Meta has unveiled a new generation of in-house artificial intelligence chips, which signal a broader shift that AI is no longer just a software battle but a hardware war.
The company’s latest chips, which are part of its Meta Training and Inference Accelerator (MTIA) family, are designed to power everything from content recommendations to generative AI across its platforms, including Facebook and Instagram.
At the core of this strategy is a decisive move away from reliance on traditional chipmakers such as Nvidia and AMD, toward purpose-built processors optimised for Meta’s own workloads.
Despite the momentum, building AI chips is complex and costly, as Meta has faced setbacks, including abandoning a more advanced training chip design due to technical challenges.
For now, the company still depends heavily on partners like Nvidia and AMD for training large AI models, even as it expands its in-house capabilities.
Why Meta is building its own chips
Meta processes billions of AI-driven interactions daily across its platforms. Even marginal efficiency gains can translate into millions of dollars in savings annually.
Custom chips allow the company to cut infrastructure costs by optimising performance-per-watt. It will also reduce dependence on Nvidia, whose GPUs dominate the AI market.
It will also help to scale faster as AI demand surges globally, control its full AI stack, from models to hardware. This mirrors a broader industry trend, with Google, Amazon, and Microsoft all developing their own AI accelerators.
New generation of AI-first hardware
Meta has introduced multiple chips, which are MTIA 300, 400, 450, and 500, with a rapid release cycle of roughly every six months, far faster than the traditional multi-year semiconductor timeline.
These chips are tailored specifically for AI inference, the stage where trained models respond to user queries, generate content, or rank feeds. This is critical because inference workloads now dominate real-world AI usage at scale.
Reports have shown that some of these chips are already deployed in production, supporting recommendation systems and ad delivery across Meta’s apps.
A high-end version of the system is said to deliver up to 30 petaflops of performance with massive high-bandwidth memory, highlighting how far specialised AI hardware has evolved.
Implications
There is a rise of ‘AI-native’ chips because companies are now designing chips specifically for AI from the ground up instead of adapting existing hardware.
Meta’s six-month chip roadmap shows how quickly AI capabilities are evolving and how hardware must keep up. Rather than relying on a single supplier, companies are adopting multi-vendor and in-house strategies, combining custom silicon with external GPUs.
Meta alone expects to spend up to $135 billion on infrastructure in 2026, which highlights the scale of the AI hardware race.
What this means for the future of AI
Meta’s strategy shows a reality that the future of AI will be shaped as much by hardware as by algorithms.
As AI systems become more powerful and widespread, control over chips will determine cost, speed, and scalability.
Big Tech firms will increasingly act like semiconductor companies, and countries and regions may push for ‘AI sovereignty’ through local chip development.
For emerging markets such as Nigeria and across Africa, the shift could have downstream effects, thereby impacting everything from cloud pricing to access to AI-powered services.
Join BusinessDay whatsapp Channel, to stay up to date
Open In Whatsapp
