Meta's MTIA 400 chip splits duties between LLM training and ad recommender inference
jonathanmendez · x · 2026-09-14
At Hot Chips, Meta detailed its MTIA 400, its first proper generative AI accelerator aimed at LLM training — reported faster than Nvidia's Blackwell, though still not a full replacement for AMD or Nvidia parts.
Unusually, the chip will also run inference for Meta's ad recommender systems (DLRM), the business that actually pays Meta's bills. Mixing LLM training with recommender inference is rare since the two workloads have vastly different performance demands.
The author argues advertising deserves more credit: Meta, Google, and Amazon all built their first custom AI chips to serve ads, and much of today's AI infrastructure traces back to ad-tech.
More from Companies & People
- Next Token podcast ep2: iPhone Duo, GPT 6 Astra, DeepSeek Flash and model degradations — oran_ge · 2026-09-14
- Boomer startup legends: 90% of their advice still brill, 10% harmful — jonchu · 2026-09-14
- Cohere CEO Aidan Gomez: AI Needs Evidenced Standards, Not a Big-Lab Cartel — cohere · 2026-09-14
- Alignment researcher's 2022 "moment of terror": GPT-4 done training, mentors quit OpenAI — JacquesThibs · 2026-09-14
- Artificial Analysis signs MOU as official evaluator for Korea's Sovereign AI program — ArtificialAnlys · 2026-09-14
- Zuckerberg Says AI Development Needs to Move Faster — inductionheads · 2026-09-14