6.9 billion smartphones, 1.5 billion laptops and tablets, hundreds of millions of gaming rigs and consoles sit idle 90% of the day. Their owners paid for hardware that earns nothing. The work most products actually run, speech in, speech out, summarising, translating, embedding, fits on a phone's NPU.
A requester submits a job with a tier and a price ceiling. The router scores devices by capability, measured latency, reputation, price and redundancy, then dispatches to n of them, encrypted to those devices only. The first result back is the one the requester gets.
Batch in minutes on anything that is charging, watches and glasses included. Standard under a second, twice redundant, on phones, laptops and rigs with reputation ≥ 0.8. Realtime under 300 ms, three times redundant, on staked rigs, desktops and laptops on mains.
Devices return the result and a proof: output hash, timing, an attestation where the hardware has one. Validators check that hashes agree across the redundant devices and timing is inside the SLA. The ledger settles the job in one block and splits the payout: 70% to the device, 15% to the model provider, 10% to validators, 3% to the router, 2% to the treasury.
A job that pays $0.0004 cannot carry a $0.02 gas fee. UHI is designed as a sovereign Substrate L1, launched once job volume needs it: one block every 1–2 seconds, settlement fees near zero and paid by the protocol, forkless upgrades so the economics change by vote, not by hard fork, and five parties who govern the price of the work.
A requester submits a job with an SLA. The router dispatches it to devices that already hold the model warm. The devices return the result and a proof. Validators approve. The ledger settles the split in one block, in seconds. Then the next job.
INDICATIVE. Price = 15% of cloud list × the job size × a tier factor (Batch 0.6 · Standard 1 · Realtime 1.8). The device share is split across the n devices that returned a valid result. Settled in one block, fee paid by the protocol.
Enterprises, developers, consumer apps; retail users via apps.
What they doSubmit AI jobs (TTS, STT, summarise, translate, embed, vision, small-LLM) with an SLA: latency, throughput, redundancy, region, privacy. Pay in dollars or UHI.
Why they show upAbout 5× cheaper than cloud GPU for small-model work; elastic capacity with no queue; data can stay in-region or on the user's own device.
Anyone with a phone, tablet, laptop, desktop, gaming rig, console, watch or glasses; later fleets, telcos, OEMs.
What they doInstall the UHI node, set a budget (battery floor, thermal cap, hours, data), keep the device online. Earn per job.
Why they show upIncome from hardware already paid for. A phone pays its own plan; a gaming rig pays its own electricity and then some.
70% of every jobML teams who make models small and efficient: quantised, distilled, device-tuned.
What they doPublish a model with a device spec (RAM, NPU TOPS, OS) and an SLA envelope (p95 latency, tokens/s per device class). Earn a royalty on every job that runs their model.
Why they show upDistribution to millions of devices without running infrastructure, and recurring income. Loyalty: a provider's royalty share rises with sustained SLA performance and job volume.
15% of every jobStaked operators running the chain.
What they doVerify work (sampled redundant execution, output hashing, canary jobs, attestation where the device has a TEE), approve payouts, produce blocks.
Why they show upBlock rewards and a validation fee on every settled job. Slashed for approving bad work.
10% of every jobThe matching layer; UHI Labs at launch, progressively decentralised to staked routers.
What they doScore devices by capability, measured latency, reputation, price and redundancy; dispatch; retry; meet the SLA.
Why they show upA routing fee. Routers compete on SLA attainment; the protocol rotates traffic toward the best.
3% of every jobGoverned on-chain. Funds model-provider grants, device-onboarding incentives and audits.
Pick a tier, set a price ceiling, keep data in-region or on the user's own device. You pay in dollars per job; the network earns in UHI. Elastic capacity with no queue, because the fleet is already switched on.
| Tier | p95 latency | Redundancy | Devices | Region | Price |
|---|---|---|---|---|---|
| Batch | minutes | 1× + spot check | any, including watches and glasses when charging | any | lowest |
| Standard | < 1 s | 2× | phones, laptops, rigs with reputation ≥ 0.8 | continent | mid |
| Realtime | < 300 ms | 3× | staked rigs, desktops, laptops on mains, reputation ≥ 0.95 | metro | highest |
INDICATIVE. Cloud = published pay-as-you-go list prices for neural TTS, streaming STT and a small hosted LLM, Oct 2026. UHI = 15% of cloud list, the launch target for the Standard tier; the Market track sets the final number.
Install the node, pick a budget, keep the device online. Every job your device runs is paid in seconds. Try the estimate for a busy device at reference utilisation: every assumption it uses is in the table under it.
A busy device at reference utilisation, net of power. Indicative.
| Class | Online | Busy | Jobs / busy h | $ / job | Power / day | Tiers |
|---|
INDICATIVE. A device only earns while it is busy, so admission is paced to demand: devices register, sit on a waitlist and are admitted as paid jobs grow, holding active devices near the reference utilisation (a phone busy ~20% of its online hours, a rig ~30%). $/job is the operator's 70% of a typical job in that class at 15% of cloud list; power at $0.17/kWh. Launch proposals, set by the Market track. Not a promise of income. Derived in uhi-net/uhi-model.
Distribution to millions of devices without running infrastructure. Quantise, distil, tune for a device class, publish with a spec and an SLA envelope, and the router sends your model the jobs it can meet.
Upload the weights and a model card. Declare the device spec it needs: RAM, NPU TOPS, OS.
Measured on the network's benchmark devices: p95 latency and tokens/s per device class. That envelope decides which tiers your model is eligible for.
15% of every job that runs your model, settled in UHI in the same block as the device's share. Hold it or cash out.
The share rises with sustained SLA attainment and job volume and falls back if attainment drops. The curve is a launch proposal set by the Market track, where providers vote weighted by served jobs.
Grants from the treasury. 2% of every job funds grants for quantising and distilling models to device spec, audits, and the onboarding of the first providers. Apply with a model card and a benchmark on one device class. A provider council holds a veto on provider grants.
A job that pays $0.0004 cannot carry a $0.02 gas fee, and a network doing thousands of jobs a second cannot wait for someone else's block space. So the ledger is ours: six pallets, 1–2 second blocks, fees near zero and paid by the protocol, upgrades without a hard fork.
jobsThe market: a job, its SLA tier, its price ceiling, its encrypted dispatch.
slaThe tier registry and attainment: who met p95, who did not, per device and per router.
reputationA score per device from attainment, agreement and uptime; the gate for Standard and Realtime.
modelsThe provider registry, device specs, SLA envelopes and royalties.
settlementThe split, 70 / 15 / 10 / 3 / 2, settled in one block; USD converted to UHI at settlement.
validationSampling, canaries, attestation checks, and slashing for approving bad work.
A mock of the explorer, running in your browser. Testnet at M6.
OpenGov-style, time-locked conviction voting on every track. The token is the vote on protocol and treasury matters. On the market track the vote is weighted by what each party actually did on the network.
| Track | Decides | Who votes |
|---|---|---|
| Protocol | runtime upgrades, validator set size, slashing rules | staked validators and token holders |
| Market | SLA tier definitions, payout split, routing-fee cap, model-provider loyalty curve | every party, weighted: operators by earned jobs, providers by served jobs, requesters by spend, validators by stake |
| Treasury | grants to model providers, onboarding incentives, audits | token holders, with a provider council veto on provider grants |
UHI is the unit of settlement, the stake behind validators and realtime devices, the loyalty asset for providers, and the vote. A share of every job fee is burned; usage, not speculation, sets the floor.
The first customer is us: QuickDial's AgentBox runs our own speech, language and voice models on commodity CPUs, at a measured $0.0015 of infrastructure per call-minute. Two device counts, on purpose: registered is the waitlist, everyone who installed the node; active is the devices admission has let in because demand keeps them busy. Every active device earns at the reference rate; the waitlist is the growth engine. The numbers are the plan.
We already own the first workload. Between us: the speech, language and voice models and the platform under them, fifteen years each in regulated enterprise software, and a company built together before this one.
Built speech recognition, language and voice models and the platform under them at QuickDial AI (USA) and Vartalaap (India). Fifteen years of software for GE HealthCare, S&P Global, Bristol Myers Squibb and State Street. Author of The Inference Mechanic.
Fifteen years in financial-services technology with Tata Consultancy Services, contracted to GE Capital, Silicon Valley Bank and USAA, where he is a Product Manager today. Cornell Johnson MBA 2020; ITIL and Lean Six Sigma. Co-founder and investor at QuickDial AI.
Thirty minutes is enough to walk through the loop on your workload, the tier you need and the economics on your volumes. Investors and accelerator partners: the same call, with the plan.
With Puru Chaudhary, founder & CEO. Pick a slot and the invite lands in your calendar with a video link.
UHI.NETEvery year the industry spends hundreds of billions of dollars on data centres to run models that, for most products, would fit on the phone in your pocket. Speech in, speech out, a summary, a translation, an embedding: small models do this work well now, on chips already shipped to billions of people. That compute is paid for. It sits idle 90% of the day.
I have spent the last three years making inference cheap: training speech and language models and building the platform under them for QuickDial in the United States and Vartalaap in India. Those products run on commodity CPUs at a measured $0.0015 of infrastructure per call-minute, and every time I look at that bill, then at the idle fleet, I see the same gap. UHI closes it. An enterprise buys inference with an SLA it can sign. A phone, a laptop or a gaming rig does the work and is paid for every job, in seconds, on a chain designed for it because settlement is the product.
We are not a GPU marketplace and not a token with a roadmap. We are an inference provider whose data centre is everyone's pocket, and whose price is set by the people who do the work. Our first customer is us. The next three are design partners. After that, the fleet.