Meta expands in-house AI chips

Meta unveiled four in-house AI chips in its MTIA family and said the MTIA 300 has been deployed while MTIA 400, 450, and 500 will follow about every six months.

Yee Jiun Song, vice president of engineering, told CNBC the chips give Meta more diversity in silicon supply and help insulate the company from price changes.

Meta said MTIA 300 is being used to help train smaller models that support ranking and recommendation tasks across Facebook and Instagram, and Song said MTIA 400 through 500 target generative AI inference rather than training giant LLMs.

Meta said MTIA 400 has completed testing, and that one rack will include 72 MTIA 400 chips, while MTIA 450 and 500 are slated to enter service in 2027.

The chips are manufactured by Taiwan Semiconductor Manufacturing Co. (TSMC) and will include more high-bandwidth memory, which Song said Meta believes it has secured for its planned buildout despite industry shortages.

Meta has also signed large deals to buy millions of Nvidia GPUs and up to 6 gigawatts of AMD GPUs over multiple years as it expands data center capacity in the US.

Meta said most of the engineers who worked on the silicon are based in the US and that 26 of its 30 operational or planned data centers are in the country.

🔗 Source: CNBC

🧠 Food for thought

Implications, context, and why it matters.

Meta’s chip strategy carries high costs, shaped by setbacks and spending

Meta expects up to $119 billion in total expenses for 2025, with a large share going to AI infrastructure 1.
Meta wants to cut the billions it pays each year for third-party hardware. One report compared this to Meta “renting the keys to its AI future” from Nvidia 2.
Meta has pursued custom silicon (custom-designed chips) before, then scrapped an earlier chip. Executives now describe the plan as “walk, crawl, run” to keep the rollout cautious 2.
Meta already runs an inference chip at scale through MTIA (Meta Training and Inference Accelerator). It is tuned for ranking and recommendation inference workloads, and it powers recommendations across Facebook and Instagram 2.

A split is emerging in AI hardware

Meta fits a broader shift where hyperscalers (the biggest cloud and internet platform companies that run massive data centers) build specialized chips, rather than trying to replace Nvidia across every workload 3.
These custom parts trade some general GPU (graphics processing unit) performance for efficiency on a narrow set of jobs, such as serving recommendation algorithms to billions of users 3.
Meta runs a two-track approach with large, multi-year AI infrastructure and chip procurement deals involving Nvidia and AMD 4.
Cutting-edge commercial chips train the largest models, including work on 24,000-GPU clusters. Meta-designed silicon aims to reduce the cost of inference at scale 5.