Meta plans $135B compute bet
An analysis projects Meta is entering a 'giga‑capex' era with roughly $135 billion planned and a focus on compute for personal AI initiatives. The write‑up frames the spend as an infrastructure bet that will prioritise backend engineering areas like model serving, observability and cost control. (markets.financialcontent.com)
Meta is preparing to spend at a scale usually reserved for utilities and chipmakers, as it builds the computing backbone for its next wave of artificial intelligence products. (investor.atmeta.com) The company reported $72.22 billion in capital expenditures for full-year 2025, up from $39.23 billion in 2024, and Chief Executive Officer Mark Zuckerberg said on January 28 that 2026 would be the year he advances “personal superintelligence.” (investor.atmeta.com 1) (investor.atmeta.com 2) In Meta’s telling, the product goal is a more personal assistant that knows a user’s history, interests, content and relationships. Zuckerberg said on the January 28 earnings call that Meta is also trying to merge large language models with the recommendation systems behind Facebook, Instagram, Threads and its ads business. (s21.q4cdn.com) That requires more than flashy chatbots. It requires the back-room machinery that trains models, runs them every time a user asks a question, and keeps the system fast enough and cheap enough to serve billions of people. (about.fb.com 1) (about.fb.com 2) Meta has spent the past two months laying out that machinery in public. On February 17 it announced a multiyear infrastructure partnership with NVIDIA for data centers built for both training and inference, and on March 11 it said four new generations of its Meta Training and Inference Accelerator chips are due within two years. (about.fb.com 1) (about.fb.com 2) On March 24, Meta said it was partnering with Arm on a new class of central processing units for data centers and large-scale artificial intelligence deployments. Meta said its data centers were outgrowing the limits of traditional central processing units as it pushed more artificial intelligence workloads into production. (about.fb.com) The product side is starting to show up. Meta launched its standalone Meta AI app in April 2025 as “a first step toward building a more personal AI,” and on April 8, 2026, it introduced Muse Spark, a new model now powering the Meta AI app and website, with rollouts planned for WhatsApp, Instagram, Facebook, Messenger and artificial intelligence glasses. (about.fb.com) (about.fb.com) The financial tradeoff is already visible in Meta’s numbers. Revenue rose to $200.97 billion in 2025 from $164.50 billion in 2024, but costs and expenses also climbed to $117.69 billion from $95.12 billion as the company poured money into infrastructure and technical hiring. (investor.atmeta.com) (investor.atmeta.com) Meta’s pitch is that the spending supports both new artificial intelligence products and its existing ad machine. If that holds, the company is not just buying more chips and buildings; it is trying to turn compute into the core asset behind its next business cycle. (s21.q4cdn.com) (about.fb.com)
