Company file
Meta
A systems builder treating custom silicon, networking, optics, power, sites, and internal workloads as one coordinated deployment challenge.
Meta's Corning agreement supports domestic optical connectivity, while its Broadcom partnership connects custom silicon to a multi-gigawatt deployment plan. Those signals broaden the map without becoming a commitment-intensity row.
Operating profile
How Meta builds access
- Operating model
- Compute platform and custom-silicon builder
- Controls through
- Internal workloads, custom-silicon programs, supplier partnerships, data-center investment, and deployment planning
- Physical stack
- Custom accelerators, merchant GPUs, optics, networking, power, cooling, construction, and operations
The fleet is the product
Meta's infrastructure work starts with a simple but powerful idea: a useful AI capability depends on a fleet, not just a chip. That fleet includes custom silicon, merchant accelerators, networking, optical connectivity, data centers, power, cooling, and the software and operating teams that turn all of those components into a reliable internal service. The Company File follows that integrated system rather than isolating one supplier announcement.
Custom silicon adds another control layer
Meta and Broadcom announced a multi-generation custom-AI-silicon partnership through 2029, beginning with an MTIA deployment commitment above one gigawatt and described as the first stage of a multi-gigawatt rollout. That is a forward-looking deployment signal, not a disclosed manufacturing commitment. Its relevance is architectural: Meta is building more of the compute system around its own workloads and operating requirements.
Supplier diversification reaches across the stack
Meta also announced a multiyear agreement for up to six gigawatts of AMD Instinct GPUs and an up-to-$6 billion Corning agreement for optical fiber, cable, and connectivity. These agreements are not interchangeable. One is a hardware-capacity roadmap and the other supports a physical networking layer. Together they illustrate a systems builder creating several routes through the constraints that can slow a large deployment.
Capital spending funds a wider operating machine
Meta's 2026 CapEx guidance rose to $125 billion to $145 billion. That is aggregate infrastructure spending, not a semiconductor supply commitment. The useful reading is broader: deployment at scale requires money for sites, servers, networking, power, and the surrounding infrastructure. CapEx tells one part of that story, while supplier agreements and custom-silicon programs reveal other layers.
A portfolio approach makes the fleet more adaptable
Meta's custom-silicon work, merchant GPU agreements, and optical-connectivity investments create several paths for expanding the fleet. Each path has a different technical and operating role, which is exactly what makes the portfolio useful. The company can align a workload with the most suitable system while preserving the ability to learn from multiple suppliers, architectures, and deployment patterns as its infrastructure grows.
What to watch
The central question is whether custom silicon, merchant capacity, optics, power, sites, and operating teams can expand on compatible timelines. Meta is an aspirational example of treating compute as a physical system with many moving parts. The test is not a single headline commitment. It is whether the complete fleet becomes more capable, more efficient, and more reliable as those parts come online.