AI Chips · Brand guide

Meta MTIA: The Complete Guide

Meta's own AI chips, built to rank the posts and ads billions of people see, and now moving into training and generative AI. What MTIA is, how it works, who builds it, the Broadcom deal, and the facts buried in Meta's filings. Three reading levels, every number sourced.

Stuck at any point? Ask an AI about this page →
Meta MTIA 16 sections · 3 levels 41 linked sources Checked September 2026
How to read this page. Every brand in our AI Chips section uses the same 16 sections in the same order. Each section starts Simple, then goes Deeper, then Expert: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where a company has not published something, we say so rather than estimate.

1.At a glance

SimpleStart here

MTIA (Meta Training and Inference Accelerator) is Meta's own family of AI chips 1. They began as chips for the recommendation systems that decide which posts, reels and ads each person sees on Facebook and Instagram 1 2. Meta says it has "deployed hundreds of thousands of MTIA chips in production" 3. In March 2026 it announced four new generations, one roughly every six months 3.

What it isAI chips designed for Meta's own ranking, recommendation and, now, generative AI work 4
Latest chipMTIA 300, in production for training; MTIA 400 next 3
Who builds itCo-developed with Broadcom; TSMC made earlier generations 5 6
Can you buy one?No; Meta uses them only in its own data centres 4
DeeperThe detail

Meta runs several kinds of AI chip side by side: its own MTIA plus millions of NVIDIA GPUs and up to 6 gigawatts of AMD GPUs 7 8. On the April 2026 call Mark Zuckerberg said Meta is "rolling out more than 1GW of our own custom silicon that we're developing with Broadcom" 9. Meta's 2025 paper said its second-generation chip cut total cost of ownership by 44% compared with GPUs 2.

ExpertFor specialists

Meta's June 2026 quarter: revenue $60.8 billion, up 28%; capital spending $31.08 billion including finance-lease payments 10. It guides 2026 capital spending of $130 to $145 billion and total expenses of $165 to $169 billion 10. MTIA was not named on the call; CFO Susan Li called the "internal custom silicon effort" a source of "long-term strategic flexibility and supply chain leverage" 11.

2.Company card

SimpleStart here

Meta, formerly Facebook, was incorporated in July 2004 12. Mark Zuckerberg is chief executive 13.

Legal nameMeta Platforms, Inc. 12
Headquarters1 Meta Way, Menlo Park, California 94025, USA 12 (map)
Chief ExecutiveMark Zuckerberg 13
Chief Financial OfficerSusan Li 11
Employees75,472 (30 June 2026) 10
StockMETA, Nasdaq 12
Investor siteinvestor.atmeta.com
DeeperThe detail

In January 2026 Meta set up "Meta Compute" to run its AI infrastructure, led by Santosh Janardhan and Daniel Gross; Janardhan's remit includes the in-house chip effort 14. Meta's annual report does not name MTIA; it says Meta designs and builds "key portions of our technical infrastructure" 12.

ExpertFor specialists

Meta had 78,865 employees at the end of 2025 12. Its 2025 revenue was $200.97 billion and its capital spending $72.22 billion 15. Meta is incorporated in Delaware 12.

3.The story

SimpleStart here
  • 2020 Meta starts the MTIA programme 1.
  • 2023 Reveals MTIA v1, now called MTIA 100, and its video chip, MSVP 1 16.
  • 2024 Next-generation MTIA, now called MTIA 200, in production in 16 regions 6.
  • 2025 Publishes a detailed paper on the chip 17.
  • 2026 Announces MTIA 300, 400, 450 and 500 in March, and a multi-gigawatt Broadcom deal in April 3 5.
DeeperThe detail

MTIA began as an inference chip for recommendation models, the systems that rank content and ads 1. Meta renamed the first two chips in 2026: v1 is now "MTIA 100" and the second generation (also called MTIA 2i) is "MTIA 200" 3. In January 2026 Susan Li said MTIA would add ranking and recommendation training from the first quarter of 2026 13, and MTIA 300 is now in production for that 3.

ExpertFor specialists

The road was not smooth. Reuters reported that Meta scrapped an in-house inference chip after a failed test in 2022 and then ordered billions of dollars of NVIDIA GPUs 18; in February 2026 The Information reported Meta had scrapped its most advanced training chip after design problems 19. Both are press reports. In October 2025 Meta bought Rivos, a RISC-V chip start-up; a Meta spokesperson said "this will further accelerate our efforts" 20.

4.The lineup

SimpleStart here
NameWhat it doesStatus
MTIA 100 (v1)Inference for recommendationsRevealed 2023 1
MTIA 200 (v2 / 2i)Inference for ranking and recommendationsIn production since 2024 6
MTIA 300Training for ranking and recommendationsIn production 3
MTIA 400Generative AI and ranking"On the path" to deployment (Mar 2026) 3
MTIA 450Mainly generative AI inferenceMass deployment early 2027 3
MTIA 500Generative AI inference2027 3
DeeperThe detail

Meta says the 400, 450 and 500 share the "same chassis, rack, and network infrastructure", so each new chip drops into existing racks 3. From MTIA 300 to MTIA 500, Meta says memory bandwidth rises 4.5 times and compute 25 times, partly by moving to smaller number formats 3. It plans new chips "every six months or less" 4.

ExpertFor specialists

Meta's official step-ups: MTIA 400 has four times more FP8 compute than MTIA 300 (Meta says "400% higher") and 51% more memory bandwidth; MTIA 450 doubles MTIA 400's bandwidth and has 75% more 4-bit compute; MTIA 500 adds 50% bandwidth, up to 80% more memory and 43% more 4-bit compute 3. Reuters reported from an internal memo that a chip code-named "Iris" enters production in September 2026 21.

5.How it’s made

SimpleStart here

Meta designs MTIA with Broadcom's help. For the first two generations Meta named TSMC as the factory 1 6.

  1. Design: Meta, co-developed with Broadcom "across chip design, packaging, and networking" 5.
  2. Chip factory: TSMC 7 nm for MTIA 100 and 5 nm for MTIA 200 1 6. Meta has not named the factory for the 2026 chips; a Reuters report of an internal memo says TSMC 21.
  3. Memory: the first two generations used cheap LPDDR5 memory; from MTIA 300, stacked HBM3E (216 GB on MTIA 300) 2 22.
  4. Racks: up to 72 chips per rack, built to Open Compute Project designs 6 4.
DeeperThe detail

Broadcom says the partnership runs "through 2029", starts with more than 1 gigawatt, and includes "the industry's first 2nm AI compute accelerator" 23. Broadcom's CEO Hock Tan left Meta's board to become an adviser on its chip roadmap 5. Neither company disclosed a dollar value 5.

ExpertFor specialists

The newer chips are built from chiplets: MTIA 300 has a 3 nm compute die and a 5 nm input/output die, as presented at Hot Chips 2026 24, plus two network chiplets, each with six 800 Gbps network interfaces 22. The same Reuters memo names Samsung for memory, SanDisk for flash and Sumitomo Electric for fibre optics; Meta has not confirmed these suppliers 21.

6.Where it’s made: factories and addresses

SimpleStart here

Meta has not named the factory for its newest chips and does not say which data centres hold MTIA, so there is no factory address to give 4. What Meta does publish:

SiteLocationWhat we know
MTIA 200 deployment16 Meta data-centre regions (2024)Serving models in production 6
HyperionRichland Parish, Louisiana, USAMeta's largest AI cluster, 5 GW planned; chips not specified 25
PrometheusNot specified in Meta's post1 GW when complete, "tens of thousands of GPUs" 26
JamnagarGujarat, India168 MW, built by Reliance, leased by Meta; chips not disclosed 27

Meta lists its data-centre locations by town, without street addresses 28.

DeeperThe detail

Meta has more than 30 data-centre sites, including Singapore, Odense (Denmark), Clonee (Ireland), Luleå (Sweden) and many US towns 28. Hyperion is financed through a joint venture in which Blue Owl funds 80% and Meta 20%, about $27 billion of buildings and infrastructure 29; a similar venture with BlackRock will build a 1 GW site in El Paso, Texas, due in 2028 30.

ExpertFor specialists

Meta's 10-Q, as summarised by StockTitan, shows $349.31 billion of non-cancellable commitments at 30 June 2026, mostly for cloud capacity, servers, networking and data centres 31. Reuters reported from the internal memo that Meta plans about 7 gigawatts of computing capacity in 2026 and 14 gigawatts in 2027 21.

7.How it works

SimpleStart here

An MTIA chip is a grid of small processing elements, each with two RISC-V cores and fixed-function maths units, joined by an on-chip network to shared memory 1 3. The early chips kept most of their working data in large on-chip memory and used cheaper LPDDR memory instead of costly HBM, to "reduce cost and power consumption" 2.

DeeperThe detail

MTIA 100 and 200 used an 8 by 8 grid of 64 processing elements 6; MTIA 300 uses a 12 by 6 grid of 72, plus 16 "message engines" for moving data 22. MTIA 300 puts the network cards on the chip: Meta says this made communication 3.9 times faster than an equivalent GPU cluster on a 150-billion-parameter recommendation model 22.

ExpertFor specialists

Software is PyTorch-native: MTIA 200 used PyTorch 2.0 with a Triton-based compiler 6, and the 2026 chips support PyTorch, vLLM, Triton, MLIR and LLVM 3. Meta's recommendation engine for ads retrieval, Andromeda, "can now run on Nvidia, AMD, and MTIA" 13.

8.Specs explained

SimpleStart here

Meta's own figures for the first two generations. "Dense" and "sparse" are explained in our How AI chips work guide.

ChipAI computeOn-chip memoryMain memoryPowerProcess
MTIA 100 (v1)102.4 TOPS INT8; 51.2 TFLOPS FP16128 MBLPDDR5, 176 GB/s25 WTSMC 7 nm 1 6
MTIA 200 (v2)354 TOPS INT8 dense (708 sparse); 177 TFLOPS FP16 dense256 MB128 GB LPDDR5, 204.8 GB/s90 WTSMC 5 nm 6
MTIA 300——216 GB HBM3E—3 nm compute die 22 24
DeeperThe detail

Meta published the newer chips' figures only as an image; press reports transcribed them as below. Treat these as reported, not official text 32 33.

ChipFP8 computeHBMBandwidthPower
MTIA 3001.2 PFLOPS216 GB6.1 TB/s800 W 32
MTIA 4006 PFLOPS288 GBAbout 9.2 TB/s1,200 W 32
MTIA 4507 PFLOPS288 GB18.4 TB/s1,400 W 32
MTIA 50010 PFLOPS 32384–512 GB 3227.6 TB/s 331,700 W 32
ExpertFor specialists

At Hot Chips 2026 MTIA 400 was described as a 3 nm chiplet design with eight stacks of HBM3e (288 GB), 12 petaflops of 4-bit (MXFP4) compute at 1.7 GHz, four chips per compute blade and 72 per rack 34; one report gives its bandwidth as 9.4 TB/s rather than 9.2 24. The power jump is striking: from 25 W for MTIA 100 to a reported 1,700 W for MTIA 500 1 32.

9.Official pricing

SimpleStart here

MTIA is not sold or rented. Meta designs the chips for its own data centres only, and publishes no price 4.

ProductPrice
Any MTIA chip or rentalNot published by the company 3
Broadcom partnershipValue not disclosed 5
DeeperThe detail

The only cost figure Meta has published is relative: its ISCA 2025 paper says MTIA 2i "reduces the TCO by an average of 44% compared to GPUs" on Meta's workloads 2. Meta's newsroom calls its custom chips "much more cost efficient" than general-purpose chips for its own work 4.

ExpertFor specialists

Meta's big chip purchases from others carry no public prices either: NVIDIA ("millions" of Blackwell and Rubin GPUs, February 2026) and AMD (up to 6 gigawatts of Instinct GPUs, with a warrant for up to 160 million AMD shares) 7 8. See our NVIDIA and AMD guides for their published cloud prices.

10.The software lock-in

SimpleStart here

Meta's advantage is knowing its own workload. Recommendation models behave differently from chatbots, and Meta designs MTIA around the models that rank content for billions of people 2. It also controls the racks, network and software, and aligns its hardware with Open Compute Project standards 4.

DeeperThe detail

Meta's strategy is flexibility, not replacing NVIDIA. It uses MTIA, NVIDIA, AMD and Arm-based processors side by side 7 8 35, and its ads retrieval engine runs on all three accelerator types 13. Susan Li called custom silicon a source of "long-term strategic flexibility and supply chain leverage" 11.

ExpertFor specialists

Pace is the other lever: four generations in two years, each fitting the same racks 3. Meta has also bought chip talent directly, acquiring RISC-V start-up Rivos in 2025 20.

11.Who buys it and why

SimpleStart here

Meta is MTIA's only user. The chips serve inference for organic content and ads across Meta's apps, and MTIA 300 trains ranking and recommendation models 4 3.

DeeperThe detail

MTIA 2i is "deployed at scale and serving billions of users", according to Meta's 2025 paper 2. Meta has not published what share of its AI work runs on MTIA rather than GPUs 3.

ExpertFor specialists

For training its large language models Meta relies on GPUs: the Hyperion site is described as delivering "over two gigawatts of compute capacity to train future open-source large language models" 25. Press reported in February 2026 that Meta also agreed to rent Google TPUs; neither company confirmed it 36.

12.Hidden in plain sight

SimpleStart here

Public facts buried in Meta's filings and releases. Each links to the exact document.

  1. Meta's annual report never names MTIA 12.
  2. Meta's earnings calls barely mention it either: the July 2026 call referred only to the "internal custom silicon effort" 11.
  3. The chip-design partner's CEO sat on Meta's board: Broadcom's Hock Tan left the board in April 2026 to become an adviser on Meta's chip roadmap 5.
  4. Meta's cheaper chips skipped HBM: the first two generations used laptop-style LPDDR memory to cut cost and power 2.
  5. Capital spending guidance rose during 2026, from $115–135 billion to $125–145 billion, citing "higher component pricing", then to $130–145 billion 15 37 38.
DeeperThe detail
  1. Meta can earn AMD shares for buying AMD chips: a warrant for up to 160 million AMD shares, vesting with GPU shipments 8.
  2. $349 billion of non-cancellable commitments at 30 June 2026 31.
  3. Meta owns only 20% of its biggest AI data centre venture, Hyperion; Blue Owl funds the other 80% 29.
  4. The first MTIA drew just 25 W; press reports put MTIA 500 at 1,700 W 1 32.
  5. Broadcom calls its Meta chip the industry's first 2 nm AI accelerator 23.
ExpertFor specialists
  1. Headcount fell from 78,865 at the end of 2025 to 75,472 in June 2026 while spending soared 12 10.
  2. Depreciation grew 46% year on year in the June 2026 quarter, to $6.356 billion 38.
  3. No 2027 spending guidance yet: "we aren't providing a specific outlook for 2027 CapEx at this time" 11.
  4. Meta's first chip, MSVP, was for video, not AI 16.
  5. Meta's first India data centre is leased from Reliance, not built by Meta 27.

13.Weak spots

SimpleStart here

Limits, from Meta's own statements and reporting:

  • Still dependent on NVIDIA and AMD for training large language models 7 8 25.
  • Setbacks: press reports of a scrapped chip in 2022 and a cancelled training chip in 2026 18 19.
  • Internal only: no outside customers to share the cost 4.
  • Few official numbers for the newest chips; Meta published them only as an image 32.
DeeperThe detail

Meta's newsroom says MTIA 400 to 500 are "capable of handling all workloads" but focused on generative AI inference, with generative AI training secondary 4. Chris Cox described the approach to press in 2025 as "walk, crawl, run" 18.

ExpertFor specialists

The newest chips move from cheaper LPDDR to HBM memory 22 2, giving up one of MTIA's early cost advantages. Meta cited "higher component pricing" when it raised its 2026 spending guidance 37.

14.India angle

SimpleStart here

In June 2026 Meta announced its first data centre in India: a 168 MW site in Jamnagar, Gujarat, built by Reliance and leased by Meta, cooled with desalinated seawater and powered by renewable energy 27. Meta has not said which chips it will use there 27.

DeeperThe detail

Meta has contracted about 1 gigawatt of clean energy in India, from CleanMax (837 MW) and Fourth Partner Energy (88 MW) 39. In October 2025 Reliance and Meta formed an enterprise AI joint venture with an initial ₹855 crore, 70% Reliance and 30% Meta, according to Reliance's exchange filing 40.

ExpertFor specialists

With the IndiaAI Mission, Meta set up the Srijan Centre for Generative AI at IIT Jodhpur and a skilling programme for 1 lakh young people with AICTE 41. No MTIA deployment in India has been disclosed 27.

15.What’s next

SimpleStart here

Only what Meta has announced:

  • MTIA 400: moving toward deployment 3.
  • MTIA 450: mass deployment early 2027 3.
  • MTIA 500: 2027 3.
  • Broadcom: more than 1 gigawatt first, multi-gigawatt later, through 2029 23.
DeeperThe detail

Meta plans a new chip "every six months or less" 4. Reuters reported from an internal memo that production of a chip code-named Iris begins in September 2026 21.

ExpertFor specialists

The test is generative AI. MTIA proved itself on recommendations; the 450 and 500 are aimed at running large generative models 3. Meta's parallel multi-gigawatt GPU orders suggest it is hedging 7 8.

16.Sources

41 sources, all checked September 2026. Official = the company or organisation’s own page; Filing = a document filed with a regulator; Press = news coverage, used only where no official source exists and labelled in the text.

  1. Meta AI Blog: MTIA v1: Meta's first-generation AI inference accelerator (18 May 2023)Official
  2. Meta authors, ISCA 2025: Meta's second-generation AI chip: model-chip co-design and productionization experiencesPaper
  3. Meta AI Blog: Scaling MTIA: AI chips for billions (11 Mar 2026)Official
  4. Meta Newsroom: Expanding Meta's custom silicon to power our AI workloads (Mar 2026)Official
  5. Meta Newsroom: Meta partners with Broadcom to co-develop custom AI silicon (14 Apr 2026)Official
  6. Meta AI Blog: Our next-generation Meta Training and Inference Accelerator (10 Apr 2024)Official
  7. NVIDIA Newsroom: Meta builds AI infrastructure with NVIDIA (17 Feb 2026)Official
  8. AMD (SEC exhibit): AMD and Meta announce expanded strategic partnership (24 Feb 2026)Filing
  9. Meta Investor Relations: Q1 2026 earnings call transcript (29 Apr 2026)Official
  10. Meta Investor Relations: Meta reports second-quarter 2026 results (29 Jul 2026)Official
  11. Meta Investor Relations: Q2 2026 earnings call transcriptOfficial
  12. Meta Platforms: Form 10-K, fiscal year 2025Filing
  13. Meta Investor Relations: Q4 2025 earnings call transcript (28 Jan 2026)Official
  14. SiliconANGLE (press): Meta creates new organization to lead AI infrastructure buildout (12 Jan 2026)Press
  15. Meta (SEC exhibit): Fourth-quarter and full-year 2025 results, Form 8-K exhibit 99.1Filing
  16. Meta AI Blog: MSVP: Meta's first ASIC for video transcoding (18 May 2023)Official
  17. ACM Digital Library: ISCA '25 paper record: Meta's second-generation AI chipPaper
  18. Investing.com / Reuters (press): Meta begins testing its first in-house AI training chip (11 Mar 2025)Press
  19. Dataconomy (press, citing The Information): Meta scraps advanced AI training chip after design roadblocks (Feb 2026)Press
  20. DatacenterDynamics (press): Meta acquires RISC-V chip startup Rivos (Oct 2025)Press
  21. Investing.com / Reuters (press): Meta to put AI chip into production in September, memo showsPress
  22. Engineering at Meta: MTIA 300: Meta's first training chip with built-in NICs (24 Aug 2026)Official
  23. Broadcom (GlobeNewswire): Broadcom announces extended partnership with Meta for multi-gigawatts of MTIA (14 Apr 2026)Official
  24. ServeTheHome (press): Meta's MTIA custom AI silicon at Hot Chips 2026Press
  25. Meta Data Centers: Richland Parish data center (Hyperion)Official
  26. Engineering at Meta: Building Prometheus: gigawatt-scale AI clusters (9 Feb 2026)Official
  27. Meta Newsroom: Meta partners with Reliance on AI-enabled data center in India (10 Jun 2026)Official
  28. Meta Data Centers: All locationsOfficial
  29. Meta Newsroom: Meta and Blue Owl Capital to develop Hyperion data center (21 Oct 2025)Official
  30. Meta Newsroom: Meta announces new venture with BlackRock for El Paso data center (28 Jul 2026)Official
  31. StockTitan (press): Summary of Meta Q2 2026 Form 10-QPress
  32. DatacenterDynamics (press): Meta unveils next four generations of its MTIA chip (Mar 2026)Press
  33. ServeTheHome (press): Meta outlines new MTIA accelerator roadmap (Mar 2026)Press
  34. The Register (press): Meta's new MTIA 400 chip has a split personality (26 Aug 2026)Press
  35. Arm Newsroom: Arm and Meta strategic partnership (15 Oct 2025)Official
  36. Dataconomy (press, citing Reuters): Meta signs multibillion-dollar deal to rent Google TPUs (Feb 2026)Press
  37. Meta (PR Newswire): Meta reports first-quarter 2026 resultsOfficial
  38. Meta (SEC exhibit): Second-quarter 2026 results, Form 8-K exhibit 99.1Filing
  39. Meta Data Centers: Bringing AI-enabled infrastructure to India (Jun 2026)Official
  40. Business Standard (press): Reliance and Meta AI joint venture, ₹855 crore (25 Oct 2025)Press
  41. Press Information Bureau, Government of India: IndiaAI and Meta: Srijan Centre at IIT Jodhpur (25 Oct 2024)Government

On this page

1. At a glance2. Company card3. The story4. The lineup5. How it’s made6. Where it’s made: factories and addresses7. How it works8. Specs explained9. Official pricing10. The software lock-in11. Who buys it and why12. Hidden in plain sight13. Weak spots14. India angle15. What’s next16. Sources

Quick facts

HeadquartersMenlo Park, California
Founded2004
CEOMark Zuckerberg
Makes its own chips?Designs them; partners build them
Main chip partnerBroadcom
Latest quarter revenue$60.8 billion

AI Chips

All chip brandsAmazon TrainiumAMDAppleCerebrasGoogle TPUGroqHuawei AscendIntelMicrosoft MaiaNVIDIAAI memory (HBM) and why it is scarceExport rules and the US–China chip contestHow a chip is made: sand to serverHow AI chips workIndia’s semiconductor pushHow to read a chip spec sheetToday’s AI news

In the news

Newest stories mentioning Meta MTIA, straight from the companies’ own newsrooms and the wider press. Links open the original.

All of today’s AI news →

Ask an AI about this page

Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.

ChatGPTClaudeGeminiPerplexityGrok

Nothing is sent from here. The link carries only this page’s title and address.