Meta outlines its RSC supercomputer and work on two chips: MTIA to train and run AI models, set for release in 2025, and MSVP, to accelerate video workloads
At a virtual event this morning, Meta lifted the curtains on its efforts to develop in-house infrastructure for AI workloads …
Context & Ripple Effects
Meta is laying out its full in-house compute stack on one day: the AI Research SuperCluster it unveiled in early 2022 as the training backbone, plus two custom ASICs announced at this event — MTIA for AI training and inference, slated for 2025, and MSVP for video workloads. The through-line is that Meta wants silicon designed around its own ranking, recommendation, and video pipeline rather than around generic merchant parts.
The announcement reads as a roadmap reveal rather than a shipping product — no benchmarks, no deployment numbers — which makes its significance about intent and cadence. The related coverage shows that cadence held: MTIA v1 and a next-generation version were both in production within a year, and by 2026 Meta had expanded the line into four numbered MTIA chips led by the MTIA 300 for content ranking.
First-order effects
- Meta gains two internal design efforts targeting its largest cost centers directly — model training/inference (MTIA) and video transcoding/serving (MSVP) — reducing dependence on any single merchant accelerator vendor for its core feed workloads.
- Because neither chip ships immediately, near-term compute capacity still rests on existing infrastructure like RSC and purchased GPUs, so the 2025 MTIA target becomes a hard deadline for Meta's silicon teams.
Second-order effects
- Accelerator vendors' largest customers are now also competitors-in-waiting: if Meta's own ranking and video silicon proves out, every hyperscale buyer of AI chips has more leverage to demand custom options or better pricing.
- A dedicated MSVP signals that video processing is becoming its own silicon category — peers running heavy video pipelines face pressure to justify why they still run those workloads on general-purpose accelerators.
Third-order effects
- The subsequent record — next-gen MTIA in production by 2024 and the MTIA 300–500 family by 2027 — shows the pattern compounding: hyperscalers shift from buying finished accelerators to owning their own accelerator product lines.
- That structure splits the AI hardware market into merchant vendors selling platforms to everyone else and vertically integrated giants whose best workloads never touch the open market — with content ranking, not frontier model training, emerging as the beachhead workload for in-house chips.
The trend: Hyperscalers are moving from renting AI accelerators to designing their own silicon lines, starting with high-volume internal workloads like ranking and video before extending toward training.