how can we find an efficient space to deploy a local LLM?

Alot depends on Joe, the LLM, your chosen pricing model, hosting location which are all choosable by you. Is it an OpenAI model or a custom model? OpenAI models are on Azure OpenAI, custom models are on Azure Machine Learning.

Here is a link to Azure OpenAI pricing page: https://azure.microsoft.com/en-us/pricing/details/azure-openai/

There you will find the pricing models:

Standard (On-Demand): Pay-as-you-go for input and output tokens. Provisioned (PTUs): Allocate throughput with predictable costs, with monthly and annual reservations available to reduce overall spend. Batch API: Language models are also now available in the Batch API for global deployments and three regions, that returns completions within 24 hours for a 50% discount on Global Standard Pricing.

If Joe, the LLM is an OpenAI model that is the pricing you should be looking at, and the costs vary alot on the specifics. The generic quote being $5/$30 per million input/output tokens(deploying in other regions has an impact too), or if you choose provisioned, then about $15 per hour and upwards, with monthly and yearly reservations as an option. You should check the pricing page for details and up to date information.

Here is a link to Azure Machine Learning pricing page: https://azure.microsoft.com/en-us/pricing/details/machine-learning/
There you will find the pricing for a great number of virtual machines available for deploying custom LLM models. There, also, there is a pay-as-you-go, Azure savings plan for compute and Reservations pricing options. The pricing depends on the type and specs of the Virtual Machine chosen, and, since you ware not specific at all about what you might need, the lowest price listed right now it "$53.29/month" without using a savings plan, which does lower it significantly in return for a long term commitment. If your use case fits, you should definitely check out that pricing page where you will also find other details and up to date information.

So, yes, there is a place on Azure where you could continue this project. The processing costs are not unresonable and are down to your choice.

Yes. The best Azure option depends on whether "Joe" is an opensource/custom model that you are running locally, or an Azure OpenAI model already available as a managed service.

If Joe is an open-source or custom LLM, I would generally avoid moving it straight to a permanently running GPU VM, as that can become expensive for experimentation and occasional use.

A cost-conscious approach is to use Azure Container Apps with serverless GPU support, which provides GPU-backed containers with scale-to-zero behaviour. This allows the environment to consume GPU resources only when requests are being processed, making it a practical option for proofs of concept, testing, and infrequent workloads.

If you need a persistent development environment, consider a GPU-enabled Azure VM such as the NCasT4_v3 family and shut it down when not in use. Using Azure Spot VMs can further reduce costs if occasional eviction is acceptable.

Before choosing infrastructure, check the model's:

• Parameter size (for example, 7B, 13B, 20B)
• Quantisation format
• GPU memory requirements
• Expected number of concurrent users

These factors have a much larger impact on cost and performance than the hosting platform itself.

If Joe is actually an Azure OpenAI-supported model, then you may not need to host the model yourself at all. In that case, Azure OpenAI's consumption-based pricing could be simpler and potentially more economical than managing GPU infrastructure.

For a POC, I would start with Azure Container Apps and a serverless GPU-backed model container. If the workload becomes steady or requires an always-on development environment, then evaluate a dedicated or Spot GPU VM.

Help make this community better for everyone: if this answer resolved your issue, please accept it or leave an upvote. If not, share more details in a comment so we can continue the discussion and find the right solution.

I think the cheapest Azure spot for a local LLM is Container Apps with serverless GPU, it stays at zero cost til you hit it. to pick the right GPU path, tell me Joe’s size, how it’s packed, and how much mem it eats when you run it.