An open network of home computers for AI
Idle computers work for language models. Owners get paid.
Lumia takes batch jobs from LLM companies, routes them to vetted home nodes and pays owners for accepted results. We start with a 14-day pilot, and the stop conditions are written down in advance.
Why now
Demand for compute is growing, and idle home capacity is plentiful
Home GPUs already earn money
Paid text inference on consumer graphics cards exists today. This confirms the technical model and shows a real competitor.
Providers sell deferred processing
Major providers sell batch processing at about half price for customers willing to wait. Jobs where timing is not critical are already separated from urgent ones.
Hardware got more accessible
Desktop accelerators, Macs and compact PCs can run small open models. For narrow tasks that is enough.
Clients
Who pays: nine demand segments
There are no signed contracts yet. Below are all the channels we are testing and the phase in which each one joins. The base of demand is continuous, non-urgent flows of open data.
How demand stacks up
Base: steady flows of open data where timing is not critical. They keep nodes busy during ordinary hours. Peaks: large batches with a deadline on the Priority tier. Top-up: private users in phase 2; their demand is uneven, so it is not the base.
LLM providers and inference platforms
What they buy: queues of deferred processing on open models.
Why it fits: providers already sell batch APIs with a deadline, and their backlog of checks and synthetic jobs is waiting for spare capacity.
Status: hypothesis, a pilot goal.
Labs and startups on open models
What they buy: synthetic data generation, distillation, reinforcement learning rollouts.
Why it fits: huge volumes, flexible deadlines, open data.
Status: hypothesis.
Researchers and universities
What they buy: model evaluation on open benchmarks, ablations.
Why it fits: recurring runs, small budgets, quick decisions.
Status: hypothesis.
Data and labeling companies
What they buy: pre-labeling, filtering and deduplication of open text corpora.
Why it fits: a constant stream of text, part of the work needs no human.
Status: hypothesis.
Public-information monitoring
What they buy: sentiment analysis, classification of publications, summaries of open sources (reputation, OSINT).
Why it fits: the stream runs around the clock and an instant answer is not needed.
Status: the founders' own workload (SERM.online, CGS), being assessed. This tests the mechanics; it is not a market signal.
Content agencies and e-commerce
What they buy: translation, product descriptions, catalog classification on public data.
Why it fits: bulk jobs with no confidential data.
Status: hypothesis.
Enterprise clients on the corporate plan
What they buy: reserved capacity windows for their own jobs: Lumia resells qualified capacity.
Why it fits: a steady payment for the reservation. Mid-sensitivity data needs trusted nodes.
Status: after trusted nodes launch.
Compute marketplaces and aggregators
What they buy: resell our capacity to their own customers: a partner channel.
Why it fits: a ready audience of buyers.
Status: hypothesis. They are also possible competitors.
Private users
What they buy: node owners spend earned credits on AI, including models that do not fit on their own PC. Ordinary users buy an interactive mode.
Why it fits: adds demand and keeps node owners engaged.
Status: after phase 1: demand is uneven, so it is a top-up, not the base.
Phase 1: the pilot on open and synthetic data. Phase 2: trusted nodes, the corporate plan, private users. "Hypothesis" means there is no letter of intent yet: the goal of the first 14 days is to get a job profile and the price of the current solution from one buyer.
How it works
One order: one model, one data profile, one deadline
- 1OrderThe buyer submits a batch through an API: model, data, time window, budget cap, acceptance criteria.
- 2MatchingThe scheduler distributes independent chunks to qualified nodes.
- 3ExecutionAn agent on the owner's computer runs within the hours and load limits the owner sets.
- 4VerificationControl examples and sampled re-runs. If a node drops, its chunk goes to another.
- 5SettlementThe buyer receives accepted results and a report; the owner is paid for accepted work.
Tiers
Three speeds: pay for the deadline and node class
Speed comes from resources and priority, not from artificial slowdown. The main price lever is flexibility on the deadline.
ECONOMY
Up to 24 hours
- Idle nodes of classes C and D
- No reservation; jobs can be preempted
- Price goal: in line with consumer-GPU market rates
STANDARD
2–6 hours
- Class B nodes and above
- Automatic retry if a node fails
- Baseline modeled price
PRIORITY
Hard window, in minutes
- Best nodes, latency limits
- Reserved pool, job splitting, duplication on two nodes
- Higher price for the deadline guarantee
Try it yourself
The project mechanics in three working tools
The calculator is calibrated to the concept's financial model. The simulator shows the scheduler's logic; its numbers are illustrative. The computer check runs in your browser.
Other direct costs cover result verification, re-runs, infrastructure and payments. Defaults are tuned so the base case gives $1,161 revenue and $343 contribution on 20 devices, and the stress case gives −$139, as in the concept.
Contribution is before salaries, taxes and development. Owner electricity: $0.0615 per active hour. This is a model, not a forecast; the pilot will confirm the rates.
An illustration of the scheduler's logic. Node throughput and node-hour prices are notional: Economy $0.30, Standard $0.64, Priority $1.10. These rates are hypotheses to be tested in the pilot, not a price list.
A quick in-browser check: graphics card (WebGPU), performance, latency to the page's server. Nothing leaves your computer.
- Graphics card and environment
- Performance (short test)
- Network: latency to the server
This is a preliminary estimate. Full qualification runs in the agent: hardware inventory, a benchmark on the reference model, correctness checks, a network test and a 48–72 hour trial period.
Market
We need a small part of a large market
We are not after all of inference. Our addressable segment is batch jobs on open models and open data. Analyst estimates differ widely because the market definitions differ, so we give ranges.
Entire AI inference market, 2026
$118–137B
Context, not our segment. Forecast to 2034: about $313B.
Synthetic data, 2026
$0.7–1.0B
Growth of roughly 25–37% a year in most estimates. The segment closest to us.
Data annotation and labeling, 2026
$2–6B
From tools-only estimates to tools plus services. Part of the work is shifting to models.
Our working target: about $3–7B a year
This is synthetic data plus labeling. The estimates overlap and use different definitions, so treat it as an order of magnitude, not an exact figure. $1M in annual revenue at the calculator's base assumptions needs about 1,400 devices, roughly 0.01–0.03% of this segment.
Demand for deferred processing is priced in
Google, OpenAI and Anthropic sell batch processing at about half the regular price. Together AI advertises discounts of up to 50%.
Decentralized GPU networks are growing
Annualized protocol revenue is estimated above $200M. These are secondary, unaudited sources: use as a rough guide, not as fact.
Competitors exist
SaladCloud, GPU rental marketplaces, decentralized networks. We compete on the price of an accepted result by a deadline, not on hardware.
Sources checked on November 21, 2026. Market definitions differ between research firms; figures are given as ranges.
Risks
The main objections and our answers
| Risk | How we address it |
|---|---|
| No confirmed buyer yet | We proceed only with a signed paid order. This is the gate for stage two. |
| Data privacy | Home nodes receive only public and explicitly approved synthetic data. Closed data comes later through trusted nodes and confidential computing. |
| Closed models | We work with open-weight models only. |
| Price above market | Tiers by deadline and node class, with the price of an accepted result compared on the buyer's own batch. |
| Unreliable nodes | Five-stage qualification, control jobs, re-runs, duplication. |
| Node utilization | Demand rests on steady batch orders. A consumer mode is added later as a top-up. |
| Regulation | No token at launch. Lawyers review the structure and jurisdictions before entering a country. |
Plan and stop conditions
90 days with exit points
Continue if a buyer with a budget and a clear path to payment exists.
In preparation
Continue if a paid pilot is signed: price, window, acceptance terms.
Continue at ≥99% accepted on time and ≥25% margin after direct costs.
Scale only the node classes with proven economics.
Team
Founders and contributors
Confidence Guard Solutions
Founding company. Reputation management, communications, partner negotiations. Office in Dallas, USA (per the company's website).
SERM.online
Founding company. Reputation consulting by "Mikhaylin & Partners"; the website states experience since 1993 (per the company's website).
Head of BD
Open role: negotiations with compute buyers.
Inference engineer
Open role: testbed, agent, node qualification.
The ask and the investor list
Approve a 14-day stage with a $5–12K cap
After that we continue only on results: a signed paid order, 99% on time, margin of 25% or more. The full 90-day range of $40–92K remains a guide if the stages are passed.