Meta has expanded its relationship with CoreWeave through a new AI infrastructure agreement valued at approximately $21 billion and valid until December 2032. The deal, announced by CoreWeave and detailed in a document submitted to the SEC, cements the company as one of the leading providers of specialized cloud capacity for large-scale AI workloads.
The agreement is not starting from zero. According to the regulatory notification, the new order was signed on March 31, 2026, under the umbrella of an existing framework agreement between the two companies, dated December 10, 2023. The $21 billion figure encompasses both access to new computing capacity until December 20, 2032 and the exercise of a previous option to access additional contracted capacity in a prior order, valid until April 10, 2032.
From a technical perspective, the most notable aspect is that the dedicated capacity will be deployed across multiple locations and will include some of the first deployments of NVIDIA Vera Rubin platform, the upcoming major generation of the company’s accelerated infrastructure. CoreWeave has not publicly detailed how many systems or what exact configurations will be part of the agreement, but the fact that Meta secures early access to Rubin clearly indicates how much the race for the next wave of AI computing is already underway long before the hardware hits the mass market.
The message is clear: inference is now king
The CoreWeave statement itself emphasizes a concept that is increasingly important in the industry: inference. Meta will use CoreWeave’s cloud platform to scale inference loads, which means the actual execution of already trained models in production environments. This shift matters because, for a long time, AI discussions focused almost exclusively on training large models, whereas now the bottleneck is shifting to when these models need to respond, generate content, or serve millions of users.
This change in focus alters the economics of infrastructure. Training a foundational model remains very expensive, but serving it at scale may require even greater operational uptime, energy efficiency, lower latency, and a more stable GPU supply chain. For Meta, which integrates AI into products like Facebook, Instagram, WhatsApp, Messenger, and their smart glasses, ensuring inference capacity is not a luxury but an industrial necessity. Reuters previously reported that the company anticipates capital expenditures between $115 billion and $135 billion in 2026, precisely to accelerate its AI infrastructure.
CoreWeave gains ground against traditional cloud providers
This agreement also reinforces an increasingly visible trend: the rise of AI-specific clouds competing for a portion of the market that was previously dominated by hyperscalers. CoreWeave presents itself as an “AI cloud” built specifically for such workloads, rather than a general cloud platform adapted post hoc to the AI boom. According to the company’s own narrative, this specialization is enabling it to attract major labs, startups, and tech giants.
In recent days, CoreWeave has announced several high-profile deals. First was the reinforcement of its agreement with Meta, and shortly after, the company disclosed a multi-year partnership with Anthropic to support the Claude model family. The market’s interpretation is straightforward: CoreWeave is positioning itself as a key component in AI infrastructure, not just as an opportunistic provider of rented GPUs.
Alongside this, a significant technological takeaway is the distributed deployment “across multiple locations” highlighted by CoreWeave. It suggests an architecture designed to optimize performance, resilience, and scalability for Meta’s operations. This aligns with the current industry reality: having large capacity in a single data center is no longer enough. Production AI workloads demand geographic distribution, redundancy, and a network capable of handling vast data volumes without compromising response times.
Vera Rubin enters the scene ahead of schedule
The mention of NVIDIA Vera Rubin is arguably the most strategic detail in the announcement. NVIDIA introduced Rubin as the next major AI platform after Blackwell, designed to push training and inference performance even further. That Meta has secured early access to some of these initial deployments via CoreWeave hints at two things: first, that access to next-generation hardware is already being secured through multi-year agreements; second, that providers like CoreWeave are serving as preferred channels to bring this silicon into production before many traditional players.
This also explains why AI cloud services are evolving into a distinctly different business from traditional cloud. Success is no longer solely about regions or managed services, but about who can secure prime access to coveted hardware first, with energy, networking, cooling, and software optimized for immediate deployment. The Meta deal sends a powerful signal of technical trust towards CoreWeave in this highly competitive landscape.
In summary, Meta isn’t just purchasing capacity; it’s securing part of the near-term future of its AI infrastructure at a time when inference is scaling faster than ever and when upcoming GPU generations are already spoken for even before mass deployment. For CoreWeave, this deal is another validation of its model. For the industry, it demonstrates that the AI wars are fought as much over who controls the computing infrastructure as over the models themselves.
Frequently Asked Questions
How much is the new Meta-CoreWeave agreement worth?
The initial committed amount is approximately $21 billion, according to CoreWeave and its SEC documentation.
How long does the Meta-CoreWeave contract last?
The new capacity access extends until December 20, 2032, and includes an option that remains valid until April 10, 2032.
What will Meta use this infrastructure for?
CoreWeave explained that Meta will primarily use it to scale AI inference workloads, meaning executing models in real-world production environments.
What role does NVIDIA Vera Rubin play in this deal?
Part of the contracted capacity will include some of the first deployments of the NVIDIA Vera Rubin platform, the upcoming major generation of accelerated AI infrastructure.