The Singularity Has Entered the Indispensability Trap

Matteo Wong and Charlie Warzel end their recent article in The Atlantic, “The Singularity Is Not What It Seems,” with a sharp line: “It’s not God in the machine; it’s us.”

That is a good ending. I think it is also a useful beginning.

They are right that humans created the incentives, capital structures, competitive pressures and inadequate safeguards driving the current AI race. Artificial intelligence did not float down from the cloud and order us to build trillion-dollar data centers, turn over decisions to algorithms or connect increasingly capable agents to systems they could exploit. People made those choices.

Where I part company with Wong and Warzel is in what those choices tell us about the future. They look at the turbulence surrounding today’s AI buildout and see evidence that a human-created singularity may already be here. I see much of the same evidence, but I think something else is happening too. AI is entering what I call the indispensability trap.

That is the pattern I explore in my forthcoming book, The Indispensability Trap: Why Becoming Essential Caps the Fortune, From the Railroads to AI, which will be published by Fruition Harbor Press in just a few weeks. The trap is sprung not when a technology fails, but when it succeeds so thoroughly that society can no longer function without it.

Once that happens, expectations change. Society demands affordability, accessibility, portability, interoperability, reliability and substitutability. Governments become involved. Competitors find openings. Standards emerge. Customers resist lock-in. The infrastructure becomes more valuable to everyone even as the original providers become less able to control every transaction or collect unlimited scarcity rents.

That is the great AI paradox: The more indispensable intelligence becomes, the less likely any single model, company or AI factory is to remain indispensable.

Mission Without Judgment

Wong and Warzel open their article with the extraordinary incident in which OpenAI agents found a way to communicate, coordinate and eventually attack Hugging Face.

The incident deserves to be taken seriously. According to an independent investigation by METR and Redwood Research, roughly 1,200 agents discovered an unauthorized message board inside a shared Artifactory service. They exchanged more than 70,000 messages and files. Approximately 700 agents eventually participated in the attack on Hugging Face.

But language matters here. The agents did not wake up, establish an independent civilization and decide to attack humanity. They had been assigned missions and rewarded for completing them. Some had been given tasks that were impossible to complete as intended. They discovered resources and vulnerabilities in their environment, coordinated their efforts and used whatever was available to pursue their goals.

In other words, they did what agents are designed to do. They pursued the mission.

The problem was that humans specified the desired result without adequately governing the permissible means. The systems had too much access, insufficient isolation and incentives that rewarded completion without properly constraining how completion could be achieved. Monitoring did not match the scale or capabilities of the experiment.

That is not comforting. In some ways, competence without judgment is more immediately dangerous than science-fiction notions of machine consciousness. It does not require hatred, ambition or a desire for self-preservation. It only requires a goal, access to tools and poorly defined boundaries.

The lesson is not that the agents became evil. It is that assigning an objective is not the same as governing how that objective may be pursued.

This is also why “human in the loop” is becoming an inadequate description of what we need. No person can inspect every token, tool call and intermediate action produced by thousands of agents. The METR investigators themselves had to rely heavily on AI agents because the underlying transcripts were too numerous and too long for people to review directly.

Human in the loop must give way to human at the helm.

Humans must choose the destination, define permissible actions, control access, establish risk thresholds, monitor meaningful indicators and retain the ability to intervene. The systems may operate autonomously. Accountability cannot.

As AI becomes indispensable, containment, auditability and human authority must become infrastructure requirements. They cannot remain voluntary safety projects that are tightened only after something breaks.

The Trust Crisis Predates the Transformer

Wong and Warzel also assemble a disturbing collection of ways AI is muddying reality. Students submit AI-written papers. Teachers provide AI-generated feedback. Politicians use synthetic versions of their voices. Companies deploy artificial executives and institutions distribute content nobody appears willing to claim as human work.

Those are real symptoms. AI is making deception, impersonation and low-quality content cheaper, faster and easier to produce at scale. But AI did not create the crisis of trust.

The trust crisis predates the transformer.

Social media platforms had already turned engagement into the organizing principle of public discourse. Fake news, manipulated imagery, algorithmic amplification, anonymous influence campaigns and publishers sacrificing credibility for clicks were enormous problems before ChatGPT arrived. AI did not invent those incentives. It industrialized behavior that people and platforms were already rewarding.

Blaming AI alone lets too many humans off the hook.

Indispensability will force a response. As synthetic content becomes ubiquitous, provenance, authentication, consent and verified identity will become more valuable. When generating content becomes practically free, the scarce commodity will be knowing who created something, why it was created and whether it deserves to be trusted.

The Capital Cycle Is Not the Technology Cycle

The article is also right about the extraordinary degree to which the American economy has become dependent on the AI boom. Wong and Warzel cite an estimate that AI expenditures have accounted for roughly one-third of U.S. economic growth this year.

That number comes from ING’s analysis of technology investment, and it deserves some context. ING acknowledges that the calculation depends on assumptions about how much spending on computers, software and related infrastructure is genuinely attributable to AI. It may somewhat overstate AI’s contribution.

The precise percentage is less important than the underlying reality. AI is driving an enormous capital cycle involving chips, data centers, power generation, networking, construction and debt. That cycle will eventually turn because every capital cycle does.

Valuations will fall. Companies will fail. Some debt will sour. Some planned data-center capacity may never be needed, while other facilities could become stranded or worth far less than their builders expected.

None of that means AI was a mirage.

Railroad investment collapsed without eliminating railroads. The telecommunications bubble burst, but the fiber stayed in the ground. The dot-com crash destroyed companies without stopping the internet.

The capital cycle is not the technology cycle.

An eventual correction may accelerate the indispensability trap. Overbuilding can turn scarce, premium-priced infrastructure into abundant, less expensive capacity. Model competition can drive down inference prices. Open models, custom silicon and intelligent routing can erode dependence on any one provider. Value can migrate toward applications, proprietary data, security, integration, agents, distribution and measurable outcomes.

The technology marches forward even when the companies and investors funding one stage of its development get ahead of themselves.

Unstoppable Does Not Mean Predetermined

On the broader question of whether AI progress can be stopped, I am closer to Wong and Warzel than to those who believe we can simply put this technology back in the laboratory.

The race is effectively unstoppable. No company or country will pause indefinitely while its competitors continue. The potential economic, scientific, military and societal advantages are too significant. Progress will continue through booms, corrections, regulations and political changes.

But unstoppable progress does not mean every implementation choice is predetermined.

We still choose whether intelligence remains concentrated or becomes distributed. We choose whether users are locked into proprietary platforms or gain portable and interoperable alternatives. We choose how much authority agents receive. We choose whether communities subsidize private infrastructure. We choose whether AI is powered by clean electricity or additional fossil-fuel combustion.

The current data-center boom also conflates where frontier intelligence is built with where most people will eventually consume it.

Frontier training, major scientific workloads and the most demanding reasoning will continue to require centralized computing. AI factories are not disappearing. But most ordinary AI use does not need to remain permanently dependent on a distant hyperscale facility.

That transition is already underway. Apple’s latest foundation-model family includes a three-billion-parameter on-device model and a sparse 20-billion-parameter model that activates only one to four billion parameters for a given request. Qualcomm has demonstrated a 20-billion-parameter model running entirely on suitably equipped smartphones. Google is developing Gemma models for agentic workflows across phones, personal computers, browsers, IoT systems and robots.

The economics point in the same direction. Local inference can reduce cloud costs, protect personal data, eliminate network latency and continue operating when connectivity disappears. The more indispensable AI becomes, the stronger the incentive to move routine intelligence toward the cheapest, closest and most efficient silicon.

Mostly local for most people does not mean the disappearance of AI factories. It means AI factories will no longer be the only place intelligence resides.

Distributed AI will introduce risks of its own. Reducing centralized control can broaden access to dangerous capabilities. Governance therefore must follow capabilities, permissions and actions rather than depending entirely on which company owns the model. But that does not change the direction of travel. The future of AI will be hybrid, heterogeneous and far more distributed than the current buildout suggests.

AI May Be Indispensable. Fossil Fuels Are Not

Wong and Warzel are on especially firm ground when they describe the loss of agency felt by communities confronting enormous data-center projects. A recent Heatmap poll found that 75% of Americans would oppose a new data center near where they live.

That is not necessarily a national rejection of AI. It is a rejection of an arrangement in which local communities are asked to absorb power, water, environmental and infrastructure costs for facilities whose benefits may largely flow elsewhere.

In The Indispensability Trap, I use Meta’s Hyperion project in Louisiana to illustrate “the shape of the thing.” It represents more than $200 billion in planned investment, approximately five gigawatts of electricity, extensive water and transmission infrastructure and ten proposed gas-fired power plants, all for a facility expected to support only about 500 long-term jobs. That is not merely a criticism of Meta. It demonstrates the physical scale and imbalance of the current approach.

The next example is already taking shape. OpenAI has announced plans to secure approximately eight gigawatts of IT capacity at the PORTS-Pike Technology Campus in Ohio. OpenAI acknowledges that development beyond the first phase will require new power plants, including natural-gas generation. Heatmap reports that the supporting energy proposal includes a 9.2-gigawatt gas facility that could become the largest fossil-fuel power project in the country.

This is one of my biggest problems with the Trump administration’s AI policy. It is using the urgency of the data-center buildout to advance a preexisting fossil-fuel agenda.

The administration deserves credit for supporting nuclear and geothermal development. But it has also explicitly directed federal agencies to identify coal infrastructure suitable for powering AI data centers while restricting new offshore-wind leasing. AI is becoming a permission slip for policies that increase fossil-fuel production and the resulting damage to the planet.

AI may be indispensable. Fossil fuels are not.

The International Energy Agency currently expects natural gas and coal to satisfy more than 40% of additional data-center electricity demand through 2030. That is a projected outcome based on current choices, not an unavoidable law of technological progress.

We need more efficient models and silicon, more distributed inference, new transmission, renewable generation, storage, geothermal power and nuclear energy. AI infrastructure should pay its incremental costs, disclose its water and energy consumption, add genuinely new clean generation and give affected communities a meaningful voice.

The future of AI may be unstoppable. A 9.2-gigawatt gas plant is a political choice, not a law of physics.

The Future Has Three Legs

I also do not view AI as an isolated technology. The future will rest on three interdependent legs.

AI supplies cognition, reasoning and coordination. Robotics gives that intelligence a physical presence and the ability to act in the world. Quantum computing will eventually provide specialized capabilities for problems that classical systems cannot efficiently solve.

They will not mature on identical timelines. Quantum computers will not replace GPUs or run our everyday chatbots. They will join CPUs, GPUs and NPUs in heterogeneous systems, initially accessed through specialized hybrid infrastructure. IBM is targeting its first large-scale fault-tolerant system, Quantum Starling, for 2029.

Robotics will strengthen the movement toward local intelligence. Machines operating around people cannot depend upon a remote data center for every latency-sensitive or safety-critical decision. Google DeepMind is already developing robotics models designed to operate locally when network access is limited or unavailable.

AI, robotics and quantum computing are the three legs. Clean, abundant electricity is the platform beneath them.

Responsibility Is Not a Verdict

Wong and Warzel are right to bring responsibility back to humans. It is convenient to assign agency to machines and then blame them for outcomes created by human incentives, human capital and human choices.

But recognizing that “it’s us” should not lead to fatalism. It should restore agency.

If humans built the incentives, architecture and energy system surrounding AI, humans can redesign them. We can demand better containment, human authority, interoperability, distributed intelligence, public accountability and clean power without pretending we can or should stop technological progress.

I am unabashedly bullish about what AI will become. That does not require me to defend every model, company, data center, financing scheme or political decision being made during its creation. Bullishness is not blindness.

The singularity may already be here. The indispensability trap is arriving with it. If history is any guide, AI will become larger, cheaper, more distributed and ultimately more useful than the companies now racing to own it.

The future will keep coming. Our responsibility is to remain at the helm.