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 200 is Meta’s second AI chip, revealed 10 April 2024 as “MTIA v2” (also called “2i”) 1. Meta says it went “from first silicon to production models running in 16 regions in less than nine months” — the clearest “in production” claim of any MTIA generation 1.
Deep diveThe technical detail
Meta renamed it “MTIA 200” in its March 2026 roadmap post, alongside renaming MTIA v1 to “MTIA 100” 2. Meta’s own words at reveal: “MTIA has been deployed in our data centers and is now serving models in production” 3 — this page treats that as a first-party production claim, not a marketing generality, because it names a specific region count and timeline.
2.Launch and history
SimpleStart here
MTIA 200 picked up where MTIA 100 left off: the same job — ranking and recommending content — but faster, and deployed far more widely, far more quickly.
Deep diveThe technical detail
Meta’s ISCA 2025 paper, describing MTIA 200 under its interim name “MTIA 2i,” frames it as a co-design exercise between chip and model: Meta says it “deployed at scale and serving billions of users” 4. On its 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 — the job MTIA 300 (this site’s sibling page) was built for.
3.What’s inside it
SimpleStart here
Like MTIA 100, an 8-by-8 grid of 64 processing elements — but Meta redesigned the individual elements and doubled the on-chip memory to push more work through the same layout 1.
Deep diveThe technical detail
Meta’s own description keeps the 64-element grid of MTIA 100 but restructures each processing element and raises on-chip memory from 128 MB to 256 MB 1. Software stayed PyTorch-native: MTIA 200 shipped with PyTorch 2.0 and a Triton-based compiler 1. Meta has not published a die size or transistor count for this generation either.
4.Full spec table
SimpleStart here
354 TOPS at INT8 (708 with sparsity), 177 TFLOPS at FP16, in a 90-watt power budget — roughly 3.5× the raw compute of MTIA 100 at 3.6× the power.
Deep diveThe technical detail
| Spec | Meta MTIA 200 |
|---|
| AI compute (INT8, dense) | 354 TOPS 1 |
|---|
| AI compute (INT8, sparse) | 708 TOPS 1 |
|---|
| AI compute (FP16, dense) | 177 TFLOPS 1 |
|---|
| On-chip memory | 256 MB 1 |
|---|
| Main memory | 128 GB LPDDR5, 204.8 GB/s 1 |
|---|
| Power | 90 W 1 |
|---|
| Process | TSMC 5-nanometre 1 |
|---|
All figures are Meta’s own, published at the chip’s April 2024 reveal 1.
5.Where it’s made
SimpleStart here
Meta designed MTIA 200 itself and named TSMC as the foundry, on a more advanced 5-nanometre process than MTIA 100’s 7nm 1.
Deep diveThe technical detail
As with MTIA 100, Meta’s reveal names no outside co-design partner for MTIA 200 beyond the foundry — TSMC, on 5nm 1. The Broadcom co-design partnership Meta later announced (April 2026) applies to MTIA 300 onward, not to this generation 2.
6.Which systems use it
SimpleStart here
Deployed across 16 Meta data-centre regions within nine months of first silicon, serving production ranking and recommendation models 1.
Deep diveThe technical detail
Meta’s 2025 ISCA paper adds that this generation is “deployed at scale and serving billions of users” 4, and racks hold up to 72 chips built to Open Compute Project designs 1. Meta has not published what share of its total inference workload MTIA 200 carries versus GPUs.
7.Official pricing
SimpleStart here
Not for sale. Like every MTIA generation, MTIA 200 was built only for Meta’s own infrastructure and was never offered for sale or rent 2.
Deep diveThe technical detail
The only cost figure Meta has published for this generation is relative, not absolute: its ISCA 2025 paper states MTIA 2i “reduces the TCO by an average of 44% compared to GPUs” on Meta’s own recommendation workloads 4 6 — a total-cost-of-ownership comparison, not a chip price.
9.What came before, what came next
SimpleStart here
Came before: MTIA 100, Meta’s first AI chip, covered on its own page on this site.
Came after: MTIA 300, Meta’s first training-capable MTIA chip, also covered on its own page.
Deep diveThe technical detail
MTIA 200 kept MTIA 100’s inference-only role but scaled it up; MTIA 300 (this site’s sibling page) is the first MTIA generation Meta built for training as well as inference, and the first built with HBM memory and Broadcom co-design rather than Meta’s own earlier LPDDR-based approach 2 4.
10.Hidden in plain sight
SimpleStart here
Meta’s clearest “in production” language for any MTIA chip is for this one — “first silicon to production... in less than nine months” — yet Meta’s annual report still never names MTIA once.
Deep diveThe technical detail
Meta’s 10-K describes infrastructure spending only in aggregate — it says Meta designs and builds “key portions of our technical infrastructure” — with no chip named, even in the year MTIA 200 reached 16 production regions 7. The gap between how confidently Meta discusses MTIA 200 on its own engineering blog and how completely it omits the chip from financial disclosure is consistent across every MTIA generation on this site 1 8.