Microsoft is preparing to add China's Moonshot AI "Kimi K3" to Azure and is evaluating whether it can be used for some Copilot features, The Information reported. If adopted, it could reduce inference costs currently spent on OpenAI and Anthropic models by up to $600 million, according to the report. However, Microsoft has not disclosed either the figure or whether it will adopt the model, and Kimi K3 is not yet listed in Azure's current official catalog. Framing this move as a "full replacement of GPT and Claude" would miss the model mix-and-match approach that Microsoft is already pursuing.
The $600 Million Figure Reflects an Evaluation Stage, Not a "Replacement"
According to The Information's July 20 report, Microsoft is working to offer Kimi K3 through the Azure cloud. Copilot team members plan to test whether some features currently handled by OpenAI and Anthropic models could be shifted to Kimi K3. The company reportedly tested DeepSeek and earlier versions of Kimi in the past as well.
The up-to-$600-million figure is said to be an internal estimate arising from this evaluation. It remains unclear what time period it covers, which Copilot products or features are included, or what percentage of workload is assumed to shift to Kimi K3. Microsoft has not officially confirmed the estimate, so the savings figure cannot be treated as finalized.
Publicly available Azure information is also consistent with the idea that this is still at the evaluation stage. Microsoft Foundry's pricing page lists Kimi K2 Thinking, K2.5, and K2.6, but as of July 21, K3 is not included. Since Microsoft already offers Moonshot AI models for enterprise use, adding K3 itself would be an extension of an existing trajectory. However, there remains a process of verifying quality and safety between simply selling a model on Azure and actually incorporating it into Microsoft's own Copilot.
The Mechanism for Running 2.8 Trillion Parameters Cheaply
Kimi K3 has a total of 2.8 trillion parameters and supports a 1-million-token context window along with image input. In terms of scale alone, it is not a lightweight model. The key to reducing inference costs lies in its Mixture of Experts (MoE) architecture, which doesn't activate all parameters for every input, and caching that reduces recomputation for long inputs.
According to Moonshot AI, Kimi K3 effectively activates 16 out of 896 expert networks. The company's official API pricing is $0.30 per million tokens for cached input, $3 for non-cached input, and $15 for output. The company also states it achieved a cache hit rate exceeding 90% for coding tasks.
However, Microsoft's $600 million figure cannot be recalculated from this pricing table. The published pricing reflects the cost of using Moonshot AI's own API, and does not represent the hardware costs of running the model on Azure infrastructure or the actual input/output token volume Copilot would consume. Furthermore, Moonshot AI recommends supernodes combining 64 or more accelerators to extract efficiency. Even when the model operates sparsely, infrastructure is still needed to hold 2.8 trillion weights and move data between the 896 expert networks.
Even so, Microsoft has a strong incentive to fine-tune per-inference costs. In fiscal Q3 2026, paid seats for Microsoft 365 Copilot exceeded 20 million. The company states that through software and hardware optimizations, it has increased inference throughput for its primary Copilot models by 40%. As the user base grows, the percentage of workload that can be safely routed to cheaper models becomes a key determinant of profitability.
Copilot Already Selects Models Per Task
Official documentation for Microsoft 365 Copilot Coworker lists GPT 5.5 alongside Claude Sonnet 5, Claude Opus 4.8, and Claude Fable 5, among others. The default "Auto" setting automatically selects a model suited to the request. Microsoft also states that it may add other AI models that it hosts and operates itself to Copilot, in which case data will not leave Microsoft's environment.
The same principle is implemented even more explicitly in Microsoft Foundry, which is aimed at developers. The Model Router's Balanced mode compares candidates within roughly 1-2% of the model judged to offer the highest quality for a given input, then selects the one that minimizes cost. Cost mode widens the acceptable quality gap to roughly 5-6%. Quality mode, conversely, disregards cost entirely and routes to the model judged highest quality on evaluation metrics.
Given this product design, it's easy to envision a scenario where Kimi K3 does not wholesale replace GPT and Claude. Instead, the likely path is reallocating tasks such as classification and routine Q&A—where K3 meets quality thresholds—while leaving more difficult reasoning tasks to pricier models. The validity of the $600 million estimate, too, will hinge less on K3's published unit pricing and more on what percentage of Copilot's overall workload can actually be shifted.
Quality and Policy Hurdles Remain Beyond July 27
Moonshot AI announced Kimi K3 on July 16, but the full public release of model weights is scheduled for July 27. Obtaining the weights will be a milestone for Microsoft as it examines the model in detail on its own infrastructure and optimizes it for Azure. While the model is currently accessible through Kimi's official service, the conditions for deploying it as an open-weight model are not yet fully in place.
Moonshot AI itself acknowledges quality-related limitations. Kimi K3 may produce significantly unstable output in execution environments that fail to properly carry over reasoning history, or in scenarios where the model is switched from a different model mid-conversation. The company also notes the model has a tendency to act beyond user intent when given ambiguous requests, leaving a gap in user experience compared to Claude Fable 5 and GPT 5.6 Sol. In architectures that switch models mid-conversation, adoption will hinge less on benchmark scores and more on the handoff of conversational state and permission management.
Given that a Chinese company's model would be embedded in a core product of a U.S. company, policy scrutiny is unavoidable. That said, publicly available federal policy does not appear to contain provisions directly banning the use of Chinese-made models in private-sector products like Copilot. The June 2026 National Security Presidential Memorandum permits the adoption of commercial and open-source AI while requiring verification of reliability and controllability for national security applications, along with contractual and technical conditions ensuring adversaries cannot cause shutdowns or degradation. A separate White House document also explicitly states that no blanket licensing or pre-approval requirements will be imposed on the development or release of AI models.
Therefore, the possibility of political backlash and the existence of rules that would actually block adoption need to be considered separately. If Microsoft can host Kimi K3 within its own environment, maintain customer data boundaries, and ensure it cannot be shut down externally, open weights could actually reduce dependency on the supplier. Whether K3 gets added to Azure's catalog after July 27, and whether Copilot's official model list and quality metrics subsequently change, will be the conditions that transform the $600 million estimate into an actual business decision.