AWS-backed test: ION Video beats AI agents on cost, speed and storage
An AWS-funded proof of concept found ION Video used 99.6% less processor time and 99.2% lower AI costs than Cursor and Gemini on identical EC2 hardware, delivering edits 38 times faster.

Why it matters
- ION Video's system used 99.6% less processor time and 99.2% lower AI model costs than Cursor and Gemini on identical AWS EC2 servers.
- ION delivered playable edits for five AFL requests in 1 minute 47 seconds; the AI agents took 1 hour, 8 minutes and 34 seconds.
- At one million edits, ION estimated costs of USD $8,950 versus USD $1.13 million for Cursor.
- Swinburne University of Technology is independently testing whether ION can deliver 30% more data centre capacity and 30% lower operating costs.
- Nordic operators including Telia and Fossefall have confirmed interest, subject to the first validation phase.
An AWS-funded proof of concept found that ION Video's editing system used 99.6% less processor time, wrote 99.8% less new storage and incurred 99.2% lower AI model costs than two off-the-shelf AI agents. That is the headline result of a benchmark that ION Video, an Australian company, ran against Cursor and Google Gemini on identical AWS servers — and it frames a concrete commercial question for the industry: does generative AI need to be this expensive for video work?
The stakes are real. Video processing is one of the heaviest workloads in modern data centres, and the test was designed explicitly to measure how a different approach to editing would affect that load. If the numbers hold up under independent validation, they point to a way of cutting both capital and operating costs in facilities that are straining under AI-driven demand.
How the test was run
The benchmark was carried out with CloudMates, an AWS Advanced Tier Partner, which worked on the benchmarking design and model selection. To keep the comparison honest, both systems processed the same footage, used the same scene index, the same AI model and the same Amazon EC2 hardware.
The workload was a two-hour, 10-minute AFL game between Carlton and Hawthorn. Each system received five requests:
- Hawthorn's shots at goal
- Highlights of Hawthorn player Jai Newcombe
- Carlton's tackles in the third quarter
- First-quarter centre bounce hit-outs
- Hawthorn's inside-50 entries
ION completed all five requests and delivered playable edits in one minute and 47 seconds. The AI agents took one hour, eight minutes and 34 seconds. In other words, ION delivered playable video edits 38 times faster.
The numbers behind the claim
The raw resource figures are stark. On processor use, ION reported 6.2 seconds of CPU time, against 1,742 seconds for the AI agent. It wrote 2.8MB of new storage; the agent wrote 1.65GB. AI model cost came to USD $0.045 for ION versus USD $5.61 for the rival approach.
Accuracy did not suffer for the speed. ION achieved a 97% match rate for relevant clips, compared with 89% for the AI agents.
ION also extrapolated the results to one million edits. At that scale, it estimated:
- AI and computing costs of USD $8,950 for ION, versus USD $1.13 million for Cursor
- Monthly storage costs of USD $13 against USD $7,600
- Processing time of 342 hours for ION and 96,760 hours for Cursor
- New storage of 560GB against 330TB
- AI token usage of 4.2 billion for ION and 6.75 trillion for Cursor
Extrapolations are estimates, not measurements. But even with generous error margins, the gap between the two architectures is orders of magnitude wide, which is what makes the comparison interesting to data centre operators rather than just to editing teams.
How does ION's system actually work?
The difference, according to ION, comes down to how each system creates video. Rather than producing a newly rendered output file for every edit, ION's software assembles edits from footage already in storage and creates a virtual file that points back to the original content. No new video is rendered; the edit is a set of instructions referencing existing material.
The system first mapped the source video once and built a searchable catalogue of 2,799 scene entries. Those entries included information on players, actions, quarters and scoring events. Once the catalogue exists, finding and assembling a highlight reel becomes a lookup and pointer operation rather than a generation task — which explains why the CPU, storage and token numbers are so lopsided.
Anthony Baker, Chief Executive Officer of ION Video, described the result as a sign of a broader change in how video could be handled. "Video is one of the most widely consumed forms of content in the world, yet the fundamental way we create, edit and store it has barely changed in decades. ION is changing that. We are effectively taking video from being a static file to something that can be dynamically assembled, edited and personalised without creating an entirely new video," Baker said.
That framing matters beyond sports highlights. If edits can be assembled as virtual files rather than rendered as new assets, the storage economics of video libraries change — and so does the processing footprint of every downstream workflow built on top of them.
Why data centres are watching
The benchmark feeds directly into ION's argument that its approach could reduce the strain video processing places on data centres. ION said the latest results exceed thresholds set for a separate independent validation trial with Swinburne University of Technology. That trial is testing whether the system can deliver a 30% increase in data centre capacity and a 30% reduction in operating costs.
Baker said the AWS-backed test supports that argument. "This is an important milestone for ION and demonstrates the potential of purpose-built video infrastructure to fundamentally change the economics of video processing. It also delivers further evidence of the potential for this technology to dramatically improve data centre efficiency," Baker said.
The results also feed into ION's wider validation work linked to data centre operators in the Nordic region. Baker said several operators, including Telia and Fossefall, had confirmed interest, subject to the outcome of the first validation phase.
CloudMates' contribution centred on making the search layer reliable, not just fast. Sumit Menaria, Co-Founder and Director of CloudMates, said: "ION had already built a way to assemble video without rendering it. Our role was to make finding the right moment as fast and dependable as delivering it. Amazon Bedrock let us test leading models side by side and choose each one on the evidence."
What happens next?
Two caveats are worth keeping in view. The proof of concept was AWS-funded, and ION published the results itself. The Swinburne University of Technology trial is the independent check, and its 30% capacity and 30% operating-cost targets give observers a concrete yardstick to judge the claims against.
If that validation lands where the AWS benchmark points, the Nordic operators already lined up give ION a commercial path. And for the broader AI industry, the test poses an uncomfortable question in reverse: how much of today's generative AI compute spend on video is solving a rendering problem that purpose-built infrastructure never needed to have in the first place?
Source: itbrief.com.au
More from Elena Vasquez
Show full bio
Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.
708 articles
Related articles
- Adobe's EditAgent Makes AI Judge Its Own Photo Edits With Hard Numbers
- AI Video in 2026: Wan 3.0 Leads at 1175 Elo, MiniMax H3 Is the Open-Weights Base
- Google Cloud ships Gemini Omni 1.2 Flash at $0.07 per second
- GPT-6.1 Terra halves API prices and closes the gap on OpenAI's biggest models
- Nol8 launches AI Data Plane with 1 GB/s deterministic inspection