Allocate
PricingSign inBook a call
New · RecursiveDB, the learning layer

The inference cloud that turns your company’s data into models you own.

Start on frontier models in minutes. Every completed task trains your own model. Export the weights any time.

Start trainingBook a call

Trusted by teams who own their AI.

Jump Point
Hype Digital
Colas

Infrastructure that learns is different. Every task your agents complete becomes training signal. The signal trains your model. The next task runs on a sharper one. The loop is the product, and the weights it produces are yours.

fig. 01

Recursive harness

The agent runtime and SDK. Ships your agent’s code, tools, and learning with every call.

fig. 02

RecursiveDB

The learning layer. Every completed task becomes training signal.

fig. 03

Agent computers

Dedicated machines per tenant, metered per second. Idle time costs nothing.

1.0 · runtime

Recursive harness. Ships your agent’s code, tools, and learning with every call.

Your agents
Fraud reviewClearing overnight card disputes
Review every dispute that lands overnight and clear what’s legitimate.
Received 3,120 disputescore-banking feed · overnight batch
Learned from 18 months of rulingschargebacks.parquet · your analysts’ calls
Cleared 2,981 automatically139 flagged with evidence

139 flagged cases carry evidence. Every analyst ruling lands in RecursiveDB as training signal.

Fraud model, version 4+18% precisionDispute agentDeployed
A human approves every write actionCredentials never enter the sandboxRunning turns always finish
2.0 · learning layer

RecursiveDB. Every completed task becomes training signal.

Signals, live
Dispute cleared, analyst agreedground_truth+signal
Booking confirmed, no editsground_truth+signal
Ticket resolved, no reopensexecution+signal
Escalation resolved, outcome loggedexecution+signal
Document filed, human agreedground_truth+signal
Triage decision upheld on reviewjudged+signal
Training as compaction
Every promotioneval gated, reversible
Checkpoint ownershipyours, export any time
Signals come from real outcomes, not labelingEvery promotion is eval gated and reversibleExport your weights and data any time
3.0 · machines

Agent computers. Dedicated machines per tenant, metered per second.

api.allocate.networkresolves to the nearest cell

Live in Johannesburg, London, Frankfurt, N. Virginia, Dubai, and Singapore. São Paulo and Sydney soon.

Agent computer hoursincluded per month
Free5 h
Starter50 h
Pro150 h
Growth300 h
Team500 h

Metered per second. Idle time costs nothing.

Live in 6 regions, 2 more soonMetered per second, idle costs nothingYour own containers at every tier
4.0 · models

Models. 158 models behind one endpoint, routes to weights you own.

alc_••••••••7f2eprod/fraud-reviewdeepseek/deepseek-v4-proOne key. The route stays when the model changes.
158 models behind one endpointOpenAI compatible, swap one base URLThe served model returns in x-allocate-model
weights

You own the weights.

Every training run produces a checkpoint that is yours. Export your fine-tuned weights whenever you want.

How training works

Pay for work delivered, not capacity reserved.

Free plan with 5 agent computer hours and a one-time $25 token credit. No card.

See pricing

Allocate for startups

A free plan for everyone, credits for early-stage teams.

Free plan, no cardSeed to Series AApply in 5 minutes
Apply now

Make your first call in minutes.

Free plan with 5 agent computer hours and a one-time $25 token credit. No card.

Start trainingBook a call