Meta to Deploy Third Generation MTIA 450 Arke AI Chip in Data Centers by 2027

Meta Platforms first announced plans to develop homegrown artificial intelligence chips in 2023. The company is now testing the third generation of that line, called MTIA 450, or Arke. Meta plans to begin deploying the chip in data centers during the first half of 2027, a move it says will save money and energy when running AI models. The effort is part of a broader push to reduce reliance on Nvidia Corp.'s industry-leading processors.

The chip is designed with Broadcom Inc. and manufactured by Taiwan Semiconductor Manufacturing Co. (TSMC). Twelve of the new chips arrived at Meta from TSMC on Sept. 1, 2026, and performance is within 2% or 3% of the company's simulations. On the first day, the team used the processors to run Meta models as well as ones from DeepSeek and Alibaba Group Holding Ltd. Yee Jiun Song, Meta's vice president of engineering, said each generation 'takes on a little bit more risk technologically, and gets us better performance,' including better performance per watt of energy and per dollar spent.

Meta has committed to more than a gigawatt's worth of the chips over a 12-month period, measured by energy use, after which 'we expect it to accelerate,' Song said. The next chip, MTIA 500 or Astrid, will complete design work in about a month and go into data centers at the end of 2027, and Meta expects to use that one even more widely. The company previously planned a chip code-named Olympus for both training and inference that would have arrived in 2028 or 2029, but canceled it to focus on inference, partly for cost reasons.

The custom silicon push comes as Meta rapidly expands its broader AI infrastructure. According to Reuters, the company aims to double its overall computing capacity to 14 GW by 2027. All four generations of the MTIA line rely heavily on high-bandwidth memory (HBM) and are designed for general-purpose inference rather than workloads requiring ultrafast response times. Meta says the chips can run AI models more efficiently than 'whatever Nvidia is currently shipping,' Song said, 'just because we're doing a lot of the engineering ourselves.'

Key Facts

Meta Platforms plans to begin deploying its latest in-house AI accelerator, the third-generation MTIA 450 (code-named Arke), in data centers during the first half of 2027. Bloomberg reported on September 15, 2026 that the company is testing the third generation of the line. The chip is designed with Broadcom and manufactured by TSMC on leading-edge process nodes.

Twelve prototype Arke chips arrived from TSMC on Sept. 1, 2026. Their initial performance came within 2% to 3% of Meta's pre-silicon simulations. PCMasterInsider reported on September 16, 2026 that early physical samples delivered from fabrication performed within two to three percent of pre-silicon digital simulations, running internal workloads on day one.

Meta has committed to more than one gigawatt of these chips over a 12-month period, measured by energy use, and expects the rollout to accelerate afterward. The successor, MTIA 500 or Astrid, should complete design work in about a month and enter data centers by the end of 2027, with significantly wider deployment expected. Meta also aims to double its overall computing capacity to 14 GW by 2027, as Reuters reported in July.

The company canceled a planned dual-purpose processor code-named Olympus, which would have handled both training and inference and would have arrived in 2028 or 2029. Executives noted that combining training and inference into a single package would have increased unit costs by roughly 30 percent, which proved impractical at gigawatt scale. Meta Vice President of Engineering Yee Jiun Song said that if a combined training and inference chip is potentially 30% more expensive, at gigawatt scale that seems 'completely unacceptable.'

Both MTIA 450 and MTIA 500 incorporate HBM and focus on general-purpose AI inference, handling tasks such as content recommendation and running generative AI models. Earlier MTIA iterations focused primarily on ranking and recommendation workloads. TrendForce reported on September 16, 2026 that Meta claims an efficiency edge over Nvidia's current shipping products. Benzinga reported on September 15, 2026 that META shares were down 0.05% at $665.23 at the time of publication, with a relative strength index of 71.02.

Analysis

What this really means is that Meta is trying to vertically integrate its AI stack, moving from a buyer of Nvidia GPUs to a designer of custom inference silicon. The strategic logic is cost and energy. At gigawatt scale, even modest efficiency gains compound. Meta's commitment to more than 1 GW of these chips over a 12-month period shows the scale it envisions. The company says the chips deliver better performance per watt and per dollar spent, which matters as AI infrastructure costs continue to climb.

The decision to cancel Olympus and focus on inference is telling. Inference is the workhorse of deployed AI models, from content recommendation to generative AI. By narrowing scope, Meta can optimize for the metrics that matter most in its data centers. The roughly 30% cost penalty for a combined training and inference chip was 'completely unacceptable' at gigawatt scale, according to Song. This suggests that Meta sees inference as the larger and more predictable workload for its custom silicon.

Still, the effort is not without risk. Each generation 'takes on a little bit more risk technologically,' Song said. The 2% to 3% deviation from simulations is small but real. Volume manufacturing, software maturity, and ecosystem support remain challenges. Meta's reliance on Broadcom for design and TSMC for manufacturing means it is not fully independent, but it does reduce reliance on Nvidia. The company is working with Broadcom on physical implementation, IP integration, and high-speed networking co-design, with production handled by TSMC.

The competitive framing is direct: Meta claims its chips run AI models more efficiently than 'whatever Nvidia is currently shipping.' That is a bold claim, but Nvidia's advantage is its ecosystem and continued innovation. The bigger picture here is that Meta is betting on custom silicon to control costs, and its progress will be watched closely by the industry. If Meta can deploy more than a gigawatt of custom inference chips, it could pressure Nvidia's pricing and accelerate the shift toward specialized AI accelerators.

Why It Matters

For Meta, custom silicon could lower the cost of running AI models on Instagram, Facebook, WhatsApp, and Meta AI. The company's AI infrastructure is expanding rapidly, and energy use is a constraint. More than 1 GW of custom compute over 12 months would represent a significant share of its fleet. The chips are designed to handle general-purpose inference, which is the bulk of Meta's AI workload.

For the supply chain, Broadcom and TSMC gain a major customer for leading-edge nodes and HBM. For Nvidia, it signals that its largest customers are developing alternatives, though Nvidia remains the leader in training and high-performance AI. The market for inference chips is becoming more competitive. Meta's move also highlights the growing importance of HBM in AI accelerators, as all four generations of MTIA rely heavily on it.

For the broader industry, Meta's approach of focusing on inference rather than training is a bet that most AI compute demand will be for running models, not training them. If successful, it could validate a new playbook for companies building large AI infrastructure. It also raises questions about the balance between custom silicon and merchant chips, and how quickly custom designs can iterate. Meta's first day of testing ran models from Meta, DeepSeek, and Alibaba, showing the chip is not limited to Meta's own software.

Next Up

Meta expects MTIA 500 (Astrid) to complete design work in about a month and enter data centers by the end of 2027. Meta expects to use that chip even more widely. The company aims to double its overall computing capacity to 14 GW by 2027, according to Reuters. That target underscores how aggressively Meta is scaling its AI infrastructure.

Meta has committed to more than a gigawatt of the MTIA 450 chips over a 12-month period, and Song said 'we expect it to accelerate.' The first half of 2027 will be the critical test for the third-generation chip. If deployment goes smoothly, Meta could expand its custom silicon program further, potentially reducing its reliance on Nvidia and changing the economics of AI inference at scale.