One Year After Sam Altman Declared the Singularity, Who Actually Benefited?

Technology Investigation · One Year Later · Reviewed September 23, 2026

On June 10, 2025, Sam Altman announced that humanity had crossed an invisible threshold. “We are past the event horizon; the takeoff has started,” the OpenAI chief executive wrote in an essay called The Gentle Singularity.

Altman did not claim that robots had taken over the streets or that a machine had suddenly become an all-powerful superintelligence. His argument was quieter and, in some ways, more consequential: the decisive transition had already begun even though ordinary life had not yet caught up with it.

AI agents, he predicted, would perform real cognitive work. Software development would never be the same. Scientific discovery would accelerate. AI would help researchers build better AI. Intelligence would become inexpensive, personalized and broadly available. Society would increasingly be limited by the quality of its ideas rather than by the difficulty of executing them.

A year later, Altman’s declaration deserves more than a benchmark scorecard.

Artificial intelligence has clearly improved. Models are more capable at coding, research assistance, document analysis, media generation and structured multi-step work. Used by a capable person, they can reduce the time required for many tasks and make some forms of expertise more accessible.

But the larger short-term revolution sold to companies, workers, investors and governments is much harder to find.

The broad productivity surge remains uneven. Sustainable enterprise returns remain difficult to establish. Organizations have removed employees in the name of AI while adding model costs, consultants, infrastructure, governance, human review and security exposure. Even where individual workers save time, companies have not consistently proven that those savings become greater output, stronger profits or better service after every new cost is counted.

Meanwhile, something more durable has happened. The companies building AI have gained extraordinary access to capital and influence over public policy. Their expansion increasingly shapes decisions about electricity generation, transmission capacity, water, land, chips, cloud infrastructure, tax incentives and national security.

What Altman actually declared

Altman’s original essay blended observations, predictions and political argument. Some elements were measurable. Others were deliberately difficult to disprove.

He wrote that the “takeoff” had started and that humanity was close to building digital superintelligence. He expected 2025 to bring agents capable of real cognitive work, 2026 to bring systems capable of producing novel insights and 2027 to bring robots capable of useful physical work.

The phrase “past the event horizon” carried enormous rhetorical force because it framed AI development as irreversible. If the takeoff had already begun, delay became dangerous. Governments and businesses were no longer choosing whether to reorganize around AI; they were choosing whether to keep up.

What the people leading frontier AI are actually saying now

The conversation is no longer coming from one executive or one company. By September 2026, several of the people leading frontier AI are publicly describing a future in which today's systems are only an early stage.

They do not agree on exactly how fast to move, how much to regulate, or how severe the risks are. But they broadly agree on one thing: they believe much more capable systems are coming.

Other influential leaders disagree about the prescription. Nvidia CEO Jensen Huang has argued against broad slowdown framing and treats safety more as an engineering and deployment problem, while Microsoft AI CEO Mustafa Suleyman has pushed for stronger containment, independent evaluation and practical regulation.

A forecast is not a measurement

This is where I want the reader to slow down.

These people are not random commentators. They have access to models, research teams, internal evaluations and technical information most of us will never see. Their forecasts deserve attention.

But expertise does not turn a future estimate into an observed fact.

A forecast can be informed, sincere and technically sophisticated while still being wrong on timing, scale or consequence.

So the public should treat leader forecasts as important evidence about what the builders themselves believe—not as proof that the predicted future has already arrived.

Why the forecast itself has power

When the leaders of frontier AI say that AGI or superintelligence may be only a few years away, the prediction does more than describe the future.

It can influence capital spending, data-centre construction, chip demand, electricity planning, national-security policy, hiring, layoffs, regulation and public expectations.

That creates a feedback loop:

This does not make the forecast false. It means the forecast can help create some of the conditions that make further progress possible.

The useful evidence test for leader claims

• What has already been demonstrated? Separate current capability from projected capability.
• What depends on a timeline? “Possible eventually” and “likely within three years” are different claims.
• What depends on scaling continuing? Ask whether the forecast assumes more compute, data, energy or capital.
• What depends on breakthroughs that have not happened yet? Identify the missing technical steps.
• What changes if the forecast is wrong? Who has already committed money, infrastructure or policy based on it?
• Who is making the claim? Their expertise matters. So do their institutional incentives and responsibilities.

The point is not to distrust AI leaders because they run AI companies. The point is to avoid giving any expert class the ability to convert its own forecast into unquestioned public fact.

What clearly improved

AI-assisted coding became more useful. Models became better at navigating files, editing code, calling tools and completing structured workflows. Research systems became more capable at searching large bodies of information and proposing lines of inquiry. Creative tools improved in music, voice, image and video generation. Translation, accessibility and document processing became faster and less expensive.

For an independent creator, small company or researcher without institutional resources, these gains can be extraordinary. Tasks that once required a team can sometimes be prototyped by one person.

But a capability improvement is not the same thing as an economic transformation.

AI has repeatedly demonstrated the first category. Evidence becomes more complicated as we move toward the third. A faster draft may still require extensive verification. More generated code may create a larger review burden. Automated customer responses may reduce first-line handling while increasing difficult escalations.

The promised productivity revolution remains difficult to prove

Worldwide AI spending is accelerating much faster than clear evidence of enterprise return. Gartner forecast in May 2026 that worldwide AI spending would reach approximately US$2.59 trillion during the year, a 47% increase over 2025. The firm said AI infrastructure would represent the largest spending segment and noted that enterprises still faced challenges proving tangible business outcomes.

Spending at that scale proves that companies believe AI is strategically important. It proves competitive fear and that technology providers are building capacity. It does not prove that customers are receiving proportional economic value.

The cost that rarely appears in the AI demo

• Cloud computing and model usage
• Data preparation, retrieval and storage
• Application development and integration
• Consultants and implementation partners
• Identity, permission and access controls
• Evaluation, monitoring and quality assurance
• Employee training and workflow redesign
• Legal, privacy and regulatory review
• Cybersecurity controls and incident response
• Human verification, correction and escalation

These investments may ultimately be worthwhile. The point is that they must be counted before an organization declares victory.

Layoffs were treated as proof before the proof existed

The most aggressive short-term AI promise was not that workers would receive better tools. It was that companies would need substantially fewer workers.

But downsizing is not the same as productivity. A company can eliminate thousands of jobs and lower payroll without demonstrating that AI performed the missing work. It may have cancelled projects, accepted slower service, transferred work to customers, increased pressure on remaining staff or shifted responsibilities to contractors.

Adecco Group chief executive Denis Machuel told Reuters in July 2026 that some companies were blaming AI for job reductions actually driven by weaker performance, restructuring or other problems. He also warned that eliminating too many junior roles could damage the pipeline that produces future experienced workers.

At the same time, companies have explicitly linked cuts and resource shifts to AI. The layoffs are not imaginary. The unsupported leap is assuming that every removed employee represents a successfully automated function.

Before accepting a layoff as evidence of AI success, ask:

• What measurable output increased after the reduction?
• What happened to service quality, errors and customer satisfaction?
• Were consultants, contractors or specialist hires added elsewhere?
• How much did model access, infrastructure and security cost?
• How much human review remained necessary?
• Did the company later restore any lost capability?

The work did not disappear. It moved.

Generative AI arrived through an interface that concealed complexity: type an instruction and receive an answer. That experience made automation appear cleaner than it is.

The visible production task becomes cheaper. The surrounding system becomes more demanding.

Security risk belongs inside the ROI calculation

AI systems increase the number of ways sensitive information can move through an organization. Employees may paste confidential material into unapproved services. Generated code may contain vulnerabilities. Connected agents may receive excessive permissions. Prompt-injection attacks can manipulate systems into revealing information or taking unintended actions.

The World Economic Forum’s Global Cybersecurity Outlook 2026 reported that 87% of respondents identified AI-related vulnerabilities as the fastest-growing cyber risk over the course of 2025. Data leakage connected to generative AI and increasingly capable adversaries were among the leading concerns for 2026.

The clearest AI achievement may be political

While businesses struggle to prove consistent returns, the AI industry has already convinced governments that advanced computing capacity is a strategic national requirement.

That claim is not entirely hype. Countries without dependable access to advanced computing could become dependent on foreign cloud platforms and foreign models. Domestic capacity may matter for scientific research, healthcare, defence, public administration, cybersecurity and control of sensitive data.

But once AI is framed as an unavoidable national competition, the companies building it gain leverage.

The race to control AI begins with the data centre

Artificial intelligence does not exist in a weightless cloud. It depends on buildings filled with processors, networking equipment, storage, cooling systems and backup power. Those facilities depend on electricity, water, fibre, land, financing, skilled labour and government approval.

That physical reality is the foundation of the series Who Controls the Intelligence?

The central issue is not whether AI data centres should exist. They will be essential to many useful services and scientific advances. The issue is whether private companies receive lasting control over strategic infrastructure while the public absorbs a disproportionate share of the cost, risk and resource pressure.

Public money can reduce private risk

Governments have legitimate reasons to invest in computing capacity. Public infrastructure can support universities, hospitals, smaller companies, researchers and government services. Domestic capacity can reduce dependence on foreign suppliers.

But public participation must produce a measurable and enforceable public return. Support can include tax exemptions, discounted electricity, new substations, transmission infrastructure, roads, water systems, prepared land, grants, loans, accelerated approvals and reserved capacity.

If revenue and productivity gains arrive more slowly than predicted, taxpayers and ratepayers may remain committed to infrastructure built around forecasts they did not create. Private operators may still retain valuable land, power agreements, buildings and strategic position.

An AI correction could leave the largest companies stronger

A commercial correction would not necessarily make the technology disappear or weaken the largest firms. Smaller developers face high model costs, cloud commitments, expensive talent and constant pressure to finance the next generation of systems. Many applications remain easy for larger platforms to copy or bundle.

The AI boom could disappoint economically while succeeding politically and structurally.

Cheap access is not democratic control

Millions of people can now use capabilities once reserved for large institutions. A creator can produce a song demo, visual concept, marketing draft, software prototype or research plan at remarkably low cost.

But access is not ownership. The company providing the model can change the price, generation limits, permitted uses, commercial terms, content policy, export options or account status. It can discontinue the model or remove a feature on which a business depends.

What this means for independent creators

Creators should not reject AI because the corporate story was overstated. The tools can be genuinely empowering. The goal is to use powerful systems without surrendering your files, audience, evidence or decision-making capacity.

One year later: the verdict

Sam Altman may ultimately prove correct about the long arc of artificial intelligence. AI could accelerate medicine, scientific research, accessibility, engineering, education and creative production.

That possibility should not be confused with proof that the immediate commercial story was accurate. Future breakthroughs do not establish that companies were right to remove workers before dependable replacements existed. They do not prove that current AI spending will produce proportional returns. They do not erase integration and security costs or justify giving private companies lasting influence before the public receives measurable value.

One year after Altman declared that the singularity had begun, the clearest transformation is not yet a world overflowing with broadly shared productivity and prosperity. It is the construction of a new infrastructure order.

Sources and verification notes

1. Sam Altman, “The Gentle Singularity,” June 10, 2025 — primary source for the event-horizon and takeoff claims.
2. Gartner, worldwide AI spending forecast, May 19, 2026 — source for the US$2.59 trillion forecast and 47% growth estimate.
3. Reuters, “AI will not trigger employment collapse, staffing company Adecco Group says,” July 23, 2026 — source for Denis Machuel’s workforce comments.
4. World Economic Forum, Global Cybersecurity Outlook 2026 — source for the AI-vulnerability findings.
5. Axios, DeepMind CEO says society is in the “foothills of the singularity,” May 26, 2026 — source for Hassabis's AGI timeline and agent comments.
6. Axios, Hassabis calls for a U.S.-led frontier-AI watchdog, July 14, 2026 — source for the proposed testing and slowdown framework.
7. Dario Amodei, “We Must Pace the Frontier,” September 2026 — primary source for Anthropic's CEO on benefits, risks and pacing frontier development.
8. Reuters, Anthropic CEO urges AI companies to slow model development, Sept. 12, 2026 — current reporting on Amodei's proposal and industry response.

Editorial standard: AI assisted with source discovery, organization and HTML production. Human editorial judgment controls the thesis, sourcing, conclusions and publication approval.