Bloomberg: Meta Consumes Trillions of Tokens Weekly, Quietly Becoming Microsoft's Major AI Customer

Meta Platforms (META-US) has become one of Microsoft's (MSFT-US) largest artificial intelligence (AI) customers, underscoring how demand for this emerging technology remains heavily concentrated within the tech sector.

According to Bloomberg's report on Thursday (20th), sources revealed that Meta spends hundreds of millions of dollars annually using AI models via Microsoft Azure's cloud service, with its weekly AI computing volume through Azure already reaching trillions of tokens (word units).

A key part of Microsoft's AI strategy is Foundry, a model marketplace offering AI models from various providers. As of July this year, Foundry had reached 100,000 customers. Microsoft frequently showcases clients from traditional industries such as manufacturing and transportation in its marketing campaigns.

However, sources indicate that Microsoft’s largest AI customers are still predominantly other technology companies, including Adobe, AI firm Perplexity, and Sierra—a customer service AI startup co-founded by OpenAI chairman Bret Taylor.

Microsoft particularly relies on a small number of tech industry clients. Its long-term partner OpenAI contributed approximately 70% of Microsoft’s total AI revenue in the most recent fiscal year. In addition to providing access to AI models, Microsoft also offers the AI assistant Copilot and rents out AI-optimized computing resources.

In recent years, over-concentration of revenue and circular transactions among enterprises have been persistent concerns in the tech industry. For AI to meet market expectations, the technology must achieve broad adoption across the entire economy—not just within powerful tech companies.

Sources say Meta is making significant investments in AI to assist software development. Meta acquires model usage rights across multiple platforms based on model availability and cost.

One source noted that Meta developers previously used OpenAI’s technology via Foundry to help evaluate the outputs of their in-house developed models.