Meta Platforms has locked in a five-year, $12 billion package of dedicated AI computing capacity from Amsterdam-based Nebius, with a contingent option that could lift the contract’s value to as much as $27 billion by 2027. The deal adds another jolt to the AI infrastructure trade as Meta META leans further into spending on training and inference. Nebius, which uses Nvidia NVDA processors to deliver capacity, framed the agreement as a scale test it is prepared to meet. “Both contracted tranches related to Meta were successfully delivered on time, and we are now fully in the servicing stage,” CEO Arkady Volozh said in a recent letter to shareholders, underscoring execution on earlier phases as the new build ramps.
The agreement gives Meta $12 billion of dedicated AI capacity across multiple locations by 2027, according to Nebius. On top of that, Meta will purchase up to $15 billion of additional capacity earmarked by Nebius over the coming five years if the provider cannot sell it to other customers—an embedded backstop that effectively sets a floor under Nebius’s buildout and secures Meta overflow access in a tight market. Monday’s disclosure follows an initial $3 billion arrangement the two companies signed in November, expanding an alliance that shifts some of Meta’s compute sourcing away from the hyperscale incumbents. The clear message: Meta intends to insulate its model roadmap from supply shocks in GPUs, data center power, and rack space by diversifying suppliers and pre-buying scale.
Meta has been open about ramping AI capital outlays to support generative features across Facebook, Instagram, WhatsApp, and its core ad stack. With Microsoft MSFT and Alphabet investing heavily and OpenAI’s model cadence accelerating, delay risk is now a material business risk. Locking up compute with a specialist “neocloud” player reduces dependence on shared public-cloud clusters, can improve cost visibility, and, crucially, tightens delivery timelines with dedicated facilities. For Meta, time-to-train is money: shaving weeks from a training run can translate into faster product iteration and improved ad relevance. The structure of the Nebius deal—dedicated capacity plus a take-or-pay style backstop—signals Meta is prioritizing certainty over optionality. It also hints at continued scarcity in high-end accelerators and power-constrained data center slots through 2027.
Nebius, branding itself as a neocloud provider, sells hardware and AI cloud capacity as services to large tech customers. It runs on Nvidia silicon and is positioning as a scaled alternative to hyperscalers for workloads that demand guaranteed GPU pools and custom interconnects. The Meta pact builds on Nebius’s earlier multiyear contract with Microsoft that totaled $17.4 billion, and an initial $3 billion phase with Meta. According to Nebius investor materials, GPU deployments tied to Meta were staged in December 2025 and February 2026, a cadence that both validated its construction pipeline and de-risked this larger commitment. Management says cash flow from the Meta work is earmarked to help finance capital expenditures needed for execution, a reinvestment loop that could strengthen the company’s footprint as additional customers lean in. Investors have treated the Meta win as a credibility test; the backstop clause provides demand visibility, but the scale-up still hinges on flawless delivery.
The risk side of the ledger is straightforward. GPU supply remains lumpy even as Nvidia’s next-generation parts come online, with long lead times for networking gear and liquid cooling. Power procurement—and the ability to energize campuses on schedule across multiple jurisdictions—has become the gating factor for new AI data centers. European hubs are facing grid constraints and permitting friction; the U.S. is seeing utility interconnection queues stretch project timelines. Nebius’s multi-location promise spreads risk but also multiplies the variables: construction partners, local incentives, and regulatory compliance. The contract’s termination protections, while standard in deals of this size, are a reminder that minor slippage can cascade in an environment where a few weeks can cost millions in lost model training cycles. For Meta, that argues for overbooking to ensure continuity; for Nebius, it demands ruthless project management.
For Nvidia, the agreement is another signal that demand for high-bandwidth, scale-out GPU clusters remains well ahead of general supply. Each incremental dedicated build locks in networking (InfiniBand or Ethernet variants), memory, and power delivery kit, pulling through a broader ecosystem of switch vendors, server assemblers, and data center operators. The deal also hardens expectations that high-end accelerator capacity will be persistently tight into 2027, supporting pricing and allocation leverage for Nvidia’s data center segment. AMD’s accelerator ambitions will keep it in the conversation for future tranches, but today’s contract language and Nebius’s existing fleet suggest Nvidia remains the default choice for Meta’s near-term training needs. That dynamic matters for investors handicapping how long the AI capex supercycle can run before normalization.
The Meta-Nebius structure—dedicated, multi-year, and backstopped—underlines a broader shift in how large platforms source compute. Instead of renting elastic capacity from general-purpose clouds, companies are carving out bespoke pools with specialists to secure performance, cost predictability, and guaranteed access windows. It is a model closer to wholesale long-term power purchases than pay-as-you-go cloud, and it moves more AI cost into quasi-fixed commitments. For Meta, that could temper short-term free-cash-flow variability while raising the bar on utilization discipline. For hyperscalers, the competitive response may be more private cloud-like offerings and tighter SLAs on reserved GPU clusters. Either way, the days of relying solely on spot access to top-tier accelerators are over for any platform trying to train frontier-scale models on a fixed cadence.
Milestones are clear. Nebius says it will deliver the $12 billion of dedicated capacity by 2027, adding to deployments already completed for Meta in late 2025 and early 2026. Watch for updates on site energization, rack arrivals, and the transition from construction to steady-state operations at each location. Power contracts and grid interconnections will be the tell on schedule risk. On Meta’s side, capex guidance and commentary on AI training throughput will indicate whether the added capacity is translating into faster model turnarounds and new product launches. The embedded $15 billion backstop becomes pivotal if industry demand cools and Nebius needs Meta to absorb unallocated capacity; in a still-scarce world, it may never be triggered. For now, the deal extends the AI infrastructure arms race and puts another dedicated supplier on the field, with Nvidia and a web of data center vendors riding shotgun.
