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 100 was Meta’s first AI chip — a small, low-power part built to help decide which posts, reels and ads show up in a Facebook or Instagram feed. Meta revealed it 18 May 2023, under the name “MTIA v1” 1.
Deep diveThe technical detail
Meta renamed its early chips in its March 2026 roadmap post: MTIA v1 became “MTIA 100,” and the second generation (MTIA v2 / 2i) became “MTIA 200” 2. That same post states Meta has “deployed hundreds of thousands of MTIA chips in production,” referring to MTIA 100 and 200 together 2 — Meta has not published a production figure for MTIA 100 alone.
2.Launch and history
SimpleStart here
Before MTIA, Meta ran its recommendation and ranking systems on general-purpose GPUs. MTIA 100 was Meta’s first purpose-built alternative — a chip designed only for the specific pattern of math those systems use, not general AI work.
Deep diveThe technical detail
Meta’s MTIA programme began in 2020, aimed specifically at inference for recommendation models — the systems that rank content and ads 1. It reached the public as MTIA v1 in May 2023. MTIA 100 is the small, cheap end of Meta’s chip family: later generations (MTIA 200, then MTIA 300, this site’s sibling pages) took on larger models and, eventually, training rather than just inference 2.
3.What’s inside it
SimpleStart here
An MTIA 100 chip is a grid of 64 small processing elements, each with two RISC-V cores and fixed-function math units, connected by an on-chip network 1.
Deep diveThe technical detail
Meta’s own description: an 8-by-8 grid of 64 processing elements, each pairing two RISC-V cores with dedicated matrix-multiply and accumulate hardware, sharing a large on-chip memory pool rather than relying on costly off-chip HBM 1 3. Meta’s ISCA 2025 paper — about the second-generation chip, MTIA 200, not this one — states the same design logic in its own words: “a large SRAM... backed by LPDDR DRAM, avoiding HBM to reduce cost and power consumption” 4. Meta has not published an equivalent sentence specifically about MTIA 100, but the two generations share the same on-chip-memory, no-HBM approach by design — a deliberate trade-off against the HBM-heavy training chips (see the MTIA 300 page on this site).
4.Full spec table
SimpleStart here
102.4 TOPS at INT8, 51.2 TFLOPS at FP16, in a 25-watt power budget — small enough to pack densely into a data-centre rack, not a chip built to chase peak throughput.
Deep diveThe technical detail
| Spec | Meta MTIA 100 |
|---|
| AI compute (INT8) | 102.4 TOPS 1 |
|---|
| AI compute (FP16) | 51.2 TFLOPS 1 |
|---|
| On-chip memory | 128 MB 1 |
|---|
| Main memory | LPDDR5, 176 GB/s 3 |
|---|
| Power | 25 W 1 |
|---|
| Process | TSMC 7-nanometre 1 3 |
|---|
| Processing-element grid | 8×8, 64 elements, two RISC-V cores each 1 |
|---|
All figures are Meta’s own, published at the chip’s 2023 reveal and restated in its 2024 successor announcement 1 3.
5.Where it’s made
SimpleStart here
Meta designed MTIA 100 itself and named TSMC as the chip’s foundry, on a 7-nanometre process 1.
Deep diveThe technical detail
Unlike the later, Broadcom-co-developed generations (MTIA 300 onward, covered on this site’s sibling pages), Meta’s original 2023 reveal names no outside co-design partner for MTIA 100 beyond the foundry itself — TSMC, on a 7nm process 1. Meta has not published a die size, transistor count or yield figure for this chip.
6.Which systems use it
SimpleStart here
Built only for Meta’s own data centres, running the ranking and recommendation systems behind Facebook and Instagram — never sold or rented to any outside customer 2.
Deep diveThe technical detail
Meta has not published which specific data-centre regions ran MTIA 100, or what share of Meta’s recommendation-inference workload it handled versus GPUs. The clearest public statement covering both MTIA 100 and its successor, MTIA 200, is Meta’s 2026 claim of “hundreds of thousands of MTIA chips in production” 2 — a combined figure, not broken out by generation.
7.Official pricing
SimpleStart here
Not for sale. MTIA 100 was built only for Meta’s own use and was never offered for sale or rent to any outside customer 2.
Deep diveThe technical detail
Consistent with every MTIA generation on this site, Meta has never published — or even implied — a per-chip price for MTIA 100. Its cost is internal to Meta’s own infrastructure spending, not a line item any outside buyer could see.
9.What came before, what came next
SimpleStart here
Came before: nothing — MTIA 100 was Meta’s first AI chip.
Came after: MTIA 200, in production since 2024 and covered on its own page on this site.
Deep diveThe technical detail
Meta’s 2026 roadmap post names four further generations after MTIA 200: MTIA 300 (in production, covered on its own page on this site), and MTIA 400, 450 and 500 — all still pre-production as of this page’s September 2026 check date and not yet given their own pages on this site (see the Meta MTIA brand page for the full roadmap) 2 5.
10.Hidden in plain sight
SimpleStart here
Meta’s own annual report has never named MTIA once, in any year — including the year this chip launched.
Deep diveThe technical detail
Meta’s 10-K describes its infrastructure spending only in aggregate terms — it says Meta designs and builds “key portions of our technical infrastructure” — without naming MTIA specifically 6, a pattern that has continued through every filing since MTIA 100’s 2023 reveal. The chip’s existence is public only because Meta chose to publish a blog post about it, not because any financial disclosure required it.