The Moving Bottleneck: What Actually Limits AI Progress?

AI progress is supposed to be running into limits almost everywhere at once. Frontier models are consuming much of the readily available high-quality public text; giant training clusters strain semiconductor supply chains and electricity infrastructure; more elaborate reasoning systems consume increasing amounts of computation after training; and AI agents can perform impressive individual tasks while remaining unreliable over long workflows.

Yet frontier AI is not behaving like a technology at its ceiling. Stanford’s 2026 AI Index reports a roughly 30 percentage-point improvement in a single year on Humanity’s Last Exam, a benchmark deliberately designed to remain difficult for advanced models. On OSWorld, which tests an agent’s ability to operate computers, performance rose from roughly 12 percent to 66.3 percent. The same report captures the other side of the picture: agents still fail roughly one in three structured benchmark attempts, while systems capable of solving extremely difficult mathematical problems remain surprisingly unreliable at apparently simple tasks.

This looks less like a technological plateau than the end of an unusually simple phase of AI development. For much of the past decade, one scaling recipe worked remarkably well: more parameters, more data and more compute generally produced better models. Now attempts to escape one constraint increasingly consume another scarce resource. Compute can manufacture synthetic training data, but only if the results can be checked. Inference compute can compensate for a model that cannot immediately solve a difficult problem, but only if the extra reasoning is worth its cost. Sparse architectures can save computation while demanding more communication between processors. Simulation can provide experience without operating in the real world, but only to the extent that its lessons survive contact with reality. AI may therefore face not one scaling wall but a moving bottleneck, with the harder question being how long these substitutions remain economical.

AI scaling increasingly depends on infrastructure beyond the server rack. Editorial image generated by the author.

Scaling Was the Unusually Easy Phase

The remarkable feature of the transformer era was not simply that models became better. Researchers found a comparatively reliable way of making them better: increase model size, increase the amount of training data and increase training compute. Scaling laws made the returns predictable enough that an AI company could spend vastly more on a training run with some confidence that the resulting system would improve. For a research field normally characterised by uncertain breakthroughs, that predictability was enormously valuable.

The rational response was to build larger clusters, collect larger datasets and train increasingly large models. But scaling laws never implied that every required input could expand indefinitely at the same cost. Eventually one component becomes more difficult to increase than the others.

Computing has encountered this pattern before. For decades, processors became faster partly by increasing clock speeds. When heat and power consumption made that route increasingly unattractive, progress did not stop; more work moved towards multiple cores, GPUs, specialised accelerators, larger caches and distributed systems. The architecture of computing changed around the constraint. AI may now be undergoing a similar transition, from an era in which one dominant scaling recipe solved many capability problems to one in which several constraints interact.

Pretraining Is Running Into Information and Infrastructure

Training data is perhaps the most obvious example. A large language model benefits from exposure to more human knowledge and behaviour, but there is a finite amount of high-quality human-generated public text. Researchers at Epoch AI estimated the effective stock of quality- and repetition-adjusted public human-generated text at roughly 300 trillion tokens, with a very wide uncertainty range. Their modelling suggested that, if historical dataset-growth trends continued, training datasets could fully utilise that stock sometime between 2026 and 2032. Epoch’s estimate of public human-generated text does not imply that AI is about to run out of data altogether. Its broader analysis of scaling through 2030 argues that images, video, audio and synthetic data could expand the usable pool considerably even if high-quality public human text becomes increasingly constrained.

The increasingly scarce resource is therefore not tokens as such, but useful learning signal. A trillion copies of Wikipedia would constitute an enormous dataset without containing a trillion times as much knowledge. Once models have seen many examples of something they already understand, additional repetitions become less valuable than examples exposing a genuine weakness, unfamiliar concept or difficult decision.

That changes the kind of data that matters. Expert reasoning, scientific experiments, software failures and records connecting decisions to outcomes may become increasingly valuable because they contain information not easily reconstructed from another scrape of the public web. Some proprietary datasets held by laboratories, engineering companies, hospitals or financial institutions could therefore be exceptionally useful. Useful does not mean available: privacy, confidentiality, fragmented ownership, regulation and incompatible formats may leave much of this material inaccessible to frontier training.

Compute brings a parallel set of constraints. Larger training runs require not merely more accelerators, but the industrial system around them: advanced semiconductor fabrication, high-bandwidth memory, packaging, networking, cooling and electricity infrastructure. At sufficiently large scales, simply adding processors becomes less helpful if those processors spend increasing amounts of time waiting for information. Training a frontier model across thousands of accelerators requires constant movement of parameters, gradients and activations. Epoch’s modelling of distributed training suggests that, under current transformer-like architectures, hardware and scaling techniques, data movement can progressively reduce utilisation at extremely large training scales. The numerical thresholds depend on assumptions about hardware, network design, batch sizes and training duration, but the broader direction is familiar from computing: once arithmetic becomes cheap enough, moving information starts to dominate.

Electricity creates a similar mismatch between theoretical abundance and practical availability. The International Energy Agency projects global data-centre electricity consumption to roughly double from 2024 levels to around 945 TWh in 2030, still just under 3 percent of worldwide electricity consumption. That makes claims that AI will simply consume all available global electricity difficult to sustain. The more immediate constraint is timing and location: the IEA notes that a data centre can become operational in two or three years, while generation, transmission and other supporting infrastructure often require much longer lead times.

An AI company does not need an abstract share of world electricity. It may need hundreds of megawatts or even gigawatts at one particular location, continuously, within a few years. The global percentage can therefore remain modest while the local grid consequences of AI demand become substantial. The constraint increasingly lies in the less glamorous infrastructure around the server rack: transmission lines, transformers, substations, grid connections and new generating capacity. AI has made electricity infrastructure part of the computing stack.

Synthetic Data Turns Scarcity into a Verification Problem

If high-quality human training material becomes difficult to obtain, one obvious response is to manufacture more data. AI systems can generate mathematical problems, programming exercises, hypothetical scenarios and reasoning examples vastly faster than humans can produce them. In principle the quantity is unlimited, but unlimited generated text is not unlimited knowledge. The more useful distinction is whether the learning loop remains open to independent correction.

Research published in Nature demonstrated how recursively training generative models on model-generated outputs can distort the underlying distribution when generated material is used indiscriminately. Rare patterns disappear, errors accumulate and later generations increasingly learn a machine-generated approximation of the original material. The result is not evidence that synthetic data is inherently harmful. It shows why simply recycling generated material cannot be treated as an infinite substitute for fresh information.

Synthetic data becomes much more powerful when something independent can judge it. Generate a million equations and verify their answers mathematically. Generate software and execute the tests. Generate candidate strategies in a game and evaluate the outcomes. The model can explore enormous numbers of possibilities because incorrect examples can be discarded cheaply. This helps explain why mathematics and programming have been particularly fertile domains for reinforcement learning: they often provide something much of the world does not, an inexpensive way to determine whether an answer works.

There is no equally straightforward test suite for deciding which monetary policy a country should adopt or whether an unfamiliar scientific hypothesis deserves further investigation. There may be better and worse answers, but correctness is less easily reduced to a machine-checkable reward. Recent work on reinforcement learning with verifiable rewards describes reliance on simple rule-based verifiers as a fundamental bottleneck when expanding these techniques beyond structured areas such as mathematics and coding. Researchers are experimenting with learned and generative evaluators, but that shifts part of the problem into the evaluator itself.

This may become one of the deeper constraints on AI development. Human feedback is most useful when people can reliably judge what the model has produced. A system capable of proposing genuinely novel mathematics, engineering or scientific theories may eventually generate work that very few people can evaluate. Asking another model to judge it helps only insofar as there is some reason to trust the judge. The scarce resource then stops being candidate intelligence and becomes confidence that the candidate is correct.

Reasoning Moves Scaling from Training to Inference

The rise of reasoning models has already demonstrated one route around the limits of conventional pretraining. Traditional scaling concentrated most computation before deployment: a company spent heavily training a model, after which the finished system produced relatively inexpensive answers. Increasingly, models can also spend substantial computation after receiving a question, trying several approaches, using software tools, searching for information, testing intermediate results and backtracking when an approach fails.

This creates another dimension of scaling. Instead of asking only how much computation should be spent building the model, we can ask how much should be spent solving this particular problem. Compute can substitute, at least partly, for immediately available capability: a model that cannot produce the correct answer on its first attempt may still reach it after exploring twenty alternatives.

The exchange rate is not constant. A 2026 study of test-time scaling found diminishing marginal returns as reasoning budgets increased and documented cases of “overthinking”, where additional reasoning led models to abandon answers that had previously been correct. The optimal amount of reasoning also varied with problem difficulty. Simply allowing a model to think longer is therefore not an unlimited source of capability.

The economic question changes with it. It is no longer sufficient to ask whether an AI can solve a problem; the additional intelligence purchased through more inference has to justify its cost. A difficult scientific or engineering problem might warrant thousands of times the computation of an ordinary query, while most everyday tasks will not. Efficient systems will need to allocate reasoning selectively, spending computation where the expected return is high and stopping when additional thought is unlikely to improve the answer.

Test-time computation thus relaxes the constraint imposed by what the pretrained model can immediately do while increasing demand for inference hardware, electricity and cheap computation. A route around one bottleneck becomes additional pressure on another.

Agents Move the Bottleneck from Intelligence to Reliability

AI agents expose a different problem. A model can perform extremely well on isolated tasks and still be unsuitable for running a complicated process autonomously. Suppose an agent performs each consequential step correctly 99 percent of the time. Across twenty independent steps, the probability of completing all of them without error is about 82 percent. Across two hundred, it falls to roughly 13 percent.

Real agent failures are not independent in such a neat way, and many errors can be detected and corrected. The example nevertheless captures the underlying difficulty: long-horizon autonomy magnifies small weaknesses. An assistant researching a subject for ten minutes can make a mistake and recover. An agent operating for days or weeks has to survive thousands of decisions, tool calls, external changes and conflicting pieces of information without gradually drifting away from its objective.

This helps explain the apparently contradictory benchmark picture. Stanford’s 2026 AI Index reports enormous progress on OSWorld, yet agents still fail roughly one third of structured benchmark attempts. A system can therefore become much more capable at individual actions without becoming correspondingly dependable over long sequences. The engineering challenge begins to shift towards memory, planning, error detection, recovery and goal stability.

A major future improvement in AI might consequently look surprisingly modest on conventional intelligence benchmarks. A model that is only slightly better at solving individual problems but far less likely to lose track of its objective, misuse a tool or fail to notice its own mistake could be much more economically useful. For autonomous systems, dependability may matter more than another spectacular benchmark score.

Eventually the Scarce Resource Is Experience

There is a deeper limit behind the training-data problem. Language models mostly learn from records of experience rather than experience itself. Text describes what people observed; images preserve snapshots; video captures sequences of events. These contain enormous amounts of information about the world, and multimodal systems can learn aspects of physical behaviour that are difficult to infer from text alone.

An acting system can acquire another kind of information: what happens after it intervenes. Change something and observe the consequences. Try a strategy and see whether it works. Make a prediction and compare it with the outcome. The cycle of action, consequence and adjustment supplies information that passive observation cannot fully reproduce.

Robotics, scientific experimentation and increasingly sophisticated world models could give future AI systems more of this feedback. The gap between simulation and reality, however, remains large. Stanford’s 2026 AI Index reports 89.4 percent success on the simulated RLBench manipulation benchmark, compared with only 12 percent success on real household tasks.

Reality runs in real time.

A computer can generate millions of mathematical exercises cheaply. A physical robot cannot conduct millions of real experiments at anything like the same speed or cost. Hardware wears out, experiments consume materials, failures can damage equipment or injure people, and many physical processes simply take time.

Simulation offers an obvious substitute. Large numbers of agents can act in simulated worlds and acquire experience much faster than physical systems could. The value of that substitution depends on whether the simulation captured the parts of reality that matter. An agent that becomes brilliant at exploiting a simplified virtual world may merely have learned the omissions of the simulator. The further AI moves from manipulating symbols towards manipulating physical reality, the less cheaply experience can be scaled.

New Architectures Change the Exchange Rate

None of this implies that transformers must remain the dominant architecture. They may eventually be supplemented or displaced by systems that use computation, memory and data more efficiently. A new architecture, however, does not abolish resource constraints. It changes the trade-offs between them.

Mixture-of-experts systems are a useful example. Instead of activating an entire enormous network for every input, they route each token through only a subset of the model. This can dramatically increase total model capacity without increasing computation proportionally. The price is a more complicated routing and communication problem: the relevant experts may reside on different processors, and the hardware must keep them supplied without leaving expensive accelerators idle.

Other solutions create similar exchanges. Very large context windows reduce the need to retrieve information from external memory but increase the amount of material the model must process. External memory reduces context requirements but creates problems of selection: what should be remembered, what should be forgotten, which memory should be retrieved and whether it is still correct.

The analogy with CPUs is useful as long as it is not pushed too far. Modern computers did not abandon central processors when processors alone proved insufficient; layers of memory, networking, accelerators and software grew around them. Large language models may similarly become one component of larger cognitive systems combining reasoning, retrieval, memory, software tools, verifiers, planning and world models. That does not mean transformers are permanent. It means that discovering shortcomings in a bare LLM is not the same thing as demonstrating that the architecture has exhausted its usefulness.

What Would an Actual AI Wall Look Like?

The pattern so far can be understood as a series of substitutions.

Constraint	Possible substitute	What becomes scarce next
Public human training data	Synthetic and multimodal data	Verification and useful learning signal
Immediate model capability	Test-time reasoning	Inference cost and latency
Dense-model computation	Sparsity and specialisation	Communication, memory and routing
Short-task intelligence	Agents and longer workflows	Reliability and error accumulation
Passive observations	Simulation and physical interaction	Real-world experience, time and safety

This framework is not an argument that AI progress is guaranteed. There is no law requiring every bottleneck to have an affordable substitute. Indeed, that is where the strongest version of the scaling-wall argument lies: a genuine slowdown may occur not when one resource literally runs out, but when several substitutions become expensive at the same time.

Synthetic data might become increasingly difficult to verify outside narrow domains just as additional training compute delivers smaller gains. Longer reasoning might continue to improve performance but at costs that make many applications uneconomic. Simulation might provide enormous amounts of experience while transferring poorly enough to the physical world that real experiments remain indispensable. At that point, removing one constraint would no longer merely expose a manageable next bottleneck; the exchange rate between resources would have turned against further scaling.

We do not appear to be there yet. Current capability improvements make claims that LLMs have already reached their peak difficult to reconcile with the evidence. But the success of the past decade can mislead in the opposite direction as well. The fact that scaling has repeatedly found a route around previous constraints does not guarantee that every future constraint will yield as cheaply.

The easy scaling era taught us to ask how much larger the next model could become. The harder question is what it costs to move the bottleneck. Compute can compensate for scarce data only if useful synthetic data can be generated and checked cheaply enough. More inference can compensate for imperfect immediate reasoning only while additional computation continues buying enough accuracy. Simulation can substitute for real experience only insofar as its lessons transfer back into reality. AI may never encounter one dramatic wall; its limits may instead emerge as a worsening series of trades, with more money, energy, computation or complexity required for each additional increment of capability.

If that happens, the decisive question will not be whether another bottleneck can be moved. It will be how many times we can afford to move it before the exchange rate finally turns against us.