Stop Spending Tokens. Start Investing Them.
We're about to spend an extraordinary amount of money on AI. And I suspect most of it will evaporate.
Not because the models aren't good enough. Because we're treating intelligence like electricity: buy some, use some, switch it off, buy some more tomorrow.
A prompt goes in. A response comes out. The meter ticks up. Everyone congratulates themselves on being productive.
Then, the next morning, we pay the machine to think about exactly the same problem again.
That's not an investment in intelligence. That's renting amnesia by the token.
A billion tokens. Nothing to show for it.
Imagine two companies each spend $100,000 on AI.
The first generates presentations, rewrites emails, brainstorms strategies and holds endless conversations about projects that never quite ship. It has an impressive usage graph. Perhaps even an enormous archive of chats. But next month it needs to explain its business to the models all over again.
The second spends the same amount creating something that remains after the conversation ends. Every useful answer becomes part of a durable knowledge system. Every repeated task becomes a tool. Every failure tightens a process. Every customer interaction improves the way the business operates. Decisions, permissions and outcomes are recorded so the next task starts from somewhere better.
One company has consumed intelligence. The other has turned token consumption into an asset.
Same invoice. Completely different economics.
The token isn't the product. What survives it is.
I keep coming back to one question: what did the business retain from that last million tokens?
Not how many words did we generate. Not how many agents did we run. Not how impressive was the demo.
What does the business now own that it didn't own before?
A better process? A working application? A reusable capability? A reliable customer record? A decision rule? A tool that makes the next hundred decisions cheaper?
Or just a transcript?
I don't believe the future belongs to whoever can consume the most tokens. That sounds like measuring the success of a factory by how much electricity it uses.
The factory exists to make things. So should our AI systems.
Every token should make the next one more valuable.
Compounding is usually discussed in terms of money. But its underlying principle is much broader: today's output should improve the economics of tomorrow's output.
The first time an agent encounters a problem, it might need hours of reasoning, research and trial and error. Fine. Let it spend the tokens.
But if it solves the problem and then throws away the solution, we've bought the same expensive lesson twice. Or a hundred times.
The correct response isn't necessarily to write a shorter prompt. It's to make the work reusable.
Turn the solution into a tested workflow. Turn the discovery into searchable knowledge. Turn a brittle instruction into a proper tool with a contract. Turn the error into a regression test. Turn a human correction into a lasting change in how the system behaves.
Then the next agent doesn't begin from zero. It inherits the progress.
And when the next agent improves the process again, the one after that inherits both improvements.
That's where the return starts compounding.
Cheap tokens can be very expensive.
There's an almost comical obsession with the price of a million tokens. Models get compared on how cheaply they can produce text, as though text were the thing businesses ultimately wanted.
A model that costs half as much but makes you repeat the work five times isn't cheaper.
Nor is the smartest model necessarily the best investment. If a costly reasoning run creates a dependable tool that eliminates thousands of future runs, it may be the bargain of the century.
The question isn't how few tokens can we use. It's how much future value can we create per token spent?
Sometimes the right move is to spend ten times as many tokens on a task today so that nobody—human or machine—ever needs to do that task the hard way again.
That's a very different optimization problem.
Memory is not the same as accumulated intelligence.
Of course, every AI company is going to tell us its assistant has memory.
That's useful. But remembering a conversation isn't the same as building a capability. Knowing that you once fixed a broken sales handoff doesn't guarantee the next sales handoff works.
The interesting progression is from remembering what happened to changing what happens next.
A genuine intelligence asset needs more than stored words. It needs structure, ownership, verification, retrieval and the ability to do something with what it knows.
If an agent learns that a particular process fails, does the process change? If an answer becomes a business rule, who approved it? If a workflow is improved, does the improvement survive a model upgrade? Can another agent use it? Can the company take it elsewhere?
Otherwise we're not compounding intelligence. We're hoarding context and calling it progress.
An agent that repeats its mistakes is a terrible investment.
This has become one of my biggest frustrations with AI-driven work.
An agent can look incredibly busy. It can burn through vast amounts of context, create documents, make elaborate plans, run checks, produce status updates and declare victory.
Then you discover the thing you actually wanted still hasn't happened.
Worse, the next run starts again: another audit, another plan, another explanation of the same blocker.
The token counter goes up. The business goes nowhere.
I want systems that become harder to fool with every completed task. Systems that remember what didn't work, stop repeating pointless activity and develop a better instinct for where to spend their next unit of intelligence.
An agent shouldn't merely finish a task. Where possible, it should leave the environment better equipped to finish the next one.
Completed work is the dividend. Improved capability is the capital.
The companies that own the compounding will win.
Here's the bigger strategic question.
If all that accumulated knowledge lives inside someone else's closed conversation history, do you really own the asset? Or are you paying rent on a capability you can't move?
For me, the most valuable parts of an AI-enabled business should survive a change of model. Its customer relationships, operating knowledge, tools, permissions, processes and institutional memory should remain intact even when the underlying intelligence provider changes.
Models will improve. Prices will fall. New winners will emerge.
The lasting advantage won't be the model you rented this quarter. It'll be what your business built with it.
And as that asset grows, something wonderful should happen: the next unit of useful work becomes easier, faster and more reliable. Less time spent rediscovering. Less human attention spent supervising. More capacity freed up to build the next thing.
This is why I've stopped thinking about token consumption as simply an operating cost.
Every token should be an opportunity to increase the value of every token that follows.
If all we're doing is generating output that disappears, we're going to spend a fortune and have remarkably little to show for it.
If we're converting that expenditure into durable capabilities, we're doing something entirely different.
We're building a business whose intelligence compounds.
Stop spending tokens. Start investing them.