Microsoft Azure AI Revenue Runs on Three Tech Giants: OpenAI, TikTok, Meta

Microsoft's Azure AI cloud, despite serving 100,000 customers on its Foundry model marketplace, generates the bulk of its AI revenue from three companies: OpenAI, ByteDance — the Chinese parent of TikTok — and Meta Platforms. Bloomberg's Tech In Depth newsletter, published Monday, named all three explicitly for the first time in a single framing, drawing together disclosures accumulated over the past three weeks. The synthesis arrives two days before Nvidia reports its own quarterly results on August 26, the industry's most closely watched indicator of whether AI infrastructure spending reflects durable demand or a narrowing customer pool.

How Foundry Turns Tokens Into Revenue — and Why Scale Concentrates at the Top

To understand why three customers can dominate a platform serving 100,000, the billing model matters. Microsoft Foundry charges for AI model access in two ways: pay-as-you-go, billed per thousand input and output tokens, or Provisioned Throughput Units (PTUs), which reserve dedicated GPU-backed capacity at a fixed hourly rate. A token is roughly four characters of text; every prompt and every generated word is metered and charged at rates that vary across Foundry's catalog of more than 11,000 models from OpenAI, Anthropic, Meta's Llama family, Mistral, and others.

At the scale Meta reportedly operates — trillions of tokens processed through Azure each week — the arithmetic implies annual spending in the hundreds of millions, consistent with Bloomberg's characterization. A trillion tokens weekly at one dollar per million tokens equals roughly fifty-two million dollars annually; at five dollars per million for GPT-class models, the figure climbs above two hundred fifty million. ByteDance, which Bloomberg identified in June as Foundry's single largest spender and on track for more than one billion dollars annually on Microsoft AI and cloud services, operates at greater scale still. Microsoft does not sell these customers its own models — it resells third-party models and collects a margin or revenue share on every token consumed, regardless of model origin.

The OpenAI Loop: $24.1 Billion and a 70% Share

The most concentrated piece of the story involves Microsoft's closest partner. A Microsoft filing disclosed on August 5, 2026 revealed $24.1 billion in revenue from commercial arrangements with OpenAI for the fiscal year ended June 30 — encompassing both OpenAI's own Azure compute consumption and the revenue-sharing payments that flow back to Microsoft. Bloomberg estimated that figure represents approximately 70% of Microsoft's total AI business, which Satya Nadella had placed at a $37 billion annualized run rate. Microsoft has not published a formal segment breakdown confirming that ratio.

The circularity matters. Much of the $24.1 billion is OpenAI's own compute bill — the cost of running ChatGPT inference on Azure infrastructure — routed back to Microsoft as revenue. Microsoft invested more than $13 billion in OpenAI; OpenAI uses that capital to pay Microsoft for compute; Microsoft records those payments as AI growth. Against Microsoft's total FY2026 revenue of $331.8 billion, the OpenAI contribution is roughly 7.3% — significant but not dominant. Narrowed to the AI business, it is the defining fact.

The ex-OpenAI indicators tell a different story. On the Q4 FY2026 earnings call, CFO Amy Hood disclosed that commercial remaining performance obligation grew 25% excluding OpenAI, and commercial bookings rose 18% on the same basis, versus headline Azure growth of 43%. Those stripped-down figures are the more meaningful gauge of how broadly enterprise AI adoption is actually spreading.

ByteDance at the Top: The Billion-Dollar Policy Problem

ByteDance's position as Foundry's largest non-OpenAI spender carries a dimension that neither Microsoft's growth metrics nor its earnings calls illuminate. OpenAI and Anthropic have declined to offer direct API access in mainland China over intellectual property and security concerns. Microsoft fills that gap — selling access to OpenAI's GPT model series through Azure to ByteDance and other Chinese firms including Ant Group, Meituan, and Tencent, with routing typically outside mainland Chinese data centers.

The exposure is asymmetric. Chinese government agencies have been accelerating the removal of Microsoft products from state systems — five of six Chinese government procurement guides reviewed by Reuters through May 2026 either do not recommend or restrict Microsoft products. Yet ByteDance is on track to spend more than one billion dollars annually on Microsoft's AI and cloud services — an arrangement that critics have described as monetizing the gap between Washington's stated AI containment posture and what commercial cloud contracts permit. Any tightening of export controls on advanced AI model access to Chinese entities would fall directly on this revenue line.

Meta's Paradox: Buying the Rival It Plans to Compete With

Meta's position carries a structural tension of a different kind. The company raised its full-year capital expenditure forecast to $130–$145 billion, builds its own frontier Llama models, and has confirmed plans to launch an AI API service that would compete directly with Foundry. And yet it is simultaneously spending hundreds of millions annually on Foundry model access, according to Bloomberg. Bloomberg reported that Meta developers have used OpenAI models purchased through Microsoft to evaluate outputs from Meta's own in-house models — paying Microsoft, which resells OpenAI, to grade their own homework. CTO Andrew Bosworth confirmed publicly in July on the Big Technology podcast that Meta rents leading external AI models as part of its development process alongside building its own.

The historical pattern is instructive. Microsoft's Bing powered web search on Facebook for years; Meta built its own replacement by 2014. Microsoft supplied Meta with AI compute before ChatGPT launched. The cycle — buy it, internalize it, stop — has repeated before. Meta's third-party cloud spending is an operating expense sitting outside its capex figures, so the Azure bill runs in parallel with its own infrastructure buildout rather than instead of it.

What Azure's 43% Growth Rate Is Actually Measuring

Together, these three customers define a circular AI economy: Microsoft invests in OpenAI; OpenAI pays Azure for compute; Microsoft books the payment as AI revenue; ByteDance buys OpenAI models through Azure that OpenAI won't sell it directly; Meta buys the same models to evaluate its own competing systems. Microsoft collects on every leg.

The implication for interpreting Microsoft's numbers is direct. Azure cloud services revenue rose 43% in the most recent quarter — a real growth rate, but one whose quality differs depending on which customers are driving it. A number anchored primarily by a single partner whose compute bill is funded by Microsoft's own investment capital, plus a Chinese firm facing live export-control risk, plus a customer building a competing marketplace, describes a structurally different business than 43% growth spread across independent enterprise buyers in manufacturing, healthcare, and finance — the sectors Microsoft features prominently in Foundry marketing. Deutsche Bank analysts warned in August of "some concentration risk" tied to the OpenAI relationship, even while recommending the stock.

Three Forces That Will Determine Whether the Concentration Holds

Three developments will test whether the customer base broadens over the next twelve months. First, OpenAI exclusivity has already loosened: a restructuring completed in April 2026 allows OpenAI to route certain workloads to other cloud providers, modestly reducing the lock-in that has kept nearly all OpenAI inference on Azure. Second, US export control policy toward China remains a live variable; any restriction on advanced AI model access for Chinese entities would directly affect ByteDance's spending trajectory. Third, Meta's progress on its own API marketplace will determine whether it remains a major Foundry customer or follows the pattern it has set with Microsoft services before.

For investors and enterprise architects, the most informative number is one Microsoft now provides but does not headline: commercial RPO growth excluding OpenAI. At 25%, it is real but materially lower than the headline rate. Nvidia's earnings on August 26 will add a second data point — whether the hyperscalers buying GPUs to serve these token workloads are still expanding or beginning to absorb what they have built. That answer will say as much about whether AI demand is broad-based as any platform customer count.