Why Microsoft is Pulling back on AI?
- 2 days ago
- 4 min read
The Economics Nobody Talks About in AI
Last week, I came across a headline that made me pause.
It was not about a new AI model. It was not about another breakthrough in reasoning. And it was not about another company raising billions of dollars.
It was about Microsoft.
Reportedly, Microsoft had to withdraw advanced AI coding tool licenses from many of its own engineers because internal teams were consuming cloud token budgets at a pace that surprised the company's finance division.
I found that fascinating.
If the company helping power the AI revolution is feeling pressure from the cost of using AI internally, then something important is changing. This isn't just a Microsoft story. It's an early signal of how AI is beginning to challenge one of the assumptions that has shaped software for decades. I found that intrigued and wanted to understand the economics behind it.
Let's do some simple math.
That is the real shift.
For the last twenty years, we have lived in the era of predictable software pricing. You buy a subscription. Your team uses the software ten times or ten thousand times. The bill remains largely the same.
AI changes that equation.
Every prompt has a cost. Every document uploaded, every question asked, and every response generated requires computing power. The more AI is used, the more those costs accumulate.
Before we go further...
If you've ever wondered what a token actually is, this is probably the easiest way to understand it.
I built a simple interactive explainer using Claude that lets you type your own text and watch how AI breaks it into tokens. It also visualises how those tokens are converted into vectors - the numerical representations that help AI understand language behind the scenes.
Spend two minutes experimenting with it. The rest of this article will make much more sense afterwards.
Now let's come back to the economics.
You are no longer paying for a single answer.
You are paying for the AI's entire internal trial-and-error process.
This is where the industry is starting to encounter what some are calling the Token Tax.
The Token Tax Explosion
The challenge isn't the simple chatbot sitting in a browser window.
The challenge is that organisations are increasingly moving toward AI Agents. These systems are designed to operate with greater autonomy. Instead of producing a single response, they break problems into multiple tasks, gather information, evaluate options, check results, and continue working until they reach an outcome.
That capability is powerful.
It is also expensive.
Because agents repeatedly process information and revisit context, they can consume many times more tokens than a traditional chatbot. When these systems are connected to enterprise workflows, databases, documents, and software environments, token usage can rise dramatically behind the scenes.
What looks like a simple request from a user may trigger dozens of AI operations that nobody ever sees.
That is how budgets intended to last for months can disappear much faster than expected.
How the Software Playbook Is Changing
To deal with this new reality, software companies are beginning to rethink how AI-powered products are priced and delivered.
For years, the software industry has relied on the seat-based model. A customer pays a fixed amount per user and receives access to the platform. AI makes that model harder to sustain because the cost of serving one user can vary enormously depending on how much AI they consume.
As a result, three shifts are becoming increasingly visible.
From Seats to Meters - Many companies are moving toward hybrid pricing models. Customers will continue paying for access to the software, but AI-intensive actions may be charged separately based on usage. In many ways, AI is beginning to resemble a utility where consumption matters.
Token-Efficient Models - Model providers are racing to build smaller and more efficient systems that can deliver similar results while using fewer tokens. Efficiency is quickly becoming as important as raw capability.
Smarter Inputs - Organisations are also discovering that reducing unnecessary data before it reaches an AI system can significantly lower costs. Cleaning documents, removing irrelevant content, and improving information quality can have a meaningful impact on token consumption.
What We Can Take Away From This
Even if you never manage a cloud budget or build AI systems yourself, this shift matters.
Most conversations about AI focus on capability. We talk about what AI can do, how fast it can work, and how much productivity it can unlock.
Far fewer conversations focus on the economics behind those capabilities.
As AI becomes embedded into everyday workflows, understanding those economics will become increasingly important.
Most importantly, we need to stop thinking of AI as an unlimited resource.
The companies that succeed in the next phase of AI won't simply be the ones using the smartest models.
They will be the ones that learn how to manage the economics behind them.
Stay grounded. Stay curious.
-----
This is part of my ongoing work helping managers build an AI mindset—practical ways to think, decide, and work in the age of AI. It's not about mastering tools, but understanding how to work intelligently with intelligence. Looking forward to your feedbacks
Comments