Model-as-a-Service (MaaS)

Model-as-a-Service is the cloud computing pattern applied to AI models: instead of buying and operating the hardware required to run a large language model, an enterprise calls a REST API and pays per token, per request, or per hour of compute consumed. The model provider handles GPU provisioning, scaling, redundancy, security patching, and model updates — the enterprise simply integrates the endpoint.

The MaaS market has bifurcated into two tiers. Hyperscaler MaaS (AWS Bedrock, Azure OpenAI Service, Google Vertex AI Model Garden) bundles model access with enterprise cloud commitments, IAM integration, VPC support, and compliance certifications — making them the default choice for regulated industries. Specialist MaaS providers (Together AI, Replicate, Fireworks AI) compete on model variety, price-per-token, and inference speed, often running open-source models (Llama, Mistral, Mixtral) at fractions of the hyperscaler cost.

The enterprise decision framework centers on three trade-offs: **control vs. convenience** (managed API vs. self-hosted model), **cost vs. compliance** (specialist providers offer lower prices but fewer certifications), and **flexibility vs. lock-in** (provider-specific APIs vs. OpenAI-compatible endpoints that simplify switching). Organizations running high-volume inference workloads should model total cost of ownership carefully — at sufficient scale, dedicated GPU instances often undercut per-token MaaS pricing.