Meta to deploy custom AI chips in 2027 to cut Nvidia reliance

Meta Platforms ↗ 0.84% plans to begin deploying its third-generation custom artificial intelligence chip, MTIA 450 or Arke, in data centers during the first half of 2027, followed by a broader rollout of its fourth-generation chip, MTIA 500 or Astrid, by the end of that year, according to a Bloomberg report published Tuesday.bloomberg

The move marks a major step in the social media giant's effort to reduce its dependence on Nvidia ↗ 2.78% hardware while cutting energy consumption and infrastructure costs across its sprawling AI operations.

Custom Silicon Takes Shape

Meta first announced plans to develop homegrown AI chips in 2023, and the Arke processor represents the third iteration of that initiative. The chips are being developed in collaboration with Broadcom ↗ 0.74% on design and TSMC Taiwan Semiconductor Manufacturing Company Limited ↗ 6.33% on manufacturing.investing+1

The first shipment of 12 test processors arrived from TSMC on September 1, with performance falling within 2% to 3% of design simulations. Engineering teams immediately began running Meta's proprietary AI models as well as third-party tools from DeepSeek and Alibaba ↗ 1.28% on the new hardware.blockonomi+1

Yee Jiun Song, Meta's vice president of engineering who oversees the custom silicon program, told Bloomberg that each chip generation delivers better performance per watt of energy and per dollar spent compared to commercially available alternatives. The chips can perform inference operations — the phase where AI systems generate real-time responses — more efficiently than current Nvidia offerings, Song said.investing+1

Scale and Strategy

Meta has committed to more than a gigawatt's worth of the proprietary chips over a 12-month period. Song emphasized that at such scale, absorbing even a 30% cost increase becomes financially impractical, underscoring the rationale for the in-house approach.investing+1

The company previously pursued a chip code-named Olympus, planned for 2028 or 2029, that would have handled both model training and inference. Meta canceled that project citing cost considerations, choosing instead to focus its semiconductor strategy exclusively on inference workloads. Meta's Superintelligence Labs contributes to the effort by communicating upcoming model specifications to the chip engineering division, allowing processors to be designed around practical deployment scenarios.blockonomi+1

Market Reaction

Meta shares rose as much as 2% Tuesday morning before settling at roughly 0.75% gains. The move extends recent positive momentum following the company's introduction of the Muse AI agent.investing+1