Token Economics: The Hidden Driver Behind Model Choice
Enterprise AI costs refer to the total spend businesses incur to process tokens, run agentic workflows, and integrate large models into everyday software, and these costs are increasingly reshaping which AI systems companies deploy at scale. As models handle longer prompts, multi-step tools and code-heavy tasks, token usage explodes, turning pricing and quotas into hard constraints rather than fine print. Microsoft’s Copilot Cowork is moving from flat fees to metered billing, tying customer bills directly to the volume of tokens consumed instead of simple seat counts. Against this backdrop, OpenAI token pricing and Anthropic’s tariffs have risen while some plans add stricter token caps, leaving enterprises squeezed between enthusiastic internal adoption and fast-burning AI budgets. The result is a powerful incentive to seek cheaper, good-enough alternatives that can support large, continuous workloads without constant spending shocks.
DeepSeek V4 Model: A Cheaper Engine for Copilot Cowork
Microsoft is now testing a self-hosted version of the DeepSeek V4 model as a lower-cost engine for Copilot Cowork, the enterprise agent that currently leans on OpenAI and Anthropic. DeepSeek’s stack builds on open-source foundations, offering a flexible path for Microsoft to tune and operate it on Azure infrastructure while keeping data and governance inside its own cloud. The move responds directly to soaring token bills and stories of companies draining annual AI budgets in months as employees experiment with long prompts and agent loops. With Copilot Cowork priced per token, switching to DeepSeek V4 promises meaningful savings for coding and workflow-heavy scenarios where context windows are large and outputs are lengthy. This shift does not eliminate OpenAI or Anthropic from Microsoft Azure AI, but it introduces a cheaper default that can handle many day-to-day tasks without premium model costs.
Azure as a Two-Way Bridge for US and Chinese Models
Microsoft Azure AI has quietly become a two-way bridge, distributing American models into China and Chinese models into Western enterprises. Bloomberg reporting shows that Microsoft is the main supplier of GPT models to major Chinese internet firms through Azure, even though OpenAI and Anthropic do not sell into that market themselves. ByteDance has emerged as Microsoft’s largest AI customer, expected to spend more than USD 1 billion (approx. RM4.6 billion) a year on OpenAI-based services delivered via Azure. At the same time, Azure AI Foundry now lists DeepSeek R1 and is testing a fine-tuned DeepSeek V4 model for Copilot Cowork. According to AI News, Azure’s AI revenue in China has tripled in one financial year after growing about 400% the year before, underscoring how token-hungry workloads and Azure’s unique resale contract with OpenAI have made Microsoft the central broker of cross-border AI capacity.
How DeepSeek Pricing Rewrites the AI Model Hierarchy
The arrival of DeepSeek V4 as a cheaper, capable option forces enterprises to reconsider which tasks really require top-tier proprietary models. Many coding assistants, document workflows and knowledge-management tools care more about context length, reliability and price than about squeezing out marginal gains in reasoning benchmarks. By offering DeepSeek V4 inside Microsoft Azure AI, alongside GPT and other frontier systems, Microsoft is turning model choice into an economic decision rather than a pure performance race. For customers under pressure from soaring enterprise AI costs, this means building architectures where high-end OpenAI models are reserved for critical reasoning tasks, while cheaper models handle summarisation, routine chat and bulk content generation. Over time, this tiered approach could erode the perceived dominance of any single model family and encourage an ecosystem in which open-source-derived models capture a growing share of practical enterprise workloads.
Policy Contradictions and the Future of Enterprise AI
Microsoft’s dual-market strategy exposes an awkward tension between policy alarms and commercial incentives. Legislators worry about advanced AI flowing into rival ecosystems, while OpenAI and Anthropic avoid direct exposure by refusing to sell their models in certain markets. Yet Microsoft not only resells GPT models abroad under its contract, it also hosts DeepSeek models for Western customers and experiments with DeepSeek V4 as a core enterprise engine. To limit risk, Microsoft keeps OpenAI models outside local data centres and sells primarily to established companies, but distillation and synthetic data remain hard to police at scale. As token costs climb and enterprises demand cheaper options, economic gravity pulls Microsoft toward models like DeepSeek that reduce spend. This creates a feedback loop where financial pressures, not policy preferences, shape which models dominate daily business workflows and how the AI landscape is distributed.
