在加州 Fremont 的 Meta 园区，一间名为「GUY WITH ARROW TO THE KNEE」的实验室正进行高风险的自研晶片验证；这些测试支撑 Meta 一项规模达「数千亿美元」等级的人工智慧押注。相较于常被点名的 Nvidia、Advanced Micro Devices（AMD）与 Qualcomm，Meta 过去并非顶尖晶片设计者的典型代表，但在算力需求急遽上升下，仍试图透过自制矽晶片分散供应并降低对外部加速器的依赖。
实验室目前同时跑两款新晶片工作负载：MTIA 300（Athena）已进入量产，主要用于 Facebook 与 Instagram 的排序与推荐等推论型 AI；MTIA 400（Iris）面积更大，面向生成式 AI 推论，协助如 Meta 的 Llama 这类预训练模型处理提示并输出回应（类似 ChatGPT 的互动）。Meta 表示晶片将以「相隔 6 个月」的节奏快速迭代，并在 Iris 之后推进 MTIA 450 与 MTIA 500（Arke、Astrid），且这些针对「2027 年」的产品同样聚焦推论流程。
为了同时推进至少 4 款新 AI 晶片的研发与布署，Meta 在「去年」收购新创 Rivos 以补强人才梯队；然而公司能否从推论晶片跨越到更复杂、可用于「训练」模型的自制晶片仍不明朗。尽管 CFO Susan Li 表示最终「预期」会做出自家训练晶片，但时间点、成本与达成条件未被量化说明；在硬体通常是 AI 竞赛中最昂贵环节的前提下，客制化晶片理论上可减少浪费，但先进矽开发本身仍具高不确定性与失败风险。
At Meta’s campus in Fremont, California, a lab labeled “GUY WITH ARROW TO THE KNEE” is running high-stakes validation of in-house silicon, supporting a multihundred-billion-dollar AI bet. Although top chip designers are typically associated with Nvidia, Advanced Micro Devices (AMD), or Qualcomm, Meta aims to diversify suppliers amid surging compute demand by building its own chips alongside continued purchases of external AI processors.
The lab is testing workloads on two new parts: MTIA 300 (Athena), already in production, targeted at inference for Facebook and Instagram ranking and recommendation; and MTIA 400 (Iris), a larger chip aimed at generative-AI inference that helps pretrained models such as Meta’s Llama process prompts and produce outputs similar to the call-and-response familiar from tools like ChatGPT. Meta plans rapid iteration with releases six months apart, followed by MTIA 450 and MTIA 500 (Arke and Astrid), with inference-focused chips targeted for 2027.
To execute at least four new AI chips in parallel, Meta acquired the startup Rivos last year to deepen its talent bench, but the leap from inference accelerators to homemade training chips remains uncertain. CFO Susan Li said Meta expects to eventually build its own training chips, yet timing, cost, and requirements were left unquantified; while bespoke hardware could reduce waste because Meta knows its workloads, advanced silicon development is still technically difficult, expensive, and prone to setbacks. (Key numbers: 4)