Hardware accel and in‑house compute showing up
Signals from the field: Livepeer flagged Netflix’s acquisition of an AI video startup (interpositive) as evidence of in‑house compute moves, while NETINT promoted VPU acceleration via Scalstrm to ease CPU bottlenecks in video workflows noted explained. The trend is clear — companies are blending specialized silicon and private infrastructure to cut per‑video AI costs.
Netflix’s March 5, 2026 announcement formalized the transfer of InterPositive’s 16‑person engineering and research team into Netflix as part of the deal. Bloomberg later reported the transaction could total as much as $600 million including performance‑linked payouts. Livepeer has been publicly repositioning around a real‑time AI video strategy called “Cascade,” with a November 2025 update that maps orchestrator and inference workflows for production‑grade AI jobs on its network explained. Livepeer’s technical docs also detail node types and gateway components intended to run AI inference alongside traditional transcoding in distributed deployments. NETINT’s VPU ecosystem pages describe how Scalstrm integrates orchestration and just‑in‑time (JIT) transcoding to route workloads to NETINT VPUs across cloud, hybrid, and on‑prem pipelines. Scalstrm’s own writeups and demos from NAB 2025 showcase JIT workflows that the company says cut pre‑encode storage and server counts in production tests, including a case report of replacing dozens of servers with a small VPU‑backed footprint profiled. Akamai launched cloud instances using NETINT Quadra VPUs on March 27, 2025 announced, and NETINT documentation and analyst writeups cite a Quadra T1U encoding up to 32 concurrent 1080p30 streams while consuming roughly 17W of power reported. NETINT and Scalstrm materials quantify that shifting CPU‑heavy transcoding and inference onto VPUs plus JIT orchestration reduces per‑video storage and compute overhead in both cloud and hybrid setups.
