I write a lot about the enormous cost of artificial intelligence.
The world’s biggest technology companies are spending hundreds of billions of dollars building data centers and securing enough electricity to power them.
But there’s another AI cost that businesses are starting to pay much closer attention to.
The cost of actually using it.
That’s especially true for AI agents, which can spend hours researching a topic, writing code or completing complicated assignments. Along the way, they might make hundreds or even thousands of requests to an AI model.
And somebody has to pay for all that computing power.
But what if I told you that one AI model recently completed an assignment for roughly 98% less than one of its leading competitors?
You’d probably think that kind of savings makes that model a no-brainer.
But there’s a catch.
And I believe that catch tells us something important about where the AI industry is headed next.
The Half-Penny AI
AI developer Dheeraj Sharma recently conducted an experiment where he ran the same research task through three different models using Claude Code as the agentic framework.
Claude Opus 5 completed the job for about 32 cents.
Kimi K3, a massive open-weight model developed by Chinese AI startup Moonshot AI, cost roughly 25 cents.
But DeepSeek V4-Flash cost less than half a penny.
Image: Dheeraj Sharma
Not only that, but DeepSeek was also the fastest of the three.
Of course, one test doesn’t prove DeepSeek is always 98% cheaper or that it can match Claude on every task.
But this result is part of a much bigger trend.
Over the past year, Chinese AI companies have been rapidly closing the performance gap with their American rivals. Models from DeepSeek, Alibaba and Moonshot AI now compete with some of the best American systems on a growing number of tasks.
And they’re often doing it for a fraction of the price.
A few pennies might not matter if you’re asking an AI a single question. But AI agents can make hundreds or even thousands of model calls while completing a task.
And as Uber found out this year, when companies start running AI agents at scale, that price difference can become enormous.
Microsoft is currently wrestling with this issue. Its new Copilot Cowork system can make repeated model calls while completing a task, which means usage costs can climb quickly.
That’s why Microsoft has explored the idea of using DeepSeek V4 inside Cowork, partly because it could lower those costs.
But there’s an obvious problem.
DeepSeek is based in China. And its privacy policy says information submitted through its services can be processed and stored there.
That creates a serious concern for Microsoft and any other American company that handles proprietary research, financial records, customer data or other sensitive information.
That’s exactly the kind of information you want to keep under wraps.
And there are technical concerns too.
When security researchers at NowSecure examined DeepSeek’s iPhone app, they found problems with how the app encrypted and transmitted user data.
The firm ultimately recommended that companies remove DeepSeek from employee devices.
And when Cisco researchers tried to trick DeepSeek-R1 into answering 50 harmful requests, the model failed every single test. In other words, its built-in safety guardrails were surprisingly easy to get around.
So the concern isn’t just political.
For many American businesses, sending sensitive information through DeepSeek’s own service is simply a nonstarter.
But DeepSeek has a major advantage that changes the equation.
Its models can be run somewhere else.
DeepSeek has released open-weight models that companies can install on infrastructure they control. That means American businesses don’t necessarily have to send data to DeepSeek’s servers at all.
They can run the model inside their own data centers or through a trusted cloud provider.
Microsoft is apparently considering this. Rather than connecting Cowork users directly to DeepSeek, Microsoft could host the model on Azure, its own cloud-computing platform, to keep customer data inside its own infrastructure.
Researchers at a German university hospital recently took the same approach.
They ran DeepSeek-R1 on their own computers and blocked it from sending sensitive information outside the hospital.
The system was eventually approved to work with real patient data, including names and other identifying information.
Which once again brings us back to Sovereign AI.
The idea is that companies keep control of their computing infrastructure, while choosing which AI models run on top of it.
That means an American business could potentially take advantage of DeepSeek’s lower costs without sending its most valuable information to China.
And that makes DeepSeek a lot harder to ignore.
Here’s My Take
DeepSeek is the elephant in the room for America’s AI industry.
Companies like OpenAI, Anthropic and Google are spending billions of dollars building increasingly powerful AI models.
Meanwhile, Chinese companies are rapidly closing the performance gap while offering some of that intelligence for a fraction of the price.
And if American businesses can safely run those models on infrastructure they already control, then one of the biggest reasons to avoid them goes away.
I’m not saying DeepSeek will replace America’s leading AI models.
But as AI agents consume more and more computing power, businesses are going to pay much closer attention to what that intelligence costs.
And a half-penny AI model that’s good enough to do the job could become very difficult to dismiss.
Regards,
Ian King
Chief Strategist, Banyan Hill Publishing
Editor’s Note: We’d love to hear from you!
If you want to share your thoughts or suggestions about the Daily Disruptor, or if there are any specific topics you’d like us to cover, just send an email to dailydisruptor@banyanhill.com.
Don’t worry, we won’t reveal your full name in the event we publish a response. So feel free to comment away!
