In my previous articles, I have kept coming back to the extraordinary amounts of money being invested in AI infrastructure and one fairly simple question: who ultimately pays for all of this?

That got me thinking about another question.

What exactly are enterprises going to buy?

Today, much of the AI industry still talks about tokens. Millions of tokens, billions of tokens, cost per million tokens.

I have a hard time believing tokens are ultimately the product.

A large bank doesn't really care how many tokens it consumed. It cares about how many customer inquiries AI resolved, how many KYC cases it processed, how much software it developed, how much employee time it saved and, ultimately, how much all of that cost.

In other words, I think the product gradually moves from tokens to work.

And when I look at the strategies of the major AI companies through that lens, they start to make a lot more sense.

The competitive landscape

Microsoft: It already owns much of the employee's workplace through Office, Teams, GitHub and Azure. Its opportunity is to add AI workers to an environment that enterprises already use.

Google: Similar idea. It owns Workspace, cloud and enormous amounts of enterprise data infrastructure. Add AI agents on top of it.

OpenAI: This one is more interesting. It doesn't own the traditional workplace, so perhaps the ambition is to make ChatGPT itself the new work environment. Employees interact with ChatGPT, Codex performs software work, agents perform business processes, and OpenAI's enterprise platform connects and governs all of it.

Anthropic: I see this somewhat differently. Claude Code is a good example. The customer isn't really buying tokens. The customer is buying software engineering work. Extend that idea into research, compliance, financial analysis and other professional functions and Claude increasingly starts looking like a cognitive worker.

Different strategies, but I think they are heading toward the same place.

Nobody wants to remain a token vendor.

If models become increasingly interchangeable for many tasks, tokens risk becoming a commodity. The more valuable position is higher up the stack: own the agent, the workflow, the integration and ultimately the work being performed.

And that's where the economics become interesting

Imagine an AI agent processes 400,000 KYC reviews.

I doubt the CFO is going to ask how many tokens were consumed.

The CFO is going to ask:

What did those 400,000 reviews cost us?

And then the really interesting question:

What did they cost when humans were doing them?

At that point, the unit of comparison is no longer tokens. It may not even be compute.

It becomes work and economic output.

Which brings me back to the question I have been asking since starting Inside Enterprise AI.

We are spending extraordinary amounts of money building the infrastructure to produce AI intelligence. For that investment to work, enterprises eventually have to turn that intelligence into economic value.

Perhaps the metric that ultimately matters isn't tokens per dollar.

It's how many dollars of useful work an enterprise gets for every dollar it spends on AI.

If that number works, a lot of today's enormous AI investment starts to make sense.

If it doesn't, we are going to have a very different conversation.