free while OpenAI charges $200/month? Because Meta's business is ads, not subscriptions. Llama 4 powers 350,000+ apps on Hugging Face — every one running on Meta's inference infrastructure. The strategy: commoditize the model layer, own the pipes, and feed every interaction back into a data flywheel that improves ad targeting for 3.2 billion daily users.

"By making Llama free, Meta becomes the Android of AI — the open platform that captures the majority of usage while competitors fight over flashier demos."

The 405B parameter version matches GPT-4o and Gemini Ultra on most benchmarks, and leads on code generation. Meta now operates an estimated 1.6 million GPUs — the single largest NVIDIA customer — with plans to hit 2.5 million by late 2026. But the real monetization isn't Llama model sales. It's what the ecosystem unlocks for Meta's ad stack.

Advantage+: The AI Ad Engine

This is where the capex gets its ROI. Meta's AI-powered Advantage+ platform now handles 40% of all ad impressions, up from 15% in early 2024. Advertisers see 32% lower CPA, 28% higher ROAS, and 45% less setup time. Every GPU allocated to ad ranking generates a near-immediate return: Meta's internal models show each $1B in ad AI capex drives ~$1.7B in incremental annual revenue within 6-9 months — a payback period well under a year.

  • Q1 2026 ad revenue: $47.2B, up 22% YoY — nearly double the digital ad market growth rate of 13%
  • Global ad market share: 23%, closing fast on Google's 28%
  • E-commerce advertisers: up 35% YoY, driven by AI-powered dynamic product ads

The Andromeda recommendation system — a massive transformer model — powers feed ranking across Facebook, Instagram, Reels, and Threads. Reels watch time jumped 17% after the 2026 update. More engagement = more ad inventory = more revenue. It's a virtuous flywheel that compounds with every infrastructure dollar spent.

The Balance Sheet Question

Can Meta afford this? The numbers say yes. Meta closed Q1 2026 with $58.2B in cash, $37.1B in debt (net cash: $21.1B), and $68.4B in trailing free cash flow. Even at $65B capex, FCF covers the investment without needing to borrow — and the company is still buying back $24B in shares annually. Free cash flow would need to drop 50% before Meta faces a funding gap.

The real risk is capex irreversibility. GPU useful life for AI training is 3-4 years, and NVIDIA's release cycle means H100s bought in 2024 are already outperformed 3x by B200s in 2026. Depreciation could hit $15-18B annually by 2028, weighing on GAAP earnings even if cash flows stay healthy.

The Verdict

Meta's $65B bet is aggressive but defensible. Nearly half the capex — the Advantage+ inference infrastructure — pays back within a year. The Llama play is a long-option on commoditizing the model layer while owning the infrastructure. Even the bear case (10% revenue growth, margin compression) puts fair value at $550, limiting downside. Our base case DCF supports $780, ~35% upside from current 24x forward earnings.

The key insight: Meta is the only hyperscaler whose AI spend directly and measurably enhances its core revenue engine with clear attribution. As one analyst put it: "Meta's AI capex isn't a leap of faith — it's a spreadsheet that pencils out."