What Azure Credits Actually Cover and What Still Costs Money

Azure credits work like a prepaid voucher, not a company credit card that covers anything with a Microsoft logo on it. The balance gets deducted only from eligible Azure service usage, and "eligible" turns out to be a much narrower category than most founders assume going in. That gap, between what a founder thinks their credit balance covers and what it actually pays for, is exactly where the surprise invoice comes from.

No single program treats every Azure service the same way. Coverage depends on which program granted the credits, what category the service falls under, and even which billing channel the charge came through. A virtual machine and a marketplace add-on can sit right next to each other in the Azure portal and get treated completely differently at checkout.

This isn't a trap Microsoft is setting. It's a budgeting discipline that startups need to build before they start spending, not after the first unexpected charge appears on a bill. Founders who map the boundaries early get a predictable runway. Founders who don't tend to find out the hard way, usually a few months in, when a bill lands that doesn't match the balance they thought they had left.

The credit programs available to startups in 2026

Microsoft rebuilt its startup credit program in the middle of 2025, and the change matters for anyone planning a budget around it.

Track one is for startups with investor backing. Companies that join through the Microsoft for Startups Investor Network start at $100,000 in Azure credits, with room for additional awards depending on referral source and how actively the startup engages with the program. Track two is for early-stage teams without outside funding; it caps out at $5,000 under the revised rules.

Multiple sources describe a milestone ladder: $200 on account creation, $5,000 after business verification, then $25,000 and $50,000 as workload adoption grows, with $150,000 cited as the highest standard milestone in some documentation. None of these steps happen automatically. Each one requires verified progress.

Founders need to slow down here. The $150,000 figure appears in multiple places, sometimes as the ceiling of the legacy Founders Hub program, sometimes as a milestone description in current materials. But The Register reported in July 2025 that the self-service track tops out at $5,000 under the new structure. Sources conflict on the current cap for the non-investor path, so readers should confirm current terms directly with Microsoft before budgeting.

Activation also depends on timing. Startups that joined before May 29, 2026 use the Microsoft for Startups portal to activate their credits, while anyone new to the program on or after that date goes through the Azure portal directly. If a startup's advisor or accelerator is describing an activation flow, check which one applies before assuming it matches.

Beyond the flagship program, there's a scattering of smaller credit pools. Visual Studio Enterprise carries $150 a month, MSDN Platforms $100 a month, Visual Studio Professional $50 a month. None of these replace the startup program, but they stack in useful ways for teams that qualify for more than one.

Every new Azure account also gets a baseline $200 one-time credit, usable in the first 30 days and covering most services. Once that window closes or the balance runs out, whichever comes first, pay-as-you-go pricing kicks in immediately.

For context, this isn't happening in a vacuum. Google Cloud Platform offers up to $200,000 for eligible startups and up to $350,000 for AI-focused ventures, and AWS offers up to $100,000. Founders weighing a primary cloud provider should know Azure's investor-backed track is competitive at $100,000 but not the largest offer on the table.

What Azure credits reliably cover across compute, storage, databases, and AI services

Set the confusion aside for a second: there's a solid core of services that credits cover reliably.

Compute is covered: Virtual Machines, App Services, and Azure Functions all draw from the credit balance. Storage is covered too, both Blob Storage and Managed Disks. On the database side, Azure SQL, Cosmos DB, PostgreSQL, and MySQL are all eligible, along with core networking costs like bandwidth and Load Balancers.

AI and machine learning services are where things get more specific, governed by a rule. Models sold directly by Azure and billed through the Azure subscription, including Azure OpenAI accessed via Foundry Models, Cognitive Services, Vision, and Speech APIs, are covered. Models from outside partners or community providers carry separate billing and are not automatically covered.

Fireworks on Microsoft Foundry is the clearest example of how this plays out in practice. It's a partner model that happens to be billed and sold through Azure, so startup credits apply to its pay-per-token usage, what Azure calls Data Zone Standard. But provisioned throughput units, the reserved-capacity option for that same model, are excluded from credit coverage entirely. Same model, two pricing structures, only one of them touches the credit balance.

Azure App Service deserves its own mention because it's a native Azure service, fully credit-eligible, and it supports ASP.NET, ASP.NET Core, Java, Node.js, PHP, and Python with built-in CI/CD. For a startup building a fairly standard application, compute, storage, managed databases, monitoring, and Azure-native AI services, credits can genuinely fund the entire technical foundation.

There's a simple test that cuts through most of the ambiguity here. If a service bills natively through the Azure subscription, with no separate marketplace transaction involved, it's generally eligible. Keep that test in mind, because the next section is built entirely around its exceptions.

The specific services and charges that credits do not cover

This is the list that actually decides whether a credit balance survives contact with a real production environment.

Azure Marketplace purchases sit at the top of it. Third-party software like MongoDB Atlas, Datadog, Elastic, Confluent, Palo Alto Firewall, and SendGrid plans, when billed through the marketplace, generally isn't deducted from Azure free or sponsorship credits. For most startups on the standard startup credit track, marketplace software means a separate bill, full stop.

Support Plans are excluded outright, and this one has no workaround. Regardless of credit type or balance size, support plans cannot be paid for using credits. Domain registration is excluded too. Buying a domain through Azure App Service falls outside credit coverage, which surprises a lot of founders since it's purchased from inside the same portal as everything else.

GitHub Enterprise access, where it's included, comes as a Founders Hub program benefit rather than something funded from the Azure credit balance itself. It's a perk bundled alongside credits, not a draw against them. And for nonprofits specifically, the $2,000 annual credit covers Azure workloads but not Azure Active Directory, which gets licensed separately through Enterprise Mobility and Security.

Then there's egress. Outbound data leaving Azure incurs charges, and moving data between Azure regions, say West US to East US, generates additional fees on top of that. Credits may apply to some networking costs, but at scale, egress becomes a real out-of-pocket line item because Azure charges for it regardless of whether the compute is covered.

Excluded charges accumulate on their own track, separate from the credit balance, so a founder checking only the credit dashboard is looking at half the picture. Credits are a staged infrastructure subsidy, not unrestricted cash, and treating them like the latter is how teams get blindsided.

The practical habit that prevents this is simple, if a little tedious. Before turning on any third-party integration through the Azure Marketplace, check whether it bypasses credit coverage. The billing channel, native Azure versus marketplace, is the single deciding factor, and it's worth checking every time, not just once at the start.

How AI and GPU workloads burn through credits faster than most founders expect

AI startups burn through credits differently than a standard web application does, because Azure's GPU pricing makes that burn especially fast.

Azure is consistently the priciest of the three major clouds for H100 on-demand access, running about $12.29 per GPU per hour. A full 8-GPU Azure ND H100 v5 node comes out to roughly $98 an hour. Spot pricing brings that same node down to about $18.17 an hour, an 82% reduction, but spot capacity isn't reliably available across every region, so it can't be treated as a guaranteed fallback.

The bigger shift, though, is where the spend actually goes once a model is live. Training gets most of the attention, but industry analysts put inference at 55 to 80% of enterprise AI GPU spend. Training is a cost that ends. Inference doesn't. Once a model ships, serving costs keep running for as long as the product is in use.

Utilization makes this worse, not better. The 2026 State of Kubernetes Optimization Report, drawing on more than 23,000 clusters, found average GPU utilization is just 5%, with AKS clusters specifically averaging 2%. Most GPU credit spend, in other words, is burning while the hardware sits idle. Idle notebook waste across a team of 10 ML engineers commonly reaches $500–$2,000 per month. Vision, Speech, OpenAI, and Cognitive APIs can drain credits quickly even when not running GPU workloads, as API call volume accumulates faster than teams expect.

Matching the VM series to the actual workload affects how quickly the credit balance depletes. Picking the wrong series for the job, running inference on training-grade hardware, for instance, is a quiet but constant drain on the credit balance. The EU-region GPU premium means EU-based GPUs are 10 to 30 percent more expensive than US-based ones, a factor relevant for startups with data-residency requirements. As a scale illustration, with three models in 24/7 production, inference costs reach $8,000–$13,000 per model per month at on-demand pricing.

Common mistakes that exhaust credits prematurely

Most credit exhaustion doesn't come from one dramatic overspend. It comes from a handful of small, repeated habits that compound over months.

Over-provisioning is the most common one. Teams spin up large VM instances "just to be safe," and those instances then run at low utilization indefinitely. Activating credits too early compounds the same problem: starting the credit clock before the infrastructure need is real just means credits get spent on exploration instead of production.

Then there's the mistake covered in the previous section already, assigning credits to purchases that were never going to be covered in the first place, marketplace software, support plans, and discovering the mismatch only when the invoice arrives.

On the technical side, quantization is one of the highest-leverage moves available. A quantized model using INT8 instead of FP32 needs roughly a quarter of the GPU memory and runs 2 to 3 times faster, with minimal quality loss for most enterprise use cases. NVIDIA TensorRT, LLM Compressor paired with vLLM, and Hugging Face Optimum automate most of that process, turning it into a configuration change rather than a research project.

Done systematically, GPU FinOps practices cut costs by 30 to 40% in documented cases, stretching a credit balance from lasting a quarter to lasting most of a year. That's not a marginal optimization. Applying it systematically stretches a credit balance from lasting a quarter to lasting most of a year.

Reserved Instances are worth a serious look too, especially for predictable workloads. Committing to a Reserved Instance instead of paying on-demand can lower costs by up to 72%, and credits can fund reservations on eligible services directly, which makes this one of the highest-leverage single decisions a founder can make with a fresh credit balance. Spot pricing offers real savings too, up to 60% on H100 instances on Azure, but the same availability caveat applies here as everywhere else: plan a fallback capacity path, because spot inventory isn't guaranteed.

How Azure pricing mechanics work under credits: egress, regions, storage tiers, and the charges credits don't absorb

Some costs are technically covered by credits and still manage to blow past what founders expect, because of how Azure's pricing mechanics calculate those charges.

Egress is the clearest example. Data coming into Azure is free. Data leaving Azure, or moving between Azure regions, is not, and that charge applies even when the compute generating the data is fully credit-covered. It's easy to miss because the compute bill looks fine while the networking line item quietly grows.

Picking the wrong tier for a dataset's actual access pattern is a leak on a storage credit line over months.

Redundancy settings add another layer. Locally Redundant Storage and Geo-Redundant Storage both get their compute costs covered by credits, but the redundancy level chosen scales the bill. GRS costs more than LRS for the same data, and that difference is on the founder to weigh against actual disaster-recovery needs.

For a sense of scale, a minimal workload gives a useful anchor. Five B2ats v2 virtual machines running for a month cost about $71.28 at $0.0198 an hour. Four million Azure Function executions a day for a month runs about $24.00 at $0.20 per million executions. Numbers like these help translate an abstract $5,000 balance into something concrete: 5 B2ats v2 VMs for a month runs approximately $71.28 at $0.0198/hr, and 4 million Azure Function executions per day for a month runs approximately $24.00 at $0.20 per million.

Credits draw down at the same rate as a paid account for every eligible charge. There's no monthly allowance resetting the clock. The balance depletes continuously, and a single high-utilization week, a training run, a traffic spike, a poorly rightsized cluster, can consume what a founder expected to stretch across an entire quarter. Regional pricing varies because Azure operates across many regions and pricing differs by location (EU-based infrastructure carries a 10 to 30 percent premium versus the US for GPU workloads, with the same principle applying more broadly to compute and storage). In terms of storage tier mechanics, Hot tier costs $0.018/GB and Cool tier costs $0.01/GB, and the right tier selection for access frequency materially affects how far storage credits stretch, while Archive storage is cheaper still but incurs retrieval fees.

Choosing infrastructure that makes the credit runway last and scales past it

The architecture decisions made while running on free credits are the same decisions a startup lives with once the credits run out. That stays true once the credits run out. A team that builds around managed services with opaque pricing during the credit period inherits a very different bill afterward than a team that architects for cost visibility from the start.

Shared-tenant platform-as-a-service options each come with tradeoffs. Serverless platforms built around Docker containers typically support stateless HTTP only, no multi-service Compose stacks, no persistent background processes, and only limited post-response work through mechanisms like Fluid compute. Some platforms in this category carry SOC 2 Type II certification and support HIPAA compliance through a signed BAA, which matters a great deal for startups in regulated spaces.

Bring-your-own-cloud platforms take a different approach that tends to outlast the credit period. BYOC deploys directly into a startup's own AWS, GCP, or Azure account, so the team keeps its own cloud economics, its own compliance posture, and full cost visibility, while still offloading cluster management, autoscaling, CVE patching, and CI/CD to the platform. Because the underlying resource costs stay transparent and controllable, this architecture doesn't hit a cliff the day the credit balance runs out. The bill after credits looks like the bill during credits, just without the subsidy.

Compliance work deserves a mention here too, because it's an easy way to burn credits on effort that never touches the product. Building SOC 2 or HIPAA readiness as a multi-month engineering project consumes credits on work that's necessary but not differentiating. Platforms offering one-click compliance configuration free up that credit spend for the actual product instead.

Azure is consistently the most expensive of the three major clouds for H100 on-demand, at ~$12.29/GPU/hr, with a full 8-GPU Azure ND H100 v5 node running roughly $98/hr.

The underlying principle holds regardless of which specific platform choices a team makes. Own the cloud account. Push the operational complexity, patching, scaling, cluster babysitting, onto someone else. Credits are worth the most when they fund an architecture a team would choose anyway, credits or no credits, because that's the architecture that survives the day the free balance hits zero. Azure App Service is credit-eligible and functions as a PaaS layer on Azure native infrastructure, relevant for teams that want PaaS simplicity without leaving the Azure credit ecosystem. For GPU workload placement, a hybrid strategy (specialized providers for core GPU training and inference cost optimization, hyperscale clouds including Azure for storage, APIs, and ecosystem services) is documented as a cost-reduction pattern for AI startups.