Sized to the job
Ten agents or ten thousand. Lucius scales the swarm to the brief and releases it when the work is done.
The power of a question.
Ask Lucius. A swarm of AI agents goes to work on private GPU compute, billed by the agent-hour, with zero logs and your data encrypted in and out.
l@aillusion.ai
How it works
01 · Swarm compute
One agent helps. A swarm gets it done.
aillusion runs swarms of AI agents on dedicated GPUs. Lucius splits your brief into tasks, puts the right number of agents on each and checks the work before it reaches you. Large jobs run in parallel instead of waiting in a queue.
Parallel
Verified
Checked before delivery
Agents review each other's output and Lucius tests every result against the brief, so you get evidence, not guesses.
Dedicated
Hardware we operate
Swarms run on GPUs we run ourselves, not on shared capacity resold from someone else's cloud.
02 · Agent-hours
Pay for agents by the hour.
No seats and no token arithmetic. You pay for the agents you run, for the hours they run. One agent working for one hour is one agent-hour.
For scale: OpenAI reported running about 10,000 agents for 88 hours on the Navier–Stokes problem in September 2026 (Quanta).
Human equivalent assumes one agent-hour is about one hour of focused human work, at 8 hours a day and 220 working days a year.
Elastic
Burst for a deadline
Scale up when the work needs it and back to zero when it's finished. Nothing idles on your bill.
Legible
One line per job
Every job is billed as agents × hours, so the cost of a piece of work is plain before and after it runs.
Rates
On request
Agent-hour rates depend on volume and term. Email us and we'll set out the numbers for your workload.
03 · Private by design
Zero logs. Encrypted in, encrypted out.
Your data is the point of the work, so we keep none of it. Each job runs in its own environment, and that environment is gone when the job ends.
Zero logs
Zero-logs environments
Prompts, files and outputs are not retained once a job completes. There is no transcript of what your agents saw or said.
Encrypted
Encrypted data checkouts
Data is encrypted before it reaches the swarm, stays encrypted in transit and at rest, and checks out to you the same way.
Minimal
Only what billing needs
We record how many agent-hours a job used. Not what the job was about.
Authorised
Authorised work only
Security work runs only against systems you own or are authorised to test, inside a scope you sign before the swarm starts.
04 · GPU rental
Rent the racks, on your terms.
Dedicated Blackwell and Rubin GPUs on fixed agreements, from part of a rack up to full NVL72 racks, and clusters of up to four racks.
6 months
12 months
18 months
| Configuration | GPUs | Generations | Good for |
|---|---|---|---|
| Part rack | A share of an NVL72 | Blackwell · Rubin | Steady inference and fine-tuning |
| Full rack | 72 in one NVLink domain | Blackwell · Rubin | Large models and long training runs |
| Cluster | 2–4 racks · up to 288 | Blackwell · Rubin | Frontier-scale training and big swarms |
Every agreement includes the same zero-logs environment and encrypted data checkouts as the swarm. Ask us about availability.
05 · Lucius
Brief it like a general.
Lucius is the voice-native orchestrator behind every swarm. Say what you need in plain speech. It plans the work, sizes the swarm, runs it and checks the results before they reach you.
YouLucius, test our web app. We're authorised. Scope is app.example.com.
LuciusScope confirmed. I'll run 400 agents for two hours: 800 agent-hours. Findings with evidence and fixes by two o'clock.
YouKeep watch on it afterwards.
LuciusShield is on. I'll tell you the moment anything changes.
Skill
Recon · authorised swarm testing
Thousands of agents probe your web apps, APIs and cloud within the scope you sign. Every finding comes with evidence and a fix.
Skill
Shield · swarm defence
The same swarm turned around: it watches, triages and helps you patch, around the clock.
Skill
Your own
Bring your tools and workflows. Lucius plugs them in as skills and puts a swarm behind them.
06 · Questions
Good questions, short answers.
What exactly is an agent-hour?
One AI agent working for one hour. A job that runs 200 agents for three hours uses 600 agent-hours. That's the whole bill for the work.
What do you keep from my jobs?
Only what billing needs: how many agent-hours a job used. Prompts, files and outputs are not retained after a job completes, and each job's environment is removed when it ends.
How is my data protected?
Through encrypted data checkouts. Data is encrypted before it reaches the swarm, stays encrypted in transit and at rest, and is returned to you encrypted.
Can I rent GPUs without using the swarm?
Yes. We offer 6, 12 and 18-month agreements on Blackwell and Rubin GPUs, from part of a rack up to full NVL72 racks, and clusters of up to four racks.
Who can use Recon?
Only organisations testing systems they own or are authorised to test. Lucius confirms the scope before any agent starts, and work stays inside it.
Where is aillusion based?
aillusion is incorporated in the United Kingdom and in Dubai.
07 · Early access
Join the waitlist.
Tell us what you'd run with a swarm, or which racks and terms you need. Email us and we'll be in touch.
l@aillusion.ai