Amsterdam-based Nebius Group signed a five-year agreement with Meta to deliver up to $27 billion in AI infrastructure services, including GPU capacity and specialized hardware. The deal commits Meta to $12 billion in contracted compute capacity with options scaling to the full $27 billion ceiling depending on deployment velocity and model training demands through 2030.

Nebius, a spinoff from Yandex's cloud and autonomous driving units, has positioned itself as infrastructure-as-a-service for frontier AI labs unwilling to wait eighteen months for Nvidia delivery schedules. The Meta contract represents the largest known third-party compute agreement disclosed this cycle, dwarfing Oracle's $8 billion multi-cloud deal with OpenAI last year and matching the scale of internal capex budgets at mid-tier hyperscalers. Meta confirmed the arrangement but declined to specify GPU counts or cluster architectures, noting only that Nebius will operate geographically distributed sites with direct fiber to Meta's existing backbone.

The deal solves two problems. Meta avoids $12 billion in upfront capital expenditure and the eighteen-month lead time on data center construction, while Nebius converts Nvidia H100 and forthcoming B200 allocations into contracted revenue at margins estimated near 28-32% based on comparable infrastructure leases. For Meta, this is capacity arbitrage: paying a premium to Nebius in exchange for flexibility to scale Llama 4 training runs without committing balance sheet or executive bandwidth to site permitting in Iowa. For Nebius, it is vendor financing at scale, using access to Nvidia supply as the wedge into hyperscaler budgets previously allocated to internal build teams.

The structure matters for allocators tracking AI capex visibility. Meta's $60-65 billion 2025 capex guidance, disclosed in February earnings, now includes external compute contracts, blurring the line between traditional capex and opex lease commitments. This mirrors the 2015-2018 shift when hyperscalers moved from owned data centers to long-term colocation agreements with Equinix and Digital Realty, reclassifying billions in capital outlays as operating expenses with better return profiles. If Meta's CFO can show 15-20% lower total cost of ownership through Nebius versus internal build, expect Amazon and Google to explore similar structures by mid-2026, particularly for edge inference clusters where utilization is less predictable.

Watch three follow-on events. First, Nebius must disclose financing arrangements by May when its next quarterly results publish; the company will likely tap project finance or sale-leaseback structures to fund GPU purchases without diluting equity. Second, Nvidia's April earnings call may address how much of its $120 billion trailing-twelve-month revenue now flows through infrastructure intermediaries rather than direct hyperscaler sales, a metric analysts have not broken out cleanly. Third, Meta's Q2 earnings in late July will clarify whether the Nebius capacity accelerates Llama 4 training timelines or simply offloads peak demand during multi-model parallel runs.

The contract validates Nebius as the first non-US firm to capture double-digit-billion AI infrastructure commitments from a hyperscaler. Meta gets compute without the capital. Nebius gets contracted cash flow that makes it bankable for the next GPU cycle. The rest of the market gets a template for how to turn Nvidia scarcity into infrastructure equity.