Allocate
Cloud AgentsA fleet that works, in secure sandboxesModelsEvery model behind one keyRecursiveDBWork becomes training signalTrainingYour data becomes your modelMeterEvery token tracked, liveData SecurityIsolated inside your boundary
EnterprisePricing
ChangelogDocsFree toolsGlossary
ModelsSign inTalk to sales
Allocate documentation
QuickstartsYour first inference callUse Allocate from your coding agent
PlatformGatewayRoutesModelsKeysMeter
API referenceCreate a chat completionList models
PlanesPrivate Inference CloudCloudlightCloudheavyLearningAgentsRecursiveDBTraining
Allocate

Regulated cloud inference. An AI model you own and grow, on infrastructure you control.

Product
Cloud AgentsModelsRecursiveDBTrainingMeterData SecurityPricing
Solutions
HealthcareInsuranceFinancial servicesOperationsAgenciesStartupsSouth AfricaGulf region
Resources
ChangelogDocumentationAPI referenceFree toolsGlossarySecurity
Popular models
Gemini 3.5 FlashDeepSeek V4GPT-5.5Qwen 3.5Best open-weight modelsExplore all
Company
EnterpriseAboutCareersContactTalk to sales
© 2026 Allocate. All rights reserved.POPIA ready · security overview

Cloudlight

The default execution plane: low latency, close to your users.

Cloudlight is where your workloads run unless you have pinned them elsewhere: a globally distributed edge plane that serves gateway requests, agent sandboxes, and dashboards with the lowest latency the platform offers.

What runs here

  • Gateway authentication and routing, within the 30ms p50 overhead budget.
  • Agent sandboxes: isolated execution with a warm window, roughly 4s to wake warm and 9s cold.
  • The Meter, key cache, and route resolution.

Characteristics

Cloudlight is stateless by design. Sandboxes hydrate from your data plane and hold nothing durable of their own, which is what makes them safe to recycle and fast to scale. Durable state lives in your Private Inference Cloud's data plane.

When a workload needs region pinning, heavy compute, or at-rest residency guarantees, it moves to Cloudheavy. Most tenants use both without thinking about the split: the platform places each workload on the plane that fits it.

PreviousPrivate Inference CloudNextCloudheavy
On this pageWhat runs hereCharacteristics