Idle compute, sellable by the token.
Phones, laptops, cars, drones — billions of NPUs sit idle 90% of the time. The HQ Compute Mesh is the global spot market: clients post inference jobs, workers claim leases, HQ schedules and prices, and every passing-attestation result settles per-token via the Loadit Settlement Grid. Compute is the new oil — and the user owns it.
The paradigm break
Compute is concentrating in 5 hyperscalers who capture ~95% of the margin. That isn't because hardware is scarce — every recent phone ships an NPU capable of ~50 INT8 TFLOPS, and there are billions of them. The blocker has always been scheduling, settlement, and trust.
HQ does scheduling (job ↔ worker matching with reputation weighting). Loadit + the Settlement Grid do settlement (per-token, atomic, with a verifiable receipt). Hylaq does trust (every device is a @handle/device-XX with attested capability and accumulating reputation). That triangle is what unlocks the marketplace.
What we replace
OpenAI / Anthropic / Google API margin
$60B+/yr inference revenue, ~70% margin
The hyperscaler API tax is real. The Mesh undercuts list prices by 30%+ at the floor, and approaches their cost as utilization rises. Their moat is distribution, not unit economics.
AWS / Azure / GCP GPU spot rental
$80B+/yr in cloud GPU revenue
Spot rentals require commitment and capacity planning. The Mesh is per-request: pay for the tokens you got, not the GPU-hours you reserved.
Vast.ai / RunPod / decentralized GPU markets
$2B+/yr and growing
They proved demand for marketplace compute. They lack identity (handle = device), settlement (Loadit/Grid), and a real reputation+audit layer. The Mesh has all three.
On-device inference frameworks (alone)
Free today — no economic loop
Apple Foundation Models / Gemini Nano / Llama on-device are great runtimes with zero economic flywheel. The Mesh is the flywheel: your device earns when other people's apps need a token.
Federated learning vendors
$1B+/yr in proprietary stacks
They optimize for one customer. The Mesh is open: any model, any client, any worker, with the Settlement Grid as the universal payment leg.
Telco edge compute (5G MEC)
$50B+/yr telco capex
Carriers built edge nodes nobody uses because there's no buyer protocol. The Mesh is the buyer protocol — telco nodes register as workers and auto-monetize.
Six devices. One marketplace.
Phone NPU at 3am
iPhone Pro on a charger overnight. Mesh worker registers PHONE_NPU + 50 INT8 TFLOPS. Earns ~$0.02 per 1k tokens served. ~$30/month passive income — paid for the phone itself in two years.
Laptop GPU on lunch
MacBook Pro (M-series) idle from 12–1pm. Workshop-lunch hour worldwide is a 60-minute compute glut. The Mesh schedulers route bursty inference there at LAPTOP_GPU rates.
Tesla in a parking lot
The Tesla NN compute is idle when parked. EDGE_NPU class. 100M+ vehicles globally x ~10 TFLOPS each = a planet-scale dormant cluster. The Mesh wakes it up.
Burst-pricing the AI launch
A new model goes viral. Hyperscaler queues spike for 6 hours. Mesh utilization climbs and the spot curve ramps; idle laptops join because the price is suddenly worth it. Self-balancing flash capacity.
DUAL-attestation for high-stakes jobs
A pharma client wants a bio-research inference deterministic to the bit. Mesh routes DUAL: two independent workers run it; matching outputHash pays both, divergence triggers a 3rd-runner tiebreaker.
The Settlement Grid pays everyone
Per-token: gross = outputTokens x spot price. 200 bps protocol fee. Net via Settlement Grid USD-HYLAQ -> USD-HYLAQ. One sha256 receipt. Recipient is the worker's @handle. That's it.
How a job moves
// Client posts a job:
POST /api/compute-mesh/jobs/post → price quote locked, status QUEUED
// Worker pulls work:
POST /api/compute-mesh/jobs/claim → Lease ISSUED, lockedPrice
// Worker runs the model + submits:
POST /api/compute-mesh/jobs/complete → attest → settle → receipt
// Server, atomically:
1. attestLease() // ACCEPT / AUDIT / SLASH / DUAL
2. settleLease() // builds Settlement Grid quote, executes
3. setLeaseReceipt(hash) // SETTLED, worker IDLE, +rep
The 40-year arc
- Years 1–3Power-user laptopsIndie devs running local LLMs lend NPU/GPU at night. Small developer audiences buy from the Mesh because it's cheaper than OpenAI for non-frontier models. ~10K workers, $1B/yr inference.
- Years 4–7Phones at scaleiOS/Android SDKs ship; opt-in 'earn on charger' becomes a setting. ~100M phones online overnight. The Mesh becomes the cheapest place to serve open-weight models on Earth.
- Years 8–12Vehicles + edgeTesla / Rivian / GM open EDGE_NPU APIs. Telco MEC nodes auto-register. Inference spot price decouples from hyperscaler cost curves. ~$100B/yr Mesh volume.
- Years 13–20Default for anything not frontierEverything except trillion-parameter frontier models runs on the Mesh by default. Hyperscalers pivot to specialty training. The Mesh is the Visa of compute.
- Years 20–40Citizen-owned computeThe biggest distributed cluster on Earth is owned by hundreds of millions of individuals. The dividend is meaningful. Compute is the new asset class everyone holds.
Shipped today
- src/lib/compute-mesh/ — types, capability scorer, pricing engine, reputation, in-memory registry with heartbeat-TTL OFFLINE detection, job queue with spot-priced-on-post, lease lifecycle (ISSUED/ACTIVE/COMPLETED/SETTLED/EXPIRED/FAILED/SLASHED), bidirectional scheduler, attestation + audit (STAKED + DUAL), settlement that calls into Settlement Grid, BigInt-safe wire serialize
- 8 API routes — POST /workers/register, POST /workers/heartbeat, POST /jobs/post, POST /jobs/claim, POST /jobs/complete, GET /jobs/[id], GET /prices, GET /health
- Pricing curve — piecewise-linear utilization (floor → 4x → ceiling) clearing per device class. Floor undercuts hyperscaler unit economics; ceiling undercuts hyperscaler list price.
- Capability scorer — per-class score caps (PHONE_NPU 250, LAPTOP_NPU 450, LAPTOP_GPU 700, DESKTOP_GPU 1000) with sustained-power discount + bandwidth gate
- Reputation arithmetic — bounded [0, 10000], +SUCCESS / -FAILURE / -SLASH, banned <1000, DUAL-eligible >=3000
- Worker SDK — httpTransport(baseUrl) + meshWorkerSdk + runWorkerLoop with heartbeat + long-poll claim + complete; runs in browsers, RN, Node, Bun
- End-to-end jest test — register → post → match → claim → complete → attest (forced ACCEPT) → settle via Settlement Grid → verifiable receipt + worker IDLE + reputation up
- Phase-2 swap points called out — TEE attestation on capable hardware, federated audit, persist to Postgres, multi-region scheduling with proximity, HQ-driven dynamic price calibration
Compute is the new oil. The user owns the well.
Hyperscalers concentrate the margin because they own the rails. The Mesh + Settlement Grid is a different rail — one where the hardware is already owned, the settlement is already protocol-native, and the dividend flows back to the people who own the silicon.