MakinaRocks Runway

Read Runway as infrastructure, not as a model vendor. The product page sells three things in order: GPU return on investment, a standardized development and operations environment, and governance that survives an audit. Each of those is a concrete, checkable claim rather than a capability badge.On GPU economics, MakinaRocks says cards are split at fractional granularity so training reaches up to 90% utilization and inference 60-70%, idle GPUs are detected and reclaimed in real time, and a dashboard lets you assign priority so one team cannot monopolize a card. The vendor frames the throughput gain as more than 4x. If your binding constraint is GPU spend rather than model quality, this is the part to test first — ask for a utilization baseline on your own workload before you sign.On standardization, the platform ships a catalogue of pre-configured open-source apps and an integrated development environment, so a model experiment, its training run, its deployment and its monitoring stay in one place. The vendor claims environment build time drops by 100% and overall platform build time from months to about two weeks; a July 2026 case note describes a semiconductor materials customer build cut from 12 days to 4 hours. The mechanism is the offline install package plus the mirror repository.On governance, Keycloak-based SSO applies role-based access control across every layer, and every action is written to an audit log. For defense and public-sector buyers this is often the deciding line item, because it is what makes an AI deployment defensible to a security reviewer.The agent families — expert augmentation, control, prediction and operations, machine vision — are where the industry specificity lives, and DrawX is the one you can evaluate today: drawings become a searchable standard asset, search and comparison are automated, and RFQ handling is automated on top. DrawX is also the clearest proof point, with the vendor claiming review and management time cut by up to 90%.Weaknesses worth naming. Runway assumes you already run MES, SCADA or ERP for agents to act on, and rollouts are services-led rather than self-serve, so plan for a project not a signup. Published product material, case studies and customer stories are Korean-language even though the site is served in Korean, English and Japanese; non-Korean teams should budget localization time. The vendor publishes no connector catalogue, so fit against your own stack has to be established in a scoping conversation. Recent engineering posts — deploying apps as a non-developer with a coding assistant plus Runway, and predicting robot failures from normal data only — are useful signals of where the team is investing,