Sources: Facebook is developing an ML chip for tasks such as recommending content to users and a data center chip to improve video transcoding
Wayne Ma / The Information :
Context & Ripple Effects
Facebook's in-house silicon effort has been building for years: it open-sourced its AI hardware back in 2015 while racing Google, gave Intel technical input on the Nervana Neural Network Processor in 2017, and by 2018 was staffing a dedicated chip design team, with Yann LeCun framing the goal as energy-efficient chips for filtering live video. The new report says that work has crystallized into two concrete chips — one for ML tasks like content recommendation, one for data center video transcoding.
What makes the timing notable is the workload split: recommendation and transcoding are Facebook's highest-volume, most power-hungry data center tasks, exactly the targets LeCun described in 2018. Moving them onto custom silicon is the payoff of that multi-year build-out rather than a new direction.
First-order effects
- Facebook's recommendation and video transcoding pipelines shift toward in-house chips, reducing its dependence on merchant silicon for the workloads that dominate its data center spend.
- Intel's position changes from collaborator — Facebook helped shape the Nervana chip — to incumbent supplier being designed around on Facebook's most repetitive workloads.
Second-order effects
- Chipmakers lose pricing leverage with the largest hyperscalers as those buyers internalize design for predictable, high-volume tasks, leaving merchant vendors to compete on the workloads Facebook still won't build itself.
- Given Facebook's history of publishing server designs through Open Compute, its custom silicon work could eventually surface as shared specs, pulling other operators toward the same workload-specific approach.
Third-order effects
- The pattern points toward hyperscalers permanently splitting their compute between merchant chips and workload-specific in-house silicon — a structural shift in which the biggest cloud and social platforms become their own chip vendors for the tasks that define their cost base.
- If the 2015-to-2021 arc holds, custom silicon stops being a differentiator project and becomes table stakes for any platform operating at Facebook's scale.
The trend: Hyperscale platforms are moving from collaborating with chipmakers to designing workload-specific silicon in-house, with Facebook's multi-year chip effort as one data point in that migration.