Meta raises AI capex plan to $125–145 billion

Meta on Wednesday raised its 2026 capex estimate to $125–145 billion.

Reuters reported on April 29 that Meta Platforms lifted the range from its previous $115–135 billion forecast. The company is doing this while also looking for cost savings and planning layoffs. CNBC reported the same capex range in its earnings coverage.

Fact: this is not a model launch. It is a financial signal about how much infrastructure one of the largest AI platforms believes it must control to compete. Reuters wrote that Meta is spending heavily on AI infrastructure, high compensation for AI talent and Meta Superintelligence Labs. At the same time, the company is embedding AI deeper into advertising, WhatsApp, Threads and internal workflows.

For executives outside Big Tech, the issue is not Meta’s share price. The issue is the cost curve. When hyperscalers and platform companies keep expanding data center and chip spending, capacity, energy access and model availability become strategic procurement questions, not just IT architecture choices.

The CIO consequence is direct: the AI roadmap needs a capacity strategy. Which workloads should run on a hyperscaler? What can run more cheaply on smaller models, open-weight models or specialized inference platforms? Which suppliers will still have capacity if prices or allocation priorities change?

CFOs and boards should read this as a call for stricter ROI discipline. It is tempting to copy Big Tech’s language and assume more AI capacity always creates more value. For most companies, that is the wrong starting point. Start with processes where the effect can be measured: customer service, developer productivity, case handling, market analysis, document workflows and security operations. Tie every initiative to cost per transaction, quality, risk and time saved.

Assessment: Meta’s forecast reinforces a pattern already visible at Microsoft, Google, Amazon, OpenAI and Chinese AI players: AI infrastructure is becoming a source of market power. That can produce better services, but it also raises vendor lock-in risk. Companies should require portability in architecture, clear contract exit options and ongoing measurement of real usage.

The practical advice: do not lock the 2026 budget to one AI platform without stress-testing it. Build a simple matrix for data sensitivity, latency, cost, model quality and regulatory risk. It is not glamorous work, but it is how AI moves from pilot to governed operations.