Token Cost Shock: The New Constraint on Enterprise AI
The enterprise shift from premium AI models to lower-cost alternatives such as the DeepSeek V4 model is driven by rising AI token costs that make large-scale, agentic workloads financially unsustainable for many organizations. In modern AI systems, a token is the smallest unit of text the model processes, and everything from input prompts to generated responses consumes tokens. As enterprises move from simple chatbots to complex coding assistants and multi-step agents, token consumption can spiral, especially in coding-related tasks where long context windows and iterative loops are common. Reports of companies burning through annual AI budgets in a few months have made finance teams far more sensitive to AI model pricing and usage patterns. At the same time, leading vendors like OpenAI and Anthropic are raising enterprise prices while adding caps on total tokens, pushing cost-conscious CIOs to seek an OpenAI alternative that offers predictable, scalable enterprise AI spending.
DeepSeek V4: Cheaper Model, Competitive Performance
DeepSeek V4 has emerged as a compelling option for enterprises that need capable models without runaway bills, especially as agent workflows grow more complex. While its architecture is based on open-source foundations, Microsoft is considering a self-hosted version to power Copilot Cowork, its enterprise agent that currently runs on OpenAI and Anthropic models. That move signals confidence that the DeepSeek V4 model can deliver competitive performance at a lower AI token cost profile. The appeal is not just headline AI model pricing, but the economics of sustained, metered usage: when customers pay per token instead of a flat seat fee, every reduction in cost per token directly extends how much automation they can afford. DeepSeek’s recent USD 7.4 billion (approx. RM34.0 billion) fundraise at a USD 50 billion (approx. RM230.0 billion) valuation suggests investors see room for a scaled, price-aggressive challenger in this space.
Microsoft’s Strategic Tension: Sell OpenAI, Adopt DeepSeek
Microsoft now sits in an unusual position. It is both the largest reseller of OpenAI models to enterprises and a potential adopter of DeepSeek V4 as a cheaper engine for its own Copilot Cowork service. According to AI News, Azure’s AI revenue in China roughly tripled in the financial year to June 2025 after growing about 400% the year before, driven by customers such as ByteDance, which is on track to spend more than USD 1 billion (approx. RM4.6 billion) a year on Microsoft’s AI and cloud services. Yet Axios reporting, cited in Wccftech’s coverage, indicates Microsoft is testing a fine-tuned, Azure-hosted DeepSeek-V4 as a lower-cost alternative to the OpenAI and Anthropic models that currently power its enterprise agent. That strategic tension is clear: Microsoft profits from selling premium GPT-based services while also feeling the same AI token cost pressure its customers face internally.
China Market Access Gives Microsoft a Unique DeepSeek Angle
Microsoft’s AI strategy is also shaped by its unusual position in China’s cloud market. It is the main supplier of OpenAI models there, selling GPT series systems through Azure to internet giants such as ByteDance, Ant Group, Meituan and Tencent, even though OpenAI and Anthropic do not sell directly. At the same time, Microsoft has added Chinese-origin DeepSeek models like R1 and DeepSeek-V4 to Azure AI Foundry for global customers, while keeping OpenAI models physically hosted outside China and delivered over the network. The result is a two-way trade: American models flowing into Chinese companies, and Chinese models flowing into Western enterprises, with Microsoft taking the margin on both. As AI token costs climb and distillation risks increase, that position will attract more regulatory scrutiny and could make DeepSeek either an important partner or a political liability inside Microsoft’s broader AI portfolio.
How Cost Pressures Could Reshape AI Vendor Lock-in
The DeepSeek V4 model is not only a cheaper OpenAI alternative; it also highlights a deeper shift in how enterprises think about AI vendor lock-in. Usage-based billing for AI tokens makes finance teams wary of being tied to a single high-priced provider, especially when that provider can raise prices or tighten token limits. In response, Microsoft’s tests with a self-hosted DeepSeek V4 for Copilot Cowork hint at a future in which large platforms route workloads dynamically across multiple models based on price-performance and risk. If open-source-derived models continue to narrow the quality gap, CIOs will feel more confident adopting multi-model strategies and negotiating harder on AI model pricing. Over time, this could erode the dominance of a few premium vendors and reward platforms that abstract away the underlying model, giving enterprises more room to control AI token costs and switch providers when economics change.
