The train is leaving the station. AI 2.0 will require memory, governance and revenue. Are you on board?

The first wave of AI adoption is over.

Not because the technology failed, but because users hit its limits faster than the industry expected. Models forgot. Systems hallucinated. Outputs couldn’t be trusted. And when reliability drops, adoption at scale is imp.

That is what the data now shows.

Growth has flattened across nearly every major AI platform, on desktop and mobile. Not slowed — flattened. New users try AI, encounter inconsistency, and quietly disengage. Power users remain, but everyone else moves on.

This is the moment the industry usually tries to outrun with more features, bigger models, louder marketing and better demos.

It won’t work this time.

The next phase of AI is not about intelligence. It’s about structure. Systems that can remember. Systems users can control. Systems that can be priced in a way that maps to real cost.

Without those three things — memory, governance and revenue — AI 1.0 is where adoption stalls.

The train is leaving the station.

The first failure: amnesia masquerading as intelligence

The core problem users encounter immediately is memory — or more precisely, the lack of it.

Large language models don’t remember you.
They don’t remember what you told them yesterday.
They don’t remember the rules you set, the preferences you stated, or the context you carefully built.

Instead, users are forced into an unnatural ritual:

• restating goals
• re-explaining constraints
• re-negotiating tone
• re-teaching the system how to behave

Every session starts from zero.

This is not a minor inconvenience. It is the opposite of collaboration.

To paper over this gap, the industry leaned heavily on retrieval-augmented generation (RAG) — a workaround designed to fetch relevant documents and stuff them into a limited context window. RAG was marketed as a bridge to something better.

In practice, it introduced new problems:

• brittle pipelines
• hallucinations at the seams
• misplaced confidence in stale or partial data
• failures that are invisible until they matter most

Users don’t experience this as an architectural limitation.
They experience it as unreliability.

And unreliable tools don’t get adopted — they get abandoned.

This is why growth stalls.
Not because AI isn’t powerful, but because it isn’t dependable.

The second failure: governance without agency

The industry talks endlessly about “AI safety,” but almost always in the narrowest sense:

• self-harm
• prohibited content
• edge-case abuse

That framing misses the real opportunity.

The future of AI is not about protecting users from themselves.
It’s about giving users agency.

True governance means:

• letting each user define how the system behaves
• tailoring vigilance, tone, depth, and autonomy to individual needs
• enabling AI to act as a directed collaborator, not an overgrown autocomplete engine

This is not theoretical to me.

I experience the difference minute by minute, every day.

I have collaborated professionally with some of the best designers, editors, writers, photographers and visual journalists on five continents. I know what good collaboration feels like. And I know what happens when direction, trust and control are aligned.

Today’s AI only approaches that level of usefulness when it is explicitly constrained and instructed by the user.

That is why I load a custom script before each session — a manual workaround that forces consistency, enforces rules and establishes expectations up front.

Try to find another user doing this at scale. You won’t.

Not because it doesn’t work — but because it shouldn’t be necessary.

This capability should be native. Automated. Persistent.

I’m comfortable disclosing this publicly because I filed for this functionality months ago. At this point, disclosure is not risk — it is notice.

And notice matters.

Without user-driven governance, AI remains a novelty.

With it, AI becomes something else entirely: a true collaborator across disciplines.

That leap has barely begun.

The third failure: revenue built on a false proxy

Now we come to the issue that will force change whether the industry likes it or not: money.

Over the past two sessions, Oracle shares have fallen roughly 13 percent after the company reported weaker-than-expected results and sharply higher capital expenditures tied to AI infrastructure — spending that investors were told would drive growth, but which has yet to translate into commensurate revenue.

LA Times: More drops for superstar artificial-intelligence stocks knocked Wall Street off its record heights on Friday.

The selloff helped drag down the broader tech sector, as markets began to reassess AI valuations in light of rising costs and delayed monetization.

This is not an Oracle problem.
It is a business-model problem.

AI systems are enormously expensive to run.

And the dominant pricing model — flat rates or token-based billing — does not map to real cost.

Many executives already know flat pricing doesn’t work.
They accept that metering is inevitable.

But here is where even sophisticated leaders get misled:
they assume tokens are the right unit of measure.

They are not.

To understand why, we need an analogy — one grounded in reality – which is why business schools rely on case studies. As did Santayana: “Those who forget history are condemned to repeat it.”

A lesson from education: when proxies fail

From 2008 to 2014, I worked in education on a platform that served roughly three million registered middle-school students and their English Language Arts teachers.

Those teachers relied on a system called Lexile, which attempted to quantify reading level using two easily computable proxies:

• word length
• sentence length

It was simple. It was scalable. But it was deeply flawed.

For instance, in a paper titled "Interpreting Lexiles in Online Contexts and with Informational Texts", Elfrieda Hiebert concluded that the variability of Lexile scores within the same text can be extensive and that slight changes in punctuation can result in "significant reclassification" on the Lexile scale.

Now let’s get down to Brass Tacks:

• William Faulkner was famous for multi-page paragraphs.
• Ernest Hemingway was famous for short, spare sentences, where much of the meaning lived between the lines.

Lexile routinely rated Hemingway as third-grade level and Faulkner as post-graduate – yet both are taught side by side in tenth grade. Lexile considers Catcher in the Rye a 4th grade level text, despite the appearance of an “prostitute” on Page 2.

Why? Because Lexile can’t read. It measures surface features, not conceptual complexity.

Do you see where this is going?

Why token-based billing fails

Nearly all AI billing today is based on tokens — small chunks of text used as a proxy for computational effort.

But tokens do not measure compute. They measure words.

And words are a terrible proxy for thinking.

Consider two scenarios.

Scenario 1: A young man starts a voice chat with an AI. He talks for thirty minutes about his girlfriend. He recounts detail after detail, searching for clarity.

The system dutifully transcribes every word.
It responds empathetically.
Tokens pile up.

Why?

Because of five words the system does not say — but any human would shout immediately:

SHE’S CHEATING ON YOU!

Now consider the opposite case:

Scenario 2: A three-word query:

“Is God real?”

Few questions demand more cognitive depth.
Few questions invoke more reasoning, context, philosophy and uncertainty.

Yet under token-based billing, the second interaction may cost less than the first.

That alone should end the debate.

Tokens are not compute. They are a proxy — and a bad one.

If you want to bill for compute, you must measure compute.

Anything else is Lexile all over again.

The missing pillar: optimization

There is one more divide separating what comes next – AI 2.0 – from what came before.

Optimization.

Today’s engineers came of age in a world of abundant resources:

• cloud instances on demand
• storage measured in pennies
• CPU cycles treated as infinite

That mindset collapses under the weight of large-scale AI.

More data centers are not the answer.
Ever-increasing capital expenditures are not sustainable.

Optimization is.

If you cut your teeth on DEC PDP-11s — as I did — you learned early that constraints are not obstacles. They are teachers.

Bill Gates knows this.
Steve Wozniak knows this.

They worked within severe limits and built enduring systems anyway.

But they are in their seventies now.

How many engineers at today’s hyperscalers have lived through that discipline? How many learned to treat compute as precious?

Optimization is not nostalgia.
It is necessity.

The moment of truth

This is what defines AI 2.0:

Not bigger models.
Not louder hype.

The world needs systems that:

• remember
• give users agency
• meter value accurately
• optimize ruthlessly

The first generation of AI was about possibility.
The second will be about viability.

And that transition is already underway.

The train is leaving the station.

The only real question is whether you’re on board — or standing on the platform explaining why tokens “mostly work” as it disappears down the track.