Meta has announced that it will start manufacturing a new custom AI chip from September as part of its plan to boost its overall computing power to 14 GW. This is a major milestone in the company’s goal to reduce reliance on external chip suppliers.
Code-named Iris, the new chip has completed the testing phase without any major issues, signaling that it is ready for production. The chip is expected to power Meta’s AI inference workloads more efficiently while helping lower infrastructure costs. This is expected to give Meta greater control over its rapidly expanding AI infrastructure.
The significance of launching its own custom silicon extends beyond the chip itself. At BigDATAWire, we’ve been following a broader industry trend in which hyperscalers are trying to own more of their AI infrastructure. Many view this as an important strategic advantage, and Meta appears to agree.
Perhaps the biggest benefit to having your own silicon is that it can help lower operating costs. Meta performs billions of inference requests every day across Facebook, Instagram, WhatsApp, and Meta AI. Even a 10 to 15% improvement in performance per watt or cost per inference could save hundreds of millions of dollars annually.
Another key advantage is less reliance on Nvidia, which means Meta won’t have to necessarily align its strategy based on chip manufacturers’ product roadmap, pricing or supply chain fluctuations. These are factors that Meta can’t control, and this is where having internal capabilities offers flexibility. This in turn also offers greater leverage in negotiations with suppliers like Nvidia.
Meta also gets access to chips optimized for its own workloads. Unlike general-purpose GPUs that are designed to support a wide range of AI applications, Iris can be tailored specifically for the recommendation systems, ranking models, advertising algorithms, and GenAI services that power Meta’s platforms.
Another advantage is that Meta can optimize more than just the chip itself. It can build the silicon alongside its software, networking and data centers so the different pieces work together more efficiently. At Meta’s scale, even small improvements across the entire system can have a meaningful impact.
Building custom AI chips isn’t easy, though. It takes years to design a new chip, and the costs can run into the billions of dollars. That’s why only a handful of companies have the money and engineering talent to even attempt it.
Meta also has to compete with Nvidia, and that won’t be easy. Nvidia hasn’t just built powerful GPUs. It has spent years developing the software, networking technology and developer tools that make those chips useful. Building that kind of ecosystem is a much bigger challenge than building the chip itself.
It’s also worth remembering that Meta isn’t trying to build a chip for everyone. Iris is designed for Meta’s own infrastructure and workloads. Nvidia’s GPUs still offer much more flexibility and will likely continue to handle many of Meta’s biggest AI training jobs. The idea is simply to move more inference workloads onto hardware that Meta owns and controls.
This also isn’t Meta’s first move into custom silicon. The company has already outlined its Meta Training and Inference Accelerator (MTIA) roadmap, which includes several generations of AI chips for its own infrastructure. Meta hasn’t confirmed that Iris belongs to that family, but it clearly follows the same strategy.
Meta highlighted this challenge in a recent engineering blog, “Every day, billions of people on Meta’s platforms enjoy an array of AI-powered experiences ranging from personalized recommendations to AI assistants.”
“Meanwhile, the AI models that will define the next era of computing are evolving faster than any single hardware generation can anticipate. Serving a wide range of AI models on a global scale, while maintaining the lowest possible costs, is one of the most demanding infrastructure challenges in the industry.”
Meta plans to introduce a new AI chip roughly every six months through 2027. This is much faster than the traditional pace of releasing a new chip every year or two.
To support that expansion, the company has signed long-term supply agreements with Samsung Electronics for memory chips, SanDisk for flash storage, and Sumitomo Electric for fiber-optic equipment. Meta is also expected to spend up to $145B on AI infrastructure in 2026 alone. This just highlights the scale of its AI ambitions.
Meta isn’t the only one pushing for custom chips. Google has its TPUs. Amazon has Trainium and Inferentia. Microsoft has Maia. All of them are building chips for their own AI workloads instead of relying completely on Nvidia’s GPUs.
That doesn’t mean Nvidia is going anywhere. Its GPUs will almost certainly remain the first choice for training the biggest AI models. However, as AI becomes more important to these companies, owning more of the hardware behind it could help them cut costs and have more control over how they build and expand their AI infrastructure.
The post Meta’s Custom AI Chip Could Redefine the Economics of AI appeared first on AIwire.
