AI Chips · Product page

Meta MTIA 300: The Complete Guide

Meta’s first MTIA chip built for training, not just inference — and the first with its own networking built in. Simple to expert, every number links to its source.

Stuck at any point? Ask an AI about this page →
Ask an AI about this chip:CAGPX
Meta MTIA 300 10 sections · 2 levels 11 linked sources Checked September 2026
How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.

1.At a glance

SimpleStart here

MTIA 300 is Meta’s first AI chip built for training, not just inference — the earlier MTIA 100 and MTIA 200 (this site’s sibling pages) only ran already-trained models. Meta announced it in March 2026 alongside three future chips, and says it is already in production 1.

Deep diveThe technical detail

Meta’s own words at announcement: MTIA 300 “is in production for R&R [ranking and recommendation] training” 1 — corroborated independently by The Register (“the chip is in production”) and DataCenterDynamics, both reporting the same March 2026 event 2 3. Meta’s Engineering team published a detailed architecture deep-dive on 24 August 2026, discussing MTIA 300 running production recommendation models directly 4.

2.Launch and history

SimpleStart here

Every MTIA chip before this one only ran models Meta had already trained elsewhere. MTIA 300 is the first Meta chip built to do the training itself — and the first built with help from an outside chip-design partner, Broadcom.

Deep diveThe technical detail

On Meta’s Q4 2025 earnings call, CFO Susan Li said: “In Q1, we will extend our MTIA program to support our core ranking and recommendation training workloads, in addition to the inference workloads it currently runs” 5 — MTIA 300 is that chip. Two months later, in April 2026, Meta announced a multi-year co-design partnership with Broadcom “across chip design, packaging, and networking,” running through 2029 and starting with more than a gigawatt of capacity 6. Neither company disclosed a dollar value for the deal 6.

3.What’s inside it

SimpleStart here

A chiplet design — several separate silicon dies in one package — with its own network cards built directly onto the chip, rather than added as a separate part 4.

Deep diveThe technical detail

At Hot Chips 2026, Meta described MTIA 300 as a 3-nanometre compute die paired with a 5-nanometre input/output die, plus two dedicated network chiplets, each carrying six 800 Gbps network interfaces built into the package 7. The processing grid grew from MTIA 200’s 8×8 layout to 12×6 (72 elements), plus 16 additional “message engines” dedicated to moving data between chips 4. Meta’s own claim: putting networking on the chip itself made communication 3.9× faster than an equivalent GPU cluster on a 150-billion-parameter recommendation model 4 — a Meta-published figure, not an independently verified benchmark.

4.Full spec table

SimpleStart here

Meta has not published a full spec sheet as text; the clearest public figures come from press coverage of an image Meta presented at its March 2026 announcement.

Deep diveThe technical detail
SpecMeta MTIA 300
Compute (FP8, press-reported)1.2 PFLOPS 3
HBM memory216 GB HBM3E 4 3
Memory bandwidth (press-reported)6.1 TB/s 3
Power (press-reported)800 W 3
Process3nm compute die + 5nm I/O die (chiplet) 7
Processing-element grid12×6, 72 elements, plus 16 message engines 4
Networking2 chiplets × six 800 Gbps interfaces, on-package 7

The FP8/bandwidth/power figures are press transcriptions of an image Meta showed at its March 2026 event, not published Meta text 3 — this page states that plainly rather than presenting them as an official Meta datasheet. The architecture figures (process, grid, networking) come from Meta’s own Hot Chips 2026 presentation and engineering blog, and are treated as official 7 4.

5.Where it’s made

SimpleStart here

Co-developed with Broadcom “across chip design, packaging, and networking” — Meta’s first MTIA generation built with an outside chip-design partner 6.

Deep diveThe technical detail

Meta has not named the foundry for MTIA 300 in any source this page found — unlike MTIA 100 and 200, where Meta named TSMC directly. Broadcom’s own announcement of the partnership describes the work starting with “the industry’s first 2nm AI compute accelerator” without confirming that description applies to MTIA 300 itself rather than a later generation in the partnership 6. Broadcom CEO Hock Tan left Meta’s board in April 2026 to become an adviser on Meta’s chip roadmap 6.

6.Which systems use it

SimpleStart here

Meta’s own data centres only, training the ranking and recommendation models that power Facebook and Instagram 1.

Deep diveThe technical detail

Meta’s Engineering blog frames MTIA 300’s built-in networking as a rack-scale design choice, not just a chip-level one — the on-package network chiplets are built to interconnect racks of MTIA 300 chips directly for the large, distributed training jobs ranking models require 4. Meta has not disclosed which specific data-centre sites run MTIA 300.

7.Official pricing

SimpleStart here

Not for sale. Like every MTIA generation, MTIA 300 was built only for Meta’s own infrastructure 1.

Deep diveThe technical detail

Neither Meta nor Broadcom disclosed a dollar value for their co-design partnership 6. No published price, lease rate, or cost-per-chip figure for MTIA 300 was found in any source.

8.Real-world performance

SimpleStart here

Meta’s own claim: 3.9× faster inter-chip communication than an equivalent GPU cluster, on a 150-billion-parameter recommendation model 4.

Deep diveThe technical detail

This is a Meta-published figure on Meta’s own workload, not an independently verified benchmark — no third-party test of MTIA 300 against a named GPU, on a public benchmark, was found. It measures one specific thing (inter-chip communication speed on one model), not overall training throughput or cost, and this page states that scope directly rather than generalising it.

9.What came before, what came next

SimpleStart here
Came before: MTIA 200, Meta’s inference-only chip, covered on its own page on this site. Came after: MTIA 400 — announced alongside MTIA 300 but, as of this page’s September 2026 check date, still pre-production by Meta’s own account.
Deep diveThe technical detail

Meta’s own most recent public language on MTIA 400, from its March 2026 roadmap post: “We have finished testing MTIA 400 in our labs and are on the path to deploying it in our data centers” 1 — explicitly pre-production, and no later Meta statement through this page’s September 2026 check date upgrades that language. The Register’s own Hot Chips 2026 coverage (26 August 2026) describes MTIA 400’s architecture in detail but makes no production claim for it — the only chip that piece says is “expected to enter production next year” is MTIA 450, not 400 8. A Reuters report (9 July 2026), sourced to an internal memo, said Meta planned to put an unnamed AI chip — widely referred to in press coverage by the codename “Iris” — into production in September 2026 9 10. Meta declined to comment on that report, and — important for accuracy — the more careful outlets covering it (TechCrunch, DataCenterDynamics) do not confirm which MTIA generation “Iris” refers to 9 10; this page found no reputable, named-source confirmation, before or after the fact, that MTIA 400 specifically entered production in September 2026. Rather than treat an unconfirmed, ungeneration-specific press report as settling the question, this page follows Meta’s own most recent public language and leaves MTIA 400 off this site until its production status is confirmed by Meta directly. MTIA 450 (“mass deployment in early 2027”) and MTIA 500 (“2027”) are unambiguously future and excluded on the same basis 1 11.

10.Hidden in plain sight

SimpleStart here

The chip-design partner’s own chief executive sat on Meta’s board until the month the partnership was announced.

Deep diveThe technical detail

Broadcom CEO Hock Tan left Meta’s board of directors in April 2026, the same month Meta and Broadcom announced their multi-year MTIA co-design partnership — he moved from board member to adviser on Meta’s own chip roadmap 6. Neither company’s announcement discloses whether that timing was a formal conflict-of-interest resignation or a planned transition; both companies describe it only as Tan becoming “a strategic adviser to Meta.”

11.Sources

11 sources, checked September 2026. Where NVIDIA or another maker has not published a figure, this page says so rather than estimate it.

  1. Meta AI Blog: Four MTIA Chips in Two Years: Scaling AI Experiences for BillionsOfficial
  2. The Register (press): Meta unveils next MTIA silicon, still chasing Nvidia (12 Mar 2026)Press
  3. DataCenterDynamics (press): Meta unveils next four generations of its MTIA chipPress
  4. Engineering at Meta: MTIA 300: Meta’s training chip with built-in NICsOfficial
  5. Meta Investor Relations: Q4 2025 earnings call transcript (28 Jan 2026)Official
  6. Meta Newsroom: Meta partners with Broadcom to co-develop custom AI silicon (14 Apr 2026)Official
  7. ServeTheHome (press): Meta’s MTIA custom AI silicon at Hot Chips 2026Press
  8. The Register (press): Meta’s new MTIA 400 chip has a split personality (26 Aug 2026)Press
  9. TechCrunch (press): Meta’s new AI chips will begin production in September (9 Jul 2026)Press
  10. DataCenterDynamics (press): Meta could start production of "Iris" AI chip in September (10 Jul 2026)Press
  11. Clarigital: Meta MTIA — brand guide (this site)Official

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.