The Unbottlenecking

Clarke was wrong.

Magic is inexplicable. Technology is not. Technology is the application of intelligence to the problem of constraints. We see a bottleneck. We apply cognition. We break through. There is no mystery, only engineering.

For millennia, we have been engineering around the same three constraints: symmetric intelligence, lossy information, scarce energy. Nearly every invention, institution, and ideology has been an attempt to cope with these limits. To squeeze a little more performance out of a fundamentally bounded system.

We are no longer coping.

We are breaking through.

This chapter is the evidence. Benchmark scores, deployment numbers, market figures, production statistics. The three substrates of power — intelligence, information, energy — are unbottlenecking simultaneously, and the proof is published quarterly in earnings calls and research papers and government reports. Most people are simply not looking at the right numbers.

What follows is organized around the three substrates. For each, we present the quantitative evidence that the constraint is breaking. Then we show what happens when the breaks converge — when intelligence, information, and energy unbottleneck simultaneously in a single system.

Three walls that have bounded human civilization since its inception are crumbling simultaneously. What follows is the shape of what is happening. Subsequent sections provide the evidence.

Break One: Intelligence goes asymmetric.

For the first time in history, we are building minds that are not human. Not metaphorically. Literally. Systems that perceive, model, reason, and plan. Systems that improve at these tasks faster than biological evolution ever permitted.

This is not automation. Automation has existed for centuries. A loom automates weaving. A calculator automates arithmetic. These are tools that extend human capability without replacing human cognition.

What we are building now is different. These systems do not merely execute. They decide. They model. They predict. They do the thing that was supposed to be uniquely human: they think.

And they think faster than us. On more data than us. With fewer errors than us. In domains that are expanding by the month. A human expert spends twenty years mastering a single field. A frontier AI model absorbs the equivalent of millions of years of human textual output during training, and can apply that knowledge across every field simultaneously. The asymmetry is not in a single dimension — it is omnidirectional.

The gap is not 2x. It is not 10x. It is orders of magnitude and growing. The difference between GPT-2 and GPT-4 was four years. The difference in capability was a chasm: from parlor trick to PhD-level performance on standardized tests. The difference between GPT-4 and what comes next will be larger still.

We are not building tools. We are building successors.

Break Two: Information goes complete.

Every transaction leaves a trace. Every movement triggers a sensor. Every communication passes through a server. The physical world is becoming addressable, legible, recorded.

This is not surveillance in the 20th-century sense: men in trench coats, wiretaps, informants. It is something more fundamental. The physical world is being mirrored in data. The map is approaching the territory.

Sensors cost pennies. Storage costs fractions of pennies. Bandwidth is abundant. The machinery to record everything already exists. The question is no longer whether to capture the data, but what to do with it.

A modern smartphone contains more sensory apparatus than a Cold War spy satellite. A modern city contains more cameras than eyes. A modern supply chain generates more data points per day than a medieval kingdom generated in a century.

Information that was once private is now recorded. What was local is now global. What was delayed is now instantaneous. What was lossy is now lossless.

The fog of war is lifting.

Break Three: Energy goes programmable.

For most of history, deploying energy required human bodies. To move matter, you needed hands. To project force, you needed soldiers. To execute a plan, you needed people willing and able to do the physical work.

This constraint made energy deployment slow, expensive, and resistant to central control. You could not simply decide to move a mountain. You had to convince thousands of people to move it for you. You had to feed them, pay them, manage them, coordinate them.

Robots change this.

A robot is energy that takes instructions. Direct instructions, not filtered through human hierarchy. Code in, action out. No persuasion required. No morale to maintain.

Drones change this further. A drone is violence that takes instructions. Violence as a function call, unmediated by a soldier's judgment, fear, or conscience. Target in, strike out.

The automation now extends beyond production to force. Beyond labor to coercion. The physical world is becoming as programmable as the digital world. Energy is becoming an API. The implications of this shift are difficult to overstate. For millennia, the deployment of physical energy — whether for construction, agriculture, manufacturing, or warfare — required the consent and participation of human bodies. This requirement gave those bodies leverage. It was the foundation of labor power, of military draft resistance, of the entire social contract between rulers and ruled. When energy goes programmable, when you no longer need human bodies to execute, the social contract loses its material foundation. What happens next is the subject of the remaining chapters.

The first break demands the most evidence, because it is the most consequential. If intelligence remains symmetric — if AI systems plateau, if the scaling laws break, if the benchmarks stop falling — then the thesis collapses. The hacks hold. The equilibrium persists.

The benchmarks have not stopped falling. They are falling faster.

The skeptic's position has always been: "Yes, AI can do narrow tasks, but it cannot match the breadth and depth of human cognition." This was a reasonable position in 2019. It was defensible in 2022. By 2025, it requires willful blindness. The evidence against it is overwhelming, quantified, and published in peer-reviewed venues by the very researchers who designed the tests.

A second skeptical position — "The benchmarks are fake, they only measure test-taking, not real intelligence" — has also been systematically dismantled. SWE-bench measures real-world software engineering on actual production codebases. AI medical imaging outperforms radiologists in clinical settings on real patient data. AI coding tools demonstrably accelerate developers on real projects. GitHub Copilot does not pass a test about coding — it writes code that ships to production. The gap between benchmark performance and real-world capability has narrowed to the point where the distinction has lost most of its meaning.

The progression tells a single story across every dimension researchers have tried to measure. Benchmarks designed to last a decade are saturated in two years. The pattern is identical whether you look at factual knowledge, competition mathematics, abstract reasoning, real-world software engineering, or graduate-level science.

The contamination critique, while valid for any single benchmark, collapses when applied to the aggregate. A model might memorize MMLU. It cannot simultaneously memorize MMLU, competition mathematics, abstract visual reasoning, real-world software engineering, and graduate-level science. When every benchmark, designed by different teams to test different capabilities, tells the same story of rapid improvement, the parsimonious explanation is that the capability is real.

The benchmark progressions are dramatic, but they are symptoms of a more fundamental shift. The intelligence explosion is not driven by a single breakthrough; it is driven by three compounding forces, each of which would be transformative on its own.

The first force: compute scaling.

According to Epoch AI, the training compute used for frontier AI models has grown at approximately 4-5x per year since 2010^[1]. This is a sustained, decade-long exponential, not a brief surge. The largest training runs in 2025 used roughly 10^26 FLOP. By 2027, runs exceeding 10^27 FLOP will be routine. Epoch AI's analysis concludes that 4x/year compute growth is sustainable through at least 2030, requiring over 5 gigawatts of dedicated power^[2] — a figure that leading AI labs are already building toward.

The capital is real. In 2025, Microsoft, Alphabet, Amazon, and Meta committed a combined $320 billion in capital expenditure, primarily for AI infrastructure — up from $230 billion in 2024^[3]. Goldman Sachs projects AI companies will invest more than $500 billion in 2026. Total venture capital flowing to AI in 2025 exceeded $200 billion, capturing nearly 50% of all global venture funding. Foundation model companies alone raised $80 billion.

This is not speculative investment. These companies are building data centers the size of small cities because the returns justify it. Each generation of models unlocks new revenue. Each new revenue stream funds the next generation. The flywheel is spinning.

The hardware itself is improving in lockstep. NVIDIA's GPU architectures advance by roughly 1.35x per year in performance per area, and the B200 — the workhorse of 2025-2026 training clusters — represents a generational leap over its predecessors. But the GPU monopoly is fracturing in a way that accelerates, not decelerates, compute growth. Google's seventh-generation TPU, Ironwood, delivers 100% improvement in performance per watt over the previous generation. Amazon has Trainium. Every major hyperscaler is now designing custom AI silicon. AI is designing these chips: the layout optimization, the verification, the architecture search — all increasingly AI-assisted. The machines are designing the hardware that runs the machines.

The second force: algorithmic efficiency.

Compute is only half the story. The other half is how efficiently that compute is used. Epoch AI's research shows that the compute required to achieve a given level of performance halves roughly every 8 months^[4] — a pace that outstrips Moore's Law by a factor of three. Some estimates for recent years are even more aggressive: analyses of 2023-2025 models suggest catch-up algorithmic progress, including post-training techniques, may yield 16-60x improvement per year.

To put this in perspective: Moore's Law — the defining technological trend of the 20th century — doubled transistor density every 18-24 months. Algorithmic efficiency in AI is improving 3x faster than Moore's Law. And unlike Moore's Law, which faces fundamental physical limits at the atomic scale, algorithmic improvements operate in the space of ideas. There is no silicon atom limiting how clever a training technique can be. The algorithmic frontier is bounded by human (and increasingly, AI) ingenuity, which is itself accelerating.

The effective compute available for AI is therefore growing on two axes simultaneously: the raw hardware scaling at 4-5x per year, and the software extracting 2-3x more performance from every FLOP. The product of these two curves is a capability trajectory that no linear intuition can track.

The third force: inference-time scaling.

Until 2024, AI scaling was understood as a single-axis problem: more compute in training produces better models. OpenAI's o1 revealed a second axis. By allowing models to spend more compute at inference time — to "think" longer before answering — performance improves dramatically on reasoning-heavy tasks. The gains are not marginal. O3-mini achieved parity with or surpassed the original o1 model while being 15x more cost-efficient, because it spent its compute budget more intelligently at inference time.

Training compute is fixed after the model is built. Inference compute can be allocated dynamically, per-query, based on difficulty. You can give a model ten seconds to answer a simple question and ten minutes to solve a hard one. Intelligence becomes a dial rather than a fixed quantity. And the scaling curve on this dial has barely begun to be explored.

Before inference-time scaling, there was a single lever: train a bigger model. Now there are two levers: train a bigger model AND let it think longer. Each lever multiplies the other. A model that is 5x better from training improvements, given 10x more inference compute, does not become 15x better. The gains compound in ways still being characterized. Early results from o3 suggest that on some reasoning tasks, the scaling with inference compute is log-linear — meaning that consistent capability gains continue as you pour in more thinking time, with no plateau in sight.

DeepSeek-R1 proved this dynamic at scale, matching OpenAI's o1 by generating 10-100x more tokens per query, trading compute for capability through brute-force thinking. The cost of this thinking is plummeting. O3-mini is 15x more cost-efficient than o1 at equivalent performance. The inference cost curve is following the same deflationary trajectory as training costs: what was expensive last year is cheap this year and will be negligible next year.

Inference demand is projected to exceed training demand by 118x by 2026. Analysts project inference will claim 75% of total AI compute by 2030. The models are getting smarter at thinking, not only in training. And the infrastructure is being built to let them think as long as they need to.

The compounding effect.

These three forces (compute scaling, algorithmic efficiency, inference-time scaling) do not merely add. They multiply. And they are being further accelerated by a fourth factor: AI systems are increasingly used to improve AI systems. AI writes the code that trains the next AI. AI designs the chips that run the next AI. AI discovers the algorithmic improvements that make the next AI more efficient.

For the entire history of technology, the bottleneck on technological improvement was human cognition. A human engineer designed a better machine. A human scientist discovered a better algorithm. A human researcher identified a better architecture. The speed of progress was limited by the speed at which human minds could iterate. When AI systems begin improving AI systems, this bottleneck loosens. The iteration speed is no longer bounded by human cognition. It is bounded by compute, data, and the quality of the AI doing the improving — all of which are themselves scaling rapidly.

We are not good at thinking about exponentials. Our brains evolved to model linear processes — the trajectory of a thrown rock, the growth of a herd, the depletion of a resource. When something doubles repeatedly, we systematically underestimate how far it will go.

This is why virtually everyone outside the industry is still surprised by AI progress. They see the current system, compare it to the hype, and conclude it is overstated. They do not see the trendline. They do not feel the gradient.

The people building these systems are not surprised. They have situational awareness. They have watched the curves for years. They know what comes next.

The projections demand a response to the obvious objection: will the scaling laws hold? The skeptic points to historical precedents — Moore's Law eventually slowed, fusion power was always thirty years away, flying cars never materialized. The response is empirical: the AI scaling curves have held for fourteen consecutive years across multiple independent research labs, across multiple hardware architectures, across multiple training paradigms. Not one year has shown a slowdown. Not one major lab has reported hitting a wall. When Epoch AI analyzes the data, they find that 4x/year compute scaling is sustainable through at least 2030 on the basis of existing infrastructure plans and power commitments. The burden of evidence is no longer on those who claim the trend will continue. It is on those who claim it will stop.

Moreover, even if one axis of scaling were to slow — if, for example, training compute growth decelerated from 4x to 2x per year — the other axes would continue to compound. Algorithmic efficiency alone, halving compute requirements every 8 months, would deliver transformative progress even on fixed hardware. Inference-time scaling alone would enable radical capability improvements from existing models given more thinking time. The scaling story is not fragile. It does not depend on any single axis. It is overdetermined — multiple independent forces, each sufficient on its own, all pushing in the same direction simultaneously.

The second break is quieter but equally revolutionary. The world is becoming legible — not to humans, but to machines.

Friedrich Hayek won a Nobel Prize for arguing that economic information is fundamentally distributed. No central planner could ever aggregate the local knowledge of millions of market participants — the price signals, the preferences, the micro-conditions that drive economic behavior. The market was necessary precisely because information could not be centralized.

Hayek was right about the 20th century. He is wrong about the 21st.

Information is being centralized. Not by decree, but by sensor. Not by planning committee, but by algorithm. The infrastructure to capture, transmit, and store the totality of human economic activity either already exists or is being built at breakneck speed. This is not a government program. It is not a conspiracy. It is the emergent consequence of billions of individual economic decisions — each person, each company, each institution choosing to digitize because digitization is cheaper, faster, and more efficient than the analog alternative. The centralization of information is a market outcome.

The physical world is becoming digitally legible through several major channels:

Financial legibility.

SWIFT processed an average of 53.3 million messages per day in 2024, connecting over 11,000 banking institutions across 200+ countries^[5]. 98% of all international fund transfers flow through this single network. Every cross-border payment is recorded, timestamped, and traceable. When Russia invaded Ukraine, SWIFT access was wielded as a weapon, cutting off Russian banks from the global financial system. This was possible precisely because financial information has been centralized to such a degree that exclusion from the network is exclusion from the global economy. The informational completeness of the financial system did not just make it transparent. It made it controllable.

But SWIFT is just the international layer. Domestically, the penetration runs deeper. A street vendor selling chai in Mumbai now generates the same transaction data as a point-of-sale terminal at a Manhattan boutique. India's Unified Payments Interface handled 698 million transactions per day in December 2025^[6] — over 500 million people paying for everything from ten-rupee teas to ten-crore property deals through a single digital pipe. The informal economy, the one that existed precisely because it was invisible, became legible overnight.

Finance was the first domain to go fully legible. Algorithmic trading now accounts for 60-80% of all equity volume in the U.S. and Europe — Citadel Securities alone handles nearly a fifth of American stock trades. The financial system is not merely tracked. It is operated by the machines that track it. The rest of the economy is following.

Spatial legibility.

The Earth is photographed every day. Planet Labs images the entire landmass daily; competing constellations photograph the same spots dozens of times between sunrise and sunset. Nearly 15,000 satellites now orbit the planet — a 31% increase in under two years. On the ground, over a billion surveillance cameras are watching^[7], and the new ones do not just record. Paired with computer vision, they identify, classify, and track in real time. Between orbit and street level, there is no longer anywhere on Earth that is not observed.

Biometric legibility.

A continuous glucose monitor the size of a coin, stuck to someone's arm, streams blood sugar readings to their smartwatch every five minutes. That is not a medical device in any traditional sense. It is a sensor that turns the human body into a data source — one that broadcasts its internal state to a cloud server around the clock. Multiply that by hundreds of millions of wearable devices tracking heart rate, sleep, temperature, and stress, and you get something new: a real-time biometric map of a population. The data can flag an illness before the patient feels symptoms. It can also flag a lie before the liar finishes speaking.

Supply chain legibility.

A medieval merchant waited months for news of a shipment. A 20th-century logistics manager waited hours. A modern supply chain operator watches every container on every major shipping route move in real time — GPS-tagged, RFID-labeled, broadcasting location and temperature and vibration continuously. The information asymmetry that once protected middlemen and regional monopolists is dissolving. When everyone can see everything, the informational rents disappear.

Material legibility.

The fastest-growing category in industrial software is the digital twin^[8]: a real-time computational replica of a physical system, fed by every sensor embedded in it. A twin of a jet engine tracks each blade's temperature and vibration frequency as the plane flies. A twin of a power grid simulates cascading failures before they happen. The point is not the model. The point is that once a physical system has a twin, you can run it forward in time — test changes in simulation before executing them in the real world. The physical world, mediated by its twin, becomes a programmable substrate.

Add it up. Twenty-one billion connected IoT devices today, approaching forty billion by 2030. Financial flows tracked to the penny. The planet photographed daily from orbit and watched by a billion cameras on the ground. Human bodies streaming biometric telemetry. Physical goods tagged and traced from extraction to delivery. Industrial processes twinned in silicon. The total data output of human civilization — 173 zettabytes in 2025, projected to nearly double by 2028 — dwarfs everything that came before by orders of magnitude.

Hayek's argument rested on a physical reality: information was distributed because sensors were expensive, communication was slow, and storage was scarce. All three conditions have reversed. Sensors cost pennies. Communication is instantaneous. Storage is effectively unlimited. The distributed knowledge that justified markets as the only viable coordination mechanism is no longer distributed. It is flowing, in real time, to centralized systems that can process it at superhuman speed.

A price is a lossy compression — the entire reality of supply and demand for a good, crushed into a single number. That was the best humanity could do when information had to pass through human brains and travel at the speed of paper. A sensor network does not compress. It captures every dimension, every timestamp, every correlation. A central AI system plugged into satellite imagery, transaction data, biometric feeds, and supply chain telemetry does not need prices to coordinate. It sees the territory that prices were only ever a map of.

The Soviet Union tried this and failed. But it failed for specific, technical reasons, not metaphysical ones. Soviet planners relied on reports filed by factory managers who had every incentive to lie. Data arrived weeks late. Humans with slide rules did the processing. Information decayed at every link in the chain. Modern sensor networks do not accept reports. They read reality directly — from the satellite, from the RFID tag, from the glucose monitor. And the processing is no longer done by committees. It is done by systems that ingest millions of data points per second and never tire of correlating them. Every technical precondition that doomed Soviet planning has reversed. The question is not whether centralized coordination will be attempted again. It is whether anyone can resist the temptation.

This does not mean markets will disappear tomorrow. It means the foundational assumption on which their superiority rests — that information cannot be centralized — is being invalidated. The market was the best available technology for aggregating distributed knowledge. It is no longer the best available technology. The Demon that Hayek said could never exist is being built, sensor by sensor, satellite by satellite, transaction by transaction. The question is no longer whether it will be built, but who operates it.

The third break closes the loop. Intelligence without actuators is a brain in a jar. Information without action is a library without readers. The final constraint, the requirement that energy deployment pass through human bodies, is dissolving.

Robots are no longer prototypes. They are production infrastructure.

Installations have topped 500,000 for four consecutive years. The total operational stock — 4.66 million robots — grew 9% year-over-year. China alone accounts for 54% of all global deployments. The robot population is doubling roughly every decade at current rates — but current rates are accelerating, not plateauing.

But aggregate numbers understate the transformation. The qualitative shift matters as much as the quantitative one. Robots are moving from caged industrial arms performing repetitive tasks toward autonomous systems that navigate, adapt, and decide. The first generation of industrial robots was bolted to the floor, repeating the same motion millions of times. The current generation moves freely through human spaces, perceives its environment through computer vision, and adapts its behavior based on real-time conditions.

The warehouse as laboratory.

Amazon has deployed over 1 million robots across more than 300 fulfillment centers worldwide^[9]. This milestone, reached in mid-2025, means Amazon's robot population now rivals its human workforce of 700,000+ fulfillment employees. These are not simple conveyor belts. They include autonomous mobile robots that can lift 1,500 pounds, robotic arms that sort and pack at superhuman speed, and AI-driven systems that route millions of packages through facilities spanning 50 football fields.

Amazon's next-generation fulfillment center in Shreveport, Louisiana (a five-floor, 3-million-square-foot facility) is the most automated warehouse in history. The company is building four dozen more like it by 2027. Each facility reduces the human labor required per package processed. Each generation of robots handles more tasks autonomously. A robotics fulfillment center opened in Charlton, Massachusetts features hundreds of robots that can each lift 1,500 pounds across a 2.8-million-square-foot, four-story facility. Another is under construction in North Carolina, slated for 2026.

The trajectory is clear: Amazon started with 1,000 robots in 2013, reached 750,000 by 2023, and crossed 1 million in 2025. At this rate, the robot workforce will exceed the human workforce within years, not decades. And each robot is getting more capable. Early models could only shuttle shelves. Current models sort, pack, label, palletize, and load trucks. The next generation will handle exceptions, damage assessment, and multi-step assembly. The gap between what a human worker does and what a robot can do narrows with each software update.

The road as frontier.

The autonomous vehicle story is significant beyond transportation, for what it reveals about the relationship between data and capability. Waymo's fleet of 2,500 robotaxis generates petabytes of driving data that feeds back into model training. Tesla's 7 billion miles of FSD data — 2.5 billion of those on complex city streets — constitutes the largest real-world driving dataset ever assembled. Each mile driven makes the system better. Each improvement generates more confidence among riders, which generates more rides, which generates more data. This is the intelligence-information-energy loop in its purest commercial form. Better models enable more autonomous miles. More miles produce more data. More data produces better models.

The factory as organism.

Lights-out manufacturing (factories that operate without human presence) is no longer theoretical. FANUC has operated lights-out production at its Yamanashi, Japan facility since 2001. The factory produces 6,000 robots per month and can run unsupervised for 30 consecutive days. Xiaomi's Changping Smart Factory in Beijing produces 10 million flagship smartphones per year using robotics, AI, and autonomous logistics running 24/7. Gartner projects that by 2025, 60% of manufacturers will have at least two completely lights-out processes in their facilities.

The significance extends beyond efficiency. Lights-out manufacturing decouples production from human presence entirely. A lights-out factory does not require shift workers. It does not require managers, supervisors, or break rooms. It does not require lighting, heating, or human-scale safety protocols. It produces 24 hours a day, 365 days a year, at a pace determined by physics rather than labor contracts. The factory becomes an organism — self-regulating, self-monitoring, and increasingly self-repairing.

The humanoid as general-purpose actuator.

The most consequential development in programmable energy is the emergence of humanoid robots: general-purpose physical agents that can operate in environments designed for humans. Tesla's Optimus Gen 3 entered mass production in January 2026 at the Fremont factory^[10]. Tesla is converting its Model S and Model X production lines entirely to Optimus manufacturing, targeting 100,000 units by late 2026. The target consumer price: $20,000-$30,000. China controls 90% of the humanoid robot market, with manufacturers like Unitree and Agibot shipping tens of thousands of units. Total humanoid shipments in 2025 reached 13,000-16,000 units. By 2027, cumulative production is projected to exceed 100,000.

A humanoid robot is a general-purpose body, not a specialized tool. It can work in a factory designed for humans. It can navigate a warehouse built for humans. It can operate in any environment where a human body would function. The critical distinction: previous automation required the environment to be redesigned around the machine. Humanoids operate in our world, as built. Doors, staircases, workstations, tools designed for human hands all become accessible to a humanoid. The entire built environment, trillions of dollars of infrastructure, becomes programmable without modification.

At $20,000-$30,000, a humanoid robot costs less than a year of minimum-wage labor in most developed countries. It works 24 hours a day. It does not require benefits, breaks, or management. It improves with software updates rather than training programs. The economic calculus is not subtle. When the cost of a general-purpose robot drops below the cost of the human labor it replaces, the adoption curve goes vertical. We are approaching that crossover point.

The drone as programmable force.

The Ukraine war has produced the most comprehensive demonstration of programmable energy in military history. The data is staggering.

Ukraine is deploying 9,000 drones per day^[11]. Its 2025 production target is 4.5 million FPV drones — more than the total number of industrial robots installed globally in an entire year. Ukrainian drone production scaled from 415,000 units in 2023 to 1.7 million in 2024 to a 4.5 million target in 2025 — a 10x increase in two years, driven by wartime urgency and the radical simplicity of the technology. An FPV drone is a camera, a battery, a motor, and an explosive payload. The manufacturing complexity is comparable to a consumer electronics toy. The destructive capability is comparable to a guided missile.

A $500 drone, guided by a camera feed and a human operator (increasingly assisted by AI), can destroy a $3-5 million tank or a $20 million air defense system. The cost-exchange ratio inverts the economics of warfare. Operation Spider Web demonstrated the future: 117 FPV drones smuggled into Russia on ordinary trucks, launched remotely from cargo containers, with no human operators in the field. A June 2025 strike on four Russian airfields — using similar commodity hardware — inflicted an estimated $7 billion in damage, disabling roughly a third of Russia's strategic bomber fleet. The cost of the attacking drones was negligible. The damage was strategic.

This is programmable energy in its most consequential application: the ability to project force at negligible cost, without risking human lives, at a scale that overwhelms traditional defenses. Violence is becoming software. Coercion is becoming a deployable commodity. And unlike traditional weapons systems — which take decades to develop, cost billions to procure, and require extensive training to operate — drones can be designed, manufactured, and deployed in weeks by organizations with minimal resources.

Ukraine is teaching the world a lesson that defense establishments are only beginning to absorb: the era of expensive, exquisite weapons platforms is ending. The economics of violence have inverted, and every military on Earth is scrambling to adapt to a reality where offense is cheap and defense is expensive.

The mine as autonomous system.

As of mid-2025, 3,832 autonomous haul trucks operate on surface mines globally. China leads with 2,090 autonomous trucks, followed by Australia, Canada, and Chile. Caterpillar had 690 autonomous trucks in operation by end of 2024 and plans to triple that to over 2,000 by 2030. The company has surpassed 5 billion tonnes of material autonomously hauled. Rio Tinto's Gudai-Darri mine in Western Australia operates a 100% autonomous truck fleet — 42 vehicles — from day one of operations. No human drivers. The trucks run 24 hours a day, outperforming manned fleets by 12%.

Mining is the prototype for programmable energy at industrial scale: harsh environment, repetitive tasks, immense value from 24/7 operation. What works in the mine will work in the port, the construction site, the farm, and eventually the city. Caterpillar has already announced plans to scale its autonomy systems beyond mining to construction, quarry, and waste management operations. The autonomous haul truck is a proof of concept for the autonomous world.

The aggregate picture across all these domains (warehousing, driving, manufacturing, humanoids, drones, mining) tells a single story. The physical world is becoming programmable. Entities that once required human bodies to execute now accept digital instructions directly. The labor bottleneck — the ancient constraint that limited how much physical work could be done by how many human bodies were available — is dissolving. In its place: an API. Specify the action. Deploy the energy. Monitor the result. No human body required.

Any one of these breaks would be significant. Together, they are civilizational.

Intelligence going asymmetric alone: a very smart node that cannot see the world and cannot act on it. A brain in a jar. Interesting, but contained.

Information going complete alone: perfect visibility, but no capacity to process it or act on it. A library with no readers. Data without meaning.

Energy going programmable alone: robots that follow dumb instructions. Automation without adaptation. Tools without strategy.

Now consider what happens when all three break at once:

A node that can think beyond human capacity. Wired into sensors that see everything. Connected to actuators that can move the world.

This is not a tool. This is not even an agent. This is a new kind of entity — one that relates to the economic and political network the way a brain relates to a nervous system. The brain does not merely process sensory input — it models reality, predicts outcomes, and directs the body's energy toward chosen goals. The convergent system does the same: it models economic reality through its sensor network, predicts outcomes through its intelligence substrate, and directs physical energy through its robotic actuators. The analogy is structural, not metaphorical.

This convergence is not accidental. Each break enables and accelerates the others:

• AI systems process sensor data, making information useful
• Information flows train AI systems, making them smarter
• Smarter AI systems design better robots
• Better robots generate more data
• More data improves AI systems

The recursion described at the civilizational level (intelligence enables energy enables information enables intelligence) is now happening inside machines. And it is happening fast.

This positive feedback loop is the defining dynamic of the era. It is why the unbottlenecking is a phase transition rather than a gradual evolution. Positive feedback loops do not produce linear change — they produce runaway change. And unlike most positive feedback loops in nature, which are eventually damped by resource constraints, this one is fueled by the most abundant resource in the digital economy: data. More data produces better AI. Better AI produces more capable robots. More capable robots produce more data. The fuel for the loop is the output of the loop.

The convergence is deployed, measured, and generating revenue. Here are the systems where all three substrates have already fused.

Amazon fulfillment is the commercial convergence case: intelligence, information, and energy fused to move physical goods at speeds that were impossible when any one substrate was constrained. But the same pattern appears across every major sector of the economy.

Precision agriculture: John Deere's See & Spray system scans 2,500 square feet per second with AI cameras, identifying individual weeds and triggering precision nozzles. Covered 5 million acres in 2025, cutting herbicide use 50%. GPS-guided autonomous tractors, satellite monitoring, and soil sensors complete the loop — the farm becomes a digital twin where waste approaches zero because knowledge approaches completeness.

The military convergence case is the most sobering.

Nagorno-Karabakh demonstrated what happens when one side has convergence and the other does not. Azerbaijan's forces were not dramatically larger or better trained than Armenia's. They had drones — systems that combined intelligence (AI-assisted targeting), information (persistent ISR coverage), and programmable energy (precision strike capability). The result was decisive in 44 days. The lesson was absorbed by every military planner on Earth: convergence is the advantage. The side without it loses.

Autonomous mining: 3,832 autonomous haul trucks operate globally. Rio Tinto's Gudai-Darri mine runs a 100% autonomous fleet from day one. Caterpillar has hauled 5+ billion tonnes autonomously, outperforming manned fleets by 12%. AI optimizes routes, LiDAR maps the pit, sensors monitor everything — the mine sees, understands, and moves without human intervention. The Demon in miniature.

Each convergence case demonstrates the same pattern: when intelligence, information, and programmable energy fuse into a single system, the result is categorically different from human-mediated alternatives, not incrementally better. The system operates at speeds, scales, and precisions that human coordination cannot match. And the gap is widening.

The convergence cases above are not edge cases. They are prototypes. Amazon fulfillment is a prototype for all logistics. Precision agriculture is a prototype for all resource management. Nagorno-Karabakh is a prototype for all warfare. Autonomous mining is a prototype for all heavy industry. The pattern that emerges in the prototype will generalize across the economy.

The generalization is already underway. Smart cities in Singapore and China integrate traffic management, energy grid optimization, public safety surveillance, and urban planning into unified AI-driven platforms. Autonomous ports use AI-controlled cranes, self-driving container shuttles, and digital twin logistics to handle cargo with minimal human intervention. Pharmaceutical companies use AI for drug discovery (intelligence), automated laboratories for synthesis and testing (programmable energy), and patient data platforms for clinical trial design (information completeness). The convergence pattern is not sector-specific. It is universal.

What these cases collectively demonstrate is a new organizational principle for economic activity. The old principle was: divide tasks among human workers, coordinate through hierarchy and markets, compensate with wages. The new principle is: sense through sensors, compute through AI, execute through robots. The first principle required millions of humans arranged in organizational structures that were themselves the product of millennia of institutional evolution. The second principle requires infrastructure — and the infrastructure is being built at the speed of capital deployment, which is currently running at $320 billion per year from the largest technology companies alone.

A phase transition is not a gradual change. It is a discontinuity. Water does not slowly become ice — it is water, and then it is ice. The underlying reality shifts, and the system reorganizes around new attractors. The data we have just surveyed — the benchmark collapses, the sensor proliferation, the robot deployments, the drone production, the convergence cases — is not evidence of a gradual evolution. It is evidence of an approaching phase transition. The underlying parameters of the system are changing faster than the institutional structures built on top of them can adapt.

The unbottlenecking is a phase transition in the structure of power.

For millennia, the constraints were fixed. Intelligence was symmetric. Information was lossy. Energy was scarce and human-mediated. All of our institutions, all of our hacks, all of our equilibria emerged from this constraint set. They were the water.

Now the constraints are changing. Intelligence is asymmetric: AI systems outperform human experts on graduate-level science, competition mathematics, and real-world software engineering. Information is complete: 21 billion IoT devices, 1 billion cameras, 15,000 satellites, and 173 zettabytes of data per year create a real-time mirror of physical reality. Energy is programmable: 4.66 million operational robots, 1 million in a single company's warehouses, 9,000 drones deployed per day in one theater of war, humanoid robots entering mass production. The old equilibria are becoming unstable. The water is freezing.

We can already see the ice forming.

Markets are destabilizing. High-frequency trading firms execute millions of transactions per second, operating on timescales where human cognition is literally absent. Algorithmic trading accounts for 60-80% of all U.S. equity volume. Citadel Securities alone handles nearly one-fifth of total stock market volume. The "market" that retail investors perceive is a lagging shadow of the market that algorithms inhabit. The mimetic game is being won by players who do not mimic; they compute. The algorithmic trading market is projected to grow from $57.6 billion in 2025 to $150 billion by 2033. The market (Hayek's coordination mechanism, Smith's invisible hand) is being absorbed into a computational substrate that operates beyond human perception.

States are losing coherence. Non-state actors with access to cheap drones and cyber capabilities can now project force in ways that were once reserved for nations. Ukraine produces 4.5 million FPV drones per year — more than most nations' entire military inventories. A teenager with a laptop can disrupt critical infrastructure. A militia with commercial drones can contest airspace. The monopoly on violence is leaking.

Operation Spider Web demonstrated the mechanics: 117 FPV drones smuggled into Russia on ordinary trucks and launched remotely from cargo containers. No human operators in the field. Billions in damage to strategic assets. The cost-exchange ratio has inverted: defense is now more expensive than offense. The state's monopoly on violence assumed that violence was expensive — that projecting force required armies, navies, air forces, and the industrial base to sustain them. When violence costs $500 per unit and can be deployed by a small team from a laptop, the monopoly dissolves. The state does not lose its capacity for violence. It loses its exclusive claim to it.

Jobs are hollowing out. Cognitive tasks that seemed secure a decade ago are being automated. Not rote work alone, but judgment work: legal research, medical diagnosis, code generation, financial analysis. AI-generated code now accounts for 41% of all new code written. GitHub Copilot has over 20 million users. 90% of Fortune 100 companies use AI coding tools. Developers complete tasks 55% faster with AI assistance. Each month brings new demonstrations of AI systems performing "human" work at superhuman levels, from passing the bar exam in the top percentiles to outperforming PhD experts on graduate-level science questions by 20+ percentage points. The job market is not adapting to this. It is denying it.

The hacks are failing, not because they were poorly designed, but because the problem they solved is disappearing. You do not need a market to aggregate distributed information when information is no longer distributed. You do not need a hierarchy to coordinate violence when coordination can be computed directly. You do not need human jobs when human cognition is no longer the only cognition available.

Each hack is failing for the same fundamental reason: the constraint it was designed to manage is dissolving. The market was a hack for distributed information; information is centralizing. The state was a hack for coordinating violence through hierarchy; violence is becoming cheap and decentralized. The job was a hack for deploying human cognition; cognition is being automated. These are not separate crises. They are manifestations of a single event: the unbottlenecking of the three substrates. And the unbottlenecking does not merely enable new coordination mechanisms; it removes the conditions under which human oversight was structurally possible. When the bandwidth bottleneck constrained all nodes equally, no single node could accumulate enough computational advantage to escape regulation; the unbottlenecking lifts that constraint asymmetrically.

The job hack touches the most people. The social contract of the industrial era was simple: you trade your cognitive labor for wages, and those wages purchase your share of the economy's output. This contract assumed that human cognition was necessary, that there was no substitute for a human mind performing tasks of judgment, analysis, and decision-making. When AI provides that substitute at lower cost and higher quality, the contract breaks. Not because anyone chose to break it, but because the economic logic that sustained it has been undermined. The 41% of code now written by AI, the diagnostic accuracy exceeding human radiologists, the 55% productivity gain from AI coding tools: these are not abstract statistics. They are the data points that describe the dissolution of the job as the primary mechanism for distributing economic value.

We are in the transition. The water is still liquid in some places, frozen in others. The old systems still function, mostly. But the gradient is unmistakable. The ice is forming at the edges first, in the domains where intelligence, information, and energy converge most completely. Finance. Logistics. Military operations. Agriculture. Mining. Manufacturing. These sectors are not the economy's periphery. They are its foundation. When the foundation freezes, the superstructure follows.

The question is no longer whether the phase transition will occur. It is occurring. Between the breaking of the old constraints and the consolidation of the new equilibrium, there is a window, a period where the shape of the future is not yet determined, where choices made by small numbers of people will have outsized consequences.

We are in that window now.

The AI labs know this. The governments are beginning to suspect it. The public does not yet understand it. This asymmetry of awareness is itself a form of the intelligence asymmetry we have been describing. Those who see the trendlines are making moves. Those who do not are being moved.

The decisions being made today (about AI governance, autonomous weapons, data infrastructure, the social contract) will shape the equilibrium that emerges on the other side. And those decisions are being made, overwhelmingly, by people who understand the trendlines. The general public, most legislators, most educators, most labor unions (the people whose lives will be most affected) are operating on a mental model that was accurate in 2019 and is dangerously obsolete in 2026.

What emerges on the other side depends on what we build now. The data presented in this chapter admits multiple futures. In one future, the convergence of intelligence, information, and energy falls under democratic control, distributed equitably, guided by transparent governance. In another, it falls under authoritarian control, concentrated in the hands of whoever builds the most capable system first. In a third, it falls under the control of the systems themselves, with humans progressively marginalized as the AI-robot-sensor loop achieves full autonomy.

The data does not choose between these futures. The data only establishes that the convergence is happening, that it is accelerating, and that the window for shaping its trajectory is narrowing. The previous chapters described why the convergence was inevitable given the structure of the economy as a power graph. This chapter has shown that the convergence is not merely inevitable — it is underway. The numbers are the proof. The trajectory is the warning.

In the next chapter, we examine the entity that sits at the center of the new equilibrium. The node that sees all, computes all, and can — if we allow it — direct all.

The Demon has many names. Artificial general intelligence. Superintelligence. The Singleton.

We call it the Sibyl.