Meta is building its own data centers, training its own models, and preparing to sell access to artificial intelligence to others, yet it still sends hundreds of millions of dollars to a competitor every year. Bloomberg reported that Mark Zuckerberg’s company has become one of Microsoft’s largest AI customers, purchasing access to models through the Azure cloud service and processing trillions of tokens through it every week—units used to bill for artificial intelligence workloads. The agency obtained this information from a person familiar with the matter who wished to remain anonymous, and neither Meta nor Microsoft wanted to comment.
How much Meta spends on it
Hundreds of millions of dollars a year sounds almost negligible compared with Meta’s total spending, as the company spent more than $31 billion on building its own infrastructure in the three months through the end of June alone, and in the summer it also raised its full-year estimate to $130 billion to $145 billion from the originally planned $115 billion to $135 billion. However, the Azure bill is not included in these figures because purchases of third-party cloud services are accounted for as ordinary operating expenses, so they fall outside both amounts mentioned.
Bloomberg did not specify which models were involved, the prices Meta pays for them, or the contractual terms it has agreed to, nor did it disclose what share of Microsoft’s model sales revenue comes from Meta. Both figures—the spending and the token volume—come from a single anonymous source, and the agency also did not provide the relevant accounting period.
Why Meta needs third-party models
This is not about running the assistant on Instagram or anything users can see directly, because Meta purchases access to models primarily for software development and obtains them from several platforms depending on where they are currently available and at what price. Developers therefore choose based on price and capacity rather than the logo, while according to Bloomberg, Meta’s programmers used OpenAI technology purchased through Microsoft to evaluate the outputs of Meta’s own models. Meta therefore pays Microsoft, which resells OpenAI technology, and uses it to check its own work.
Meta Chief Technology Officer Andrew Bosworth described it this way on the Big Technology podcast in July. According to him, in addition to developing its own models, the company also rents the best third-party ones because they are part of the development process. If you want to know whether your model measures up, you need something to compare it against.
Microsoft’s customers
Microsoft offers models from various providers on its Foundry marketplace, which had one hundred thousand customers in July. In promotional materials, the company likes to showcase manufacturers or transportation companies, but the reality is different. The biggest spender on Foundry over the long term has been China’s ByteDance, the owner of TikTok, and Meta has now begun to catch up. Other major customers include Adobe, the Perplexity search engine, and Sierra, a customer-support startup co-founded by OpenAI board chair Bret Taylor. Two of the largest customers are therefore social networks that are also developing their own models.
OpenAI itself accounted for roughly seventy percent of all Microsoft’s artificial intelligence revenue in the last fiscal year. For the enormous investments in AI to pay off, the technology must be purchased by the entire economy, not just by technology companies selling it to one another.
Foundry’s revenue more than doubled in the year through July. The number of customers using models from several providers simultaneously on the marketplace has increased fivefold since the beginning of the year. The number of those consuming a trillion tokens annually has quadrupled. Meta belongs to this group, with the difference that it processes trillions of tokens in a single week. The Azure division’s total revenue surpassed $100 billion at the end of June and grew by forty-one percent. Microsoft Chief Financial Officer Amy Hood acknowledged at the time that the system still had limitations and demand exceeded capacity.
No one knows exactly how much of Meta’s money ends up directly in Microsoft’s accounts. Neither company will disclose it. The only certainty is that part of the billions Meta spends on artificial intelligence appears on the other side as revenue for a competitor.
Meta is building the same thing it buys
Zuckerberg confirmed in July that a cloud-based artificial intelligence business makes sense for his company. Meta has its own computing capacity and could rent it out to others. According to Bloomberg, it is also preparing a service through which it would sell access to various AI models. Such an offering would compete directly with Foundry while allowing Meta to reduce its spending on third-party services.
A similar scenario has already played out once. Microsoft’s Bing search engine powered search on Facebook from the end of the first decade of this century, but Meta discontinued it at the end of 2014 and deployed its own solution. Microsoft also supplied the company with computing power for artificial intelligence development in the years before the arrival of ChatGPT. Today, Meta is among the world’s largest builders of AI data centers.
Sources: investing.com and finance.yahoo.com
