Why coordination, not raw speed, determines whether video infrastructure scales with confidence
The Operational Breaking Point
Large video systems rarely fail because encoding stops working. They fail because operational complexity overtakes human and system capacity. The technical components keep running, but the ability to reason about, direct, and recover the whole system erodes as concurrency grows.
As pipelines scale, teams contend with more simultaneous jobs, more pipeline stages, more regions, and more failure domains. State proliferates. Dependencies multiply. Small misalignments propagate quickly, surfacing as symptoms far from their root causes.
This is the paradox of modern video infrastructure: the hardware has never been more capable, yet the systems it powers have never been harder to operate. The bottleneck has shifted from processing power to operational control. At scale, control is the system that matters.
When Efficiency Amplifies Complexity
Video Processing Units (VPUs) have materially improved the economics of transcoding. They increase density, reduce power consumption per stream, and make capacity more predictable. These gains are significant, but they also reshape the operational landscape in ways that are easy to underestimate.
When transcoding becomes cheaper and denser, teams respond by running more concurrent jobs, distributing workloads more aggressively, and concentrating more work on fewer resources. Each of these decisions is rational in isolation. Together, they expand the blast radius of any single mistake.
Efficiency, in other words, does not remove the operational problem. It shifts it. The faster and denser the system becomes, the more critical it is to decide precisely where jobs run, when they run, how they fail, and how they recover. Without those decisions being made deliberately, efficiency becomes a liability rather than an advantage.
KEY INSIGHT
VPU efficiency is a necessary condition for scalable video infrastructure, but it is not sufficient. Efficiency without orchestration concentrates risk. Orchestration is what converts hardware gains into system reliability.
The Hidden Pipeline
Modern video pipelines are no longer linear chains of discrete tasks. They are event-driven, multi-stage systems with dynamic, often unpredictable, behavior.
Ingest triggers processing. Processing fans out across encoding ladders and codecs. Outputs feed packaging and delivery. Monitoring loops feed back into scheduling, scaling, and retry logic. Each stage operates semi-independently, and each introduces its own failure modes, timing constraints, and resource requirements.
Newer architectures add further dynamism. Just-In-Time transcoding, for example, defers encoding until a viewer actually requests a stream, creating segments on demand rather than pre-encoding an entire catalog. Scalstrm’s JIT implementation supports H.264, HEVC, and AV1 in software and hardware (VPU), and VP9 in software, generating adaptive bitrate ladders in real time. The approach can reduce storage and compute consumption by up to 90 percent, but it also means that the pipeline’s workload is inherently unpredictable, shaped by viewer behavior rather than batch schedules.
Modern Video Pipeline: From Linear Chain to Event-Driven System
Exhibit 1: A modern video pipeline viewed as an event-driven system with control plane coordination.
In this environment, static configuration is insufficient. Control logic must respond continuously to queue depth, capacity availability, error rates, and downstream pressure. Without centralized visibility and intent-driven control, teams are forced to infer system behavior indirectly, debugging symptoms rather than directing outcomes.
The result is a pipeline that looks manageable on a whiteboard but behaves unpredictably under load. Failures appear upstream while their root causes hide downstream. Operators add compute to relieve congestion, only to discover they have made it worse by overwhelming a bottleneck further along the chain.
Orchestration as the Leverage Point
Once video infrastructure incorporates specialized compute such as VPUs, orchestration determines whether that efficiency is realized or wasted.
A capable control plane allows teams to allocate VPU capacity intentionally rather than opportunistically, balance workloads across heterogeneous resources, prevent overcommitment, and isolate failures before they cascade. It transforms scheduling from a mechanical task into a strategic function.
In practice, several operational patterns emerge when orchestration is treated as a first-class system component rather than an afterthought. Intent-based scheduling places workloads according to policy, not just availability. Graceful degradation lets systems shed load or reduce ladder complexity under pressure rather than crashing outright. Failure isolation ensures that individual pipeline stages can fail without collapsing the entire workflow. Capacity transparency gives teams real-time understanding of not just how much capacity exists, but precisely how it is being consumed.
Orchestration Maturity: From Manual Control to Intent-Driven Automation
Exhibit 2: The orchestration maturity model, from fragile manual control to resilient intent-driven automation.
These patterns are difficult to achieve through scripting or manual intervention. They require continuous coordination across pipeline stages, informed by real-time system state rather than static assumptions.
Where Scalstrm Fits
Scalstrm is a Stockholm-based streaming infrastructure company founded in 2017 by Ola Bengtsson and Patrik Alm, who bring 75 years of combined experience in media technology. The company now operates across 35 countries with 57 partner organizations, and its platform is used by operators including Telia (TV4, C More), Altibox, Tele2, and NEP Group.
Within the VPU Ecosystem, Scalstrm addresses the coordination problem directly. Rather than treating encoding, packaging, and delivery as isolated steps, the platform unifies them into four tightly integrated components: an Origin Platform for cloud-based tile recording and live stream management, a CDN Platform with a real-time programmable HTTP pipeline for content delivery, the Just-In-Time Transcoding engine described earlier, and an Origin Shield that provides content-aware caching across RAM, NVMe, and disk tiers.
The architecture is built on what Scalstrm calls a scalable micro-component model. Each component can be deployed and scaled independently across on-premises, private cloud, public cloud, or hybrid environments. This modularity matters because it allows the control plane to manage heterogeneous infrastructure as a single system rather than a collection of disconnected services.
At the hardware level, the numbers are concrete. A single server running Scalstrm’s Origin Platform can manage up to 800 adaptive bitrate channels while pushing 200 Gbps of egress. The transcoding layer integrates with CPU, GPU, and dense VPU architectures, including NETINT Quadra T1U units, supporting HEVC encoding across SD, HD, and 4K resolutions. New live channels can be launched in seconds through instant scaling, with operators paying only for active capacity.
Decisions about placement, retries, and prioritization are informed by real-time system state. When a VPU node reaches its encoder session limit, the control plane deprioritizes it and redistributes work. When a downstream packaging service slows, backpressure is applied automatically to upstream stages rather than allowing queues to grow unchecked. The Origin Shield layer adds a further coordination point, automatically identifying and protecting live streams, catch-up TV, and network PVR content with redundancy for uninterrupted service.
Scalstrm treats all compute resources as abstract execution targets, exposing hardware capabilities through normalized metrics. This allows scheduling decisions to be made based on actual performance characteristics rather than static labels, keeping orchestration hardware-agnostic while still optimizing for throughput, energy efficiency, and cost per processed hour.
The Failure Equation
Even with strong orchestration, video systems remain complex. Teams must still manage imperfect observability, evolving workload characteristics, and external dependencies such as networks and third-party services. These challenges do not disappear.
What changes is how failures are experienced. Two failure modes are especially common and consistently underestimated.
The first is silent performance degradation from partial resource exhaustion. Unlike complete node failures, conditions such as VPU encoder session limits being reached, CPU thermal throttling under sustained load, or memory pressure causing increased I/O wait do not trigger immediate alarms. In un-orchestrated pipelines, jobs continue to flow to affected nodes, processing times drift upward, and SLAs are missed without a clear root cause. Operators often misdiagnose the issue as bad content or temporary load.
The second is cascading backlogs caused by downstream bottlenecks. When packaging services, storage endpoints, or CDN ingest connections slow down, an un-orchestrated system keeps producing output at full speed. Queues grow invisibly. Failures surface upstream, far from the real problem. Operators respond by adding more compute, which only deepens the congestion.
Failure Response: The Orchestration Difference
Exhibit 3: How orchestration transforms failure response across five critical dimensions.
With coordinated control, both patterns are handled differently. Scalstrm’s platform continuously monitors resource health through its WebUI dashboard, providing real-time metrics and analytics across every pipeline stage. Degraded nodes are automatically deprioritized, and backpressure is applied when downstream systems falter. The auto-scalable architecture dynamically allocates resources based on real-time traffic, eliminating the need for operators to guess at capacity during demand spikes. Instead of cascading outages, teams encounter bounded incidents with clearer signals, controlled impact, and defined recovery paths.
THE BIGGER PICTURE
Orchestration does not eliminate failures. It converts unknown unknowns into known, bounded risks. The distinction between feeling brittle and feeling resilient is not about the absence of problems. It is about the presence of predictable, controlled responses
Economics Follow Control
Operational efficiency and economic efficiency are inseparable in video infrastructure.
Without orchestration, capacity is stranded across underutilized nodes, peaks are overprovisioned because there is no mechanism to redistribute load dynamically, and failures generate hidden costs through retries and reprocessing. Efficiency gains from VPU hardware erode quietly, diluted by the operational overhead of running an uncoordinated system.
With coordinated control, the gains compound. Scalstrm reports that its live transcoding layer delivers 50 percent savings per channel compared to conventional architectures, while requiring 75 percent less server footprint and 80 percent less power consumption. These numbers reflect what becomes possible when orchestration ensures that workloads are placed where capacity actually exists, rather than distributed across partially utilized or thermally constrained nodes.
The sustainability dimension is increasingly difficult to ignore. At 80 percent lower energy consumption and a corresponding reduction in carbon footprint, the economic case and the environmental case converge. Organizations operating at scale cannot separate cost-per-stream from watts-per-stream indefinitely.
In this sense, orchestration is not an operational luxury. It is an economic multiplier. The return on investment in VPU hardware is directly proportional to the quality of the control plane that manages it.
Four Questions That Reveal Readiness
Teams evaluating their readiness to operate VPU-accelerated video pipelines at scale should begin with four questions. If these questions prove difficult to answer, efficiency gains alone will not translate into operational confidence.
First: can you see where every job is running and why? If the answer requires checking multiple dashboards, querying separate systems, or asking a colleague, the pipeline lacks the visibility that coordinated control provides.
Second: can you change scheduling behavior without redeploying the pipeline? If policy changes require code modifications and release cycles, the system cannot respond to operational realities at the speed they demand.
Third: can individual stages fail without failing the workflow? If a single component failure triggers cascading impact across the entire pipeline, the system lacks the isolation that resilient architectures require.
Fourth: can you explain capacity usage in real time? If utilization is only understood in retrospect, through post-incident analysis or monthly reports, the team is operating reactively rather than proactively.
Closing Perspective
As video infrastructure evolves, performance improvements increasingly come from how systems are controlled, not just how fast individual components run.
VPUs make video processing efficient. Platforms like Scalstrm make that efficiency usable. Together, they represent a shift in how the industry thinks about scale: not as a problem of raw throughput, but as a problem of coordination, visibility, and control. The fact that a platform born in Stockholm now orchestrates video infrastructure across 35 countries suggests the market has arrived at the same conclusion.
The organizations that navigate this transition successfully will be those that recognize orchestration not as an operational add-on, but as the foundational layer that determines whether their infrastructure scales with confidence or collapses under its own complexity.
This article is part of the VPU Ecosystem series exploring how purpose-built silicon is reshaping video infrastructure across the technology stack. Each article examines a different architectural perspective through the lens of a partner within the ecosystem.
