The new enterprise AI reality: performance is cheap, usage is not
Enterprise AI teams are moving away from exclusive reliance on expensive proprietary large language models toward fine-tuned, open source AI models that deliver similar performance at much lower cost, driven by rising agent workloads, metered pricing, and finance departments demanding predictable enterprise AI costs. This is not a minor procurement tweak; it is a power shift. Once AI stops being a novelty add-on and becomes daily infrastructure, the brand prestige of frontier models matters less than whether you can afford to run agents all day without blowing the budget. The uncomfortable truth for vendors is clear: the bottleneck is no longer model quality, it is the inference bill. Buyers have noticed, and they are starting to act like they do with any other cloud resource—by optimizing away anything that looks like an overpriced commodity.
Microsoft’s DeepSeek bet: open source as a cost effective LLM
Microsoft’s exploration of a self-hosted DeepSeek V4 is the clearest sign yet that proprietary models no longer own enterprise mindshare on quality alone. The company is considering a fine-tuned, Azure-hosted version of DeepSeek V4 (or another open-source AI model) as a cheaper option beside Anthropic and OpenAI systems. DeepSeek V4 has been delivering strong performance at much lower prices, sometimes a fraction of what frontier models cost per token. That sentence should worry every premium LLM vendor. When a cost effective LLM can be dropped into the same Copilot workflows and run entirely inside Microsoft’s cloud with enterprise protections and data controls, the argument for paying frontier prices turns into a luxury choice, not a necessity. In short: Microsoft is opening the door for customers to ask whether they really need the most expensive model for everyday work.
Usage-based AI pricing models: GitHub shows the meter is here to stay
The pricing story explains why open-source alternatives are suddenly so attractive. GitHub switched Copilot from its old premium request unit system to GitHub AI Credits on June 1, tying usage to token consumption across input, output and cached tokens, with different costs per model. For Copilot Business, the list price is still USD 19 (approx. RM87.40) per user per month with USD 19 (approx. RM87.40) in monthly AI Credits; Copilot Enterprise remains USD 39 (approx. RM179.40) with USD 39 (approx. RM179.40) in credits. According to that report, users saw an immediate impact: some developers posted screenshots showing they burned through large chunks of their monthly allowance within days, and one Reddit user projected an USD 847 (approx. RM3,896.20) bill after previously paying USD 39 (approx. RM179.40) for Copilot Pro+. Developers who treated Copilot like an all-you-can-eat tool are being asked to watch the meter. Once every long session has a visible price, managers will ask harder questions and workers will ration prompts.
Copilot Cowork and agentic AI: when every worker has a running meter
Two weeks after the GitHub change, Microsoft pushed the same logic into office work by making Copilot Cowork generally available on June 16 and shifting it to usage-based pricing. This agent can sort emails, prep meetings and pull reports across Outlook, Word, Excel and Teams, maintaining context through long chains of tasks. That power is exactly why the old flat seat-based model was unsustainable: a short email rewrite is not the same product as a multi-hour agent session that opens files, calls tools, checks its own work and keeps going until the job is done. If the model works harder, you pay more. GitHub’s internal characterization of June as its "best month ever" after the pricing change shows demand is real, but the subsidy phase is ending. Finance teams now need to know what a normal task costs, what an expensive task looks like, and which groups are quietly converting subscriptions into open-ended compute spend. In this world, any cheaper DeepSeek alternative starts to look less like an experiment and more like a necessity.
From brand loyalty to cost discipline: the next phase of enterprise AI
The pattern is clear: once AI coding and office agents turn into daily infrastructure, flat-rate plans stop matching either the cost or the demand. "That can’t hold once agents are doing real work. If you let heavy users run without limits, somebody pays the inference bill." By offering model choices and metered billing, Microsoft is trying to strike a better balance so more businesses can use these tools without sticker shock. But this is no longer about one vendor’s margins. For startups and enterprise buyers, treating it as a small pricing tweak is a mistake; it is a procurement problem. Open source AI models like a self-hosted DeepSeek V4 give enterprises a way to push back: swap in a cheaper, cost effective LLM where frontier branding adds little value, set usage caps, pool credits across teams, and make every agent justify its meter. The winners in the next phase of enterprise AI will not be the flashiest models; they will be the ones that help customers keep the lights on without fear of the next invoice.
