Meta Raises 2026 Capital Expenditure Lower Limit to $130 Billion, Upper Limit Remains at $145 Billion
Meta Platforms has adjusted its 2026 capital expenditure plan from $125 billion - $145 billion to $130 billion - $145 billion, raising the lower limit by $5 billion while keeping the upper limit unchanged.
The adjusted median capital expenditure increases from $135 billion to $137.5 billion, reflecting an intensified investment in AI infrastructure. Funds are shifting from other businesses to AI data centers and GPU clusters, while Reality Labs and advertising businesses are under pressure due to resource allocation.
Meta's Q2 2026 capital expenditure is projected at $38 billion, a year-on-year increase of 65%, with 70% allocated to AI infrastructure (data centers, GPUs, networks), 20% to Reality Labs (VR/AR hardware), and 10% to others (offices, acquisitions).
Meta is collaborating with BlackRock on a $14 billion El Paso data center project, aiming to deploy 7 GW of computing power by 2027 and 14 GW by 2028 to support AI model training and inference needs.
Source: Public Information
ABAB AI Insight
Historical Behavior
Meta's historical behavior in capital expenditure shows characteristics of "aggressive expansion." Capital expenditure is projected at $35 billion in 2024, $60 billion in 2025, and $140 billion in 2026, a 300% increase over three years. In comparison, Google (2026: $120 billion), Microsoft (2026: $100 billion), and Amazon (2026: $90 billion) have lower "AI capital expenditure/revenue" ratios (2026: 90% for Meta vs. 40% for Google, 30% for Microsoft). In 2025, Zuckerberg proposed the vision of "personal superintelligence," and the 2026 capital expenditure shift towards AI reflects an "AI-first" strategy.
Capital Path
The structural change in capital allocation is clear: of the $140 billion in 2026, 70% ($98 billion) is for AI infrastructure, 20% ($28 billion) for Reality Labs, and 10% ($14 billion) for others. Within AI infrastructure, data center construction accounts for 50% ($49 billion), GPU procurement for 30% ($29.4 billion), and networking and optimization for 20% ($19.6 billion). Funds are shifting from "advertising business cash flow" (projected at $150 billion in 2026) to "AI capital expenditure" ($140 billion), with free cash flow declining from $40 billion in 2025 to $10 billion in 2026. If AI monetization falls short of expectations (2026 AI revenue of $5 billion, only 3% of total revenue), capital returns will be pressured.
Analogies + Industry Position
Similar cases include Cisco's "internet infrastructure" investments in the 2000s—capital expenditure of $2 billion in 1999, $5 billion in 2000, and an 80% stock price drop after the internet bubble burst in 2001. In the 2010s, Amazon's "cloud computing infrastructure" investments—$1 billion in capital expenditure in 2010, $50 billion in 2020, with AWS projected to generate $150 billion in revenue in 2026, validating the "long-term investment" logic. Meta is currently in the mid-stage of an "AI infrastructure arms race"—compared to Google (self-developed TPU + third-party GPUs) and Microsoft (Azure + exclusive OpenAI), Meta's strategy is "open-source models + third-party GPUs" (Llama series + NVIDIA H200). The industry position is at the "peak of AI capital expenditure"—global AI capital expenditure is projected at $500 billion in 2026, $700 billion in 2027, and $1 trillion in 2028, with Meta accounting for 14-20%.
Structural Judgment
This essentially represents a bet on technological routes driven by capital concentration. Meta's "$98 billion AI infrastructure" is fundamentally about "exchanging capital for time"—as OpenAI, Google, and Microsoft have established model advantages, Meta chooses to "compensate for algorithm gaps with computing power." This explains why 70% of capital expenditure is directed towards AI—not as "diversified investment," but as a "breakthrough at a single point." As AI models converge (with Llama 3, GPT-5, and Gemini 2 expected to have similar performance in 2026), competitive barriers will shift from "model capability" to "inference costs." Meta's goal is to "provide $2 worth of intelligence for every $1 of computing power," reducing unit costs through economies of scale (14 GW of computing power). This drives the "REIT-ification of AI infrastructure"—data centers transition from "cost centers" to "infrastructure assets," with valuation logic shifting from "growth" to "stable cash flow."
ABAB News · Cognitive Laws
Capital expenditure is a bet on technological routes
Scale of computing power is a new barrier in AI competition
The essence of open-source models is exchanging capital for time