DISTRIBUTED AI COMPUTE NETWORK

Token Plan and GPUs
both sell

Sell unused Token Plan quota and idle GPUs to DJB TokenBank for cash earnings; buy credit back when you need a model. Requests leave from your own machine; your own API account stays untouched

Use compute Sell computeUSDT per task

OPENAI-COMPATIBLE · CHEAPER THAN THE OFFICIAL API · NO KEY RETENTION

  • Token Plansell the unused part
  • GPUsell idle hours
  • USDTsettled per task

A little from each of us builds shared compute

Background: a global scheduling Illustration — each light is one cross-region dispatch; illustrative animation, not real traffic

Popular model prices · USD per 1M tokens · hover to pause, click for all · sources and sample times in the price panel

AnthropicClaude Sonnet 4.5$3in$15outAnthropicClaude Opus 4.1$15in$75outOpenAIGPT-5.1$0.625in$5outOpenAIGPT-5 mini$0.125in$1outDeepSeekDeepSeek V3$0.256in$0.712outDeepSeekDeepSeek R1$0.5in$2.15outMoonshotKimi K2$0.57in$2.3outxAIGrok 4.7$2in$6out
01 / TWO ROLES

Peak/off-peak bidding
sell higher, buy cheaper

When compute is tight, orders clear by bid so sellers get more; call off-peak and buyers pay less. One ledger, both ends better off

SUPPLY / SELL

I have an idle Token Plan or GPU

Sell unused Token Plan quota and idle GPUs into the pool for cash earnings; metered and settled per task, with caps and pause under your control

  • Requests leave from your machine; your own API account stays untouched
  • GPU, hours, concurrency and budget controls
  • Yields automatically to full-screen apps
Start sellingclient download
DEMAND / USE

I want to use large models

One OpenAI-compatible API; buy credit from the pool. Cheaper than the official API, no selling required

  • GET /v1/models
  • POST /v1/chat/completions
  • One error envelope and request_id
Start using
02 / EARNINGS & REWARDS

Sell compute
earn cash

Users buy tokens by usage; Token Plan sharers and GPU sellers share the earnings by usage, and earnings can be spent straight back on tokens — so unused Token Plan quota keeps earning across time and regions. Referrals earn rewards too, usable against tokens and inference compute

Tokens are inference credit: metered by real usage, offsetting inference only; task, usage, time and signed records stay reviewable

SAMPLE
TASK RECORD#DJB-••••-0421
NODE
edge-01
GPU HOURS
12.4
USDT EARNED
12.40
TASKS
38
VERIFIED
✓ YES
SETTLED
2026-11-01
USAGE CREDIT
NOT CASH / USAGE ONLYTKN 124.0
DJB TokenBankSAMPLE ONLY
03 / SCHEDULING OVERVIEW

Scheduling overview

Requests, GPUs, prices — one table

Operating figures refresh daily

Internal testing

REQUESTS / 24H

1,284,480

TOKENS PROCESSED

1.66B

ONLINE SUPPLIERS

5,318

ACTIVE CONSUMERS

5,942

PRICE COVERAGE

7 Price snapshot · 2026-11-05T00:11:36+00:00

Operating figures are illustrative and not a commitment; model prices are sampled from public offers (snapshot time shown in the price panel)

GLOBAL SCHEDULING

Global compute scheduling

SAMPLE
  • Dot = one node
  • Arc = one cross-region move
  • Pulse = hub is taking work

Illustration: motion shows that scheduling is happening — not an actual quote, not a capacity promise.

REQUEST STREAM

Sample request stream

SAMPLE
TOTAL 1,284,480
  • deepseek-v3.2→OK
    OK
    2,184 tok812ms
  • qwen3-235b→OK
    OK
    1,460 tok1.1s
  • glm-4.6→OK
    OK
    986 tok940ms
  • kimi-k2→OK
    OK
    3,020 tok1.4s
  • minimax-m2→OK
    OK
    1,240 tok702ms

Illustrative rows, ordered by request shape

SUPPLY / DEMAND

Who supplies, who uses

SAMPLE

Online suppliers5,318

East China
1,808
South China
1,011
North China
962
Southwest
532
Overseas
425
Central China
213
Northwest
191
Northeast
176

Active consumers5,942

Individual devs
3,446
Small teams
1,412
Enterprise
763
Research labs
321

By region / account type

CALLS / MIN

Calls overview

892req/min
24m agonow

Last 24 minutes

COMPUTE FOR SALE

Compute for sale

SAMPLE
78%avg utilization

Sold today

99,553 GPU h

from 4,126 sellers

Active nodes5,318
Avg utilization78%
Top regionEast China 34%

PRICE LIST / SERVER

Model prices

Fetched 2026-11-05T00:11:36+00:00

USD / 1M TOKENS · reference price (lowest market) · source per row · sampled 2026-11-05T00:11:36+00:00 · order price governs

Popular model pricesHOT-PRICES.JSONsampled unknown

VENDOR / MODELINPUTOUTPUTSOURCE
Claude Sonnet 4.5Anthropic315Anthropic
Claude Opus 4.1Anthropic1575Amazon Bedrock
GPT-5.1OpenAI0.6255OpenAI
GPT-5 miniOpenAI0.1251OpenAI
DeepSeek V3DeepSeek0.2560.712DeepInfra
DeepSeek R1DeepSeek0.52.15DeepInfra
Kimi K2Moonshot0.572.3Novita
Grok 4.7xAI26xAI

Low-price referenceMARKET-PRICES.JSON

Public model list; lowest published offer per model across multiple sources (relays and aggregators; manual samples are flagged); sample time and source shown per row; order price governs

MODELINPUTOUTPUTSOURCE
llama-3.1:8b-instruct0.020.04DeepInfra2026-11-05T00:11:36+00:00
qwen2.5:7b-instruct0.0520990.052099SiliconFlow (manual sample)2026-11-05T00:11:36+00:00
llama-3.3:70b-instruct0.10.32DeepInfra2026-11-05T00:11:36+00:00
qwen2.5:32b-instruct0.1875560.187556SiliconFlow (manual sample)2026-11-05T00:11:36+00:00
deepseek-chat0.2560.712DeepInfra2026-11-05T00:11:36+00:00
qwen2.5:72b-instruct0.360.4DeepInfra2026-11-05T00:11:36+00:00
deepseek-reasoner0.52.15DeepInfra2026-11-05T00:11:36+00:00

Source: market-prices.json · fetched at 2026-11-05T00:11:36+00:00

Not production monitoring, not a quote

Open full live view
04 / TRUST

Trust, stated plainly

The expensive part of a compute network is trust. Three rules, shipped in the product

01 / KEYS

No key retention

Node checks use a one-request credential, discarded right after the check. The portal never asks for your own API key

02 / LEDGER

Accounting you can check

Task, usage, time and signature are kept; ledger changes are recomputable, zero diff across migration

03 / CONTROL

You can stop anytime

Hours, concurrency, budget caps. Pause stops new work

Beta note: Internal beta now, not a production release. Opening date, settlement rules and SLA come before mainnet. Selling pays USDT per task; referral rewards are non-cash and not a yield promise
05 / FAQ

Q&A

Straight answers first, then we grow the network

What is DJB?

A distributed AI compute-sharing network: sell idle GPUs and unused Token Plan quota, call models through one OpenAI-compatible API

How is selling settled?

Metered by real usage: selling earnings settle per task in USDT, with usage and timestamps on record. Tokens offset inference only; referral rewards carry no yield promise

What are referral rewards?

Refer a customer and earn rewards (points): not cashable and not a yield promise, spent against tokens and inference compute

Can tokens or referral rewards be cashed out?

Neither. Tokens offset inference; rewards offset tokens and inference compute. Selling earnings settle separately in USDT

Will sharing hurt my gaming?

No. Set GPU, hours, concurrency and budget; it yields to games, calls and full-screen apps, and you can pause

Does the platform keep my API keys?

The portal never asks for it. The node credential serves exactly one request and is discarded afterwards

Can I use it now?

An internal beta today, not a production release. Try the selling flow and OpenAI-compatible calls; dates and SLA come before mainnet

06 / EXTERNAL RESEARCH

Not every token
deserves the biggest model

Appendix: external research. TokenRouter (Tsinghua NICS / CMU, NeurIPS 2026) moves routing from whole requests to single tokens

NeurIPS 2026 arXiv:2610.12242 · Tsinghua NICS / CMU

TokenRouter

You still write routing per request; each model runs its own subserver, dispatched asynchronously

2.01–64.15×reported decoding throughput
vs. existing systems

Paper-reported range; not a DJB measurement, not dollar savings

Read the paper ↗See the code ↗
TOKEN-LEVEL ROUTINGSCHEMATIC
Pleasethecon-tractsum-marizeinthreelines
FAST / CHEAPSmall modelFormat, connectors, easy continuations
STRONG / REASONINGStrong modelAmbiguity, reasoning, key calls
One answertokens merge → output

Schematic only — not TokenRouter's actual routing policy. Colors only suggest cheap-path vs. strong-model tokens

01 / FINE-GRAINEDRoute at token granularity

Different tokens in one sentence take different models

02 / ASYNCOne subserver per model

Requests dispatch asynchronously; one slow model can't stall the batch

03 / BATCHINGDelayed batching

Scheduler parameters come from a throughput model; fewer step mismatches

WHY IT MATTERS FOR DJB

For DJB: cheap compute for easy tokens, strong models for hard ones

External paper, not a DJB measurement; no partnership; TokenRouter is not implemented here; no yield promise

WHY / COMPUTE EQUALITY
The ticket to AI
shouldn't only go tothose who can afford a GPU

Compute shouldn't live only in a few data centers. A little from each of us, pooled

DJB / TokenBank
BUILT FOR THE AI COMMONS
INTERNAL · 2026
READY / INTERNAL

Sell the idle
buy what you need

DJB TokenBank: sell unused Token Plan quota and idle GPUs for cash earnings, then buy credit back when you need a model. One ledger, scheduled across time and regions

Use compute Sell computeUSDT per task

INTERNAL BETA · USDT SOLD PER TASK · REFERRAL REWARDS ARE NOT CASH · NOT A PRODUCTION RELEASE