Meta plans to deploy its second-gen in-house chips, referred to internally as Artemis, to its data centers in 2024, alongside “commercially available GPUs”
Facebook owner Meta Platforms (META.O) plans to deploy into its data centers this year a new version of a custom chip aimed …
Context & Ripple Effects
Meta had already outlined its RSC supercomputer and planned MTIA and MSVP accelerators, making Artemis an operational step in a broader effort to tailor compute to its own AI and video workloads. The decision to retain commercial GPUs signals that custom silicon is being added to—not abruptly replacing—Meta's existing compute mix.
The later testing of an in-house AI training chip and subsequent MTIA roadmap show how an initial deployment can become a longer-running internal silicon program spanning multiple workloads.
First-order effects
- Meta can begin validating Artemis in production data centers while continuing to use commercially available GPUs, giving its infrastructure teams two compute paths to schedule against workload requirements.
- The deployment creates a direct proving ground for Meta's chip-design and data-center software teams: Artemis must operate within the same fleet and operational processes as third-party accelerators.
Second-order effects
- A mixed fleet raises the value of workload placement, compilers, and orchestration that can match jobs to either internal chips or GPUs rather than treating accelerators as interchangeable.
- Commercial GPU suppliers remain part of Meta's near-term capacity plan, but a successful internal deployment gives Meta more leverage and a potential second source for the workloads Artemis can serve.
Third-order effects
- If successive deployments expand across training, ranking, and other workloads, large AI operators may increasingly differentiate through vertically integrated hardware-software stacks rather than GPU procurement alone.
- The enduring architecture is likely heterogeneous rather than single-vendor: internal accelerators can handle selected workloads while commercial GPUs remain necessary where they offer better availability or performance.
The trend: This is an early example of hyperscalers building heterogeneous, in-house AI compute to supplement rather than immediately displace commercial GPUs.