AI is reshaping the economics of live production
As AI moves into live production, whether its powering instant replays, real‑time tracking, and adaptive on‑screen graphics – hidden inefficiencies will begin to surface as added cost and latency.
Most broadcast and streaming infrastructures still rely on mezzanine video formats built for human perception, not machine analysis. This inefficiency reaches beyond engineering, slowing innovation, raising costs, and limits how fast AI can turn live content into new experiences.
- While ST 2110 made live over IP scalable for broadcast, AI workflows stress the network differently. They often rely on unicast fan-out and multiple redundant encodes to feed separate models, which doesn’t scale.
- Modern low‑latency mezzanines are still human-first: even JPEGXS and JPEG 2000 require full‑frame decode or coarse tiling, forcing pipelines to transport, decode then discard most pixels for model inputs.
- Result: proxy chains proliferate. Each new AI task spawns another rendition, more transport, and more latency.
If your transport treats every AI like a viewer, you pay a full‑frame tax.
Andy Beach, Alchemy Creations
The hidden bottleneck: transport
AI performance is often measured by model accuracy and throughput, but in live environments the real limiter is delivery, or how efficiently the right pixels reach the right model. Traditional mezzanine workflows often create redundant proxy chains where the same feed is re-encoded for each task. This wastes compute, floods the network, and adds delay where milliseconds matter.
The smartest model can’t scale if every new workload spins up another proxy encode. The network itself must become an AI managed foundation of the AI pipeline.
Brent Yates, swXtch.io
A new approach: swXtch.io meets VC-6
To solve the scaling problem, the data format and the network have to think together.
SMPTE VC-6 (ST 2117-1)
swXtch.io provides the AI-native multicast overlay that makes this practical in live environments. One VC-6 stream can serve dozens of AI workloads in parallel, each subscribing to the specific LoQ and ROI it needs for its task. There’s no need for proxy encodes or duplicated feeds; the same data moves through the network once and scales effortlessly to more models.
“You encode once and serve many,” says Andy Beach of Alchemy Creations, “Each AI gets exactly the pixels it needs.”
A VC-6 powered swXtch.io architecture make the transport layer an active part of the AI pipeline — compressing less, computing smarter, and moving live production closer to true real-time intelligence.
The diagram below illustrates how the components can work together in practice, using VC-6 with NVIDIA Holoscan for Media on Microsoft Azure as an example.
Proven efficiency gains
In collaboration with NVIDIA, V-Nova benchmarked VC-6 on Blackwell GPUs, demonstrating an order of magnitude improvement over CPU processing. GPU decoding takes around a millisecond for 2K frames, confirming that VC-6’s parallel S-Tree architecture maps efficiently across GPU cores.
Designed for massive concurrency, the codec is inherently GPU-native, and optimized for real-time AI pipelines where milliseconds matter. Further independent tests by V-Nova showed that VC-6 decoding is 58% faster than JPEG and nearly 90% faster than JPEG 2000 in live conditions.
These results validate VC-6 not only as a more efficient codec but as a format optimized for compute-aware workflows.
In live multicast scenarios (1080p30), deploying VC-6 on swXtch.io infrastructure delivered measurable operational gains compared to uncompressed ST 2110:
- Bandwidth: 1.304 Gb/s → 118 Mb/s (–93%)
- Packet rate: ~89.7 kpps → ~8.1 kpps (–96%)
- Receive‑side processing: ~32 ms (ST 2110) → 16.4 ms (VC‑6 LoQ0) → 9.6 ms (LoQ2), –a 70% reduction
- End‑to‑end AI latency: ~1.5 s (ST 2110) → 731 ms (LoQ0) → 317 ms (LoQ2), roughly an 80% overall improvement.
Together, these results highlight how a hierarchical codec and an AI-native network can significantly reduce data movement and processing time while maintaining production-grade quality. In live use cases, this translates directly into lower egress costs, faster model feedback, and more scalable AI participation in the signal path.
What this enables
These performance gains aren’t just numbers on a chart. They open up a new class of real-time workflows where AI participates directly in the live signal path rather than reacting after the fact. swXtch.io network layer, paired with VC-6, makes it possible to feed models only what they need, when they need it, thus reducing redundancy and unlocking new creative and operational possibilities.
Real-time tracking
Models subscribe only to the resolution or bounding boxes they need, cutting per-frame data volume by up to 90% and freeing GPU cycles for more concurrent inference tasks.
Instant replays
Replay clips assemble directly from the live multicast feed, reducing turnaround from 10–15 seconds to under 2 seconds by decoding only the required segment and LoQ.
Personalized viewports
A single VC-6 multicast can serve many individualized feeds. Each client subscribes to its own LoQ or region of interest with no new encodes and far less egress.
Agentic AI
Through swXtch.io’s Model Context Protocol (MCP), models can dynamically adjust their own data requests, raising or lowering their LoQ based on confidence or available compute and adapt inputs in under 200 ms.
Multi-task AI pipelines
Detection, tracking, segmentation, and graphics generation can run in parallel from a single encoded source, eliminating redundant processing and keeping live workflows in sync.
Edge-to-cloud efficiency
GroundSwXtch, swXtch.io’s on-premises multicast gateway, and CloudSwXtch, its cloud-based multicast overlay, operate as a unified network fabric. This allows production and AI workloads to share the same transport path without separate routing, transcoding, or storage steps. With VC-6, ground-to-cloud transmission achieves up to an order of magnitude reduction in data transfer, significantly cutting bandwidth and egress costs.
Together, these capabilities redefine what “live” means in AI-enabled production. The system not only moves media, it also understands how that media will be used.
Standards matter
Standards are what turn innovation into something that can scale. VC-6 is already a SMPTE standard (ST 2117-1) and interoperates seamlessly with ST 2110 production environments. This alignment allows broadcasters to deploy AI-native workflows without overhauling infrastructure or re-engineering existing toolchains
Standards also bring stability and trust. They make new technologies vendor agnostic, easier to integrate, and safer to procure. Because VC-6 is a SMPTE standard, it inherits long-term support, multi vendor compatibility, and a clear path for future extensions, such as standardized metadata for regions of interest or AI-driven transport signals. swXtch.io provides a standards-based open network solution that is interoperable with any IGMP standard applications and works on any cloud.
In practice, that means teams can layer swXtch.io’s infrastructure with VC-6 into today’s IP production workflows with confidence, extending what already works instead of starting from scratch.
Efficiency, scalability and Sustainability
By moving preprocessing into the network, VC-6 and swXtch.io network layer cut redundant data movement, lower compute demand, and shrink the energy footprint of live AI workflows. A single multicast stream can serve dozens of inference tasks simultaneously with no new encodes, no duplicate routes, and no wasted pixels. The result is lower bandwidth, lower power draw, and a direct reduction in cloud and egress costs.
Because this efficiency scales, operators can add new AI workloads without multiplying infrastructure requirements. This provides a practical path to sustainable, standards-based AI adoption that aligns with both operational budgets and environmental goals.
The smartest path to live AI
VC-6 and swXtch.io’s infrastructure bring intelligence to the very fabric of live media delivery. Together they combine a hierarchical, compute-aware video format with an adaptive, multicast-driven network, making real-time AI workflows scalable, efficient, and sustainable.
“V-Nova’s VC-6 makes visual data more compute-aware,” says Guendalina Cobianchi of V-Nova. “swXtch.io makes the network adaptive. Together they make live AI practical.”
The impact goes deeper than efficiency. It transforms how media is processed, interpreted, and produced in the moment. By building preprocessing into the transport itself, live content can now be shaped, analyzed, and personalized while it moves, without adding latency or duplication. For the first time, the same signal that serves a viewer can also serve the models that understand it.
For broadcasters and platform operators, that means AI becomes a native part of production, not an external system. It enables new approaches for automation, interactivity, and personalization at scale, all within the reliability of existing SMPTE standards.
And this is only the beginning. As VC-6 and swXtch.io move into commercial deployment, they point toward a media infrastructure where every frame carries both human and machine value. Where video isn’t just transmitted, it’s interpreted in real time. It’s a practical foundation for a smarter, more sustainable era of live production, built on open standards and shared innovation.
What’s next: Commercial availability in early 2026
Commercial availability is targeted for early 2026, marking a significant step toward operational AI-native live production that is standards based, cloud integrated, and ready for enterprise deployment.
