Meta is developing a staged family of custom AI accelerators for its own data centers. A new in-house processor is planned for deployment during the first half of 2027, while MTIA 500, code-named Astrid, is expected to enter data centers at the end of 2027. Broadcom is helping with the designs and Taiwan Semiconductor Manufacturing Co. (TSMC) is manufacturing the chips.
The program is aimed at inference: the phase in which trained AI models generate answers, predictions and other outputs. Meta’s roadmap is therefore less about a consumer chip launch than about controlling the hardware behind its expanding AI infrastructure.
Meta’s reported 2027 custom-chip plan
Meta’s latest roadmap includes several milestones, but the first-half and end-of-year dates refer to separate stages of the program. The first-half target covers a new in-house processor; the later milestone concerns MTIA 500/Astrid.
| Time | Reported development | Scope |
| September 1, 2026 | Twelve model chips reportedly arrived at Meta from TSMC | Hardware for testing |
| First half of 2027 | Meta plans to begin deploying a new in-house AI processor | Data-center deployment |
| End of 2027 | MTIA 500, code-named Astrid, is expected to enter data centers | Later MTIA generation |
The current generation under testing is MTIA 450, code-named Arke. Meta is also reported to be developing MTIA 500/Astrid as its successor. A separate roadmap presents MTIA 300, MTIA 400, MTIA 450 and MTIA 500 as generations leading through the end of 2027.
What MTIA 450 and MTIA 500 mean
MTIA is Meta’s family of in-house accelerators. The chips are designed as general-purpose inference “workhorses,” rather than processors focused on ultrafast-response workloads. They also rely heavily on high-bandwidth memory, or HBM, which keeps large amounts of model data close to the processing hardware.
The design is being split between specialized jobs. Broadcom is the reported design partner, while TSMC is the reported manufacturing partner. That arrangement lets Meta define silicon for its own workloads without requiring the company to manufacture the processors itself.
Meta’s strategy has changed since the Olympus project, which was intended to handle both AI training and inference. Yee Jiun Song, Meta’s vice president of engineering, said a dual-purpose design could be about 30% more expensive when deployed at gigawatt scale. The company instead focused this roadmap on the repetitive, high-volume work of running trained models.
Why Meta is focusing on inference
Inference is where an AI service turns a trained model into a response. At Meta’s scale, small differences in hardware cost and energy use can multiply across data centers and thousands of processors.
Meta has committed to more than 1 GW of the custom chips over a 12-month period. That figure describes the planned deployment scale for the custom processors. It is separate from Meta’s broader infrastructure target of 14 GW by 2027, up from a planned 7 GW in 2026.
The distinction matters: 1 GW refers to the custom-chip deployment commitment, while 14 GW covers Meta’s overall computing infrastructure. They are not two measurements of the same hardware pool.
Scale, testing and the limits of Meta’s efficiency claim
Meta says its custom chips can run AI models more efficiently than whatever NVIDIA is currently shipping. The claim is strategically important, but it is not the same as a published performance-per-watt or performance-per-dollar result.
The numerical test result reported so far is that initial performance came within 2%–3% of Meta’s internal simulations. That comparison indicates how closely the early hardware matched Meta’s expectations; it does not measure the chip against a named NVIDIA accelerator under the same workload and power conditions.
Meta also says the first-day testing ran models from Meta, DeepSeek and Alibaba. Those tests show the range of models the company put through the early hardware evaluation, while the efficiency claim remains Meta’s own assessment.
A complement to external accelerators
Meta’s custom silicon is intended to work alongside external accelerators, not to represent an established replacement for NVIDIA or AMD hardware. The company is continuing to buy chips from other suppliers while it develops its own processors.
That makes the roadmap a supply and cost strategy as much as a chip-design project. Meta wants more control over hardware tailored to its inference workloads, while external accelerators remain part of the computing mix as the company expands its AI infrastructure.
