A small, quiet constraint

There is a hard limit on how much AI I’m allowed to use at work.

It isn’t philosophical.
It isn’t ethical.
It isn’t technical.

It’s a line item.

I have a $100 per day token cap. I can see it. I can track it. And because of it, I usually run two or three AI instances in parallel, carefully, deliberately, rationing usage the same way early internet users rationed minutes.

That cap is the only thing stopping me.

Not usefulness.
Not imagination.
Not trust in the output.
Not fear of disruption.

Just a ceiling.

What my day actually looks like now

Four months ago, I wrote code all day.

Today, I haven’t written a line of code myself in over a month.

I task AI systems. I review their output. I spot subtle errors, structural mistakes, conceptual misfires. I ask for revisions. I redirect. I orchestrate.

If I tried to sit down and write code manually the way I used to, I would be vastly outperformed by coworkers who fully embraced this shift. Not because they’re smarter, but because they’re operating at a different layer.

The job didn’t get faster.

The job moved.

And once it moves, it doesn’t move back.

The thing I’m not allowed to do (yet)

If that token cap disappeared tomorrow, I wouldn’t “use AI a little more.”

I would do something categorically different.

I would have one AI write scripts that spawn and coordinate other AI instances.
I would fan out tasks across 20 or more parallel agents.
I would refactor entire systems overnight.
I would explore multiple architectural paths simultaneously instead of serially.
I would trade time for compute at a scale that feels almost irresponsible by today’s standards.

And even that — I’m increasingly convinced — would still be barely scratching the surface.

The constraint isn’t capability.
It’s permission.

Why this matters for the bigger picture

When markets talk about AI demand, they talk about what they can see.

Revenue.
Usage metrics.
Capex budgets.
Token spend.

What they don’t see is suppressed intent.

The gap between:

“What people are allowed to do”

and

“What people would do instantly if allowed”

That gap is enormous right now.

I’m not guessing. I’m living inside it.

And this is why so many bear arguments feel strange from the inside.

The “limits” everyone points to

Most of the supposed threats to AI adoption fall into the same category:

  • token costs
  • memory constraints
  • efficiency improvements
  • internal policies
  • governance delays

These are framed as ceilings.

They aren’t.

They’re gates.

A gate doesn’t stop a curve.
It delays it.

And when a gate finally opens, behavior doesn’t increase smoothly. It snaps to its natural level.

We’ve seen this before:

  • bandwidth caps before streaming
  • server provisioning before elastic cloud
  • data storage costs before big data
  • manual workflows before automation

Every time, usage was “reasonable” right up until the constraint vanished. Then it exploded.

The same thing is happening here — except this time, what’s being throttled isn’t convenience.

It’s cognition.

Efficiency doesn’t reduce demand. It reveals it

There’s a persistent idea that more efficient models will reduce the need for compute.

From where I’m standing, that idea is almost backwards.

If models got ten times cheaper tomorrow, I wouldn’t run fewer instances.

I would remove the ceiling.

I would explore more branches.
Run more redundancy.
Ask better questions.
Try riskier refactors.
Let multiple agents disagree and reconcile.

Efficiency doesn’t end work.

It expands ambition.

And ambition is what actually consumes compute.

Why even this post is probably underestimating things

Here’s the part that’s hardest to explain.

I am not an outsider speculating about the future.
I’m not a pundit.
I’m not early-career.

I work inside one of the largest companies in the world.
I’m using these systems daily.
I’m watching expectations shift in real time.

And even so, I’m increasingly convinced that my own imagination is still orders of magnitude too small.

What I’m describing here — multi-agent orchestration, overnight transformations — still assumes human-shaped workflows. One person steering a swarm.

The next phase doesn’t.

The next phase is systems that:

  • spawn sub-agents dynamically
  • evaluate and correct each other
  • coordinate without constant human intervention
  • optimize not just outcomes, but how they collaborate

When that becomes normal, today’s usage patterns will look quaint.

Like writing code by hand already does.

Why this ties back to NVIDIA

When people argue about :contentReference[oaicite:0]{index=0}, they argue about today’s visible demand.

What they are missing is tomorrow’s unexpressed demand.

The demand that exists right now in the form of:

“I would do much more, instantly, if allowed.”

That demand doesn’t show up cleanly in charts yet.
It shows up as friction, rationing, and quiet frustration.

But once the gates open — through cost reductions, policy changes, or competitive pressure — it doesn’t trickle out.

It floods.

The real curve

From the bottom, the AI curve doesn’t look like hype.

It looks like this:

  • A small group discovers leverage
  • They are constrained by policy and cost
  • Their output becomes impossible to ignore
  • Expectations reset
  • Constraints loosen
  • Usage explodes

By the time the market agrees something fundamental changed, the people doing the work have already moved on to the next layer.

That’s where we are now.

Not at the peak.
Not at the end.
But still far below the natural ceiling.

Closing

The most important thing I’ve learned in the last few months isn’t how powerful these tools are.

It’s how much unused capacity is sitting just behind arbitrary limits.

If someone like me — already inside the system, already convinced, already pushing — is still barely beginning to understand what’s possible, then the eventual outcome won’t look like a linear extension of today.

It will look discontinuous.

The curve doesn’t need belief.
It just needs the gates to open.

And when they do, even our wildest expectations will probably turn out to be conservative.

Disclaimer:
This post reflects personal opinions and is not financial advice.
OppenFolio is not an investment advisory service. See site disclaimer for full details.