Meta Platforms will be rolling out its next-generation Astrid and Arke AI inference chips in the first half of 2027, as reported by Bloomberg. The company has now committed to more than a gigawatt of the chips over a twelve-month period, with acceleration expected after that initial deployment. The Arke (MTIA 450) is the third generation of the line and goes into data centers in the first half of the year, while Astrid (MTIA 500) finishes design work in about a month and arrives at the end of 2027, with wider use expected. The goal is to deliver better performance per watt and per dollar than whatever Nvidia is currently shipping, according to Yee Jiun Song, Meta's vice president of engineering. Song described the deployment as "much broader" than the third-generation rollout, with the fourth-generation chip serving as "our main chip for general-purpose inference."
Meta's move aims to bring down data centre costs associated with using the AI chips from Nvidia, as well as consolidate its self reliance. This development comes after the company announced that it was manufacturing of a specialised AI chip known as Iris in September to double its computing capacity to 14 GW in the coming year. The effort gives Meta greater control over how its chips are designed around the specific requirements of its AI models, as reported by Bloomberg. Song attributed the efficiency gains to Meta's decision to shoulder much of the engineering work internally, allowing the design to be tailored around the specific requirements of the company's AI models rather than the one-size-fits-all approach of merchant silicon. "When you start to build up gigawatts and gigawatts of capacity, you really care about cost," Song explained, highlighting the company's focus on cost-effective scaling.
The tech giant planned to create its own custom-built AI chip to that it will design in-house to build up its AI capabilities and run its Facebook and Instagram social media platforms. Iris is a part of the company's four-generation project for Meta Training and Inference Accelerators (MTIA). The firm took six weeks to test the chip and did not notice any prominent issues, as per the memo cited by Reuters. According to Bloomberg, 12 prototypes of the new chip arrived at Meta from TSMC on September 1, with their performance being within 2% to 3% of the company's simulations. On the first day of testing, engineers used the processors to run Meta's own AI models alongside those from DeepSeek and Alibaba Group Holding. All four generations of the MTIA project lean on high-bandwidth memory and none targets ultrafast inference.
Meta partnered with Broadcom, the US-based semiconductor maker to design the AI chip, along with enlisting Taiwan Semiconductor Manufacturing Co. to manufacture it. This follows the trend of major tech companies developing in-house AI chip capabilities, with Google also announcing an AI chip named Frozen V2 on July 20, which it is developing in-house to boost the efficiency of its indigenous AI models. The arrangement mirrors the partnership model used by other hyperscalers pursuing in-house silicon, including Google's tensor processing unit program. Meta has also cancelled a chip code-named Olympus that would have handled both training and inference, due in 2028 or 2029, because a dual-purpose part would have cost potentially 30% more. Meta's expanded agreement with Broadcom covers chip design, advanced packaging and networking, with Broadcom's April 14 announcement stating the initial commitment exceeds 1 gigawatt, with plans to support MTIA generations through 2029.
Meta shares were unchanged as the company detailed its chip deployment plans, as reported by GuruFocus. Nvidia shares edged up 0.74% in response to the news. The stock reaction reflects investor recognition that Meta's chip program could meaningfully reduce the company's capital expenditure burden over time. Meta's second-quarter results showed capital expenditure nearly doubling year-over-year to $31.1 billion from $17 billion, with the company raising its fiscal 2026 outlook to between $130 billion and $145 billion. Analysts at TipRanks maintain a Strong Buy consensus on Meta stock, based on 44 analyst reviews over the past three months, with 38 Buy and six Hold ratings. The average 12-month price target of $758.03 represents 13.5% upside from current levels.
