Microsoft Tests Routing Copilot Inference to Kimi K3｜~$600M/Year Savings, But Only for Non-Core Tasks First
TL;DR
Microsoft is internally testing routing part of Copilot inference to Moonshot Kimi K3, with internal estimates of up to $600M/year savings.
Techstartups reported on July 20, corroborated by The Information — Microsoft is internally testing Moonshot AI's Kimi K3, evaluating routing part of Copilot AI assistant inference from OpenAI and Anthropic models to Kimi K3.
The savings number — internal estimates suggest up to $600M/year in cloud infrastructure savings if a portion of requests is switched. Against Copilot's scale (Microsoft's FY26 AI revenue projected at $25B) that's not huge, but it improves margins by 2-3 percentage points.
Cadence — Microsoft expects initial technical validation in the next two months, plans to follow. No final replacement decision yet.
The real logic is task tiering — even if adopted, use will likely start with non-core, low-sensitivity tasks: email summaries, calendar clean-up, simple Q&A. Complex reasoning, multi-turn dialogues, code generation stay with GPT-5.6 Sol or Claude Opus. "Cheap model for volume, expensive model for quality" tiering.
Three real blockers — data sovereignty (Kimi inference runs primarily in China, Microsoft enterprise customers can't accept that), export controls (US bans some Chinese AI models for government and sensitive industries), safety capability (Kimi's jailbreak rate still exceeds Claude).
The workaround — download weights and self-host on Azure. Kimi K3 is open-weight. Microsoft can deploy directly on Azure GPUs without depending on Moonshot's inference service. Data stays inside Azure, export controversy minimized.
Strategic layer — Microsoft is simultaneously OpenAI's 49% shareholder, Anthropic's deep partner (Claude already on Azure Foundry), and now hugging Kimi K3. Three flagships backed up — reducing negotiation leverage against any single vendor. With OpenAI Q4 contract talks looming, this is Microsoft stacking chips.
Anthropic just took some Microsoft Copilot inference last week — the Kimi rumour also pressures Anthropic's pricing envelope one turn.
via Techstartups / The Information
The savings number — internal estimates suggest up to $600M/year in cloud infrastructure savings if a portion of requests is switched. Against Copilot's scale (Microsoft's FY26 AI revenue projected at $25B) that's not huge, but it improves margins by 2-3 percentage points.
Cadence — Microsoft expects initial technical validation in the next two months, plans to follow. No final replacement decision yet.
The real logic is task tiering — even if adopted, use will likely start with non-core, low-sensitivity tasks: email summaries, calendar clean-up, simple Q&A. Complex reasoning, multi-turn dialogues, code generation stay with GPT-5.6 Sol or Claude Opus. "Cheap model for volume, expensive model for quality" tiering.
Three real blockers — data sovereignty (Kimi inference runs primarily in China, Microsoft enterprise customers can't accept that), export controls (US bans some Chinese AI models for government and sensitive industries), safety capability (Kimi's jailbreak rate still exceeds Claude).
The workaround — download weights and self-host on Azure. Kimi K3 is open-weight. Microsoft can deploy directly on Azure GPUs without depending on Moonshot's inference service. Data stays inside Azure, export controversy minimized.
Strategic layer — Microsoft is simultaneously OpenAI's 49% shareholder, Anthropic's deep partner (Claude already on Azure Foundry), and now hugging Kimi K3. Three flagships backed up — reducing negotiation leverage against any single vendor. With OpenAI Q4 contract talks looming, this is Microsoft stacking chips.
Anthropic just took some Microsoft Copilot inference last week — the Kimi rumour also pressures Anthropic's pricing envelope one turn.
via Techstartups / The Information
