Meta Platforms to deploy gigawatt of custom AI chips

Meta's push into custom silicon is not new, but the scale and specificity of its roadmap are. The company is currently testing its third-generation MTIA 450 processor, code-named Arke, while its successor, MTIA 500 or Astrid, is expected to complete design work in about a month and reach data centers by the end of 2027. This follows a strategic narrowing: Meta canceled its planned Olympus processor, which was intended to handle both AI training and inference, partly because of cost concerns, and is now prioritizing inference—the day-to-day running of AI models. Early results appear encouraging: twelve Arke chips delivered by TSMC on Sept. 1 performed within 2% to 3% of Meta's simulations and have already run Meta models alongside models from DeepSeek and Alibaba.

The effort is closely tied to Meta's broader AI economics. By designing chips in-house with partners Broadcom and Taiwan Semiconductor Manufacturing Co., Meta aims to lower the enormous energy and computing costs attached to its AI ambitions while building more control over a critical infrastructure layer. Meta engineering vice president Yee Jiun Song said each generation takes on a little more technological risk and delivers better performance per watt and per dollar.

Whether this translates into meaningful cost savings hinges on execution and deployment speed. If MTIA deployment lowers power consumption and inference expense at scale, Meta could protect margins while continuing to expand AI across advertising, recommendations, assistants and other products. The risk is that designing competitive chips requires enormous capital, long development cycles and tight coordination with partners—leaving the ultimate impact on Meta's cost curve still to be determined.