Meta Showcases 4 Generations of Inference Chip for Deployment Over the Next 2 Years

There’s probably no catching Nvidia on AI training hardware, at least in the short term, but there’s plenty of scope for more advanced and (crucially) more efficient AI inferencing hardware. That’s why Meta is just one of the companies with its own inferencing chips, and on Thursday, it showed off multiple generations of them set for release over the next couple of years.
The four new chips are MTIA 300, 400, 450, and 500, and were co-developed with Broadcom. The first one is already in production and will go on sale soon, while MTIA 400 is currently undergoing lab testing. MTIA 450 and 500 aren’t scheduled for any real-world deployment until well into 2027, so they are much further from mass production.
They are chips AI data center runners may be interested in, though. META claims that over successive generations of these chips, it’s managed to increase HBM bandwidth by 4.5 times and general compute power by 25 times, showing massive progression in capability from early MTIA 300 through to the latest MTIA 500 designs.
Meta claims that the 450 and 500 chips offer much more bandwidth than the current class-leading GPUs like Nvidia’s H100 and H200, though they will realistically need to compete with Nvidia’s Blackwell chips to really steal the show.
Announcing multiple generations of chips at once isn’t a normal strategy. Companies like Nvidia tend to conduct annual releases, drip-feeding news about the next generation throughout the preceding months. However, by announcing MTIA 300 through 500 chips at once, Meta is showcasing the potential for progression. Since all the chips will use the same platform, they should be a relatively simple drop-in upgrade as they are released, making them an attractive option for data center developers that want to scale their operations over the next few years.
Meta promises a six-month cadence for chip improvements, offering the potential for rapid development compared to alternatives.
While AMD, Nvidia, and others may have less-defined release timelines, Meta is trying to lay out a multi-year strategy in one go, which, for some, may make its inferencing hardware that bit more attractive. The competition is certainly heating up.
© 2001-2026 Ziff Davis, LLC., a Ziff Davis company. All Rights Reserved.
ExtremeTech is a federally registered trademark of Ziff Davis, LLC and may not be used by third parties without explicit permission. The display of third-party trademarks and trade names on this site does not necessarily indicate any affiliation or the endorsement of ExtremeTech. If you click an affiliate link and buy a product or service, we may be paid a fee by that merchant.

source