Meta is reportedly in discussions with Anthropic over a potentially massive data centre arrangement that could be worth as much as $10bn, according to a report from the Financial Times. The talks, if they progress, would underline a shift that has been building quietly for months: the most important “AI deals” are no longer only about models or chips, but about the physical capacity required to run them—at scale, reliably, and with enough flexibility to handle changing demand.
At the same time, Meta’s broader infrastructure push provides context for why this kind of conversation is happening now. The company has been spending heavily—around $145bn in total on infrastructure—while working to expand the compute footprint needed for training and deploying AI systems across its platforms. In other words, Meta is not simply buying capacity; it is building an ecosystem. A deal with Anthropic would fit neatly into that ecosystem, offering a way to lock in demand and align long-term capacity planning with one of the most prominent AI developers focused on safety-oriented, high-performance model development.
What makes the reported talks particularly notable is the pairing itself. Meta is a hyperscale operator with global data centre ambitions and a deep understanding of large-scale systems engineering. Anthropic, meanwhile, is a leading AI lab whose work depends on sustained access to high-end compute and the operational discipline required to keep training and inference running efficiently. When these two meet around data centres, the conversation is likely less about “renting servers” and more about designing a pipeline: how compute is provisioned, how workloads are scheduled, how power and cooling constraints are managed, and how performance targets are met under real-world conditions.
The headline number—up to $10bn—signals that this is not a small procurement. Deals at this scale typically involve multi-year commitments, significant build-out or capacity reservation, and detailed service-level expectations. Even without the final terms, the structure implied by such a figure suggests a relationship that could resemble a hybrid between a traditional infrastructure contract and a strategic partnership. For Anthropic, the value would be continuity: predictable access to the compute needed for both training runs and ongoing deployment. For Meta, the value would be utilization and credibility: ensuring that expensive capacity investments translate into stable revenue streams or at least into stronger internal efficiency and bargaining power across the supply chain.
Why data centres are becoming the center of gravity for AI
For years, AI competition was framed around model quality and research breakthroughs. But the practical reality is that even the best model is only as useful as the infrastructure that can support it. Training large models is expensive and time-sensitive; inference is relentless and unpredictable, driven by user demand, product experimentation, and shifting optimization strategies. That means the bottleneck is increasingly physical: power availability, network bandwidth, GPU throughput, cooling capacity, and the ability to scale quickly without sacrificing reliability.
Data centres are also where the economics of AI become tangible. The cost of compute is not just the price of GPUs; it includes the entire stack required to keep them fed with electricity, connected with low-latency networking, and maintained with uptime guarantees. As AI workloads intensify, the “infrastructure layer” becomes a competitive advantage. Companies that can secure capacity early, negotiate favorable terms, and operate efficiently can move faster and reduce the risk of expensive delays.
This is why partnerships around data centres are emerging as a key part of the AI supply chain. They are not glamorous, but they are foundational. A model can be improved in weeks; a data centre build-out can take months or longer, and power procurement can be even slower. The companies that plan well can avoid the worst-case scenario: being ready with the technology but unable to run it at the scale required.
Meta’s infrastructure spending and the strategic logic behind it
Meta’s reported $145bn infrastructure spending is a reminder that the company is treating compute as a long-term asset rather than a short-term expense. Hyperscalers have always invested heavily in data centres, but the AI era changes the nature of the investment. Traditional web-scale workloads are bursty and often tolerant of incremental scaling. AI workloads, especially training, are more concentrated and can require large, coordinated clusters. Inference workloads can also be spiky, but they tend to be continuous once deployed widely.
Meta’s approach appears to be aimed at building a compute platform that can serve multiple purposes: internal training, product deployment, and potentially external customers if it expands into cloud services. The FT report’s mention of Meta considering launching a cloud business is important because it reframes what a data centre deal could mean. If Meta is moving toward offering AI infrastructure externally, then securing a major customer like Anthropic would be both a commercial milestone and a technical validation.
A unique angle here is that Meta’s cloud ambitions would not necessarily be limited to generic hosting. The AI market increasingly rewards specialized infrastructure: optimized networking, scheduling tailored to training jobs, and operational tooling that reduces friction for model developers. If Meta can offer a differentiated AI infrastructure experience—rather than just commodity capacity—it could attract labs and enterprises that want performance and predictability.
In that context, a deal with Anthropic could function as a proving ground. It would test whether Meta’s infrastructure strategy can meet the demanding requirements of a top-tier AI developer. It would also help Meta refine its service offerings, from how it provisions resources to how it handles scaling and reliability.
Why Anthropic would care about Meta’s capacity
Anthropic’s involvement suggests that the lab is seeking dependable access to high-performance compute. For AI labs, the challenge is not only obtaining GPUs but ensuring that the entire system behaves consistently under heavy load. Training runs can be sensitive to interruptions, and performance can vary depending on network topology, storage throughput, and cluster management. Even small inefficiencies can translate into large costs when multiplied across repeated experiments.
A multi-year data centre arrangement could provide Anthropic with several advantages:
First, it reduces uncertainty. Instead of scrambling for capacity when demand spikes, Anthropic could plan training schedules with greater confidence. That matters because AI development is iterative: teams run experiments, evaluate results, adjust architectures, and repeat. Predictable infrastructure access helps maintain momentum.
Second, it can improve cost control. Large infrastructure contracts often come with pricing structures that are more favorable than spot purchases, especially when capacity is reserved in advance. While the exact economics are unknown, the scale implied by a potential $10bn deal suggests that both parties would be negotiating terms designed to make the arrangement financially sustainable.
Third, it can enable better performance tuning. If Anthropic’s workloads are expected to run on Meta’s infrastructure, Meta can optimize the environment specifically for those workloads—networking patterns, scheduling policies, and operational practices. Over time, that can yield measurable improvements in throughput and reduced training time.
Finally, it can strengthen Anthropic’s position in a market where compute access is increasingly strategic. As more companies compete for limited capacity, the ability to secure reliable infrastructure becomes a differentiator. Labs that can lock in compute can move faster and iterate more effectively.
The cloud question: is Meta building a new revenue engine?
The FT report frames the talks alongside the possibility that Meta could launch a cloud business. That would be a major shift in how Meta monetizes its infrastructure. Historically, hyperscalers have offered cloud services to external customers, but Meta’s primary business has been advertising and social platforms. Entering cloud would mean competing in a market dominated by established players with mature ecosystems.
However, Meta’s advantage could be its scale and its AI-specific focus. If Meta is already investing tens of billions in infrastructure for its own AI needs, it may be able to repurpose some of that capacity for external customers. The key is whether it can do so without compromising internal performance. That requires careful capacity planning and strong operational discipline.
A deal with Anthropic could help Meta answer a critical question: can it deliver the kind of infrastructure experience that serious AI developers require? If the answer is yes, Meta could use that credibility to attract additional customers—other AI labs, enterprises building AI products, and potentially government or research organizations with strict reliability requirements.
There is also a strategic signaling effect. When a major AI lab aligns with a hyperscaler on infrastructure, it sends a message to the market that the hyperscaler’s compute platform is not just large, but usable for frontier-grade workloads. That can accelerate adoption and reduce perceived risk for future customers.
What the deal could look like in practice
While the report does not provide details, deals of this magnitude typically involve several components. One possibility is that Meta would reserve or build data centre capacity specifically for Anthropic’s workloads. Another possibility is that Meta would provide access to a dedicated portion of its infrastructure, with performance guarantees and dedicated operational support.
In many AI infrastructure arrangements, the “what” is less important than the “how.” The most valuable elements often include:
1) Provisioning speed: How quickly can resources be allocated when training runs begin?
2) Cluster reliability: What is the uptime and how are failures handled?
3) Network performance: How is low-latency connectivity ensured for distributed training?
4) Power and cooling stability: How are thermal constraints managed to prevent throttling?
5) Data handling: How are storage and data pipelines optimized for training throughput?
6) Security and compliance: How are isolation and access controls implemented?
If Meta is serious about cloud services, it would likely want to demonstrate excellence in these areas. Anthropic, for its part, would want to ensure that the infrastructure supports the lab’s workflow rather than forcing it to adapt to limitations.
A unique take: this is less about “who has the GPUs” and more about “who can run the machine”
It’s tempting to view AI infrastructure deals as a simple contest over hardware. But the deeper story is operational. GPUs are necessary, but they are not sufficient. The real challenge is orchestrating complex workloads across large clusters while maintaining performance and reliability. That orchestration includes software layers, monitoring, scheduling, and the ability to respond to changing workload patterns.
In that sense, a data centre deal between Meta and Anthropic is also a deal about systems engineering maturity