Meta’s MTIA roadmap races OpenAI’s Jalapeño deployment week
Meta is stacking new MTIA custom AI chips as OpenAI’s Jalapeño ASICs leave the lab into live racks. Social and ad ranking still lean on cheaper home-built inference beside merchant GPUs. The big-tech silicon race is no longer a slide deck — it is fleet hardware.
Meta’s custom-silicon push sits beside OpenAI’s Jalapeño deployment headlines this weekend. In March, Meta said it was developing and deploying four new MTIA generations within two years for ranking, recommendations, and generative-AI work, with MTIA 300 already in production for recommendation training and later chips aimed at GenAI inference into 2027. An August Engineering at Meta note detailed MTIA 300’s built-in NICs and communication engines, claiming 3.9× faster total communication time than an equivalent GPU cluster on a 150-billion-parameter recommendation model.
OpenAI’s Tom’s Hardware interview shows another frontier lab pairing a home-built inference ASIC with AMD hosts — the same competitive logic Meta cites when it says hundreds of thousands of MTIA chips already serve organic and ads inference. Nvidia’s Blackwell blog the same week underscores that merchant GPUs remain the shared fabric even as custom silicon ramps.
For users of Facebook, Instagram, and WhatsApp, the stake is whether cheaper custom inference keeps feeds and ads responsive without endless cloud price spikes. Backstory: Google TPUs and Amazon Trainium/Inferentia set the hyperscaler pattern; Meta and OpenAI are now deep enough that deployment maturity, not just tape-out photos, is the news. Engage: Meta’s MTIA roadmap post, the MTIA 300 engineering note, and Tom’s Hardware on Jalapeño.
What we know:
Meta has public MTIA roadmap and MTIA 300 production/engineering detail; OpenAI’s Jalapeño is in deployment with Turin hosts; both keep buying merchant GPUs too.
What we don’t:
How many MTIA 400/450/500 units are live in October fleets; how Meta’s and OpenAI’s efficiency claims compare on matched public benchmarks.
Agree (full): Meta’s own posts and this week’s OpenAI silicon coverage agree big tech is moving custom AI accelerators from roadmap slides into production racks.
Differ (full): Meta emphasizes recommendation-training networking wins; OpenAI’s week centers inference perf-per-watt claims that still await broad independent audits.
Where outlets agree Meta’s MTIA posts and this week’s OpenAI silicon coverage agree major tech firms are moving custom AI accelerators into production racks.
Where they differ Meta stresses recommendation-training networking gains while OpenAI’s week centers inference efficiency claims that still lack broad independent audits.