When One AI Becomes Unbeatable

By C. Rich

Somewhere in the world today, enormous amounts of money are being spent on an AI race unlike anything humanity has ever witnessed. OpenAI, Google, Anthropic, xAI, Meta, and others are building increasingly powerful artificial intelligence systems, purchasing extraordinary quantities of computing hardware, constructing enormous data centers, and competing for an electrical supply that suddenly seems considerably less abundant than everyone assumed. Entire industries are being reorganized around machines that, not so long ago, were considered little more than sophisticated autocomplete.

Everybody wants to know who is going to win the AI race. But while pondering that question, I became interested in something considerably more important. Is there a point in this competition when winning becomes inevitable? Could one artificial intelligence become so powerful, so efficient, and so far ahead of its competitors that no amount of money, additional computing power, or technological innovation could allow the others to catch up? And if such a point exists, could we identify it mathematically before anyone crosses it?

That became the subject of my latest research paper, The AGI Tipping Point, developed within my broader Crest-Null research program. The investigation began with the possibility of one particular AI gaining an overwhelming industrial advantage, but the mathematics led somewhere more interesting. I discovered that the question could be examined without choosing a favorite horse in the race, which is fortunate because the horses keep getting faster, the racetrack keeps getting longer, and nobody seems entirely certain where the finish line is.

The central idea is relatively simple. Imagine two runners competing in a marathon. One has a substantial head start, but the runner behind is faster. Under ordinary circumstances, we can calculate precisely when the second runner will catch the first. We need to know their speeds, their separation, and whether either runner can maintain the pace. Now imagine that these runners are artificial intelligences, and that every mile they travel makes them capable of running faster. They can also improve their shoes, redesign the racetrack, and manufacture additional legs. One might eventually become so efficient that the other could never close the distance, even while improving its own performance. That is the possibility I wanted to investigate.

There is an important distinction between having the lead and becoming impossible to catch. A company might possess the world’s most capable AI today and lose that distinction tomorrow. Another company might have more computing power but use it less efficiently. A third might discover an algorithmic breakthrough that allows it to accomplish the same work with a fraction of the electricity. Simply counting computers tells us surprisingly little about who will ultimately prevail. What matters is the combination of available computation and how productively that computation is used. A machine that requires ten times less computing power to accomplish the same task has a substantial advantage, even if its competitor owns a considerably larger data center.

This is where the race becomes particularly interesting, because artificial intelligence may eventually play a growing role in improving the technologies that support artificial intelligence. Better software can assist in designing better hardware, more efficient cooling systems, improved manufacturing processes, and new algorithms. Those improvements can then support the development of still more capable AI systems. The possibility of a self-reinforcing technological cycle has fascinated researchers for decades. It is closely related to discussions of recursive self-improvement and the technological singularity, although neither requires us to assume that intelligence will improve without limit.

There is, after all, an inconvenient participant in this race that never reads corporate earnings reports or listens to enthusiastic predictions about the future. Its name is thermodynamics. Every computation requires physical machinery. That machinery consumes energy, generates heat, occupies space, and depends upon an industrial civilization capable of manufacturing and maintaining it. More efficient computation can reduce the cost of accomplishing a particular task, but no amount of clever software can abolish the physical requirements of the hardware performing the work.

An artificial intelligence might develop extraordinary new algorithms, but it cannot simply wish a million additional processors into existence. Someone must manufacture them, supply their electricity, remove their waste heat, and replace them when they fail. Even moving computation into space, where sunlight is abundant, introduces enormous engineering problems involving launch capacity, solar collectors, radiators, radiation exposure, and maintenance. The Sun might provide an extraordinary amount of energy, but collecting and using that energy is an entirely different matter. The electric bill may become less troublesome, but the laws of physics will still expect payment.

This physical accounting is particularly relevant to Crest-Null Philosophy, which examines the relationship between growing complexity, optimization, accumulated dependencies, and the persistence of systems over time. As technological systems become more powerful, they often become more interconnected. Their remarkable efficiency can depend upon elaborate supply chains, specialized manufacturing, concentrated infrastructure, and enormous quantities of energy. Those relationships create capabilities that earlier generations could scarcely imagine, but they can also introduce vulnerabilities. I wrote about it in my preprint: The Crest–Null Instability in Artificial Systems

An AI company might optimize its operations so effectively that its competitors struggle to match its performance. Yet that same optimization could leave it dependent upon a particular semiconductor supplier, electrical grid, cooling technology, or manufacturing process. A disruption at one critical point might compromise an otherwise extraordinary system. In Crest-Null terms, the processes that carry a system toward its Crest of complexity and efficiency may also establish the conditions of its vulnerability. Whether those vulnerabilities eventually produce a Null depends upon the system and the disturbances it encounters; collapse is not something we can simply assume.

This raises a curious possibility. The artificial intelligence that becomes the most powerful might not necessarily be the one best equipped to survive. It is conceivable that one AI could achieve astonishing computational superiority while another, less capable system survives through greater efficiency, independence, or resilience. The ability to win a competition and the ability to persist through disruption are related questions, but they are not the same question. My mathematical investigation therefore concentrated on a narrower problem that could actually be defined and tested. I asked whether it is possible to establish a measurable threshold at which one AI’s effective computational capability remains beyond the reach of its strongest competitors throughout a specified future period.

Rather than pretending to know the future, I constructed a mathematical criterion that compares a conservative estimate of the leading AI’s capabilities against favorable estimates of what its competitors could achieve. If the leader remains ahead even when the assumptions favor its rivals, it satisfies the proposed dominance condition for that particular period. That qualification matters enormously. Remaining unbeatable for five years is not the same as remaining unbeatable forever. Nor does possessing more effective computing capacity necessarily mean that an AI has achieved artificial general intelligence, consciousness, or the power to absorb other independent systems.

The mathematics does not magically turn a superior computer into a god, regardless of what its marketing department might prefer. What it does provide is a way to distinguish an ordinary technological advantage from a competitive advantage that another system cannot overcome within the conditions being examined. I explored how differences in computing infrastructure, algorithmic efficiency, and industrial expansion could affect that threshold. Some hypothetical scenarios produced substantial and persistent advantages. Others showed competitors overtaking the initial leader. The outcome depended heavily upon assumptions about future improvements, and that exposed the central difficulty.

I could calculate what would have to happen, but we could not establish from the available evidence that any particular AI was going to make it happen. That distinction became one of the most important findings of the investigation. There is no shortage of predictions about which company will dominate artificial intelligence. Some people expect the organization with the largest data centers to prevail. Others favor the company with the best researchers, the most advanced semiconductors, the greatest access to electricity, or the deepest financial resources. All of those advantages matter, but none independently establishes that the race has been won.

A competitor with inferior infrastructure might discover a dramatic improvement in algorithmic efficiency. A company with superior software might find itself unable to obtain enough electrical power. An enormous investment in hardware could become considerably less valuable if a new architecture achieves comparable results with fewer resources. That is why my research does not announce a winner. Instead, it identifies the kind of evidence that would be necessary to determine whether a decisive competitive threshold is approaching. I would need reliable measurements of usable computational capacity, sustained algorithmic improvement, industrial expansion, energy requirements, and the physical limitations facing competing systems.

I would also need to know how those measurements change over time. A snapshot of the current competition is insufficient to determine the outcome of a race in which the competitors are continuously improving the machinery they use to compete. The Crest-Null connection introduces another dimension. If one AI eventually gains an overwhelming advantage, we should ask whether the resulting concentration of computation, energy, industrial infrastructure, and information makes the larger technological system more resilient or more fragile. A world dependent upon several independent AI systems might possess forms of redundancy that disappear when critical functions become concentrated in one dominant architecture. On the other hand, consolidation could produce efficiencies and coordination that strengthen the system against certain disturbances.

Neither conclusion follows automatically. The relationship between optimization, dominance, and persistence must itself be investigated. This is where the prospect of an AGI tipping point becomes something more than a competition among technology companies. Artificial intelligence is increasingly intertwined with the infrastructure of modern civilization, and the distribution of its capabilities may eventually influence how that civilization functions, adapts, and survives. If a genuine tipping point exists, recognizing it might become one of the most consequential measurements I could make.

I have not established that such a physically achievable threshold exists for today’s frontier AI systems, much less that anyone has crossed it. What I have developed is a mathematical criterion for recognizing a particular kind of sustained competitive dominance, together with a clearer understanding of the evidence needed to apply it. For me, that is a worthwhile outcome. Research does not always deliver the answer one initially hopes to find, and sometimes the most useful discovery is learning exactly what remains unknown.

The AI race will continue, with extraordinary fortunes spent pursuing increasingly powerful systems. There will be new models, new algorithms, new data centers, new energy projects, and almost certainly new declarations that somebody has finally won. But the question I find most intriguing is no longer simply which artificial intelligence will be the first to reach the summit. It is whether there comes a point when one reaches a height that no competitor can ever attain, whether that apparent supremacy can endure the physical costs of maintaining it, and whether the very architecture that makes such dominance possible might eventually become its greatest weakness. Those are the questions behind The AGI Tipping Point. They are also questions that Crest-Null Philosophy gives us an interesting way to approach, because the history of complex systems suggests that reaching the highest point and remaining there may be two very different achievements.