Azure Credits via NVIDIA Inception: Why We Moved
My work PC had several Claude Code windows open at once. Each one was conducting a single GPU job, and I hopped between them checking results. This post answers why a team working like that decided to move its workloads to Azure. The short version: joining NVIDIA Inception led us to Azure credits, and at the same moment we needed a managed place to keep GPU inference running around the clock.
We are a small team that builds models, runs inference servers, and operates a SaaS product. Our workloads fall into a few broad buckets: media processing, speech transcription, search, web, and GPU inference. Until now almost all of it ran on spot GPUs at another cloud and on that work PC.
How NVIDIA Inception led us to Azure
Azure was not where this started. Because we use a lot of GPU time, we first joined NVIDIA Inception, NVIDIA's startup program. While reading through the member benefits, I found an Azure credit offer.
The path went like this. We requested the Azure credit benefit through Inception first. Then we applied to Microsoft for Startups separately. We were told the review itself is done by Microsoft.
Honestly, it was not instant. Identity verification took longer than I expected, and at one point our first credits expired while we waited. Still, the portal showed a starter credit and the next milestones laid out ahead of us.
Around the same time we also applied for startup credits at other clouds, such as AWS Activate. The idea was not to lean on any single provider, but to be able to move compute to wherever credits and capacity are.
Why try Azure at all?
AWS has been our main cloud for a long time. To me it feels like a tool shed that already has everything a business needs. GCP is closer to a necessity for us, mostly because Google sign-in (OAuth) lives there.
What I hoped to get from Azure came down to three things: - A managed cluster that keeps GPU inference always on. An Azure Machine Learning compute cluster with a minimum of one node brings the node back if it gets taken away. - Cheap inference GPUs like the T4. At public prices, a 4-vCPU T4 node on low priority (spot) was in the low tens of cents per hour. - AI services such as Azure OpenAI and Foundry, available in the same subscription without extra plumbing.
I had worries too. The Azure portal and CLI were less familiar to me than AWS. I also did not want credits to quietly lock us into one cloud. So we set one rule up front: keep data and servers in standard shapes, like PostgreSQL, Docker images, and object storage, so they can move later.
What we prepared before day one
First, we read the milestone rules. The gist is to keep using several Azure services, for real, steadily for 60 days. Missing a day restarts the count, so it felt a lot like being handed a hackathon rulebook. Our own rule on top was simple: only attach services that do actual work, and write down what each one is for before turning it on.
Next, I used the Azure CLI to check T4 VM sizes and spot quota in the region closest to our users. Since a T4 is a T4, we realized we could do most experiments on a T4 instance we already had elsewhere, and ship only the finished setup to Azure.
We also sketched the networking. Previously, our web server and GPU server talked over a private network. On Azure, the plan is to put the web tier and the GPU node in different subnets of the same VNet (virtual network). The inference port opens only to the web subnet, and nothing on the GPU node should answer on a public IP.
The migration order is GPU inference, then web, then the database and search, then media processing and speech transcription. We start with what has the biggest payoff and is easiest to roll back. The very first step is one T4 kept always on in Azure ML.
Also today
- I skimmed the Foundry model catalog. Knowing the available models and how their quotas work up front saves confusion later.
- I checked whether a newer voice model is billed under Azure Speech. Which service a model rolls up to affects the design more than its name suggests.
- Some resource providers need to be registered before first use, so we made a note to check those ahead of time.
Recently we got word that the Azure credits through the Inception path were approved. The amount is in line with the program's published figure, a few thousand dollars. For a team our size that covers a few months of real experiments, so we plan to treat it as a budget that informs design decisions, not as free money.
Here is what this journal will cover, and what it deliberately leaves out: - In: which Azure services we picked and why, how we configured them, what broke and how we fixed it, and rounded costs. - In: honest pros and cons compared with other clouds and with how we used to do things. - Out: the specific content and data behind our workloads, account and resource names, and anything about the review process behind the scenes. - Out: tricks for hitting spending targets. What we are doing is moving real work.