Meta's AI Spending Paradox: Hundreds of Billions Committed, Usage Caps Now Imposed
Having committed tens of billions to GPU clouds, chip partnerships, and data center buildouts over 18 months, Meta now reportedly joins Amazon and Walmart in capping internal AI service consumption — a signal that even the industry's biggest builders are confronting inference economics at scale.
A report out of South Korea on June 22, 2026 captured an unlikely juxtaposition: Meta Platforms, Amazon, and Walmart — three of the world's most aggressive spenders on artificial intelligence infrastructure — have reportedly begun capping employee and internal AI service consumption to manage costs. For Meta, the disclosure arrives with particular irony. Over the preceding 18 months, the company had committed to one of the most audacious capital deployment programs in technology history, locking billions into GPU cloud contracts, multi-supplier chip agreements, and data center buildouts spanning multiple continents. That the same company is now pulling back on the demand side of its own AI ledger is not simply a contradiction; it is a disclosure: at sufficient scale, even the infrastructure builders cannot escape the unit economics of inference.
The contractual commitments were staggering in both size and velocity. In September 2025, Meta signed a reported $14 billion agreement with CoreWeave — the GPU-focused neocloud that has emerged as a critical provider for AI inference workloads — securing dedicated compute capacity. By April 2026, that agreement had been expanded to $21 billion, what multiple reports described as the largest commitment by any hyperscaler to a single cloud vendor. Simultaneously, Meta announced a multi-year partnership with AMD in February 2026, with terms reported at over $100 billion across a 6-gigawatt infrastructure footprint — a figure that, if accurate, represents a generational bet on AMD's Instinct GPU roadmap as a credible alternative to Nvidia's dominant hardware lines. Meta and Nvidia also announced what was characterized in contemporaneous reporting as a multi-year strategic alliance, covering next-generation GPUs, standalone Vera CPUs — which Meta had already begun deploying at scale by February 2026 — and AI integration for WhatsApp. By June 2026, Broadcom was confirmed to be supplying custom AI silicon to Meta alongside Google, Anthropic, and OpenAI, further expanding Meta's chip sourcing matrix.
The groundwork for this spending wave was laid in January 2025, when Meta committed to investing $60 billion or more in AI infrastructure for the full year; Mark Zuckerberg refined the figure to $65 billion, tied to a target of building data centers housing 1.3 million AI GPUs. A 5-gigawatt data center campus was announced in July 2025, and Meta publicly detailed its Catalina Pod architecture in August of that year — a system coupling Nvidia's Blackwell GB200 NVL72 with Open Rack v3 and liquid cooling, the physical expression of its AI ambitions. The same month, however, brought an early warning: Meta's Hyperion data center project was reported to be facing roughly $5 billion in cost overruns, an unverified figure that nonetheless foreshadowed the cost discipline becoming explicit by mid-2026. To diversify supply risk and reduce dependence on any single vendor, Meta pursued parallel paths: deploying millions of AWS Graviton ARM cores by April 2026, acquiring custom silicon startup Rivos in September 2025, and adopting Nvidia's Spectrum-X Ethernet switching for its largest GPU clusters alongside Oracle by October 2025. The May 2026 disclosure that Meta is among the early adopters of Nvidia's Vera CPU — alongside CoreWeave, Oracle, and Alibaba — extended that diversification into the emerging CPU-GPU convergence layer.
The reported usage caps must be interpreted carefully against this backdrop. For a company operating at Meta's compute scale, restricting employee access to AI coding assistants or internal chatbots represents operationally negligible savings relative to a $21 billion CoreWeave commitment. What such caps more plausibly signal is the institutionalization of cost attribution — imposing budget accountability on business units that had been consuming AI services without direct cost allocation, a pattern common across the enterprise technology adoption cycle. Amazon and Walmart's parallel moves reinforce that this is a sector-wide rationalization rather than a Meta-specific retreat. Meta's underlying economics complicate the calculus in a way that distinguishes it from its hyperscaler peers. Unlike Microsoft, which charges enterprise customers directly for Copilot access, or Google, which monetizes Gemini through cloud contracts, Meta's AI investments are primarily defensive: embedded in recommendation systems and content ranking that sustain advertising revenue, with no per-query pricing mechanism to offset inference costs. The open-source Llama strategy, while generating substantial ecosystem goodwill and reducing third-party model licensing expenditure, does not produce direct AI revenue in the manner of a proprietary API business.
Looking ahead, Meta's structural position in the AI infrastructure buildout remains formidable. Its multi-supplier chip strategy — threading AMD, Nvidia, Broadcom custom silicon, AWS Graviton, and future Rivos-derived accelerators — provides negotiating leverage and supply resilience that few peers can replicate. Participation in an AI fiber infrastructure venture backed by Amazon and Nvidia, reported in June 2026, suggests Meta is investing in the network layer beneath the compute stack as well. Yet the risks scale with the commitments. Long-term supply agreements totaling hundreds of billions of dollars — the AMD 6GW deal alone spans years — create balance sheet obligations predicated on sustained AI demand growth and continued advertising market health. If AI inference efficiency improves faster than current trajectories suggest — and DeepSeek's January 2025 emergence demonstrated that meaningful capability gains remain achievable even under hardware constraints — the compute capacity being assembled could outpace near-term demand, compressing returns on deployed capital. Three signals will clarify the picture: whether Meta's 2026 and 2027 capital expenditure guidance holds or is revised as cost caps deepen; the pace at which Rivos-derived custom silicon reaches production and reduces external procurement volumes; and whether the open-source Llama ecosystem generates developer lock-in sufficient to justify infrastructure costs absent a direct monetization pathway.