Meta Platforms (NASDAQ: META) shares rose as much as 2% on Tuesday after Bloomberg reported that the company plans to deploy custom artificial-intelligence chips in its data centers during the first half of 2027. The stock later pared its gains and was trading approximately 0.75% higher.

The chip at the center of the report is the MTIA 450, also known as Arke. It is the third generation of Meta’s custom-silicon program, which the company first announced in 2023. Meta is currently testing the processor and plans to begin using it in data centers in early 2027.

Meta is also developing a fourth-generation chip, code-named Astrid. Design work is expected to be completed within about a month, with the chip scheduled to enter data centers by the end of 2027.

Yee Jiun Song, Meta’s vice president of engineering and head of the company’s custom-chip program, told Bloomberg that each new generation provides better performance per watt of energy and per dollar spent than off-the-shelf alternatives. Song also said the processors can run AI models more efficiently than current Nvidia chips, particularly for inference workloads, in which AI models generate responses in real time.

Reducing Reliance on Nvidia

Meta developed the custom processors in partnership with Broadcom, which is involved in the design, and Taiwan Semiconductor Manufacturing Co. (TSMC), which is handling production. The initiative is intended to reduce Meta’s dependence on Nvidia hardware for its AI workloads while lowering energy and infrastructure costs.

Meta has committed to more than one gigawatt’s worth of its custom chips over a 12-month period. Song said that, at that scale, accepting a 30% increase in costs is not an option, helping explain the company’s focus on internally developed hardware.

TSMC delivered the first batch of 12 test chips on September 1. Their performance was within 2% to 3% of design simulations. On the first day, Meta engineers ran the company’s own AI models as well as tools from DeepSeek and Alibaba.

Focus on Inference

Meta had previously been developing a chip called Olympus, which was planned for 2028 or 2029 and designed to support both model training and inference. The company canceled Olympus because of cost concerns.

The current strategy focuses entirely on inference. Meta has said this approach keeps costs lower while meeting its requirements for operating AI tools at scale. Meta Superintelligence Labs is also contributing to the effort by sharing upcoming model requirements with the engineering team, allowing the chips to be designed around specific use cases.

The next reported milestones are the completion of Arke testing and its planned early-2027 data-center deployment, followed by Astrid’s scheduled entry into data centers by the end of 2027. Those steps will provide further tests of Meta’s custom-silicon strategy, including its stated focus on inference performance and cost efficiency.

Tuesday’s share-price move followed recent momentum after Meta launched the Muse AI agent.

According to TipRanks, META has a Strong Buy consensus rating based on 44 analyst reviews published during the past three months. The breakdown includes 38 Buy ratings and six Hold ratings. The average 12-month price target is $758.03, representing 13.5% upside from current levels. The ratings and price target are analyst estimates and do not constitute investment advice.

Source: CoinCentral