
Tech • IA • Crypto
Microsoft is expanding Copilot into a multi-model, usage-based enterprise AI platform and is considering China’s DeepSeek as a lower-cost option within its new Copilot Co-work system.
Microsoft has made Copilot Co-work generally available worldwide after a three-month preview used by over half of the Fortune 500. The system is designed for “agentic” workflows, meaning it can execute complex, multi-step tasks autonomously across files, tools, and internal data rather than responding to single prompts.
Early adopters including Accenture, Capital Group, and Zurich Insurance used Co-work for tasks such as analyzing sales pipelines, comparing thousands of files, and editing large datasets. One example involved reviewing nearly 4,000 files, work that would typically take weeks manually.
Microsoft is moving to a usage-based billing model after internal tests showed some users ran hundreds of tasks weekly, driving costs sharply higher. Pricing is now based on “Copilot credits,” costing about $0.01 per credit, with task expenses determined by model usage, tool calls, data retrieval, and runtime.
Tasks are classified as light, medium, or heavy, depending on reasoning depth, number of data sources, and outputs generated. Companies are encouraged to estimate costs based on employee roles and usage frequency, reflecting the growing economic challenge of scaling AI agents.
Co-work currently uses models from OpenAI and Anthropic, including GPT 5.5, Opus 4.8, and Sonnet 4.6. Microsoft is also preparing its own model, Co-work 1, optimized for lower-cost, everyday enterprise tasks.
Microsoft is exploring integrating a fine-tuned version of China’s DeepSeek V4 as an optional model for cost-sensitive workloads. The model would be hosted on Azure, with enterprise security, compliance, and data residency controls applied.
The potential inclusion of DeepSeek does not replace existing partners but reflects a broader shift toward a multi-model routing system, where tasks are assigned to different models based on cost, performance, and security requirements.
Microsoft has built a significant AI business serving Chinese firms such as ByteDance, Tencent, and Ant Group. ByteDance alone is reportedly on track to spend over $1 billion annually on Microsoft AI and cloud services.
While competitors restrict direct access to advanced models in China, Microsoft provides access via Azure, though models are hosted خارج China in locations like Singapore. Concerns persist over model distillation and intellectual property risks.
Despite China accounting for only about 1.5% of Microsoft’s total revenue, Azure AI revenue in the region has surged, reportedly tripling in one year after a prior 400% increase, making it one of the fastest-growing markets.
Microsoft also introduced Web IQ, a Bing-based system optimized for AI agents rather than human users. It delivers structured, low-latency web data and is claimed to be 2.5 times faster than alternatives, though it currently remains in limited preview.
Co-work integrates models, search, enterprise data, security, and billing into a unified platform. Features such as audit logs, compliance tools, and admin controls position it as a comprehensive enterprise AI operating layer, not just an assistant.
Microsoft is transforming Copilot into a flexible, multi-model enterprise platform where cost, performance, and geopolitics intersect, signaling a more complex and globally interconnected AI ecosystem.