A physicist defines intelligence as the ratio of useful work to total information processed. The number explains what organizational consultants cannot: why six hundred and fifty billion dollars in AI infrastructure produces near-zero measurable return for most companies that deploy it.
The most cited finding in the AI investment debate comes from an MIT study released last August: ninety-five percent of companies investing in generative AI pilots report no measurable impact on their profit and loss statement. The consulting industry responded with the usual diagnoses — poor change management, unclear use cases, insufficient executive sponsorship, inadequate data infrastructure. These explanations are not wrong. They are also not interesting. They describe symptoms. A physicist at the University of Edinburgh may have identified the condition.
Peter Fagan's paper "Toward a Physical Theory of Intelligence" proposes that intelligence is not a capacity. It is a ratio. Specifically: chi equals the amount of goal-directed work an agent produces divided by the total information it irreversibly processes. This is not a metaphor. It is a physical quantity derived from conservation laws and the second law of thermodynamics. The definition applies to biological neural networks, artificial ones, and any other system that converts information processing into environmental effects.
The Numerator and the Denominator
Goal-directed work is the measurable change an agent produces in its environment toward a specified objective. Irreversible information processing is the total computational dissipation — every bit the system processes that cannot be recovered. An agent that processes a trillion tokens and produces no environmental change has a work ratio of zero, regardless of how sophisticated its internal representations are. An agent that processes ten tokens and produces a measurable outcome has a nonzero work ratio. The physics does not care about elegance. It cares about effect.
The AI infrastructure cycle now under construction will consume approximately ninety-six gigawatts of power globally by the end of this year — roughly double the 2023 figure. The International Energy Agency projects data center electricity demand will exceed a thousand terawatt-hours by 2026. Eighty to ninety percent of that energy services inference — the act of processing inputs and generating outputs. The irreversible information processing is enormous and growing. The denominator is not the problem.
The MIT finding is a measurement of the numerator. For ninety-five percent of companies, the goal-directed work — the measurable business outcome produced by all that computation — registers as zero. The work ratio for the median AI deployment is not low. It is zero. Not approximately zero in the way that a small number rounds down. Zero in the way that a numerator without a value is zero. The energy dissipates. The tokens flow. The work does not materialize.
What Capacity Metrics Hide
The industry measures AI by capacity: parameters in the model, tokens in the context window, benchmarks passed, operations per second. These are measurements of the denominator — how much information a system can irreversibly process. They tell you nothing about how much of that processing converts to work.
This is not an oversight. Capacity is easy to measure. Work is hard to measure. You can count parameters by reading a model card. Measuring whether a deployment produced goal-directed change requires defining the goal, instrumenting the outcome, and waiting long enough for the effect to be distinguishable from noise. The industry has optimized for the metric it can measure, which happens to be the wrong half of the ratio.
Fagan's framework predicts this. A system with massive irreversible processing and near-zero goal-directed work has a chi approaching zero. In thermodynamic terms, it is pure dissipation. The six hundred and fifty billion dollars in AI infrastructure spending this year is building dissipation capacity. Whether it builds work capacity is a separate question that the spending itself cannot answer.
What Twenty Watts Accomplishes
A Nature Communications study published this month tested a specific prediction about biological intelligence. Barbey and Wilcox at the University of Illinois scanned 831 adults from the Human Connectome Project and confirmed that human intelligence is not localized to any single brain region or network. It arises from the efficiency of distributed coordination — the balance between local specialization and global integration across the entire brain.
The brain consumes roughly twenty watts. It does not maximize raw neural firing — that would generate seizures. It maximizes the ratio of cognitive work to metabolic cost. Regulatory hubs coordinate which networks activate and when. Intelligence, measured by the full range of cognitive tasks a person can perform, correlates with the quality of that coordination — not the quantity of neural activity. The brain with the highest work ratio is not the one burning the most glucose. It is the one converting the most glucose into directed thought.
The parallel to AI deployment is direct. The companies in the MIT study's successful five percent are not the ones spending the most on compute. They are the ones whose compute produces measurable environmental change. Block published the most concrete AI productivity metric in corporate history — gross profit per employee doubling after an eighteen-month AI deployment. The metric was not tokens processed or GPU hours consumed. It was a work ratio expressed in dollars: how much economic output per unit of organizational input.
The Maturity Transition
The shift is beginning to appear in industry language. Data center operators have started discussing a new metric: tokens per watt per dollar. Tokens per watt is a capacity measurement — throughput per unit energy. Adding the dollar converts it into something closer to Fagan's ratio: how much of that processing produces economic value. The metric is crude, but it represents the first industrial attempt to measure both halves of the fraction.
Every technology follows the same maturity curve. Early steam engines were measured by displacement — how much volume the cylinder could hold. Mature steam engines were measured by thermal efficiency — the Carnot ratio of useful work to total heat energy. Early computers were measured by clock speed. Mature computing is measured by performance per watt. The metric that defines a technology shifts from capacity to efficiency as the technology matures, because capacity without efficiency is just heat.
Early AI is measured by parameters and benchmarks. Mature AI will be measured by the work ratio. The transition point is when the industry stops asking how much a model can process and starts asking how much of that processing converts to something that changes the world outside the data center.
The Hidden Variable
Fagan's framework resolves a puzzle that the organizational explanations cannot. If the problem were change management, you would expect AI returns to distribute normally — some companies executing well, most executing poorly, a few executing terribly. Instead, the distribution is binary: five percent with measurable returns, ninety-five percent with zero. A normal distribution of execution quality does not produce a binary outcome. A threshold does.
The work ratio provides the threshold. Below a certain chi, processing generates no measurable work regardless of how much processing occurs. Above it, processing converts to work and returns compound. The transition is not gradual. It is a phase change — the same physics that governs when water becomes steam. The companies in the five percent have crossed the threshold. The companies in the ninety-five percent are dissipating energy below the boiling point, and no amount of additional heat — no additional spending — produces steam until the system reaches the transition temperature.
What determines the transition? Fagan's Conservation-Congruent Encoding framework suggests the answer: the quality of the mapping between the system's internal representations and the environment's actual causal structure. A deployment whose encodings correspond to real environmental constraints converts processing to work. A deployment whose encodings float free of the environment — impressive internal representations with no causal connection to the business problem — dissipates everything it processes.
The physics is clear. Intelligence is not how much you process. It is how much processing converts to work. The AI industry is doubling its computation every year. Whether six hundred and fifty billion dollars was well spent depends not on whether the models are more capable but on whether the work ratio is rising. So far, for ninety-five percent of the companies writing the checks, it is not.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)