ByteDance in Talks to Spend US$500 Million More on Azure for OpenAI Models

OpenAI has warned Microsoft that Chinese firms could distill its models through Azure, and a House committee has asked Microsoft for a briefing on the talks.

WHY IT MATTERS

Any limit Congress or OpenAI places on Microsoft's resale terms could cut ByteDance off from OpenAI models that OpenAI restricts for most Chinese companies.

ByteDance is in advanced talks to spend about US$500 million more on Microsoft’s Azure cloud, mostly to use OpenAI’s models, The Information reported on November 18, citing two people familiar with the discussions. The money would add to an existing Azure contract. The talks have angered OpenAI, which has told Microsoft in recent months that Chinese companies could be using its models to train their own. Neither ByteDance nor Microsoft has confirmed a deal, and KrASIA said it is unclear which ByteDance products would use the extra capacity.

The report exposes a gap between the two partners’ rules. OpenAI blocks most Chinese companies from its models, but Microsoft, which holds the rights to sell OpenAI technology through Azure, does not apply the same restriction. That leaves Azure as a route for Chinese companies to reach OpenAI’s models at commercial scale.

The talks drew a political response the same day. The House Select Committee on China asked Microsoft on November 18 for a bipartisan briefing on the potential arrangement, KrASIA reported. "Microsoft should not sell AI tools to ByteDance," committee chairman John Moolenaar, a Michigan Republican, said in a statement, adding that Microsoft’s Pentagon contracts made the deal especially troubling.

Microsoft says it found no sign that ByteDance distills OpenAI models

The core complaint is distillation, the practice of feeding a model’s answers into the training of a rival model. According to IT Home, which summarized The Information’s report, OpenAI told Microsoft that some Chinese companies might use Azure to acquire advanced capabilities cheaply this way. Microsoft executives replied that they had found no evidence ByteDance was doing so and believed the company mainly used OpenAI models for tasks such as software development.

TNW noted that The Information’s report does not say ByteDance itself has distilled OpenAI’s models. IT Home added that ByteDance founder Zhang Yiming has said in the past that the company would not use distillation as a shortcut in AI development.

Microsoft has defended its position in two statements. In a spokesperson’s statement quoted by IT Home, the company said its agreement with OpenAI lets Microsoft build OpenAI models, alongside its own MAI models, into its first-party services, and that it enforces its terms of service, follows government requirements and respects its partner agreements. Microsoft told Reuters, as cited by KrASIA, that it takes misuse of its AI services seriously and does not discuss confidential customer negotiations. ByteDance has not commented publicly.

ByteDance already spends heavily on Microsoft, and its other compute routes are narrowing

The additional US$500 million would build on a large existing relationship. IT Home cited a June Bloomberg report that ByteDance expected to spend more than US$1 billion a year on Microsoft’s AI and cloud services. ByteDance also builds its own models, including the Seedream image models, TNW noted.

The Azure talks come as US rules on Chinese access to computing power tighten. On November 16, the Commerce Department issued an interim final rule, effective November 19 according to KrASIA, that requires licenses for overseas data centers providing large amounts of computing power to China-headquartered companies. Certain existing arrangements have 90 days to obtain authorization or stop.

The pressure is already reaching ByteDance’s supply chain. TNW, citing Bloomberg, reported that banks have paused work on a US$480 million loan for Nvidia chips that ByteDance would use. Singapore-based Aolani Cloud sought the loan to buy processors for a data center in Malaysia, and bankers at the five lenders involved discussed pulling out, with some calling the loan dead. Aolani said it was working with regulators and lenders to deliver the project within the new rule; DBS, which is arranging the loan, declined to comment.

Distillation is moving from a commercial dispute to a policy question

OpenAI’s complaint lands one day after Reuters found that Chinese developers train on US model outputs at industrial scale, based on a review of more than 300 papers, patents and filings. TNW also cited an earlier report that DeepSeek and Moonshot AI had trained new models on outputs from OpenAI’s GPT-6 Astra.

Moolenaar also criticized US technology companies for supporting Chinese AI while working with Washington, according to KrASIA, and the committee’s request puts Microsoft’s licensing terms for OpenAI models under congressional review. KrASIA noted that the committee’s intervention does not by itself make the deal unlawful.

Three things would change the picture: a signed contract disclosed by either company, a change in the terms under which Microsoft resells OpenAI models to Chinese customers, or a Microsoft briefing to the House committee that leads lawmakers to propose limits on cloud access to US frontier models for Chinese firms.