Microsoft Is Building AI Without OpenAI

Microsoft just shipped three proprietary AI models through a new platform called Microsoft Foundry. MAI-Transcribe-1 for speech-to-text. MAI-Voice-1 for text-to-speech. MAI-Image-2 for image generation. No OpenAI involvement. Built entirely in-house.

Microsoft, the company that invested $13 billion in OpenAI and built its entire Copilot strategy around GPT models, decided it needed its own.

If your AI stack runs on Azure OpenAI Service, this changes your vendor risk calculus.

What Microsoft Foundry Changes

	Before Foundry	After Foundry
Microsoft’s AI role	Distribution channel for OpenAI	AI model vendor in its own right
Speech-to-text	Resold OpenAI Whisper	MAI-Transcribe-1 (proprietary)
Text-to-speech	Resold OpenAI TTS	MAI-Voice-1 (proprietary)
Image generation	Resold DALL-E / partnered with OpenAI	MAI-Image-2 (proprietary)
Vendor dependency	Single-source (OpenAI)	Multi-source (OpenAI + in-house)
Enterprise pricing control	Limited by OpenAI’s margins	Full control on MAI models
Competitive positioning	Tied to OpenAI’s roadmap	Independent product roadmap

The short version: Microsoft went from reselling OpenAI’s models to competing with them. In the same product catalog. On the same cloud platform.

Why Microsoft Built These Models Now

A few things converged in the last six months to explain the timing.

OpenAI started competing with Azure. OpenAI’s enterprise API business grew 4x in 2025. They’re selling directly to the same Fortune 500 accounts that Microsoft’s Azure sales teams call on. When your distribution partner becomes your competitor, you build alternatives. That’s not paranoia. That’s procurement 101.

Google and Anthropic closed the quality gap. I wrote about Claude winning enterprise accounts last month. Gemini Pro 2.5 is strong. The market shifted from “OpenAI or nothing” to a genuine three-way race. Microsoft sitting behind OpenAI exclusively started looking like a single point of failure, not a competitive advantage.

The OpenAI IPO is coming. Once OpenAI goes public, its fiduciary duty shifts to shareholders. Pricing decisions, roadmap priorities, and partnership terms all get filtered through quarterly earnings pressure. Microsoft clearly decided it couldn’t build a long-term platform strategy on a model provider whose incentives are about to change. I flagged this exact risk in my analysis of OpenAI’s Sora shutdown.

None of these reasons are speculative. Microsoft’s own behavior tells the story. You don’t build three production-ready models and launch a new platform brand (Foundry) as a hobby project.

What Are the MAI Models?

MAI-Transcribe-1 handles speech-to-text. This competes directly with OpenAI’s Whisper, which Microsoft has been reselling through Azure AI Services. If you’re running call center transcription, meeting summaries, or voice-based data entry on Azure, you now have two options where you had one. Early benchmarks suggest comparable accuracy to Whisper large-v3 with lower latency on Azure infrastructure (which makes sense given it was optimized for that hardware).

MAI-Voice-1 does text-to-speech. This one goes after OpenAI’s TTS models and Google’s WaveNet. Voice synthesis for customer-facing applications, IVR systems, accessibility features, content narration. The voice quality samples Microsoft published are good. Not the best I’ve heard, but solidly in the top tier.

MAI-Image-2 generates images. Competes with DALL-E 3 (which Microsoft was reselling) and Google’s Imagen. This is the most crowded category of the three, but it’s also where Microsoft had the most obvious dependency problem. Every image generated through Bing Image Creator or Copilot’s visual features was flowing through OpenAI’s infrastructure and pricing.

The naming convention tells you something too. “MAI” for Microsoft AI, numbered sequentially. This isn’t a one-off experiment. It’s a product line with a roadmap. Expect MAI-Text and MAI-Code models within 12 months.

What Microsoft Foundry Actually Is

Foundry is the platform brand for Microsoft’s proprietary AI models. Think of it as Microsoft’s answer to the question “where do your non-OpenAI models live?”

Until now, Azure AI Services was essentially an OpenAI storefront with some legacy Microsoft models (speech, vision, language) that predated the partnership. Foundry creates a separate identity for Microsoft’s new AI development. Same Azure infrastructure. Different product line. Different pricing structure. Different roadmap.

For enterprise buyers, Foundry means your Azure AI contract now includes models from two competing AI organizations. Microsoft controls pricing, SLAs, and feature development on the MAI models. They negotiate with OpenAI on everything else.

That split matters when your procurement team is trying to forecast AI costs for 2027.

How This Hits Your Enterprise AI Stack

If you’re deep in a Microsoft 365 and Azure environment (and statistically, about 70% of enterprises are), this creates three immediate decisions.

1. Model Selection Gets Harder

You used to pick Azure OpenAI Service and get whatever OpenAI offered. Now you have MAI models for speech, voice, and image tasks alongside OpenAI’s offerings on the same platform. Microsoft isn’t publishing head-to-head comparisons (of course they’re not). You’ll need to run your own benchmarks on your own workloads.

I’ve been recommending model-agnostic architectures for exactly this reason. If your application code is tightly coupled to a specific model’s API format, every new model choice means rewriting integration code. If you built an abstraction layer, you swap endpoints and run tests. The companies that listened are about to have a very good quarter.

2. Pricing Pressure Works in Your Favor

Microsoft now has an incentive to price MAI models aggressively against OpenAI’s offerings. They keep more margin on proprietary models. They control the roadmap. Every customer running MAI models is a customer whose renewal isn’t subject to OpenAI’s pricing decisions.

For your next Azure AI contract negotiation, Foundry is a gift. You can play Microsoft’s models against OpenAI’s models on the same invoice. “We’ll shift our transcription workload to MAI-Transcribe-1 unless the Whisper pricing comes down.” That conversation wasn’t possible three months ago.

If you’re spending $10K+/month on Azure AI services, your account team already knows about Foundry. Call them. The discount conversation is different now than it was in Q1.

3. Vendor Risk Just Got More Complicated

Here’s the part that doesn’t make the Microsoft press release.

You’re no longer dependent on one AI model provider through Azure. You’re now dependent on a platform vendor that’s managing a complicated (and increasingly competitive) relationship with its primary model supplier. Microsoft and OpenAI are partners, competitors, and co-dependent all at once. The Microsoft board member who sits on OpenAI’s board is watching Microsoft build models that directly compete with OpenAI’s products.

That tension doesn’t mean the partnership collapses tomorrow. It means the partnership is no longer something you can take for granted when making five-year infrastructure decisions.

The practical response: build abstraction layers. Use the governance toolkit Microsoft just open-sourced to manage model selection at the policy level. Don’t hard-code model dependencies into your application logic. If the Microsoft-OpenAI relationship changes (and at some point, it will), you want to be the company that swaps a configuration value instead of the company that rewrites 50 microservices.

Who Should Care Most About This

Enterprise Architects on Azure

If Azure is your primary cloud and you’re running production AI workloads, Foundry adds options and complexity simultaneously. Start benchmarking MAI models against your current OpenAI deployments this quarter. You don’t need to switch. You need to know your options before your next contract renewal.

CTOs Making Build-vs-Buy Decisions

The AI model market just got another major player with deep enterprise distribution. Factor Foundry into your vendor evaluation matrix. Three credible model providers (OpenAI, Anthropic, Google) just became four. That’s better for pricing. It’s also more work for your team.

Procurement and Finance

Your Azure AI line item is about to get more granular. MAI model usage and OpenAI model usage may carry different pricing, different SLAs, and different data handling terms. Make sure your tracking can distinguish between them before you lose visibility into what you’re actually paying for.

What This Means for the Broader AI Market

Microsoft building proprietary models confirms what the last six months have been signaling: the era of exclusive AI partnerships is over. Google has Gemini and distributes third-party models through Vertex. Amazon has Nova and resells Anthropic and Meta models through Bedrock. Microsoft now has MAI and resells OpenAI through Azure.

Every major cloud provider is running the same playbook. Build your own models for margin and control. Resell partners’ models for breadth and customer retention. Compete and collaborate simultaneously.

For businesses choosing where to run AI workloads, this is good. More competition means better pricing and more options. For businesses planning their AI architecture, it means the model-agnostic approach I’ve been pushing isn’t cautious. It’s table stakes.

And for anyone still building their entire AI strategy around a single model provider? Foundry is your wake-up call. Microsoft itself just told you that betting everything on one model vendor is a risk they’re no longer willing to take.

Why would you?

Your Next Steps

1. Audit your Azure AI usage. Pull 90 days of billing. Separate OpenAI model usage from other Azure AI services. Know your baseline before MAI models hit general availability.
2. Test MAI models against your workloads. When Foundry access opens up (preview access is rolling out now), run your speech, voice, and image workloads against the MAI alternatives. Document quality and latency differences.
3. Build or verify your abstraction layer. If your code calls OpenAI endpoints directly, add a routing layer. You want model selection to be a config change, not a code change.
4. Renegotiate your Azure AI contract. If renewal is within six months, use Foundry as a pricing conversation starter. Microsoft wants you on MAI models. Make them earn it.
5. Watch for MAI-Text and MAI-Code. Speech, voice, and image are the opening move. When Microsoft launches proprietary LLMs, the competitive dynamics with OpenAI shift again. Plan for it now.

The Microsoft-OpenAI partnership isn’t dead. But Microsoft just made clear that it’s not the only plan anymore. Your AI strategy should reflect the same thinking.

---

Related Reading:

• Your AI Stack Has an Expiration Date
• ChatGPT Built the Market. Claude Is Winning It.
• OpenAI Killed Sora. Your AI Budget Is Next.
• Microsoft’s Free Toolkit Fixes Your Agent Governance Problem