Microsoft and Meta declined to comment to Bloomberg. The dollar figure, token volume, Meta’s rank among customers, and the claim that Meta developers use OpenAI models in Azure AI Foundry to assess Meta’s own outputs therefore remain single-sourced reporting, rather than disclosures either company has confirmed. Still, the broad strategy is no secret: Meta CTO Andrew Bosworth said in a July Big Technology interview that Meta rents outside models where their availability, price or capabilities make sense, while retaining an in-house model as its strategic backstop.
For Windows and enterprise IT readers, this is less a story about Meta suddenly adopting Microsoft’s AI stack than a clearer look at what Azure AI Foundry is becoming. It is a high-volume procurement layer for models and inference capacity, including for organizations that may never standardize on Microsoft’s own models or applications.
Meta is buying optionality, not outsourcing its AI future
Bloomberg reports that Meta is using Azure-based models as part of software development, including model evaluation. That use case is consequential because model evaluation needs an independent or at least differently trained judge: a company testing its own model can compare answers, coding output, safety behavior and instruction-following against a frontier model from another supplier.
An enterprise can apply the same logic. A team building an internal coding agent does not necessarily need to pick one model forever. It can use one model for generation, a second for review, a smaller one for routing and classification, and a third for high-stakes evaluation. The differentiator then becomes the control plane around those calls: identity, logging, access to internal data, budget enforcement, regional availability and a mechanism to swap providers without rebuilding an application.
Microsoft has been explicitly positioning Foundry around that model-agnostic role. In its fiscal 2026 first-quarter earnings call, Microsoft said the service offered developers access to more than 11,000 models, including OpenAI’s GPT-5 and xAI’s Grok 4. The company’s later adoption material also describes model swapping and fine-tuning as a core Foundry capability.
Meta’s apparent spend validates the proposition in a particularly revealing way. A company with some of the largest AI data-center investments in the industry is still willing to pay an outside cloud for access to selected models. Owning GPU capacity does not guarantee that the best model for a given benchmark, coding task or evaluation workflow is available internally at the needed time and price.
That is a more practical lesson than the familiar claim that every business should “build its own AI.” Even Meta is reportedly mixing owned infrastructure, proprietary models and rented third-party models.
Foundry’s customer count does not answer who drives the revenue
Bloomberg says Foundry had 100,000 customers as of July and reports that ByteDance has generally been its largest spender, with Adobe, Perplexity and Sierra also among large customers. Microsoft has publicized rapid adoption and growing high-value Foundry commitments in earnings calls, but it does not disclose a customer-by-customer revenue breakdown or confirm Bloomberg’s ranking.
The distinction matters because a six-figure customer count can coexist with revenue that is heavily concentrated in a small number of very large AI buyers. Microsoft said in its fiscal 2026 second-quarter call that the number of Foundry customers spending more than $1 million per quarter had grown nearly 80% year over year. That signals meaningful enterprise demand, but it does not establish how much of the service’s revenue comes from conventional corporate deployments versus large model developers and internet platforms.
Microsoft’s own recent disclosures have made the concentration issue harder to ignore. Bloomberg Law reported on August 5 that OpenAI accounted for the majority of Microsoft’s AI revenue in fiscal 2026, based on newly available company disclosures. Bloomberg’s analysis estimated the share at around 70%; that percentage is an analysis, not a number Microsoft itself has published as a formal segment metric.
Meta would not change that picture overnight. Hundreds of millions of dollars is substantial revenue by normal enterprise-software standards, but it is small next to a relationship measured in tens of billions. What Meta does provide is a second kind of major customer: a well-capitalized platform company paying for model access while it builds its own model family and prepares to sell that family through an API.
In other words, Microsoft may be diversifying away from a single marquee partner, but much of the near-term diversification still appears to be within the same technology sector.
Meta’s API push could turn a customer into a direct competitor
Meta is not merely an internal AI consumer. In April it announced that its Muse Spark model would be available through a private API preview, and in July it expanded that effort with a public preview of the Meta Model API. Meta said developers could use Muse Spark 1.1 through that service; in August, it added Muse Spark 1.2 and broader global access through its API and its Muse Code product.
That makes the relationship with Microsoft structurally temporary in at least one respect. Meta can purchase external model access today while it develops a commercial API that competes for many of the same workloads that Azure AI Foundry hopes to broker. The two companies can be customer and competitor simultaneously because the market has separate layers:
- Meta can consume models from Azure for internal development and evaluation.
- Microsoft can charge for inference, model access and its surrounding cloud services.
- Meta can sell its own models to third-party developers through the Meta Model API.
- Both companies can compete for AI compute and API customers without either arrangement immediately displacing the other.
This is not unprecedented for Meta and Microsoft. Meta previously used Microsoft’s Bing to power web-search features on Facebook, then phased out that dependency in 2014, as Reuters reported at the time. The older example is not a prediction that Meta will abandon Azure AI Foundry, but it does demonstrate a pattern: Meta will rent a strategic capability while it is useful, then replace it if an internal alternative becomes viable.
The current relationship is more complex because Foundry can offer models Meta does not own. Even a stronger Meta API would not automatically eliminate the need to evaluate against, or occasionally deploy, OpenAI, Anthropic, Google or other third-party models. But it does weaken any assumption that today’s token spend represents a durable, captive workload.
The enterprise signal is multi-model operations
Microsoft reported in July that Azure annual revenue had surpassed $100 billion, while its capital expenditures rose sharply to support demand for cloud and AI offerings. Meta, meanwhile, raised its expected 2026 capital expenditures to between $130 billion and $145 billion in its July results. Both are spending at a scale that makes outside AI consumption look counterintuitive only if cloud competition is treated as a winner-take-all contest.
It is better understood as a capacity-and-capability market. Frontier models differ in latency, cost, tool use, coding performance, regional availability and permissible data handling. Capacity can also be constrained at the moment a development team needs it. A sophisticated buyer will keep more than one route to inference, even when it owns large infrastructure itself.
That poses a concrete administrative challenge for organizations following the same pattern on a smaller scale. A multi-model approach is useful only when governance follows the workload. IT teams should know which applications are permitted to send prompts to which provider, where inputs and outputs are retained, whether an API call may include customer data, how model versions are pinned and tested, and who owns the bill when token usage rises unexpectedly.
Azure AI Foundry can help centralize those decisions for organizations already standardized on Azure. Meta’s reported behavior is a reminder that centralization should not be confused with exclusivity. The platform’s value is in making model choice manageable, not in ensuring that one supplier wins every request.
Microsoft gains a major customer if Bloomberg’s reporting is correct. More importantly, it gains evidence that the AI market’s largest builders are also becoming its largest renters. That can drive Foundry revenue now, but the contracts will be judged by whether Microsoft continues to offer the best combination of model access, capacity, controls and price after those customers’ own APIs mature.