Meta announced four successive generations of its custom Meta Training and Inference Accelerator (MTIA) chips on March 11: The MTIA 300, 400, 450, and 500, all scheduled for deployment over the next two years. The announcement arrives as major tech firms increasingly question the economics of relying on a single chip supplier for their massive AI infrastructure investments.
The MTIA 300 is already in production for ranking and recommendations training, while the 400 is currently in lab testing ahead of data centre deployment, with MTIA 450 and 500 targeted at AI inference and scheduled for mass deployment in early 2027 and later in 2027, respectively. This represents a fundamental shift in how Meta operates its data centres. Rather than relying entirely on commercial GPUs designed for training large language models, Meta described mainstream chips, built for large-scale pre-training, as then applied less cost-effectively to inference workloads.
The technical specifications reveal substantial performance gains with each generation. According to Meta's technical blog, from MTIA 300 through to MTIA 500, HBM bandwidth increases 4.5 times, and compute FLOPs increase 25 times. Meta says MTIA 450 doubles the HBM bandwidth of MTIA 400, describing it as "much higher than that of existing leading commercial products," or, in other words, Nvidia's H100 and H200.
Meta's aggressive timeline distinguishes its approach from rivals. While the industry often launches a new AI chip every one to two years, Meta said it now has the capacity to release new MTIA generations every six months or less by reusing modular designs. This rapid iteration depends partly on chip architecture choices; all MTIA chips are developed in close partnership with Broadcom and built on the open-source RISC-V architecture, which allows chip designers to implement the instruction set without paying royalties.
The business case for custom silicon has become compelling as inference demand accelerates. Meta said it already deploys hundreds of thousands of MTIA chips for inference workloads across organic content and ads in its apps, and it argued that these chips are more compute-efficient and more cost-efficient than general-purpose silicon for the company's intended uses. Meta Vice President of Engineering Yee Jiun Song said the chips give Meta more diversity in its silicon supply and help insulate for price changes.
Yet Meta's strategy avoids the false choice between internal silicon and external suppliers. Despite unveiling new chips last week, Meta inked a massive, multiyear deal with Nvidia last month to deploy literally millions of Nvidia Blackwell and Rubin chips in its data centres, along with Nvidia central processing units (CPUs), all connected via Nvidia's SpectrumX ethernet switches. A separate deal, valued at roughly $60 billion over five years according to Reuters, will deploy custom AMD Instinct MI450 GPUs and 6th Gen EPYC "Venice" CPUs starting in the second half of 2026. This three-vendor approach reflects a pragmatic assessment: custom chips excel at specific workloads like serving recommendations and generating images, whilst frontier model training still demands Nvidia's leading performance.
Meta's move follows Google, Amazon, and Microsoft along a path that threatens Nvidia's position despite the company's continued market dominance. While the company still maintains a commanding lead in the highest-end frontier model training, its market share in the broader AI accelerator space is expected to slip from its peak of 95% down toward 75-80% by the end of 2026. The share of GPUs in AI servers is projected to shrink from 75.9% last year to 69.7% this year, while the share of custom processors like ASICs is expected to jump.
These developments reflect deeper industry economics. By 2026, Deloitte projects inference will account for two-thirds of all AI compute, with most inference still running on costly data centre chips rather than edge devices. Because inference is more cost-sensitive than training, custom chips optimised for specific inference tasks can compete effectively against general-purpose GPUs. For a company operating billions of users across Facebook and Instagram, even fractional cost savings per inference multiply into hundreds of millions of dollars annually.
The broader question remains whether Meta and its peers can execute this ambition. The rapid six-month chip release cycles carry execution risk, including integration mistakes and higher costs. Analysts also question whether the software ecosystem supporting these custom chips will mature quickly enough to make them truly portable across different hyperscaler platforms. Yet the financial incentive is undeniable. According to a recent report published by Goldman Sachs, artificial intelligence hyperscalers are forecast to spend more than $500 billion on infrastructure this year. When spending approaches that scale, even small percentage improvements in efficiency justify billion-dollar engineering efforts.
For Australian exporters and service providers selling to large tech platforms, the shift carries indirect implications. Reduced hardware costs and improved efficiency could accelerate AI service deployment, potentially expanding the market for services reliant on these platforms. Yet it also underscores how concentrated computational power remains among a handful of wealthy American firms. The infrastructure wars are reshaping who controls the foundation of AI deployment, and control over silicon is increasingly inseparable from control over the models and services it powers.