When Meta Vice President of Engineering Yee Jiun Song told CNBC that by designing custom chips, which are then manufactured by Taiwan Semiconductor, the social media giant can squeeze more price per performance across its data centre fleet, he was understating the real calculus. This isn't just about efficiency. It's about leverage.
Meta plans to deploy four new generations of its in-house artificial intelligence chips by the end of 2027 as the company turns to custom silicon to help power its rapidly expanding AI workloads, as part of an effort to diversify its hardware sources, reduce reliance on outside chipmakers and bring down costs amid a fast-moving and expensive AI race. The move arrives after years of Meta's own in-house chip ambitions encountered technical setbacks. Now, with Meta planning to spend between $115 billion and $135 billion in 2026, up from $72.2 billion in 2025, the calculus has shifted.
The New Roadmap
MTIA 300 will be used for ranking and recommendations training, and is already in production. The subsequent three chips—the 400, 450, and 500—represent the real ambition. The upcoming chips are intended for more cutting-edge generative AI-related inference tasks like creating images and videos based on people's written prompts. While the industry typically launches a new AI chip every one to two years, Meta has developed the capacity to release its own every six months or less, enabling it to quickly adapt to evolving AI techniques, adopt the latest hardware technologies, and minimize costs associated with developing and deploying new chip generations.
The speed is aggressive, but the economics are compelling. Industry analysts estimate custom chips can reduce inference costs by 30-50% compared to commercial GPUs once deployed at sufficient scale. For Meta, which runs AI inference trillions of times daily, those savings translate to hundreds of millions in annual operational expenses.
Yet here lies the paradox: even as Meta pushes custom silicon, Meta is rolling out the latest generation of its MTIA chips to power its expanding AI data centre infrastructure, just weeks after announcing massive GPU procurement deals with Nvidia and AMD. In recent weeks, Meta inked deals to fill its data centres with millions of Nvidia GPUs and up to 6 gigawatts of AMD GPUs over multiple years.
The Hedged Bet
This dual strategy isn't contradiction; it's rational risk management at hyperscale. Meta's move signals a long-term bet on vertical integration for AI hardware, even as it continues spending billions on third-party accelerators to meet immediate compute demands. Meta is betting it can engineer its way out of GPU dependency while still hedging with massive Nvidia and AMD purchases. If the custom chips deliver on their promise, Meta gains both cost advantages and strategic control over its AI roadmap. If they stumble, the company has billions in commercial accelerators as backup. Either way, the move pressures competitors to accelerate their own chip projects and signals that the era of Nvidia's unchallenged dominance in AI hardware may be entering a new, more complicated phase.
The concern isn't performance but supply chain realities. The upcoming MTIA chips will contain more high-bandwidth memory, or HBM, to help power GenAI-related inference tasks. The tech industry's mega AI push has led to a shortage of memory chips in the broader market, which means that Meta's ambitious silicon roadmap could face future supply chain constraints. Song acknowledged the risk; "We're absolutely worried about HBM supply," he said. "But we think that we have secured our supply for what we're planning to build out."
The Broader Shift
Meta isn't alone. Google pioneered this approach with its TPU (Tensor Processing Unit) chips, now in their fifth generation. Amazon Web Services followed with Graviton processors for general compute and Inferentia chips for AI inference. Apple revolutionised consumer devices by ditching Intel processors for its M-series chips. Now Meta is applying the same vertical integration logic to data centre AI workloads.
The financial stakes are enormous, but so is the opportunity cost of not investing. At Meta's current spending trajectory, even marginal improvements in cost-per-inference translate directly to the bottom line. For the models Meta has launched into production, MTIA 2i reduces the total cost of ownership by an average of 44% compared to GPUs.
What becomes clear is that the era of chipmaking as a specialised, standalone business is fracturing. Hyperscalers with the scale, capital, and engineering talent are building their own silicon not out of ideology but economics. Nvidia still dominates, still profits massively, and will continue to do so. But the terrain is shifting. Meta's move signals that the question is no longer whether custom chips work, but whether any company can afford to ignore them.