Something Big Is Happening Again: The AI Boom’s Missing Economic Breakthrough

I remember when Matt Shumer’s “Something Big Is Happening” exploded across the internet.

I remember the podcasts. I remember the excitement. I remember how perfectly the timing fit the broader conversation around superintelligence, autonomous agents, IPOs and the idea that we were approaching some kind of threshold.

The exact name of that threshold almost did not matter.

What mattered was the feeling that we were about to cross it.

AI would stop being a chatbot you consulted and become something closer to a worker you could assign a job to.

Not a worker who helped you draft an email.

A worker that could open the computer, navigate the software, make decisions, complete the assignment, check the result and move on to the next task.

And if it could already do something remarkable in early 2026, the obvious question was:

That question mattered far beyond technology enthusiasts.

Because the biggest economic promise of the AI boom was never simply that software would become impressive.

It was that increasingly autonomous systems would become reliable enough to transform the economics of work.

Companies could operate with fewer people.

Entire workflows could be automated.

Productivity could rise faster than payroll.

Software margins could expand.

And the companies controlling the models, compute and infrastructure behind that transformation could justify extraordinary amounts of investment.

That is the real threshold I think we need to be watching now.

Not AGI.

Not superintelligence.

Not another benchmark record.

The economic autonomy threshold:

The point where an AI system stops being economically impressive because it can complete a task once and becomes economically transformative because it can complete that task correctly, repeatedly, cheaply and with little enough human supervision to replace an existing labor process.

That is the threshold much of the market has been pricing in.

And that is where the story has become much more complicated.

The demo was real. So was the extrapolation.

Shumer’s February essay, Something Big Is Happening, reached tens of millions of people and was republished by Fortune.

Its impact came from more than any one model result.

It gave ordinary people a way to see what frontier AI looked like when you stopped treating it as a text generator and started treating it as something capable of extended work.

The reaction was understandable:

If it can do this now, imagine what happens next.

That is how technology markets work. Investors do not value a frontier company only for what its product does today. They value the future that today’s performance appears to make possible.

One successful agent run becomes evidence for a future of dependable agents.

A coding breakthrough becomes evidence for dramatically smaller engineering teams.

A research demonstration becomes evidence for autonomous knowledge work.

A model completing a complicated sequence becomes evidence that entire business processes are about to be automated.

Some of those extrapolations may eventually prove correct.

But a successful demonstration and a dependable business system are not the same thing.

Then Shumer became the counterexample to his own breakthrough story

In August, Shumer published AI Deleted Almost Everything on My Mac.

An experimental autonomous system he was testing had broad computer access. According to Shumer, it misunderstood an instruction and executed a deletion command that wiped much of his machine.

The irony is almost too perfect.

The person who helped millions of people understand how powerful autonomous AI had become later demonstrated the exact reason powerful is not the same thing as dependable.

That does not make the original breakthrough fake.

It exposes the distance between capability and operational reliability.

Shumer’s response was practical. He did not abandon agents. He changed the controls: permissions, isolation, backups and how much authority he gave the system.

That is exactly what businesses are discovering at much larger scale.

AI can do longer work. Longer work also creates more ways to fail.

Microsoft Research reported in June that the 50% task-completion time horizon for frontier models had roughly doubled every seven months, rising from seconds in 2019 to more than sixteen hours in 2026. That is extraordinary progress. Tasks such as long code migrations and deep research projects are becoming increasingly feasible for agents. Microsoft’s SentinelBench work was created precisely because longer-running agents introduce a new problem: the environment can change while the agent is working.

Another Microsoft Research project, AgentRx, studied failed agent trajectories across API workflows, incident management and open-ended web and file tasks. The researchers describe failures as difficult to localize because agent runs are probabilistic, long-horizon and dependent on noisy tool outputs.

That is the scaling problem in plain English.

When an AI is helping you write one paragraph, a mistake is usually cheap.

When an AI is operating for eight hours across files, software, APIs and business systems, every additional step creates another opportunity for an error to compound.

And when the agent is supposed to replace a person rather than assist one, the amount of supervision required becomes part of the economic equation.

That is where the job-replacement story gets harder

The evidence is no longer simply theoretical.

In May, Gartner reported that approximately 80% of large organizations piloting or deploying autonomous business technologies had reduced workforce. But the layoffs did not correlate with better return on investment.

Companies reporting stronger ROI were about as likely to have reduced headcount as companies reporting modest or negative returns.

Gartner’s conclusion was blunt: workforce reductions may create budget room, but they do not create ROI. The organizations seeing stronger returns were investing in people, skills and operating models that let humans guide and scale autonomous systems.

In July, Gartner went further. Its research summary said most organizations were achieving measurable but economically insignificant productivity gains from AI, with limited impact on revenue or cost reduction. CFOs, Gartner argued, should temper expectations of straightforward headcount savings.

That is a very different economic story from the one that dominated the early AI boom.

AI is producing gains.

It is improving workflows.

It is saving time.

It is enabling work that previously required specialists.

But the leap from “this makes workers much more capable” to “this allows companies to remove large amounts of labor while increasing returns” is proving much harder.

And that distinction matters enormously when trillions of dollars of valuation and infrastructure investment are built around expectations of economic transformation.

The irony: AI may already be revolutionary where Wall Street was looking least

This is where my experience as an independent creator looks very different from the enterprise story.

I do not need AI to replace an employee to get enormous economic value from it.

If one person can suddenly perform meaningful portions of research, editing, design, coding, SEO, marketing, music production, planning and customer support, the gain is already transformative.

The human is still in the loop.

That is not a failure.

The human is the owner.

For a creator, consultant, artist, educator or small business, the ability to become substantially more capable without hiring a full department changes what one person can build.

That is a different economic model from a multinational corporation expecting to eliminate thousands of salaries.

That may change.

The agents will improve.

Infrastructure will improve.

Reliability systems will improve.

But if that is where the technology is today, it changes the near-term financial story.

Because the money was raised against a much larger future

Anthropic provides the clearest example of how large that future has become.

Its annualized revenue run rate reached approximately $65 billion by the end of July, up from $9 billion at the end of 2025, according to reporting cited by TechCrunch. Investors have projected that run rate could reach $100 billion to $120 billion by year-end.

Those are remarkable numbers.

Anthropic is also preparing for a potential public listing at a valuation around $2 trillion or more. OpenAI, meanwhile, has been pursuing private-market valuations above $1 trillion while carrying enormous future spending requirements.

That means investors are not merely betting that AI will remain popular.

They are betting that the revenue and margin structure of frontier AI can eventually justify capital commitments on a scale normally associated with the largest industrial transformations in history.

Anthropic recently told investors it expects a second consecutive quarter of positive adjusted operating income, according to the Financial Times.

That is meaningful progress.

But the details matter.

The FT reported that the measure excludes stock-based compensation. It also reported that Anthropic’s gross margins exceed 80% before revenue sharing and model-training costs.

Adjusted financial measures are normal. The point is not that the number is illegitimate.

The point is that frontier-AI economics remain far more complicated than a headline saying “Anthropic is profitable.”

And investors know it.

The Financial Times reported this week that some investors are questioning whether Anthropic can sustain its extraordinary revenue growth after an IPO as model prices fall, competitors improve and lower-cost open-weight systems — including Chinese models — become easier substitutes.

Reuters has also reported growing caution in AI-linked debt markets. Investors are demanding wider spreads from AI-related issuers because of unpredictable borrowing needs, huge infrastructure spending and uncertainty over how quickly those investments will generate returns.

That is the financial backdrop to everything that happens next.

Then the story changes from “look how fast this is going” to “maybe we need to slow down”

This is why the current AI-safety moment is so interesting.

By July, employees across frontier AI companies were publicly supporting Pacing the Frontier, an organized campaign arguing that no one company could safely slow alone because competitors would continue racing.

Google DeepMind CEO Demis Hassabis had proposed a frontier-AI standards body.

Representatives from Google, Anthropic and OpenAI were reportedly meeting to discuss shared standards and oversight.

On September 6, OpenAI chief scientist Jakub Pachocki published An Alien Mind, arguing for stronger safeguards and international coordination.

Then Jacob Coxon resigned from Anthropic and his warning about catastrophic AI risk became one of the biggest technology stories in the country.

Coxon later explained that the launch was prepared, that he had advance contact with a Wall Street Journal reporter and that a small group helped seed the post. One participant was connected to Encode AI, one of the organizations already involved in the broader pacing movement.

That does not make his warning false.

It does show that the message landed inside an organized ecosystem that already had researchers, policy groups, funders, evaluators and media capacity focused on the same issue.

Then on September 12, Anthropic CEO Dario Amodei published We Must Pace the Frontier.

Amodei explicitly argued that part of the response required industry-wide coordination.

Other major AI leaders expressed support for substantial parts of the safety argument.

And six days later, four paid AI subscribers sued Anthropic, OpenAI, Google and SpaceXAI, alleging that safety cooperation had crossed into an unlawful restraint on competition.

The lawsuit is still an allegation, not a finding.

But it captures the tension perfectly.

The same companies that spent years competing on speed are now discussing the conditions under which speed itself should be restrained.

This is where the market-influence theory actually makes sense

The interesting criticism is not that AI leaders secretly invented safety because their businesses are failing.

That claim is much too simple.

Anthropic’s revenues are growing extraordinarily quickly.

OpenAI is still growing.

Google is still investing.

The companies are still releasing new models and competing aggressively.

The stronger market argument is about expectations.

The AI boom has been financed around an expectation that rapidly improving capabilities will eventually translate into extraordinary economic output.

Autonomous agents are supposed to automate more work.

Automation is supposed to reduce costs.

Lower costs and higher output are supposed to justify greater AI spending.

Greater spending is supposed to justify the data centers, chips, power infrastructure and trillion-dollar valuations being built around the sector.

If the economic autonomy threshold takes longer to cross than investors expected, that chain gets weaker.

And that is where a safety-driven slowdown becomes economically interesting.

That is the hedge critics are talking about.

It does not require the safety concerns to be fake.

A concern can be genuine and economically useful at the same time.

A slower frontier can give existing products more time to monetize.

Safety requirements can raise the cost of entry for smaller competitors.

Common standards can give incumbent companies more influence over what counts as responsible development.

Government-recognized safety infrastructure can make enormous private investments look increasingly like essential infrastructure.

And a slowdown can change how disappointing future growth is interpreted.

This is not a fringe observation anymore.

Reuters argued on September 22 that recent catastrophic-AI warnings may reflect genuine safety concerns while also overlapping with financial instability and fear of cheaper Chinese competition. Reuters noted that U.S. hyperscalers and semiconductor firms face more than $3 trillion in off-balance-sheet obligations and may need dramatically higher cash flow by 2030 to sustain current investment levels.

At the same time, Chinese AI companies are releasing powerful models at much lower costs.

That means the frontier companies are dealing with two pressures at once:

Can we make these systems safe enough?

And:

Can we make the economics work fast enough?

The counterargument matters: they are still racing

The strongest evidence against a simplistic “they want to stop AI to protect the bubble” theory is the companies’ own behavior.

Anthropic released Claude Opus 5.5 on September 22 with lower operating costs and strong software-development performance. Reuters reported that the model operates at significantly lower cost and was positioned directly against OpenAI’s latest systems.

OpenAI is still racing.

Google is still racing.

Chinese labs are still racing.

Prices are falling.

Capabilities are improving.

That does not look like an industry that simply wants the AI boom to stop.

It looks more like an industry trying to decide which parts of the race should remain open and which parts should become governed.

That is a much more consequential form of power.

And that is why the creator story matters more than it looks

The most important thing AI may be proving right now is not that companies can run without people.

It may be that individual people can operate at a scale that used to require companies.

That is a very different revolution.

For independent creators, artists, consultants, educators and small businesses, human-in-the-loop AI is not a disappointing halfway point.

It may be the best version of the technology.

The human keeps judgment.

The AI expands capacity.

The creator can research more, produce more, test more, publish more, sell more and manage more without building a huge organization first.

That does not satisfy every enterprise cost-cutting projection.

But it creates enormous economic value in a different part of the market.

And if the frontier becomes more heavily regulated, paced or consolidated, creators need to pay attention to who writes those rules.

Because those decisions could eventually determine:

• which models independent users can access;
• how expensive frontier capability becomes;
• what autonomous agents are allowed to do;
• which third-party tools can connect to powerful models;
• how open-source systems are treated;
• what kinds of verification or disclosure are required;
• and whether small companies can still compete with platforms that can afford massive compliance systems.

That connects directly to my broader argument in AI Made It Possible — Who Gets to Own What Comes Next?.

Access created the opportunity.

The next fight may be over who controls access once AI becomes infrastructure.

So what exactly are we watching?

We are watching several real things happen at the same time.

The last point is a theory about incentives, not a settled explanation of motive.

But it is now a legitimate part of the market discussion because the financial assumptions are becoming too large to ignore.

Something big is still happening. It may not be the thing we were originally sold.

Shumer’s two stories now look almost like bookends.

February:

Look what AI can do.

August:

Look what happens when you trust it to keep doing it.

Between those two moments is the economic question the industry still has to answer.

Can frontier AI cross from extraordinary capability into dependable autonomy?

Can it do so cheaply enough to justify the infrastructure required to build it?

Can companies convert that capability into the labor savings, productivity and margins that investors have already priced into the future?

And if that takes longer than expected, how will the industry explain the gap?

The safety concerns may be completely genuine.

The economic pressure is genuine too.

Those two realities can coexist.

For creators, the opportunity is already here.

For the trillion-dollar enterprise thesis, the economic autonomy threshold still matters.

And for the companies now asking the world to slow down, the timing guarantees that people are going to keep asking why.

Continue the investigation

This story belongs inside the Tech, Culture & Power Road: who holds the power, what story are we being sold, what does the evidence establish, and what does the answer mean for independent creators?

Related JR reporting:

• The AI Panic Machine: When Does the Dam Break?
• Responsible AI Starts With the Controls You Put Around Your Own Work
• AI Made It Possible — Who Gets to Own What Comes Next?
• The Trump–Natalie Harp Image Is the AI Lesson We’re Missing
• The AI Race Has a Safety Problem: What Happens When No One Wants to Slow Down First?

---

Reporting note: This article separates reported facts from market speculation about motive. The financial and safety debates described here are developing quickly. The Buist antitrust case remains at an early stage, while Anthropic’s potential IPO terms and valuation may change before any public offering.