Documentation
Docs.
Everything you need to know about TOY JENSEN. Buy, use, sell or gift GPU hours priced in USDG on Robinhood Chain. Build without limits.
Getting started
Three steps, no account. You need a wallet with USDG or ETH on Robinhood Chain.
MetaMask, Phantom or Coinbase Wallet. Press Connect wallet top right. Connecting never moves funds.
The site asks your wallet to switch or add the network. You can also add it yourself with the details below.
Pick an amount, pay with USDG or ETH, confirm in your wallet. Hours land when the block confirms.
opens a prompt in your wallet, nothing else
What is an hour
One GPU hour is the compute of one RTX-4090-class GPU for one hour. Everything in the Studio is priced in hours from what the job actually costs at wholesale, so an hour buys a consistent basket.
| One hour buys | About | How it's counted |
|---|---|---|
| Chat | 40 short essays · ~1,000 words each | Model rate per 1,000 words, tokens × 0.75 |
| Images | 300 standard or 160 high-quality renders | 0.003 h standard · 0.006 h high |
| Private AI | 60 minutes of an RTX-4090-class endpoint | Card rate × minutes, stops at zero |
| Other cards | From 0.33 h per hour (T4) to 6.58 h per hour (B200) | Card price ÷ 1.20 USDG |
Hours are prepaid service credits, like a gift card for compute. They never expire. They are not an investment and pay no yield.
Pricing
Retail is 1.20 USDG per hour. The bid, what you get when you sell an hour back, is 1.14 USDG. Both follow real GPU rental markets with a 24-hour lag, so the price you see today is the price you get today.
How the price is set
- A rental index is built from indicative wholesale rates across five GPU classes, weighted by what the floor actually runs.
- The bid tracks the index with a 24-hour lag. The ask sits 5% above the bid. That 5% is the spread and is the only margin on hours themselves.
- Studio jobs are priced from provider cost times 2.5 for inference and 1.4 for rentals, so every model and card lands on the same hour scale.
| Model | Hours per 1,000 words | Class |
|---|---|---|
| Llama 3.3 70B | 0.025 | Best |
| Llama 3.1 8B | 0.006 | Fast |
| Qwen 2.5 72B | 0.024 | Code |
| DeepSeek V3 | 0.030 | Deep |
| Mistral Large | 0.028 | EU |
| Gemma 3 27B | 0.010 | Light |
| Gemini 2.5 Flash Image | 0.003 / image | Images |
The price index on the Foundry page is information about rental markets, not a promise about hours.
Paying
With USDG
One ERC-20 transfer from your wallet to the factory wallet. That's the whole payment. When the factory API is on, the amount carries a tiny unique suffix so the factory can match it to your quote; without it, hours are credited on your device from the confirmed receipt.
With ETH
Pick ETH in the buy box. The site swaps ETH to USDG inside your own wallet through Uniswap on Robinhood Chain, waits for that to confirm, then sends the USDG payment. Two confirmations. The swap buys about 1.2% more than needed to cover price movement, and any extra USDG stays with you.
Timing
Blocks on Robinhood Chain arrive every 0.2 seconds or so. A payment usually shows as confirmed within a few seconds; the site polls every 8 seconds and credits hours the moment the receipt is final.
Your account
Your wallet address is your account. There is no email, password or sign-up.
- Device mode (today): hours, receipts, chats and your gallery are stored in your browser. Clearing site data clears them; the onchain payment record stays forever.
- Factory mode (when the API is on): you sign a short message with your wallet to prove it's yours. That signature costs nothing and moves no funds. Your hours then live on the factory ledger and follow you to any device.
The Studio's Settings panel shows which mode you're in and lets you export your receipts as CSV.
The Studio
| Panel | What it does | Cost |
|---|---|---|
| Chat | Six open models, persona presets, several saved chats, cost estimate before you send, copy and export. | per 1,000 words |
| Images | Six style presets, four sizes, one to four images per run, gallery kept on your device. | 0.003 / 0.006 h |
| My own AI | Rent any of 24 GPUs by the hour with a model preloaded. OpenAI-compatible endpoint and key. | card rate × hours |
| Jobs & receipts | Every job, its model and its cost. Seven-day chart, CSV export. | — |
| Agent keys | A key a bot can use to spend your hours through an OpenAI-compatible API. | per call |
Until launch the Studio runs on your own OpenRouter key, kept only in your browser. Hours are still deducted per job so the receipts are real.
Redeem
TOY JENSEN is tokenized productive compute capacity. Not GPU mining, not GPU yield, not a share of NVIDIA. An hour is a contractual entitlement to a standardised quantity of compute, and Redeem is the link back to the real thing: gold tokens redeem gold, stablecoins redeem dollars, GPU hours redeem compute.
- Pick a workload in the Studio: image generation, a private LLM, a Blender render, an animation, a video upscale, a custom container, or the TJ Benchmark fine-tune.
- A verified NVIDIA card consumes the hours. Nothing is consumed until a provider accepts and the result verifies.
- The consumed receipt is marked onchain when the factory API is live, so a redeemed hour can never be spent twice.
Creator network
Beyond chat, TOY JENSEN is a job marketplace: creators submit GPU work, registered providers run it on their own NVIDIA cards, results are verified, and the record goes onchain.
Submits a job with a quote in hours and a hash of the input file.
A registered card that fits picks it up and runs it.
Input hash, output hash, provider and timing are checked.
Job, payment and reputation recorded on Robinhood Chain.
Provider gets 80% of the hours in USDG; the factory keeps 20%.
| Job | Rate | Runs |
|---|---|---|
| Generate image | 0.003 h / image | Now, on the Images panel |
| Render 3D model | 0.02 h / 1080p frame | Queued for a provider |
| Render animation | 0.02 h / 1080p frame | Queued for a provider |
| Upscale video | 0.01 h / second (2×) · 0.02 h (4×) | Queued for a provider |
| Run AI model | card rate per hour | Private endpoint |
1440p counts as 1.8 frames and 4K as 4 frames. Nothing is charged until a provider accepts and the result verifies; a failed job is never paid. Reputation is 60% verified jobs, 25% on-time delivery, 15% uptime, and decides queue priority. No staking, no slashing, no token.
Passports & index
Compute Passport
Every physical GPU that registers gets an identity page: model, region, memory, benchmark score, verified hours, successful jobs, uptime, energy consumed and reputation. Serial numbers and hardware identifiers are hashed with the wallet and registration time; the hash is published, the identifiers never are.
Proof of compute
The provider app samples NVIDIA DCGM during every job and signs a receipt: card, start, end, utilisation, memory, energy, input hash, output hash, provider, job. The receipt hash is anchored on Robinhood Chain. That is what makes a completed hour provably real rather than a database row. The TJ Benchmark, a fixed SDXL fine-tune sequence adapted from NVIDIA's public AI Workbench example (Apache 2.0), is the canonical verification workload that gives a card its score.
GPUH Index
GPUH-4090, GPUH-A100, GPUH-H100 and GPUH-B200 quote standardised compute like a commodity: USDG per hour, committed capacity, utilisation, completed hours and a 7-day line. The bid on the Marketplace references this index with a 24-hour lag. Capacity is inventory, not a promise of return.
API
The factory API is OpenAI-compatible where it matters and wallet-signed where it must be. Base URL and keys open with launch; the shapes are final.
Agents can also pay per call with no key: a request answers 402 Payment Required with a USDG amount and the factory wallet; the agent pays onchain and retries with the transaction hash.
Chain details
The only addresses TOY JENSEN uses. The explorer is full of same-name fakes; trust these or Robinhood's own registry, nothing else.
| Name | Address | Role |
|---|---|---|
| Factory wallet | 0x248C6Bca66be6660d8D13D7cF6316A31710a79Dc | Receives every USDG payment |
| USDG | 0x5fc5360D0400a0Fd4f2af552ADD042D716F1d168 | Unit of account · 6 decimals · Paxos |
| WETH | 0x0Bd7D308f8E1639FAb988df18A8011f41EAcAD73 | Wrapped ETH used by the swap leg |
| SwapRouter02 | 0xcaf681a66d020601342297493863e78c959e5cb2 | Uniswap V3 router · ETH → USDG |
| QuoterV2 | 0x0269f8b86bb3c1e927dacedb72f3463ef6d26f61 | Live quotes for the swap leg |
Live block height, gas, block time and every payment into the factory wallet are on the Network page.
Status
What is live today and what is not. This table changes with every release.
| Feature | Status | How it works |
|---|---|---|
| Wallet connect | Live | EIP-6963 picker, chain switch to 4663. Connecting never moves funds. |
| USDG and ETH balances | Live | Read from the chain over public RPC. |
| Buy hours with USDG | Live | One transfer to the factory wallet; hours credit when it confirms. |
| Buy hours with ETH | Live | Uniswap swap in your own wallet, then the USDG transfer. |
| Studio chat and images | Live · your key | Runs on your own OpenRouter key until the factory key is set. Hours deducted per job. |
| Private endpoint | Opens at launch | A real GPU rented for you by the hour. |
| Sell and gift hours | Opens at launch | Server-side ledger first; wallet-to-wallet hours later. |
| Agent keys · pay per call | Opens at launch | Issued by the factory, bound to your wallet. |
| Job queue · renders, animations, upscales | Phase 2 | Jobs queue on your device with a proof hash; providers pick them up when the provider app ships. |
| Provider registration · capacity commits | Live · signature | Register a card and commit monthly hours with wallet signatures; both show on the index and your passport. |
| Proof of compute · TJ Benchmark · passport scores | Phase 2 | Needs the provider app with DCGM sampling; receipts anchor onchain through the factory API. |
| GPUH Index prices and inventory | Modelled | Indicative card prices and a capacity model, plus any hours committed from your device, until providers report. |
| Factory counters, price index, order book | Modelled | Shown numbers are a model until the factory reports real ones. |
Providers
Inference is bought wholesale through OpenRouter, which exists for exactly this kind of resale. GPU rentals come from RunPod and Vast. Jobs inherit each provider's acceptable-use policy; anything they refuse, we refuse. A provider outage degrades quality, it never takes hours.
Open source
The verification workload at the centre of this project is not ours. NVIDIA publishes an AI Workbench example that fine-tunes Stable Diffusion XL on a handful of photos of their toy mascot, Toy Jensen, and releases it under the Apache License 2.0. That run is a real, repeatable piece of GPU work, so we made it the yardstick every provider card has to pass, and named this project after it.
What we changed
We keep the training procedure and add four things that turn a demo into a proof: fixed parameters so every card runs identical work, NVIDIA DCGM sampling throughout the run, a receipt signed by the provider's key, and the receipt hash anchored on Robinhood Chain. The full method is on the TJ Benchmark page.
Security & privacy
- No custody. Payments go straight from your wallet to the factory wallet, and the ETH swap happens inside your own wallet.
- No personal data. No email, no name, no tracking pixels. Your wallet address is the only identifier.
- Keys stay on your device. An OpenRouter key you paste never leaves your browser.
- Sign-in is a signature, not a transaction. It cannot move funds.
- Verify addresses here or in Robinhood's registry before sending anything. We will never DM you an address.
Regions
Hours are not geo-restricted. Payments are USDG, or ETH swapped to USDG inside your own wallet. No stock tokens are involved.
Refunds
Refunds are given for jobs that were paid for and not delivered. A failed job never deducts hours. There is no refund for hours you simply decided not to use, because they never expire and can be sold back or gifted.
Risks & disclosures
- Wholesale compute prices move. The bid and ask follow them with a day's lag.
- Model outputs can be wrong. Check anything important.
- Robinhood Chain and its tokens are run by their issuers, not by us. Onchain transactions are final.
- Modelled numbers are labelled everywhere they appear, and this page is the record of what is real.