Meta is developing its own chips for training AI models

The company, which spends tens of billions of dollars per year on artificial intelligence infrastructure, no longer wants to depend on third-party processors even at the most critical stage — model training. Meta Platforms' Chief Financial Officer confirmed that the company is developing custom chips intended for training future AI models. This announcement comes against the backdrop of recently concluded major contracts with leading semiconductor manufacturers and signals that Meta's strategy in computational infrastructure is becoming significantly more ambitious.

To understand the scale of this step, one needs to understand the context. Until now, Meta's custom chips — primarily the MTIA family — were oriented toward inference, that is, running already-trained models. Inference is a simpler task: the model already knows what to do, it just needs to produce results quickly and efficiently. Training is a completely different beast. Training a large language model requires enormous computational power, extremely complex coordination between thousands of processors, and outstanding memory bandwidth. This is precisely why NVIDIA with its H100 and B200 GPU series remains practically the only alternative supplier for this stage. To announce the development of a custom training chip is to challenge this monopoly.

Meta's motivation is quite transparent and measured in concrete numbers. In 2025, the company spent on capital expenditures, primarily related to AI infrastructure, approximately $35-40 billion. A significant portion of these funds went toward GPU purchases from NVIDIA. At such scales, even a slight reduction in dependence on an external supplier could save billions. But it's not just about money. Custom chips provide the ability to optimize architecture for specific needs: for the architecture of LLaMA models, for the specifics of distributed training in Meta's data centers, for unique load patterns. NVIDIA's universal GPUs are excellent, but they are precisely universal — which means they inevitably carry compromises.

It is noteworthy that this announcement came after Meta concluded major deals with leading chipmakers. It might seem strange: why invest in custom developments if partnerships with Broadcom, TSMC, and other players already provide access to cutting-edge technologies? The answer lies in the strategic planning horizon. Developing a training chip from scratch is a process that takes three to five years from concept to mass production. Meta is playing the long game: current contracts cover today's needs, while custom silicon should ensure independence in the next decade. This is the same logic followed by Google with TPU, Amazon with Trainium, and Microsoft with Maia — the largest consumers of computational power cannot afford strategic dependence on a single supplier.

However, the path from ambitions to a working training chip is fraught with technical and organizational complexities. It took Google almost a decade to bring TPU to the level where the company could train its largest models primarily on its own hardware. Amazon, despite significant investments in Trainium, still faces questions about software ecosystem and compatibility. Creating a chip is half the battle. You need to build a complete stack: compilers, frameworks, debugging tools, monitoring systems. You need to convince your own researchers, accustomed to CUDA and PyTorch, to transition to a new platform. You need to ensure reliability at the level of thousands of chips working together for months without failures.

For NVIDIA, this announcement is yet another warning sign in a long series of similar ones. Every major client announcing its own chip program potentially reduces the addressable market for Jensen Huang's company. However, experience shows that real replacement happens slowly. Even Google, which possesses the most mature custom AI chips in the industry, continues to purchase GPUs from NVIDIA for certain tasks. Most likely, Meta is moving toward a hybrid model, where custom chips gradually take on an increasing share of the load, while purchases from third-party suppliers remain as insurance and supplement.

Meta's decision to develop training chips is not just corporate news but an indicator of a fundamental shift in the industry. The era when one supplier could control the computational foundation of the entire AI revolution is coming to an end. The world's largest technology companies are consistently building vertical integration — from models and data to the very silicon on which it all runs. The question is no longer whether Meta can create a competitive training chip, but how quickly it can do so — and how this will reshape the map of the multi-billion-dollar AI infrastructure market.