Key Takeaways
- Meta plans to deploy its next-generation AI chips in 2027.
- The company expects to deploy more than 1 gigawatt.
- In-house chips could reduce computing costs and energy use.
Meta Platforms plans to begin deploying its next-generation Meta AI Chips in data centers during the first half of 2027. The company expects the chips to reduce the cost and energy needed to run AI models while lowering its dependence on processors supplied by Nvidia.
Meta Expands Its Custom Chip Program
Meta began developing its own Meta AI Chips in 2023 and is now testing its third-generation MTIA 450 chip, internally known as Arke. Work on the next-generation MTIA 500, known as Astrid, is expected to finish within about 1 month.
Meta plans to use Astrid more widely across its data centers than earlier chip generations. The company expects deployment to begin in the first half of 2027, with broader use planned by the end of the year.
The chips are being co-designed by Meta and Broadcom and manufactured by TSMC. TSMC delivered the first batch of 12 new chips on September 1. Their actual performance was within 2% to 3% of earlier simulation results.
Meta engineers used the chips to run the company’s AI models on the same day they were delivered. Initial tests did not identify design flaws. The hardware team still expects several months of debugging and optimization, while manufacturing output is being increased.
Power consumption remains a key measure for data center operators because AI workloads require significant computing capacity. Meta has committed to deploying more than 1 gigawatt of its own chips within 12 months after deployment begins. The company expects deployment to increase further after that period.
Meta has also involved its Superintelligence Lab in chip tuning. The AI team provides information about the computing needs of future models, particularly during the inference stage when models process user requests.
Meta Shifts Focus Toward AI Inference
Meta previously worked on a chip called Olympus that was designed to handle both AI training and inference. The company has canceled that project and is now focusing its custom chip efforts on inference workloads.
The change reflects the cost considerations linked to deploying AI chips across large data center operations. Meta expects costs to become increasingly significant as deployment reaches multi-gigawatt levels.
According to Meta’s engineering leadership, a chip designed for both training and inference could cost about 30% more than an inference-focused chip. The company has chosen to focus on the latter as it expands deployment.
Meta will continue purchasing GPUs from Nvidia and AMD alongside its own processors. Its custom chip program is intended to add another source of computing capacity rather than replace external GPU purchases entirely.
For business owners and technology companies, Meta AI chips plans show the growing focus on computing efficiency within AI infrastructure. Data center operators face rising demand for processing capacity as companies deploy more AI models and services.
Meta’s planned deployment also puts greater emphasis on performance per unit of energy and investment. The company’s deployment forecast assumes that demand for AI infrastructure remains stable and that the AI market does not face a sharp downturn.
