Meta Platforms has become one of Microsoft’s largest AI (artificial intelligence) customers, highlighting that demand for this emerging AI technology remains largely concentrated in the technology industry for now.

An insider said Meta spends hundreds of millions of dollars every year to access AI models through Microsoft Azure cloud services. It added that the compute power Meta uses each week via Azure has reached the level of tens of trillions of tokens.

An important part of Microsoft’s AI strategy is to provide AI models from different vendors through a model marketplace called Foundry. As of this past July, Foundry had 100,000 customers. In promotional materials, Microsoft often showcases customers from traditional industries such as manufacturing and transportation.

But according to insiders, Microsoft’s largest AI clients currently are still mostly other technology companies, including Adobe, the AI company Perplexity, and the customer-service AI startup Sierra, which was co-founded by OpenAI board chair Bret Taylor.

Microsoft is especially reliant on a small number of technology-industry clients. Its long-time partner OpenAI contributed about 70% of Microsoft’s overall AI revenue in its most recent fiscal year. In addition to offering model access services, Microsoft has launched its AI assistant Copilot and rents out AI-focused computing resources.

In recent years, overconcentration of revenue and the “roundabout transactions” among companies have been troubling the tech industry. Media analysis believes that for AI to truly meet market expectations, this technology must be widely adopted throughout the broader economy, rather than relying only on tech companies with strong financial resources.

Insiders say Meta is investing heavily in AI to help with software development. The company will purchase model usage rights through multiple platforms, depending on the availability of different models and their costs.

One insider said that Meta’s developers had used OpenAI’s technology via Foundry to help evaluate the output quality of their in-house models.

Meta not only develops its own AI models, but also rents other leading models during the R&D process. This was described by Meta’s chief technology officer, Andrew Bosworth, on a podcast earlier this July.

At the same time, Meta is also building its own AI model services business—so-called API services—which may in the future compete with Foundry. This could help Meta reduce spending on external services such as Microsoft.

In the past, Meta also relied on Microsoft’s technology, and later replaced it with an in-house solution. In the late 2000s, Microsoft’s search engine Bing provided Facebook with web search support. But by the end of 2014, Meta had stopped using Bing.

In the years before ChatGPT was released and sparked the current AI boom, Microsoft also provided Meta with the computing capacity needed for AI development. But now, Meta has become one of the world’s largest builders of AI data centers.