The Ten Seconds You Need, Found with a Single Sentence: ION and CloudMates

October 2026CloudMates

Somewhere in every video library is the exact moment someone is looking for: a winning goal, a near-miss at an intersection, the line a CEO said on air. Today, finding it means someone watching hours of footage. ION decided it should take one sentence.

ION worked with CloudMates, an AWS Advanced Tier Partner, to build AI search on Amazon Bedrock that finds any moment across a library in seconds. ION’s patented technology then turns those moments into a personalised video, without re-rendering a single frame. The work began as a proof of concept, delivered with support from AWS between August and September 2026.

Today: scrub through footage, cut and re-render, store a new file. With ION: ask in plain English, get the exact moments in 4–7 seconds, assemble virtually in seconds with no new file.
ION’s new AI search, built with CloudMates on Amazon Bedrock, replaces the manual hunt. ION’s patented virtualisation replaces the re-render.

Why it matters

Video is everywhere. The moment you need is buried in it.

Video carries most of the world’s internet traffic, from films and sport to news, training and security footage. But it is still stored and shipped as finished files. Turning one moment into something new, a highlight, a promo, a personalised replay, means finding it by hand, then cutting, re-rendering and storing yet another file.

Multiply that by every viewer, every market and every format, and most of the value locked inside a video library never gets used. The work is too slow and too expensive to do at scale.

ION had already solved the delivering. To solve the finding, it worked with CloudMates.

Two pieces, one flow

ION could always deliver the moment. Now it can find it too.

ION (ASX: IOV), headquartered in Melbourne, holds six patent families covering virtual video. Instead of copying and re-encoding files, ION turns a video into a lightweight map of where every moment lives, measured in kilobytes rather than gigabytes. Any application or AI agent can then ask ION to assemble any combination of moments, for any viewer, in seconds.

That leaves one question: which moments? A person can say what they want in a sentence, but a computer has to know what is happening in every second of footage to answer it. To answer it, ION worked with CloudMates to build an AI search layer using generative AI on Amazon Bedrock.

ION’s platform in three steps: Virtualise, Select (new, built with CloudMates on Amazon Bedrock), Assemble and deliver
ION’s platform in three steps. The middle step is new, built with CloudMates on AWS.

Why virtual assembly matters

Delivering the moment without rendering it

ION has published its own benchmark of the delivery step. Across five plain-English requests on a 2-hour-10-minute AFL broadcast, ION’s production system was compared with a general-purpose AI agent that builds each edit by rendering new video files. Both ran on identical Amazon EC2 instances against the same index.

38×
faster to a playable edit
1 min 47 s vs 1 h 8 min
590×
less new storage
2.8 MB vs 1.65 GB
125×
lower model cost
$0.045 vs $5.61
97%
of clips correct
vs 89% for the agent

Projected at 100,000 edits

ION’s projection from the same benchmark, at list prices.

IONAI agent
Total cost$895$112,675
Total work time596 h22,856 h
New storage added56 GB33 TB
Storage cost per month$1.29$760

Source: ION, “ION vs. AI agent” benchmark. Five requests on one broadcast, tested 1–2 October 2026. Totals are for all five requests combined. The 100,000-edit figures are ION’s extrapolation from this single broadcast and exclude ingest and serving costs; results on other footage will vary. docs.ion.video/benchmark.

Who benefits

What changes when you can just ask

The same system was tested across four very different kinds of video. Each example below is the kind of request it was built to answer.

Studios & streamers

Promos and personalised previews, in minutes

Find me 5 fighting scenes with the lead actor

Marketing teams pull the right scenes from trailers and catalogues in seconds, and every viewer can get a cut tailored to them without a single new file being rendered.

Sport

Every fan’s own highlight reel

Collingwood goals

Instead of one highlights package for everyone, each fan can get the moments they care about: their club, their player, assembled on demand.

Safety & security

Incidents found in seconds, not shifts

A car runs a red light at an intersection at night

Operators describe what happened instead of reviewing hours of CCTV or dash-cam footage, and get the exact clips to act on.

Corporate & news

The quote you need, straight from the archive

Where does the CEO talk about AI chips?

Communications and media teams jump to the exact moment a topic is discussed across interviews, briefings and broadcasts.

ION has always been able to assemble any moment in seconds. What we needed was a dependable way to find the right moment from a plain-English request. We worked with CloudMates to build that on AWS, tested it against real user requests, and connected it straight into our platform.Umair QureshiHead of Engineering

How it works

Understand once. Search in seconds.

The heavy lifting happens once, when a video is added. After that, every request is fast, because the system already knows what is seen and said in every moment.

Four stages: add videos, describe moments with generative AI on Amazon Bedrock, ask for a moment, and ION delivers the clip
Fully serverless on AWS. Models from Amazon, Anthropic, Cohere and TwelveLabs were tested side by side on Amazon Bedrock, and only those that measurably improved results were kept. No custom model training and no GPUs to manage.

Results

Put to the test with real requests

Users wrote 100 everyday requests across trailers, sport, CCTV and corporate video, and the system had to find the right moments for each one.

87/100
user requests answered correctly
across 4 content types
4–7 s
from a sentence to ranked moments
typical search response
340
videos made searchable in under 17 minutes
every one processed successfully
15,500+
individual moments ready to find
trailers · sport · CCTV · corporate

Requests answered correctly, by type of video

Movie trailers
96%
Corporate & news
92%
Sports
80%
CCTV & dash-cam
80%
25 requests per type. Answers that met only part of their criteria are counted as not correct.

The economics of virtual video

Cost. Storage. Capacity.

Finding the moment is only part of the story. ION’s virtual assembly avoids rendering a new video file for every edit—changing the resources needed to deliver it.

125×

lower model cost

Less model spend to produce the requested edits.

Model cost across five benchmark requests
590×

less new storage

A lightweight virtual edit instead of another rendered video file.

New storage across five benchmark requests
34 CPU-h

versus 9,676 CPU-h

Less compute tied up in rendering, leaving more room for other work.

ION’s projection for 100,000 similar edits

ION’s published virtual-assembly benchmark, not a measurement of the CloudMates search proof of concept. Five requests on one broadcast; model cost is not total platform cost, and storage figures cover newly created edit data. CPU figures are extrapolated to 100,000 similar edits. Ingest and serving costs are excluded; results depend on content and configuration.

Read ION’s benchmark and methodology
ION has solved the hardest infrastructure problem in video. Our role as ION’s AWS partner was to help make finding the moment as fast and dependable as delivering it. Amazon Bedrock let us test leading models side by side and choose each one on evidence, not assumption.Sumit MenariaCo-Founder & Director, CloudMates

About CloudMates

The right partner for media and generative AI

CloudMates is an AWS Advanced Tier Partner and holds both the AWS Media & Entertainment Competency and the AWS Generative AI Competency. That combination brought together the two things this project needed: an understanding of how media companies store, manage and deliver video, and hands-on experience taking generative AI on Amazon Bedrock from experiment to measurable results.

What’s next

Toward video that assembles itself for every viewer

With finding and delivering now connected, the next step is production: scaling to full libraries and deepening each content type. The longer-term direction is the one ION was built for. AI agents will understand a request, find the right moments across an archive, and deliver a personalised video for every viewer, in seconds, without rendering a new file.

To learn more about ION, visit ion.video.

ION and Video Superintelligence are trademarks of ION. AWS, Amazon Bedrock and related marks are trademarks of Amazon.com, Inc. or its affiliates. Results describe the proof-of-concept environment and an independent set of 100 user requests; production results will depend on content and configuration.