Meta to start producing its own AI chips in September 2026

On July 9, 2026, Meta announced that its new AI chips will enter mass production in September 2026. The company is taking a modular approach to development, which will let it adapt the architecture as AI requirements change.

Why modularity makes sense

When a company designs a specialized chip, it has to predict which computations will be critical one to two years out. With artificial intelligence that is nearly impossible: requirements for bandwidth, memory capacity and neural network architecture change monthly. Instead of building a single monolithic die, Meta chose a practical solution: the chips are designed modularly, so individual blocks (compute cores, caches, memory interfaces) can be redesigned or replaced without reworking the entire architecture. It is like the difference between a monolithic building and a LEGO set: the latter is far easier to adapt to new conditions.

Why it matters in 2026

AI models are growing exponentially, and computing demands are now completely different from two years ago. Where the emphasis used to be on memory bandwidth, the need now is for low latency and specific types of operations for transformer architectures. Meta is investing billions in its own infrastructure for training large language models and recommendation systems. Had the company designed its chips statically, to 2024 requirements, it would be shipping hardware that was obsolete from the start by the end of 2026. Modularity allows fast iteration: ship the first batch in September, find bottlenecks and overheating points in practice, and then redesign the necessary blocks next year without redoing the whole project.

Independence from suppliers

For the big AI labs (OpenAI, Google, Meta, Anthropic), access to powerful GPUs and specialized chips is a strategic resource. If you depend on NVIDIA or other suppliers, you are tied to their delivery schedules, pricing and commercial decisions. Meta, like Google, which already uses its own TPUs, wants independence from external monopolists. Its own chips provide:

• Full control over costs and delivery times
• The ability to optimize the architecture for its own models
• Less vulnerability to export restrictions on AI hardware

Vertical integration as the new standard

The production start in September 2026 is a milestone. Meta is signaling that it competes with OpenAI, Google and Anthropic not only at the level of models and services, but also at the hardware level. The industry is moving toward vertical integration: companies that control the whole stack (models + chips + platforms + software) gain a decisive advantage in development speed and overall economics.

What this means

When a heavyweight like Meta announces it is starting its own chip production in September, this is not just an internal project. It signals that AI competition is reaching a new level: a race for control over hardware, which is now often more important than control over algorithms.

*Meta is designated an extremist organization and banned in Russia.