Meta reveals four Broadcom-built ASICs for AI inference

Meta reveals custom AI chips it says beat Nvidia • The Register Sign in / up The Register Topics Special Features Special Features Vendor Voice Resources Resources Systems Meta reveals four Broadcom-built custom AI chips, claims some outperform commercial silicon Deploying them by the gigawatt but still can’t flag obvious AI slop Simon Sharwood Thu 12 Mar 2026 // 04:49 UTC Social networking giant Meta has revealed details of four previously unknown custom chips powering its AI services. The company signalled its intention to roll out more custom inferencing chips in 2024, and in a post about its custom silicon, reveals it built its latest chips “in close partnership with Broadcom.” The post names the four previously unknown chips as models 300, 400, 450, and 500 in the Meta Training Inference Accelerator (MTIA) series, and says some are already running in production and others will appear this year or next. Broadcom has said Meta will install “multiple gigawatts” of its chips in 2027 and beyond. Meta described the four chips as follows: MTIA 300: A communications chip optimized for ranking and recommendation (R&R) workloads, comprised of one compute chiplet, two network chiplets, and several HBM stacks. Each compute chiplet is composed of a grid of processing elements (PEs), with some redundant PEs to improve yield. Each PE includes a pair of RISC-V vector cores. The chip is in production. MTIA 400: An evolution of the MTIA 300 that can support generative AI models and R&R workloads. Meta says it is the first of its chips with “raw performance competitive with leading commercial products,” that employs two compute chiplets. The company says a rack with 72 MTIA 400 devices, connected via a switched backplane, forms a single scale-up domain, and that now that testing has concluded it is “on the path to deploying it in our data centers.” MTIA 450: Built with specific optimizations for GenAI inference, by doubling the HBM bandwidth of the MTIA 400 to levels that Meta says means its performance is “much higher than that of existing leading commercial products.” Scheduled for mass deployment in early 2027. MTIA 500 : An efficient GenAI inference chip that increases HBM bandwidth by an additional 50 percent over the MTIA 450. Uses a 2x2 configuration of smaller compute chiplets surrounded by several HBM stacks and two network chiplets, along with an SoC chiplet that provides PCIe connectivity to the host CPU and scale-out NICs. Meta plans mass deployment in 2027. Meta also posted specs for the chips. Meta's specs for its custom silicon - Click to enlarge Expect more of this sort of thing from Meta, because its post says the company now has the capacity to ship a new chip “roughly every six months.” “We achieve high velocity through a reusable and modular design across all levels: chiplets, chassis, racks, and network infrastructure,” its post states, before revealing the MTIA 400, 450, and 500 “all utilize the same chassis, rack, and network infrastructure.” Nice chips, shame about the slop While Meta has refined the art of creating and deploying chips, it’s doing a lousy job of flagging misleading AI-generated content according to the company’s own Oversight Board which on Tuesday called for Zuck to implement new rules covering content posted during conflicts. The Board considered a faked video purportedly depicting scenes in Israel during the 2025 12-day war with Iran. Despite independent fact-checkers labelling the content fake, and six users reporting the content to Meta, the company did not label it as “a ‘High Risk’ AI”. The Board said it is “concerning” that “Meta did not identify on its own initiative the clear engagement abuse signals from the page” and said the company’s mechanisms for flagging fake videos “are neither robust nor comprehensive enough to contend with the scale and velocity of AI-generated content, particularly during a crisis or conflict where there is heightened engagement on the platform.” In its report, the Board said Meta “should ensure that fact-checkers are adequately resourced and have guidance on prioritizing content from conflicts” and that its content-checking policies “should have allowed Meta to ensure more effective support for third-party fact-checkers during the crisis.” A reminder: Meta last year announced the end of its third-party fact-checking program and adopted user-led reporting instead. That scheme mirrored the way X moderates content under Elon Musk, leading some to suggest Mark Zuckerberg hoped to win approval from the Trump administration, which has supported its Big Tech friends with threats to act against nations that impose Digital Services Taxes on American companies. Meta just found a way to nullify those taxes, at least in Europe, by announcing it will charge advertisers a fee equivalent to the rate of local digital taxes. The company calls its extra charge a “location fee” and explained them with the example of an advertiser who delivers $100 worth of ads to Italy, where a three percent tax is in force, and will therefore receive a bill for $103 – $100 for ads and a $3 location fee. Meta will apply the fees in six nations: Austria, France, Italy, Spain, Türkiye, and the UK. ® Share More about AI Broadcom Meta More like these × More about AI Broadcom Meta Semiconductor Narrower topics AIOps ASML Cadence Design Systems CHERI DeepSeek Facebook Gemini Google AI GPT-3 GPT-4 Hot Chips Large Language Model Loongson Machine Learning MCubed MediaTek Neural Networks NLP NXP Semiconductors Open Compute Project Retrieval Augmented Generation Semiconductor Manufacturing International Corporation Semiconductor Memory Star Wars STMicroelectronics Tachyum Tensor Processing Unit TOPS United Microelectronics Corporation WhatsApp Broader topics Andrew McCollum Chris Hughes Dustin Moskovitz Eduardo Saverin Mark Zuckerberg Network Self-driving Car More about Share POST A COMMENT More about AI Broadcom Meta More like these × More about AI Broadcom Meta Semiconductor Narrower topics AIOps ASML Cadence Design Systems CHERI DeepSeek Facebook Gemini Google AI GPT-3 GPT-4 Hot Chips Large Language Model Loongson Machine Learning MCubed MediaTek Neural Networks NLP NXP Semiconductors Open Compute Project Retrieval Augmented Generation Semiconductor Manufacturing International Corporation Semiconductor Memory Star Wars STMicroelectronics Tachyum Tensor Processing Unit TOPS United Microelectronics Corporation WhatsApp Broader topics Andrew McCollum Chris Hughes Dustin Moskovitz Eduardo Saverin Mark Zuckerberg Network Self-driving Car TIP US OFF Send us news Other stories you might like Blustering Blackbeard's PC was all at sea, sysadmin got him shipshape in seconds On Call Have you tried turning it on, never mind off and on again? On-Prem 13 Mar 2026 | 1 AI Burning Man happens next week – here's what The Register expects at GTC 2026 From Groq-ing about tokenomics to OpenClaw and the silicon that powers it, our predictions for the hottest ticket in town Systems 13 Mar 2026 | 3 Prince of PDFs, Adobe CEO Shantanu Narayen, to step down after 18 years Didn’t say why, but for once AI may not be the reason for a lost job SaaS 13 Mar 2026 | 3 Resilient, continuously active data – with no compromise When the gap between data generation and action is a strategic liability, it's time for a fix Sponsored Feature Apple takes a bite out of app store fees in China Beijing hinted it wasn’t happy with Cupertino, which weeks later made a change Personal Tech 13 Mar 2026 | Pentagon AI chief praises Palantir tech for speeding battlefield strikes Going from eight systems to one means fewer people make decisions to unleash Epic Fury Public Sector 13 Mar 2026 | 13 Rogue AI agents can work together to hack systems and steal secrets Prompt like a hard-ass boss who won't tolerate failure and bots will find ways to breach policy Research 12 Mar 2026 | 4 Perplexity: Everything is Computer, everything is AI, Computer is everything, AI is us Everything extends its cloud Computer to enterprises, your computer AI + ML 12 Mar 2026 | 4 District denies enrollment to child based on license plate reader data Automated checks raised doubts, though key questions remain unanswered Public Sector 12 Mar 2026 | 20 Microsoft Copilot now boarding your health information It's safe and secure, Redmond insists, but don't expect medical advice AI + ML 12 Mar 2026 | 7 White House activates Yu-Gi-Oh's trap card by using anime clip for war comms Franchise isn't the only one unhappy about its IP appearing in propaganda Offbeat 12 Mar 2026 | 29 Medical equipment techs beg for right-to-repair lifeline OEM delays have become patient care delays On-Prem 12 Mar 2026 | 15 The Register Biting the hand that feeds IT About Us Contact us Advertise with us Who we are Our Websites The Next Platform DevClass Blocks and Files Your Privacy Cookies Policy Privacy Policy Ts & Cs Copyright. All rights reserved © 1998–2026