이미지: METAL LAB 생성
Summary
- Bloomberg reports Meta Platforms has become one of Microsoft Azure's biggest AI customers
- Meta spends hundreds of millions of dollars annually and processes trillions of tokens per week, benchmarking its own models against OpenAI's models on the Foundry marketplace
- At the same time, Meta is building its own API service to compete with Foundry, while about 70% of Microsoft's AI revenue comes from OpenAI alone
- 핵심 보도
- 메타가 마이크로소프트 애저의 최대 AI 고객 중 하나로 부상 (블룸버그)
- 연간 지출
- 수억 달러 규모로 애저 AI 모델 접근
- 토큰 처리량
- 매주 수조 개 토큰을 애저에서 처리
- 벤치마킹 방식
- 메타 엔지니어들이 파운드리 마켓플레이스의 오픈AI 모델로 자사 모델 성능 비교
- 메타의 대응
- 파운드리와 경쟁 가능한 자체 API 서비스 구축 중
- 파운드리 고객 순위
- 바이트댄스 1위, 어도비·퍼플렉시티·시에라 순
- 마이크로소프트 AI 매출 구조
- 약 70%가 오픈AI의 애저 컴퓨팅 구매에서 발생
A company running trillions of tokens a week on Azure
Bloomberg reports that Meta Platforms has become one of the largest AI customers of Microsoft's cloud service, Azure. Meta reportedly spends hundreds of millions of dollars a year to access AI models through Azure, processing trillions of tokens on the platform every week. The two companies have long been intertwined in a buyer-seller relationship over AI infrastructure, but this is the first time the scale of that relationship has been disclosed in such concrete terms.
Benchmarking its own models against OpenAI's
According to The Decoder, Meta's engineers pull OpenAI models from Microsoft's AI model marketplace, Foundry. The purpose is to benchmark the performance of Meta's own in-house models against OpenAI's. At the same time, Meta is building its own API service that overlaps with Foundry. The Decoder noted that this pattern resembles Meta's past history of relying on Microsoft's Bing search engine before eventually replacing it with its own technology.
In other words, Meta is currently both a Microsoft customer and a potential competitor in the making. According to The Decoder, this pattern — borrowing something while building your own version to eventually switch over — is one Meta has repeated in both search and cloud.
ByteDance is Foundry's top customer
Some customers on Microsoft Foundry's list spend even more than Meta. According to the ranking cited by The Decoder, ByteDance ranks first, followed by Adobe, Perplexity, and Sierra. Most of these companies, like Meta, share the trait of having their own in-house AI models and services.
| Rank | Customer | Note |
|---|---|---|
| 1 | ByteDance | Foundry's top customer |
| 2 | Adobe | - |
| 3 | Perplexity | - |
| 4 | Sierra | - |
| - | Meta | Among top spenders, with hundreds of millions annually |
Still, Foundry accounts for only a small share of Microsoft's overall AI revenue. About 70% of that revenue comes from OpenAI alone, largely driven by OpenAI's massive compute purchases through Azure. Spending from Foundry customers like Meta and ByteDance is minor compared to that OpenAI share.
Why circular-deal concerns are resurfacing
Bloomberg noted that this disclosure revives concerns about circular dealing in the AI industry. Microsoft is both a major shareholder in and cloud provider for OpenAI, while Meta borrows OpenAI's models on Microsoft's platform to compare against its own. Money, models, and infrastructure are effectively cycling among a handful of companies. The report also highlighted how AI demand remains concentrated among a very small number of major tech firms.
Editor's View
What matters here isn't the mere fact that Meta uses OpenAI's models — it's why. Borrowing a leading model to serve as a benchmark for validating your own model's performance isn't unusual. What stands out is that Meta is simultaneously preparing a product to compete with the very company supplying that yardstick — though this too is a familiar pattern in the cloud industry. It's much like continuing to buy NVIDIA GPUs while building your own chips: renting what you need now, while working to no longer need to rent it later.
From Microsoft's perspective, this setup cuts both ways. Money that a major customer like Meta spends on Foundry counts as revenue today, but what that customer learns along the way is ultimately how to operate without Microsoft. The fact that OpenAI accounts for 70% of Microsoft's revenue reads more heavily in that light. Customers passing through Foundry — ByteDance, Adobe, Perplexity, Meta — are all companies that could shift to their own infrastructure at any time, and have strong incentives to do so. This report underscores just how dependent Microsoft's AI business remains on a single company, OpenAI.
There's a separate lesson here for domestic companies. Borrowing another company's model for benchmarking purposes is common practice among AI teams in Korea as well. But once that spending scales into the hundreds of millions of dollars a year, a point comes where the break-even calculation for shifting to in-house infrastructure has to be made. For companies without Meta's scale, that point arrives much sooner — the threshold being the moment benchmarking costs exceed the cost of in-house development.
In the coming months, the shape of Meta's in-house API service is likely to become clearer. If that service reaches a level where it genuinely competes with Foundry, Microsoft will find itself facing one of its largest customers as a rival.
Comments