Pre-seed · October 2026 · Confidential

Every device is a worker. Every job is a wage. Universal high income.

AI inference on the devices people already own, routed by SLA and paid per job on its own chain.

0Bsmartphones in use

~$0fee per settlement

0%of every job to the device

< 300 msp95 · realtime tier

UHI · Pre-seed · Confidential

Pre-seed · October 2026 · Confidential

Every device is a worker. Every job is a wage. Universal high income.

AI inference on the devices people already own, routed by SLA and paid per job on its own chain.

6.9Bsmartphones in use

~$0fee per settlement

70%of every job to the device

< 300 msp95 · realtime tier

UHI · Pre-seed · Confidential

02 · The problem

The compute exists. The demand exists. Nothing connects them with a guarantee an enterprise can sign.

01 · The billInference is now two-thirds of all AI compute. The data-centre build-out behind it passes $400B a year in 2026 and $1.7T by 2030. GPU capacity is rationed and queued.

$0B+AI data-centre capex · 2026

02 · The work is smallWhat products actually run is speech in, speech out, summarise, translate, embed and small-model chat. At 1–3B parameters those match cloud quality, and they fit on a phone's NPU.

0 TOPSNPU in every flagship phone

03 · The hardware is idle6.9 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.

0%of the day, idle and unpaid

The gap is not compute and it is not demand. It is the contract. No one sells inference from the devices in people's pockets with a latency, redundancy and region an enterprise can sign, and no one pays those devices per job.

UHI · Pre-seed

03 · The insight · why now

The supply side is proven. The demand side is proven. The corner where they meet is empty.

what is sold · raw hours → inference with an SLA

Hyperscaler inference APIsOpenAI · Google · AWS · Azure

An SLA, from racks they own. Priced for the margin of a data centre. No consumer device will ever be in their fleet.

UHIthe empty corner

SLA-backed inference from consumer devices, sold to enterprises, settled per job on our own chain. Nobody is here.

GPU DePINAkash · io.net · Render · Bittensor

Proved people will lend hardware for tokens: $7–10B combined cap. But they sell raw GPU hours to developers, with no SLA and no enterprise buyer.

On-device SDKsApple Core ML · Qualcomm AI Hub

Run a model on one device, for that device's owner. No market, no routing, no income.

where the compute lives · racks → the devices people already own

Why now · four things became true at once

40+ TOPSNPUs ship in every flagship phone and every new laptop. 10B-parameter models run locally by 2028.

1–3BThe models that matter are small. TTS, STT, summarisation, translation and embeddings at 1–3B parameters match cloud quality on a phone.

$7–10BDePIN proved the supply side. io.net has settled $20M+ on-chain; Render carries a ~$1.2B cap. All of it sells raw GPUs to developers.

1st jobWe own the first workload. QuickDial AgentBox runs our own STT, LLM and TTS on commodity CPUs at a measured $0.0015 of infrastructure per call-minute. UHI's first customer is us.

Market caps and on-chain revenue: public dashboards, September–October 2026, see SOURCES.md. The quadrant places each category by what it sells and where its compute lives.

UHI · Pre-seed

04 · The product · the loop

A job pours in with an SLA. The result returns in milliseconds. The wage settles in seconds.

1 · SubmitA requester sends a job with an SLA tier and a price ceiling, encrypted to the chosen devices.

2 · RouteThe router picks devices holding the model warm, inside the latency budget, with the tier's reputation, and dispatches to n.

3 · RunDevices run the model locally on NPU, GPU or CPU and return the result with a proof: output hash, timing, attestation.

4 · ValidateHashes agree across the redundant devices, canaries pass, timing is inside the SLA.

5 · SettleOne block splits the payout to device, model provider, validator and router. The requester had the result already.

the requestworkload tts model pocket-tts-en · 0.3B input 400 chars sla.p95 300 ms redundancy 3 region us-east · metro privacy on-device · encrypted to n ceiling $0.0025

the receipt · block #1,042,118 · 1.6 sfirst result 212 ms · 3 / 3 agree price $0.001920 device 70% $0.001344 model 15% $0.000288 validators 10% $0.000192 router 3% $0.000058 treasury 2% $0.000038

UHI · Pre-seed

05 · The network · five parties, one ring

Five parties show up for five reasons. The job travels the ring, the wage rises along the spokes, and the work votes.

Requestersenterprises · developers · consumer apps

Buy inference with an SLA, 4–7× under cloud list, with the latency on the contract and data in-region or on the user's own device.

Device operatorsphones · laptops · rigs · consoles · watches · glasses

Income from hardware already paid for: a phone pays its plan, a rig pays its electricity and then some. They set a budget; the node never exceeds it.

Model providersquantised · distilled · device-tuned

Distribution to millions of devices without running infrastructure, and a royalty on every job that rises with sustained SLA and volume.

Validatorsstaked operators running the chain

Block rewards and a fee on every job for verifying the work (redundant execution, hashes, canaries, attestation); slashed for approving bad work.

RoutersUHI Labs at launch · then staked routers

A routing fee for matching each job to the devices that meet its SLA; the protocol rotates traffic toward the best attainment.

Treasurygoverned on-chain

Model-provider grants, device-onboarding incentives and audits, governed on-chain. USD converts to UHI at settlement, so the supply side earns in the native coin.

UHI · Pre-seed

06 · For enterprises · the SLA

Three tiers an enterprise can sign. Four to seven times cheaper than cloud list, with the latency on the contract.

Tier	p95 latency	Redundancy	Devices	Region	TTS · per 1M chars
Batch	minutes	1× + spot check	any, including watches and glasses when charging	any	$3.00
Standard	< 1 s	2×	phones, laptops, rigs · reputation ≥ 0.8	continent	$4.80
Realtime	< 300 ms	3×	staked rigs, desktops, laptops on mains · reputation ≥ 0.95	metro	$7.50

UHI price = 15% of cloud list × the tier multiplier (Batch 1.0× · Standard 1.6× · Realtime 2.5×): 6.7× under list at Batch, 4.2× at Standard, 2.7× at Realtime. Indicative; from the model. Every tier states p95, redundancy and region on the contract and the ledger records attainment per job.

Cloud list vs UHI Standard tier · list mid-points, indicative

Text to speech · per 1M characters

Cloud neural TTS · OpenAI, Google, Azure, Amazon

$15 – 30

UHI · Standard

$4.80

A voice minute · STT + TTS + 0.6 LLM turn + 0.2 summary

Cloud, at list

$0.0185

UHI · Standard

$0.0044

Small-LLM chat · per 1M tokens

Cloud small-LLM APIs · 4o-mini, Flash, 8B-class

$0.10 – 0.60

UHI · Standard

$0.096

4 – 7×cheaper than cloud listStandard 4.2× · Batch 6.7× · see SOURCES.md

on contractp95 · redundancy · regionattainment recorded per job on the ledger

in-regionor on the user's own devicejobs encrypted to the n devices that run them

UHI · Pre-seed

07 · Unit economics · one job, in cents

One TTS job, 400 characters. The requester pays a fifth of a cent. The device keeps 70%.

One job · Standard tier · $0.00192 · splits into five wages, in one block

Device operatorruns the model

70%$0.001344

Model providerroyalty

15%$0.000288

Validatorsverify + settle

10%$0.000192

Routermatches the SLA

3%$0.000058

Treasurygrants · audits

2%$0.000038

Requester pays · Standard100%$0.001920

Per million jobs$1,920 → $1,344 · $288 · $192 · $58 · $38

A busy device at reference utilisation · per day, net of power · indicative

Phone16 h online · busy 20%

~$2

Laptop12 h online · busy 20%

~$3

Gaming rig · desktop18 h online · busy 30% · 350 W paid for

~$15

Console10 h online · busy 25%

~$2

Watch · glassescharging hours · Batch tier only

cents

The protocol's takeTreasury 2% (a quarter burned) + routing 3% while UHI Labs runs the routers, + Labs' royalties on the launch models. Declines as routers decentralise.

5 – 6%of job value · M36

$4.80UHI · per 1M characters of TTS · Standard= 15% of list × 1.6 · $3.00 at Batch

$15 – 30cloud TTS · per 1M characterspublished list prices, October 2026

0×cheaper, and the device is paidevery figure from uhi-model · assumptions in SOURCES.md

UHI · Pre-seed

08 · The chain · a Substrate L1

Settlement is the product. A job that pays $0.0003 cannot carry a $0.02 gas fee, so we own the ledger.

pallet_jobsThe market: submission, price ceilings, encryption to the chosen devices, batching.

pallet_slaTier registry and attainment: p95, redundancy and region recorded per job.

pallet_reputationPer-device score from measured latency, agreement rate and uptime; gates each tier.

pallet_modelsProvider registry, device specs, SLA envelopes and the royalty curve.

pallet_settlementThe split: 70 / 15 / 10 / 3 / 2, USD-to-UHI conversion, the fee burn.

pallet_validationSampled redundant execution, output hashing, canaries, attestation, slashing.

Blocks settle while the requester is still reading the result

1 – 2 sblock time · parallel job batching

~$0weight-based fee · paid by the protocol

forklessruntime upgrades by governance

bridgesPolkadot · Ethereum · later, for liquidity

Substrate gives a sovereign L1 with custom pallets, weight-based fees we set near zero for job settlement, forkless upgrades so the economics change by governance rather than by hard fork, and a path to bridges without moving settlement off our chain.

UHI · Pre-seed

09 · The token and the governance

One token: settlement, stake, loyalty and the vote. Demand is priced in dollars, so a customer never thinks about it.

SettlementEvery job settles in UHI; USD converts at settlement.

StakeBehind validators and realtime-tier devices. Slashed for bad work.

LoyaltyA provider's royalty share rises with sustained SLA and volume.

VoteConviction voting on three tracks, time-locked.

The market track · who votes on the price of the work, and by what

The people who do the work govern the price of the work.

Track	Decides	Who votes
Protocol	runtime upgrades, validator set size, slashing rules	staked validators and token holders
Market	SLA tier definitions, the payout split, the routing-fee cap, the 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

1.0Bgenesis supply · UHI

0.08×emissions ÷ job fees · M36usage, not emissions, pays

25%of the treasury's 2% burned

7%of supply staked · M36

OpenGov-style, all on-chain, time-locked conviction voting. Supply-side emissions decline on a four-year schedule from mainnet; at M36 they are 0.08× job fees. Token figures from uhi-model (Dashboard).

UHI · Pre-seed

10 · Market · USA first

Inference is a $283B spend this year. We need 0.29% of the small-model slice bought by API to hit the year-3 plan.

$283BTAM · AI inference compute · 2026

$17.5BSAM · US + EU · 2029

$51MSOM · year 3
ARR

01 · Top-down · the spend

$400–450BAI data-centre capex · 2026Deloitte · verify

⅔of AI compute is inference→ TAM $283B

$25.8–30.9Bedge AI · 2026 → $66B · 2030GM Insights · Mordor · Roots · corroboration

$1.7Tdata-centre capex · 2030Dell'Oro · verify

02 · Bottom-up · the supply we route

6.9Bsmartphones in use

1.5Blaptops and tablets

1.9BPC gamers · 650M console

< 20%AI-capable today · end 2026the ramp is in front of us

03 · SAM, bottom-up · the chain of assumptions · from uhi-model, every line to verify

TTS$4.5B+STT$6.0B+translation$2.5B+small-LLM · summarise$8.0B+embeddings$1.2B=API markets · 2026$22.2B× 62% US+EU × 70% by APISAM 2026$9.6B× 22% CAGRSAM 2029$17.5B× 0.29%SOM · M36 ARR$51M

The USA is 60% of job value at M36 and is where the rigs, the gamers and the voice-AI buyers are. DePIN proved supply shows up for tokens: $7–10B combined cap, io.net $20M+ settled on-chain, Akash ~$4.3M annualised.

TAM is the inference share (⅔) of 2026 AI data-centre capex ($425B mid-point), a proxy for what is spent running models. SAM is five small-model API workload markets × the US + EU share × the API-purchasable share. SOM is the model's own year-3 ARR. Every figure and its flag is in SOURCES.md.

UHI · Pre-seed

11 · Competition · who sells what, to whom

Four kinds of competitor. None sells SLA-backed inference from consumer devices, and none pays the device per job.

4 – 7×cheaper than cloud list15% of list × tier · Standard 4.2× · Batch 6.7×

secondsto settle every job1–2 s blocks · ~$0 fee · our own L1

70%of every job to the devicethe hardware's owner is paid, not the data centre

Who	What they sell	Who buys	Enterprise SLA	Consumer devices	Settlement	TTS · per 1M chars
Hyperscaler inference APIs OpenAI · Google · AWS · Azure	Managed inference from their racks	Enterprises, developers	yes	no	monthly invoice	$15 – 30
GPU DePIN Akash · io.net · Render	Raw GPU hours; bring your own model	Developers, studios	no	rigs and racks, not phones	token, per hour	—
Bittensor subnets	Model mining; emissions for ranked outputs	Miners, validators	no	no	TAO emissions	—
On-device SDKs Apple Core ML · Qualcomm AI Hub	A runtime for one device	App developers	no	one device, no market	none	free, your own device
UHI universal high income	SLA-backed inference from consumer devices	Enterprises, developers, apps	three tiers	phones · laptops · rigs · consoles	own L1 · per job · seconds	$4.80

Published list prices, October 2026; cloud TTS is the range across OpenAI, Google, Azure and Amazon neural voices. DePIN and Bittensor sell compute or emissions, not a priced inference job, so no per-character price exists.

UHI · Pre-seed

12 · Roadmap · 36 months

Testnet on our own workload at six months. A million devices on the waitlist and $50M of jobs a year at thirty-six.

M6

Testnet

Node app on gaming rigs and phones
First workload: our own voice AI

1kregistered—active · testnet

—internal jobs

M12

Mainnet

Three design partners · Standard tier
USD billing · token generation

10kregistered~500active · earning

$0.6MARR · job value × 12

M18

Realtime tier

Staked rigs and desktops, metro routing
Model-provider program, treasury grants

50kregistered~3.5kactive · earning

$5MARR · job value × 12

M24

Routers decentralised

Staked routers compete on attainment
Bridges for liquidity

200kregistered~8kactive · earning

$13MARR · job value × 12

M36

USA → global

Fleets, telcos and OEMs as operators
Every small-model workload class

1Mregistered~50kactive · earning

$50MARR · job value × 12

Admission is paced to demand. Every active device earns at the reference rate (~$2 a day a phone, ~$15 a rig); the waitlist is the growth engine behind it; demand is the lever. ARR and the active count are both linear in requester volume and in the discount to cloud.

ARR plan · gross job value × 12 · $M · from uhi-modeljobs served a day · 1.3M at M12 · 21.8M at M24 · 86.1M at M36

M6M12M18M24M36

UHI · Pre-seed

13 · Traction and the raise

The first workload already runs on commodity compute at a quarter of a cent a minute. We are raising $1.5M to put it on the devices.

01 · The workload

Our own STT, LLM and TTS, running on commodity CPUs

QuickDial AgentBox (USA) and Vartalaap (India) run our own speech, language and voice models in production at a measured 0.25¢ a conversation minute. UHI's first requester is us, so the network has paying jobs on day one.

0.25¢per minute · measured cost

$0.0044UHI price per voice minute

~250 msendpoint to first audio

02 · Built, and next

The brand and the node app are designed; the Substrate runtime is specified: six pallets, the split, the tiers, the validation sampling, the governance tracks; the model is built (uhi-model, 36 months, zero formula errors). Next: testnet at M6 on rigs and phones with our own TTS and STT jobs; three design partners from the buyers we already sell to.

03 · Plainly

Ten people by M12; year-1 opex $1.33M; 13 months of runway on the pre-seed alone, so the seed (assumed $6M at M13) closes before then. The protocol's take is 5–6% of job value and declines as routers decentralise, so the company is not EBITDA-positive inside 36 months (year 3: −$1.8M). The value accrues to the token and the treasury, as with every DePIN comparable.

The raise · pre-seed · SAFE18 months of plan

$1.5Mpre-seed

$4Mpost-money

10%Alliance's $400k

Use of funds

45%Engineering · runtime, node apps, router$675k 20%Model-provider grants · device incentives$300k 15%Design partners · go-to-market, USA$225k 10%Security audits · runtime and pallets$150k 10%Operations$150k

The first cheque

Alliance: $400k at $4M post-money via SAFE with a 1:1 token side letter, plus $400k follow-on at seed. The balance of the round from angels and funds who back infrastructure with its own settlement layer.

Book 30 minutes roshan.purush@gmail.com

UHI · Pre-seed · Confidential

14 · Team

Two founders. One built the models and the platform under them; the other has shipped products inside the banks that will buy the inference.

Purushottam (Puru) ChaudharyFounder & CEO · AI inference engineer

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, 1,400 pages on inference measured on real hardware. Three companies founded.

LinkedIn The book QuickDial

Ankur KapoorCo-founder · Product & Operations

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.

QuickDial

3 / 3speech · language · voice models, ours

0.25¢the minute QuickDial reached on CPUs

1,400pages on inference, measured

3companies founded before this one

UHI · Pre-seed

15 · A message from the founder

Intelligence is the new electricity. Today a handful of companies own the power plant. The grid is already in everyone's pocket.

Every flagship phone ships with a chip that can run the models most products actually need. Every gaming rig in a bedroom outruns the server a voice company rented five years ago. These machines sit idle 90% of the day, and the people who paid for them are paid nothing while the same work is sold back to them from a data centre.

UHI connects the two with the only thing that was missing: a guarantee an enterprise can sign, and a ledger that pays the device the moment the work is checked. The requester gets a cheaper, closer, more private job. The person whose hardware ran the intelligence gets 70% of the wage. The models are small, the fees are near zero, and the price of the work is governed by the people who do it.

We start with a job. We end with an income.

Puru ChaudharyFounder & CEO, UHI · uhi.network

UHI · Pre-seed · Confidential