Microsoft Moves Bing Image Creator In-House, Citing PowerPoint GPU Savings

Microsoft’s cost claim has a defined but incomplete comparison

Microsoft says MAI-Image-2.5 reduces GPU costs by up to 84% in PowerPoint compared with GPT-Image-2. That is a company claim about GPU costs, not a disclosed total-cost or public quality benchmark: the announcement gives no workload mix, utilization profile, time window or matched quality standard for the comparison.

Mustafa Suleyman, executive vice president and chief executive of Microsoft AI, made the operational argument more directly in a report of his interview, describing Microsoft’s models as faster, cheaper and higher quality and putting the PowerPoint cost advantage at about 85%. Suleyman co-founded DeepMind and Inflection before joining Microsoft in 2024 to lead the new Microsoft AI organization, as Microsoft’s organizational announcement records. Bing, Edge, Copilot and Microsoft’s GenAI team moved under that organization, giving him authority over the consumer-AI products now using these models.

The economic incentive does not depend on Microsoft paying an ordinary per-model license fee. The report says Microsoft has long had free access to OpenAI’s models under its partnership, while still needing to provide costly computing infrastructure to run them. That makes Microsoft’s stated GPU comparison relevant even though it is not independently reproducible from the published material.

Other deployment figures add context but not independent validation. Microsoft says MAI-Image-2.5 raised OneDrive image-editing save rates by 26%, cut P95 latency by about 25% and delivered 2.5 times greater efficiency under medium-utilization production workloads. It also says MAI-Voice-2-Flash, a separate text-to-speech model, is powering Dynamics 365 Contact Center with GPU costs reduced by up to 89%. Those are company-reported results with different products and measures; they cannot be combined into a single saving rate for Microsoft’s AI operations.

The OpenAI relationship still sets the boundary

This is not evidence of a clean Microsoft–OpenAI split. In its October 2025 partnership update, OpenAI said Microsoft’s investment in OpenAI Group PBC was valued at about $135 billion and represented roughly 27% on an as-converted diluted basis after recapitalization. The agreement kept OpenAI as Microsoft’s frontier-model partner and preserved Microsoft’s exclusive IP rights and Azure API exclusivity until AGI, while also allowing each company more room to pursue innovation independently.

The same agreement says OpenAI contracted to buy an additional $250 billion in Azure services, although Microsoft no longer has a right of first refusal to be OpenAI’s compute provider. That infrastructure commitment illustrates why an in-house image deployment and a continued OpenAI relationship can coexist: the sources show a change in selected product workloads, alongside an ongoing investment, technology-rights and cloud-services arrangement.

One secondary account calls the announcement a minor shift away from OpenAI, but it also incorrectly describes MAI-Image-2.5-Pro as the model powering Bing. Microsoft’s primary announcement instead names MAI-Image-2.5. The primary account is the stronger evidence for the exact model in production; neither source supports a claim that all Bing, PowerPoint or Microsoft AI features have moved in-house.

What the next deployment needs to show

The next useful evidence is a like-for-like account of the PowerPoint comparison: the image tasks measured, the quality thresholds, the utilization range and whether the stated GPU savings survive at broader demand. Microsoft has supplied a directional, product-specific number, not the information needed to independently test the claimed 84% reduction.

For users, the immediate test is whether Bing and PowerPoint’s generation and editing experience holds up as Microsoft changes the underlying model. For Microsoft, the strategic question is narrower than an OpenAI divorce: which additional workloads can an in-house model handle at the desired quality, speed and cost while the company remains financially and operationally tied to OpenAI.