이미지: 마이크로소프트
Summary
- The Fireworks AI open-model inference service has officially launched on Microsoft Foundry, letting startups tap high-performance inference within Azure without building their own GPU clusters.
- It supports pay-per-token serverless inference, with cost savings available through duplicate-request caching via Azure Cache for Redis.
- Microsoft for Startups credits apply to Data Zone Standard deployments but not to PTU (provisioned capacity) deployments.
Fireworks AI officially launches on Azure Foundry
Microsoft has officially launched the Fireworks AI open-model inference service on Microsoft Foundry (Azure Foundry). With this launch, startups can access high-performance, low-latency open-model inference within Azure without building or operating their own GPU clusters. Fireworks AI handles the inference layer, while Foundry manages deployment, governance, and billing through a single control plane.
Alongside the announcement, Microsoft also released a deployment design blueprint for AI-native startups. The blueprint is designed to help small teams move from idea to MVP to product-market fit (PMF) in a repeatable, Azure-native way.
Why now: moving beyond reliance on a single closed model
Many startups start with a single closed model to build early prototypes quickly. The problem is that this decision has a cumulative effect on cost and differentiation strategy throughout the product lifecycle. Before Fireworks AI was integrated into Foundry, serving open models directly on Azure required building separate inference infrastructure.
With this launch, model selection, fine-tuning, and behavior tuning are now all in the hands of development teams. Teams can choose models suited to their workload characteristics and reduce the rework involved in switching models—through a consistent API and deployment workflow—when performance or cost requirements change.
Getting it running: architecture and billing
The basic stack runs entirely within an Azure subscription, and teams can get started with just a single model endpoint in their application infrastructure. The starting point is deploying a single model and routing traffic through Azure API Management. From there, teams can scale by caching duplicate inference requests with Azure Cache for Redis, tuning performance based on workload characteristics, or A/B testing multiple model variants.
Billing is based on pay-per-token serverless inference by default. Teams enrolled in the Microsoft for Startups program can apply startup credits toward Fireworks model usage on Data Zone Standard deployments, with supported credit amounts of up to $150,000. However, startup credits do not apply to PTU (Provisioned Throughput Units), i.e., reserved capacity deployments. Cost per million tokens is recommended as a core engineering metric to track.
The main components that make up the architecture are as follows. Azure Container Apps hosts the application or API that sends inference requests, while Azure Container Registry stores container images. Azure Key Vault securely stores credentials and endpoint information, and Azure Monitor optionally provides observability and performance insights.
Fine-tuning and bringing in custom model weights
Once a workload stabilizes, teams can fine-tune models using evaluation suites, prompt libraries, and graded production traffic as training data. Fireworks Training supports fine-tuning followed by import into Azure via a Bring Your Own Weights approach. Importing custom model weights directly is also supported.
Limitations and undisclosed details
The fact that startup credits do not apply to PTU deployments is a constraint that teams considering scaling on reserved capacity should be aware of in advance. Observability through Azure Monitor is offered as an optional feature, not a requirement. The specific list of supported open models and per-token pricing were not specified in the source.
