A growing number of startup founders are raising concerns about unexpected charges incurred while experimenting with AI models through Microsoft’s Azure AI Foundry platform, turning what began as an isolated complaint into a broader debate over billing transparency.
At least 20 participants in the Microsoft for Startups program have signed a Change.org petition calling on Microsoft CEO Satya Nadella to address what they describe as a “billing trap” inside Azure AI Foundry, arguing that the platform’s interface makes it difficult to distinguish between services covered by startup credits and third-party models that incur direct charges.
“Azure AI Foundry displays both Microsoft-native models (such as Azure OpenAI) and third-party Marketplace models (such as Anthropic Claude) in a completely unified interface — with no visual distinction, no warning, and no confirmation step before charges are incurred,” the petitioners wrote.
The petition, which goes to the extent of claiming that Microsoft had breached the founders’ trust, was drafted by Takyua Tominaga, founder of Tokyo-headquartered startup Leach, who was one of the first to report the billing issue in a detailed blog post.
In the post, Tominaga wrote that he was unaware of any billing for model use until his credit card statement arrived containing a charge of about $1,600 for the use of one of Anthropic’s models.
The founder further wrote that contacting Microsoft support about the issue via the Azure portal was an arduous task, as the portal wouldn’t let him get directly in touch or report the issue.
After he did get through to Azure Support on X via direct message, he was directed to fine print in Microsoft’s documentation that says that startup credits cannot be used for Microsoft Azure support plans, third-party branded products, products sold through Microsoft Azure Marketplace, or products otherwise sold separately from Microsoft Azure.
When Tominaga pushed further, he wrote that he was offered a partial refund via credits worth $1,000, which he rejected, and was directed to contact Anthropic with any further refund request.
Anthropic responded to Tominaga by saying that it does not have visibility into usage through Microsoft Foundry and was unable to process a refund.
Tominaga is not alone. Riyaj Shaikh, a systems architect at EPAM Systems in Pune, said in a post on X that he had encountered a similar situation, and that attempts to resolve the billing issues appeared to bounce between the two companies, with each pointing to the other as the appropriate party to handle refunds.
In fact, Shaikh, in the same thread, pointed out that Microsoft’s own moderators are not certain about how large language models are billed as part of the Microsoft for Startups program, and pointed to a post on Microsoft Learn’s official Q&A forum.
In response to a question from a forum user, a moderator had confirmed that startup credits could be used for deploying Claude Opus 4-5 via Azure AI Foundry.
The post was later amended to indicate that startup credits don’t apply.
Shaikh told InfoWorld in an email that he and his team have yet to receive any refund, and have instead been bounced back and forth between Anthropic and Microsoft, with each pointing them to the other for resolution.
Bogdan Sevriukov, founder of AI-based workforce training startup Comprenders, confirmed that he, too, was facing a similar issue over a charge of €999.60 (about $1,147).
Microsoft, which offers up to $150,000 in Azure credits to early-stage ventures through the Microsoft for Startups initiative, did not respond to an email seeking comment on the complaints and asking whether it plans to modify the way third-party AI models are presented and billed within Azure AI Foundry.
The petitioners say that relatively small design changes to Azure AI Foundry’s UI could prevent similar incidents, and are urging Microsoft to introduce clearer labeling, explicit billing warnings, and confirmation prompts before developers deploy third-party models.
These changes, they argue, would help ensure that startups experimenting with AI prototypes do not inadvertently incur unexpected charges and exhaust their budgets.
Go to Source
Author:
