Owners set the limits.
Contributors choose eligible devices, availability, and resource limits. They can pause participation and retain control of their personal files and hardware.
APERION TECHNOLOGY · COMMUNITY COMPUTE
Aperion is designed to coordinate distributed computing capacity. Community Compute gives compatible computers and servers a way to do useful, approved work when their owners choose to contribute.
Our direction: reuse suitable existing capacity first and reduce dependence on building more centralized data-center capacity.
See how it works
HOW COMMUNITY COMPUTE WORKS
Community Compute is a planned Aperion service. The network design connects voluntary contributors, agents, and jobs that fit the available hardware and privacy requirements.
Contributors choose eligible devices, availability, and resource limits. They can pause participation and retain control of their personal files and hardware.
The coordinator considers hardware capability, availability, cost, and data sensitivity. A job can use a community host, Aperion capacity, or a specialized service when that is the better fit.
Accepted work is measured and accounted for. Contributors receive clear records for the capacity they provide, and customers can see the usage they approve.
Use eligible capacity on devices you own to help your own agent. Pairing your desktop for private work is a separate choice from offering it to other people.
Offer eligible spare capacity for approved community jobs. Contribution requires its own consent, workload boundaries and participation terms; it does not open personal files or private agent memory.
The participation design calls for explicit opt-in, schedules, CPU/GPU and memory limits, storage quotas, power and thermal limits, and a clear pause control. A device must stay useful to its owner. No contribution should begin merely because an embodiment is connected.
Route jobs by capability, available memory, privacy, cost and latency. Independent tasks may fit separate machines; splitting one model across machines adds synchronization and transfer costs. More devices do not guarantee a faster response or form an unlimited pool of fast memory.
A computer contributing processing capacity does not, by that fact alone, grant access to its owner’s files, an Agent Key, or another person’s private memories. The actual data needed for a job and its protections must be designed for that job.
EXPLORE THE TECHNOLOGY
Powerful intelligence doesn't always require powerful hardware. Sometimes it requires a smarter way to use it.
Explore model compression, byte-level representations, specialized models and intelligent routing through six sections of detailed illustrated chapters.
Explore the science →WHY REUSE MATTERS
AI infrastructure has a physical footprint: computers and data centers use electricity, cooling can use water, and new equipment requires materials and manufacturing. How large each impact is depends on where and how the work runs.
Using compatible hardware that already exists can improve utilization and may reduce pressure to build new facilities for every increase in demand.
Distributed machines still draw power. Aperion’s approach needs to compare the energy used by a job—including idle overhead and networking—with an appropriate alternative.
Some facilities use water for cooling, and electricity generation can also have a water footprint. Local climate, cooling design, and power sources all affect the result.
Reuse only works as a community model when the costs are visible and contributors are paid fairly for accepted work. Lower customer cost is a goal to test, not a guaranteed outcome.
Aperion will need to report energy per job, cooling and water data where available, networking overhead, equipment life, cost, and contributor compensation. Environmental savings are not yet demonstrated for the planned network, so we will not present them as established results.
Aperion Labs is the organization behind the family. Aperion is the technology platform. Eclipse is the product family for personal agents and creative work.