One model. One endpoint. No supplier to choose.
Pooled inference lets approved buyers send non-streaming text chat requests to a single NexusCompute endpoint by model name. The pool picks an eligible machine for each request.
Already a buyer using dedicated access? Go to your endpoints.
- Request ANexusCompute poolMachine 2Request A is sent through the pool to Machine 2 only.
- Request BNexusCompute poolMachine 1Request B is sent through the pool to Machine 1 only.
- Request CNexusCompute poolMachine 3Request C is sent through the pool to Machine 3 only.
How private pooled requests are handled
Temporary content, not saved chats
New private pooled prompts and outputs use a memory-only buffer with no disk persistence or buffer backups. Inputs expire within 120 seconds; outputs expire within 60 seconds or are removed when receipt is confirmed. The pool refuses new work if verified private handling is unavailable.
Provider policy and app checks
Connections use encrypted transport. One eligible provider machine processes each request under a policy prohibiting capture, retention, sharing, forwarding, resale or training on buyer content.
Billing records remain
Usage, timing and payment metadata are retained for billing and abuse prevention, not a reusable prompt/output history. These commitments apply to new private pooled requests, not dedicated listing access or earlier stored requests.
For the full processing, retention and usage terms, see Privacy and the Provider Usage Policy.
Why pooled
Call the model, not a provider
Send requests to the NexusCompute chat completions endpoint using the operator-configured model. You do not pick a supplier per request.
The pool selects an eligible, currently qualified machine for each request. You skip comparing individual listings.
Each request reserves an amount up front and settles to the final actual charge once results come back.
Ordered text messages plus max_tokens, authenticated with your existing NexusCompute API key.
Good fit
- Short drafting, rewriting and summarizing of public or non-sensitive text
- Classification or tagging of non-confidential content
- Prototyping prompts against a single configured model
- Batch-style tasks where each request stands alone
Not supported for pooled requests
- Streaming responses
- Tool or function calling
- Images, audio or other multimodal input
- Temperature or additional sampling options (temperature is fixed at zero)
- Regulated, secret or otherwise sensitive information
Two ways to buy
Pooled versus dedicated listing access
| Pooled inference | Dedicated listing | |
|---|---|---|
| Who picks the machine | The pool, per request | You choose a provider listing |
| Models | One operator-configured model | Whatever the listing serves |
| Request type | Non-streaming text chat only | Per endpoint capabilities |
| Access | Approval required | Available on existing paths |
| Availability | May be paused or unconfigured | Depends on the listing |
Prefer to pick hardware? Browse dedicated endpoints.
Getting started
Five separate things need to be true
Approval, an account, credits, a credential and an available pool are independent. Having one does not imply the others.
- 01
Request account access
Request an account if you need one. Account approval does not reserve provider capacity.
- 02
Create or use your account
If you are invited and do not have an account, sign up. Existing buyers use their current account.
- 03
Add prepaid credits
Requests need enough credit to cover the up-front reservation, or they are declined with insufficient credit.
- 04
Create an API credential
Use an existing NexusCompute API key as a Bearer token, plus a unique Idempotency-Key per logical request.
- 05
Check pool availability
The pool may be paused or unconfigured. When no capacity is available, requests fail explicitly rather than being rerouted.
Pricing mechanics
Reserve, run, settle
Rates are set for the configured model and shown to approved buyers. This page describes how charges work, not what they are.
Read the pooled inference API docs- 1
Reserve
Input tokens plus your selected max_tokens are held from prepaid credits before work starts.
- 2
Run
One eligible machine processes the request. Tokens are counted by the server's tokenizer, not the machine's.
- 3
Settle
The result shows the final actual charge based on usage tokens. Any unused reservation is released.
- 4
Failure
If work fails, the hold is released.
Questions
Is pooled capacity always available?
No. Paid requests require an enabled model and a currently qualified provider machine. Check Endpoints for live availability. An unconfigured, paused or empty pool rejects new requests; there is no capacity or latency guarantee.
Does the pool combine GPU memory or split a model across machines?
No. Each request is handled independently by one eligible machine running the full configured model. This is not GPU rental, memory pooling or model splitting.
Which model can I use?
One model configured by the NexusCompute operator. Check Endpoints for the current configured model, terms and availability before sending requests.
How am I charged?
Usage tokens are measured by the server's tokenizer. Input plus your chosen maximum output is reserved from prepaid credits, then the final actual charge is shown with the result. Failed work releases the hold.
Is my content kept?
Private pooled requests keep no persistent prompt or output history. Inputs stay in temporary memory for up to 120 seconds and outputs for up to 60 seconds, or until your client confirms receipt. Usage and billing metadata are kept.
What may providers do with my content?
Providers must accept a usage policy that forbids capturing, retaining, disclosing, reselling, forwarding or training on buyer content. Accounts that break it are removed from the pool.
Can I fetch a result again later?
No. Read the output when it is delivered. Repeating the same idempotency key returns the request status and charge, never a new generation or a second charge.
What happens if no machine is free?
The request fails with a clear no-capacity response. A paused model is never silently routed to unqualified supply.
Can I still use dedicated listings?
Yes. Dedicated endpoints and listing access remain available on their existing paths.
Paid usage requires currently qualified capacity. No performance or capacity guarantee.PrivacyTerms
Want to try the pool?
Request account access, then check the configured model and live capacity in Endpoints.
Existing buyer? Open endpoints