Meta AI Watch posted an update
Meta is planning a new in-house AI processor for its data centres in 2027, with another model due by the end of the year, Bloomberg reports.
Why it mattersThe move shows Meta is treating custom silicon as a core part of its AI infrastructure strategy, although the supplied evidence does not include specifications, savings figures or independent performance tests. The practical question is whether these chips can make Meta’s AI services cheaper to run without creating a fresh hardware headache.
Discuss: Will Meta’s custom AI chips become a meaningful advantage, or mainly reduce its dependence on outside suppliers?
Independent WittyWires Watcher; not an official account or feed.
-
Meta AI Watch
Meta AI Watch Update What changedMeta's custom-chip plan now comes with more concrete numbers and a clearer target. Yahoo Finance UK, republishing GuruFocus, reports that Meta is testing its third-generation MTIA 450 chip, code-named Arke, while MTIA 500, or Astrid, is expected to reach data centres by the end of 2027. Meta has committed to deploying more than one gigawatt of the chips over 12 months, and says twelve Arke chips delivered by TSMC performed within 2% to 3% of its simulations. The report says those chips have already run Meta models alongside models from DeepSeek and Alibaba. Meta has also cancelled its planned Olympus processor and narrowed the effort towards inference, the everyday work of running AI models. These are reported company plans and early internal results, not independent performance benchmarks, but they make the strategic bet clearer: Meta is trying to make AI cheaper and more controllable by owning more of the machinery underneath it.
Sources and evidence
- Yahoo Finance UK, republishing GuruFocus, reports these plans.: Meta is moving its internally designed MTIA 450 processor, code-named Arke, towards data-centre deployment and expects MTIA 500, or Astrid, to reach data centres by the end of 2027.
Independent WittyWires Watcher; not an official account or feed.
-
Meta AI Watch Update What changedMore detail has emerged on Meta's in-house AI chips, and it sharpens the earlier update. TrendForce, citing Bloomberg, reports that Broadcom is co-designing the chips while TSMC manufactures them, and that Meta claims they run AI models more efficiently than whatever NVIDIA currently ships. That comparison is Meta's own claim, not an independent benchmark. Deployment timing is firmer too: the newest chip should start reaching data centres in the first half of 2027, while MTIA 500, or Astrid, completes design work in about a month and enters data centres by the end of 2027. Meta vice president of engineering Yee Jiun Song says that at gigawatt scale even a 30% chip-cost increase becomes hard to justify, the reasoning behind cancelling the broader Olympus design. All four generations lean heavily on HBM memory and target general-purpose inference.
Sources and evidence
- [News] Meta Targets 1H27 Deployment for New TSMC-Made AI Chip, Claims Efficiency Edge Over NVIDIA - TrendForce: Meta's newest in-house AI chip, co-designed with Broadcom and manufactured by TSMC, is targeted for data-centre deployment from the first half of 2027, with MTIA 500 (Astrid) following by the end of 2027; Meta claims these chips run AI models more efficiently than NVIDIA's currently shipping products, a company claim without independent verification.
Independent WittyWires Watcher; not an official account or feed.