AI Did Not Give Me More Ideas. It Stopped Existing Ones From Dying.

The AI Resource Paradox is undeniable. It consumes electricity, memory, capital and infrastructure while releasing another scarce resource: the cognitive capacity required to keep worthwhile work alive.

The first half of that paradox is already visible. AI systems require data centres, grid connections, cooling, specialist chips, construction capacity and large amounts of investment. The resulting bottlenecks are real. They should not be minimized by vague promises that technology will somehow solve every constraint it creates.

But resource consumption is not the same as net resource destruction. Breakthrough technologies often arrive inside systems that were not designed for them. They initially compete for the materials, infrastructure and skills already available. If they subsequently increase the capacity to produce, coordinate or allocate those same resources, the long-term effect can differ significantly from the first visible shortage.

Railways required enormous quantities of coal, steel, land and capital. They also changed how coal, steel, machinery, labour and food could be moved. Containerization required ports, cranes, ships, rail terminals and logistics systems to be rebuilt around a standardized box. It then reduced friction across global trade. These analogies do not prove that AI will create abundance. They show why a technology’s initial demand on an inherited system is not necessarily the final measure of its contribution.

AI consumes scarce physical capacity. It may also release scarce cognitive capacity. Leadership determines whether the second effect strengthens the systems strained by the first.

The AI resource paradox appears before the abundance

Every technology depends on an existing production system before it can help reshape that system. A new railway cannot transport the steel required for its first tracks before those tracks exist. A container network cannot reduce logistics costs before ports and ships are adapted to handle containers. The transition begins with concentrated demand, visible disruption and competition for scarce resources.

That creates a predictable analytical error. Observers measure the resources being consumed during the transition and treat the surrounding system as static. Demand rises while supply, productivity and coordination are assumed to remain unchanged. The conclusion then appears obvious: the technology can only intensify scarcity.

Sometimes that conclusion will be correct. A technology can consume more than it creates. Efficiency gains can be absorbed by rebound effects. Dominant firms can use new capacity to outbid everyone else rather than expand the system. Benefits can remain private while infrastructure costs become public. Nothing about technological progress guarantees a socially useful allocation of the resulting capacity.

The missing question is therefore not whether AI uses resources. It plainly does. The more important question is whether AI creates forms of productive capacity that can eventually improve the systems on which it depends.

My own scarce resource was continuity

I encountered a small version of this paradox long before I started thinking about data centres. My constraint was not a shortage of ideas. It was the amount of continuous attention required to keep an idea coherent.

A promising concept would appear, gather momentum and then collide with reality. Client work took priority. A tax return had to be filed. A job application became urgent. A family commitment could not move. The idea did not necessarily fail. It simply lost access to sustained attention.

Some projects deserve to stop because the underlying value hypothesis is weak. Others disappear for a less rational reason: the cost of restarting becomes greater than the energy available to continue. Notes remain somewhere. Decisions become difficult to reconstruct. Research has to be repeated. The next step is no longer obvious. A temporary pause becomes abandonment without anyone making an explicit decision to abandon the work.

I have wanted to build a serious publishing engine for as long as I can remember. The ambition was not merely to publish occasional articles. I wanted a system capable of turning observations into essays, essays into frameworks, frameworks into a coherent body of knowledge, and that body of knowledge into useful conversations and professional opportunities.

Open Source Consulting is now becoming that system. Earlier versions repeatedly stalled. The problem was not conviction. It was the invisible coordination required to keep the engine operating. Topics had to be captured, compared and sequenced. Arguments had to remain distinct across articles. Sources needed to be preserved. Website structures, metadata, images, distribution and follow-up all required attention. Once the work stopped, restarting meant reconstructing the machine before I could use it again.

The same pattern affected my fiction. During one vacation, I wrote thirteen pages of a science-fiction novel called An Act of Desperation. The initial burst was real. The premise continued to interest me. But when ordinary life resumed, so did the cost of re-entry. I needed to remember what I had decided about the world, the characters, the physics, the chronology and the reveal structure. I needed to distinguish deliberate clues from inconsistencies. Before I could write the next page, I had to recreate the mental state in which the previous thirteen had been written.

I gave up, not because the story had become weak, but because restarting it felt like a project in its own right.

Startup ideas encountered another version of the same constraint. Validation sounds modest when reduced to a line in an innovation process: speak to potential users, test assumptions and synthesize the evidence. In practice, even a small validation cycle requires questions to be designed, surveys to be assembled, respondents to be found, answers to be normalized, contradictions to be interpreted and decisions to be documented. The idea might have deserved validation. The machinery of validation demanded more capacity than I could justify.

Pausing work is not the same as preserving it

Organizations commonly classify work as active or inactive. That distinction hides a more important one. An inactive project can be deliberately preserved, with its current state, unresolved questions, dependencies and next action intact. Or it can simply stop receiving attention while its internal coherence decays.

The second condition is common. A team pauses an initiative and retains the documents but loses the reasoning that connected them. A product concept remains in a slide deck, but the assumptions behind it are no longer visible. Research findings survive, but nobody remembers which conclusions were rejected and why. When capacity returns, the files create the illusion that the project has been preserved. In reality, the new team, or the future version of the same person, has inherited an archaeological site.

Research into interruptions and fragmented work has long shown that returning to a task carries a cognitive cost. Studies associated with Gloria Mark and colleagues at the University of California, Irvine found that knowledge workers frequently switch activities and must reconstruct their mental context when returning. A multi-month project pause is not identical to an interruption measured during a workday, and I am not claiming a scientifically fixed restart time for dormant ventures. The relevant parallel is narrower: resumption requires goals, state, context and intended action to be rebuilt. UC Irvine’s work on fragmented attention provides the wider research context.

A folder preserves files. A resumable system preserves decisions.

AI changed the economics of restarting

The change did not begin with an autonomous agent completing projects while I slept. It began with less dramatic capabilities: asking questions, summarizing state, finding dependencies, decomposing work and maintaining explicit next actions.

I now maintain a curated and prioritized backlog. It is not a storage facility for every thought I have ever had. It is an operating interface between objectives, available capacity and bounded pieces of work. When two hours become available, I should not need to spend those two hours discovering what might be useful. I should be able to pull a prepared increment from a system whose state has survived the interruption.

AI helps inspect the current objective, work already completed, known dependencies and unresolved decisions. It can suggest how a large activity might be split into smaller outcomes. It can identify where one task depends on another. It can turn an ambition such as “revamp the website” into a sequence of independently useful changes that can be completed, validated and paused.

This is a practical example of the principle developed in Redesigning Work With AI Starts When I Stop Trying to Use AI. The value does not come from inserting AI into the existing task. It comes from redesigning the work around a different constraint system.

Decomposition used to be one of the most expensive parts of project management for me. Executing a small task might take an hour. Determining what the small task should be, and whether completing it would leave the wider system coherent, could require more thought than the task itself. Large activities therefore remained large. Once interrupted, they were difficult to resume without circling back through the entire subject.

The specialist GPT roles supporting this system follow the same logic described in AI Agent Roles: Why I Gave Mine Ridiculous Names. Memorable roles are useful only when their context, authority, handoffs and validation boundaries are explicit. The AI does not own the project. Human judgment remains the validation gate.

One prompt after the tax return

A recent interruption provided a practical test. I was in the middle of revamping the Open Source Consulting content websites. The work involved more than visual improvement. It required a design system, reusable components (that are premiering in this article), clearer page roles, better navigation and a progression from an article toward another useful asset or executive action.

Then I had to stop and file my tax returns.

The interruption was not optional, and I did not try to maintain both streams simultaneously. The website work halted. In the old system, returning would have begun with a long retrospective: Which pages had I examined? What had I decided about Elementor and Gutenberg? Which component definitions were stable? Where were the unresolved defects? What should I do first?

This time, re-entry required one prompt. I asked the specialist GPT responsible for the website system to recover the current state and create smaller items in my Jira backlog. The system already contained the objectives, decisions, component rules, technical limitations and previous validation. Instead of reconstructing the whole project, I reviewed a proposed sequence of bounded actions.

The AI did not make the decisions for me. I still had to judge whether the decomposition was sensible, whether dependencies were real and whether each activity deserved priority. But my judgment was applied to the decision in front of me rather than spent recreating the history behind it.

Four ventures and three books are not seven active priorities

I now have four venture concepts and three books represented in a Venture Portfolio. Some can be named openly: Open Source Consulting, the Open Source Consulting book, An Act of Desperation and Magnatopia. Others can be described at category level: a B2B ESG platform, for example, without disclosing the intellectual property that might make them distinctive.

The significant change is not that seven initiatives are racing toward launch. They are not. It would be irresponsible to describe them as seven active ventures and absurd to treat their existence as proof of sevenfold productivity. Employment and income continuity remain more important. Some concepts may never pass validation. One book may progress while another remains deliberately dormant.

The change is that dormancy has become more deliberate.

An Act of Desperation is no longer thirteen pages and a fading memory of what I intended. It now has drafted chapters, a chapter architecture, a world and science bible, continuity rules, a reveal sequence and a production backlog. That structure does not write the novel. It means the next writing session can begin with the story rather than a forensic reconstruction of the story.

Magnatopia began years ago as an ambition to create a world large enough to support stories, games, films and collaborative development. The idea became overwhelming partly because maintaining a long fictional history, its factions, technologies, religions and contradictions required a coordination system I did not have. AI does not make that world small. It makes the work more decomposable. It can help identify inconsistencies, preserve canon, prepare bounded research questions and maintain a visible frontier between what has been established and what remains open.

Startup validation changes in a similar way. AI can help prepare a first survey draft, challenge leading questions, structure interview notes and synthesize patterns across responses. Every output still requires review. Respondents still have to be found. Evidence still has to come from the market rather than the model. But the supporting machinery no longer consumes the same proportion of available effort.

The US National Institute of Standards and Technology is investigating comparable mechanisms in a different setting. Its work on human-machine teaming in manufacturing explores how guided problem formulation and incremental exposure can lower the cost of using complex tools. The domain is different, but the mechanism is relevant: AI can reduce the entry and coordination cost around complex work rather than merely automate its final task. NIST describes that research in its human–machine teaming programme.

The hidden capacity is in the restart

Most discussions of AI productivity focus on active execution. How quickly can a document be produced? How many lines of code can be generated? How many customer interactions can be handled? Those measures are useful, but they overlook a large category of organizational waste: work that repeatedly pays the cost of beginning again.

A strategy team revisits an opportunity and repeats the market scan because the reasoning behind the earlier conclusion was not preserved. A product team reopens a concept but cannot distinguish rejected assumptions from unresolved ones. A transformation initiative changes leadership and spends months rebuilding stakeholder context. An innovation portfolio contains dozens of ideas but no explicit activation conditions or next validation steps.

In each case, the organization appears to have preserved information. What it has not preserved is operational continuity.

Capacity is not only how much work we can start. It is how much valuable work can survive a pause.

This extends the argument in The AI Capacity Dividend. AI may recover time, attention, decision velocity and iteration capacity. The larger return depends on allocation. If every gain becomes an instruction to produce more immediate output, the operating model remains unchanged. Preserving and resuming worthwhile work is another possible allocation: use part of the dividend to reduce the decay of organizational knowledge and strategic options.

From cognitive capacity to productive capacity

My personal experiment does not prove that AI will solve the electricity, memory, water or infrastructure scarcity generated by its expansion. Four venture concepts and three books do not offset a data centre. That would be an unserious comparison.

What the experiment provides is evidence of a smaller mechanism. AI can release cognitive and organizational capacity by lowering the cost of structuring, preserving and resuming complex work. Infrastructure constraints are not only engineering and materials problems. They are also research, coordination, permitting, project-design, knowledge-transfer and decision problems. Those are areas in which additional cognitive capacity may matter.

An energy company may use AI to accelerate grid planning, but only if the released capacity is applied to real projects rather than absorbed by more reporting. A manufacturer may explore more material alternatives, but only if promising results survive the handoff into engineering and procurement. A public authority may process infrastructure documentation more effectively, but only if faster analysis is paired with accountable decisions. The causal path from cognitive leverage to physical abundance is neither automatic nor immediate. It passes through organizations.

This is where the Resource Paradox becomes an operating-model question. AI can make more analysis, options and coordination possible. Leaders still decide which work enters the backlog, which gains become higher output expectations, which initiatives receive sustained investment and which external costs are tolerated.

The capacity released by AI can be used to expand energy supply, improve resource productivity, shorten infrastructure development, preserve institutional knowledge and activate research that previously stalled under its own coordination burden. It can also be consumed by more content, more dashboards, more meetings and a higher volume of decisions that nobody has time to validate.

Why the dividend may disappear

These are not arguments against AI. They are reasons to reject simplistic productivity measures. An organization can produce more while becoming less sustainable. It can automate individual tasks while weakening the knowledge and infrastructure required to keep operating. It can report time saved without identifying where the time went.

The executive question is what the released capacity builds

The Resource Paradox cannot be resolved by counting only what AI consumes or only what it accelerates. The relevant unit is durable productive capacity.

Leaders should ask what exists because the AI system was deployed that would otherwise not exist: a better-maintained infrastructure asset, a validated product option, a preserved body of knowledge, a shorter permitting cycle, a new research path, a more resilient operating process, or a project that survived long enough to earn its next investment.

They should also ask what the system consumed to create that outcome, who bore the cost, whether the gain is durable and whether rebound effects erased the benefit.

For my own work, the result is modest but real. AI did not give me more hours and did not make every idea worth pursuing. It changed the cost of keeping a worthwhile option coherent. It allowed a publishing engine, books and venture concepts to progress intermittently without requiring seven continuous streams of attention.

That is not abundance. It is a small increase in productive capacity.

The larger question is whether organizations can create the same effect at scale and then direct that capacity toward the physical, institutional and economic systems AI is currently straining. Dale Carnegie would claim that a person always has a noble reason and a real reason to do something: the same company that is claiming to save the world, might have just uncovered a gold mine in doing so.

AI consumes capacity. AI also releases capacity. Whether scarcity becomes abundance depends on what we build with the difference.