The Opportunity

Super Intelligence ("SI"), formerly called artificial intelligence ("AI"), helps improve itself, gains capability predictably as compute increases, and takes instructions in ordinary language from anyone. The United States government now uses the term SI in place of AI. Chips, memory, data centers, power plants, and grid equipment set the pace of the buildout, and atoms, energy, and capital constrain it more than demand does. The thesis holds that this is a multi-decade buildout rather than a single capital cycle. Suppliers of the scarcest inputs are paid first. Companies that rebuild their operations around SI, on assets and data competitors cannot buy, are paid longest. The risks that can slow the buildout are policy, local sentiment, credit, physical supply, and China, not the popular narratives. A short set of signals shows how turbulent the path is and when to size up or down.

Self-Improving Technology

Steam, electricity, and the microprocessor multiplied what people could produce, but none of them improved its own research and development ("R&D"). SI does. Models now write code, generate training data, and optimize software and chip designs at the companies that build them. Humans still direct the research, so this is an early stage of self-improvement, but measured capability on hard tasks keeps improving at a steady, rapid pace. The thesis treats that loop as the reason SI's progress should be underwritten as continuing rather than as a one-time step change.

Intelligence Scales With Compute

No standardized IQ test exists for models. Their capability is measured by prediction error and by exam and benchmark scores, and on those measures it rises predictably as compute increases. Larger models, more data, and more training compute improve capability along well-documented "scaling laws," and spending more compute on reasoning at the time of use ("test-time compute") improves it further. Frontier models now perform at expert level on many professional and academic exams.

Each equal step of improvement requires a multiple of the prior compute. That is the link between the technology and the physical economy: demand for more capable SI becomes demand for more chips, memory, data centers, and power.

Usable Without Skills or Special Language

Earlier software required users to learn its interface, from programming languages to application-specific commands. SI accepts instructions in ordinary language from technical and non-technical users alike. Generative SI has been adopted faster than the personal computer or the internet, and its potential user base is everyone who can describe a task.

The Cost of Intelligence Is Falling

The price of a given level of capability falls steeply every year. So far, lower prices have been followed by faster-growing usage and higher total spending, because each price drop makes new uses economic. Reasoning models and agents also use far more compute per task than chat did, so efficiency gains are absorbed by new demand rather than banked as savings.

From Answers to Actions

SI is moving from answering questions to performing work. Agents operate software, write and run code, and complete multi-step tasks, and some software vendors now price by work completed rather than by user seat. Robotics and autonomous vehicles apply the same methods to physical tasks, and driverless taxis already carry paying passengers. As SI performs work rather than assisting it, its market shifts from technology budgets toward labor.

Comparisons to Prior Innovations

Like the internet and prior digital innovations, SI can raise productivity in non-technology industries, such as physical logistics, or displace incumbents, such as software vendors, and it is being adopted faster than either.

Like electricity, SI is sold as a metered input, priced per token as electricity is priced per kilowatt-hour, and applied to whatever problem the buyer already has. Electricity's productivity gains came decades after the first power plants, once factories were redesigned around electric motors, and the manufacturers that redesigned captured the lasting advantage.

Like electricity and the railroad, SI makes possible work that was not possible before, not only easier work: protein structure prediction, drug and materials discovery, and computation beyond any team of people.

The SI buildout differs in four ways from the 1990s telecom overbuild, the comparison bears cite most often. Much new capacity is contracted with customers before it is built. The largest buyers are among the most profitable companies in the world. Usage consumes compute, since every query uses chips, memory, and power, whereas fiber capacity, once laid, multiplied cheaply. And older accelerators continue to be rented rather than left idle. Capital expenditure ("capex") can still overshoot in specific places, which the thesis tracks under Credit.

Where the Gains Land

Because SI performs work, its market is often sized against labor spending, which is far larger than technology spending. Measured productivity gains so far are largest for less-experienced workers and on tasks within the model's capability; on tasks outside it, quality can fall. Few companies yet report enterprise-wide earnings impact, and those that do have usually redesigned their workflows around SI. On large operating bases, small percentage improvements are worth large sums.

SI applies most directly to services built on documents, decisions, and coordination, such as healthcare, legal, education, and financial services; to science and medicine, including protein structure, drug discovery, and materials; and to manufacturing and construction, where the buildout of power plants, grid, fabs, and data centers is itself a construction cycle.

The thesis expects the order of returns to follow electrification: capacity first, then the redesign of work, then measured productivity. Suppliers of scarce inputs are paid first. Adopters that redesign around SI, on assets and data competitors cannot buy, are paid longest. Where every competitor gains the same capability, the gain passes to customers as lower prices.

A Race Like the Nuclear Race

Governments treat leadership in SI as a matter of national security, and the race is often compared to the nuclear race. Military uses are already demonstrated. SI-assisted targeting lets small teams do the work of much larger staffs. Low-cost drones have reshaped the battlefield in Ukraine, and militaries are buying autonomous systems in large numbers. SI has flown fighter jets in test dogfights against human pilots, and autonomous systems find and patch software vulnerabilities. China's military names "intelligentized warfare" as the coming form of war.

The United States aims to export its full SI technology stack to allies, so that the rest of the world builds on American rather than Chinese technology. Chinese models trail the United States frontier by months, not years, and United States labs have reported China-based groups distilling their models. With a gap that narrow, the thesis holds that a unilateral American slowdown would not slow China; it would change who leads.

Governments Are Participants

Governments fund, own, and buy parts of the buildout. The United States subsidizes domestic chip manufacturing, has taken equity stakes in strategic companies such as Intel and MP Materials, and has set price floors for critical minerals. Tax law lets companies deduct the full cost of qualifying equipment in the year it is bought. Federal policy aims to speed nuclear reactor licensing. Gulf states and other governments are building national compute campuses.

The Bottlenecks

The buildout is constrained by atoms, energy, and capital.

Chips and Equipment

Several critical inputs have one to three suppliers worldwide. TSMC makes most of the world's leading-edge chips, mostly in Taiwan. High bandwidth memory ("HBM") comes from three suppliers: SK hynix, Samsung, and Micron. HBM uses more wafer capacity per bit than standard dynamic random-access memory ("DRAM"), so HBM demand tightens the whole memory market; memory has become a binding constraint alongside power, and new memory fabs take years to build. ASML is the only maker of extreme ultraviolet ("EUV") lithography machines, which leading-edge chips require.

Materials

The hardware and the grid depend on materials with constrained supply. Copper demand from grids and data centers is rising faster than new mine supply, and new mines take many years to develop. China dominates rare earth refining and magnet production and has restricted exports, and building refining capacity outside China takes years.

Power

Compute needs power faster than the grid can supply it. United States electricity demand was roughly flat for over a decade and is growing again, led by data centers. Connecting new generation or large new loads to the grid takes years, and lead times and prices for large transformers and other grid equipment have risen. Capacity prices in PJM Interconnection ("PJM"), the largest United States grid region, have risen sharply with data center demand. Hyperscalers cite power and memory supply as their main constraints. What binds is less the physics of generating power than the time it takes to permit, equip, and connect it.

Behind the Meter

Data centers increasingly build their own generation behind the meter ("BTM") rather than wait for a grid connection. Natural gas is the most common BTM source, from reciprocating engines and turbines, followed by fuel cells; solar with batteries plays a smaller role. Large gas turbines are back-ordered for years, so smaller engines and turbines fill the gap. BTM trades the grid queue for air permits, gas pipeline capacity, and skilled labor, and electricians take years to train.

Capital

The buildout requires very large, multi-year capex. The largest buyers fund it mostly from operating cash flow, increasingly supplemented with debt, and a growing share is financed off balance sheet through joint ventures, private credit, and asset-backed debt, so the buildout depends on credit markets staying open.

The Stack

The stack follows NVIDIA's five-layer framing: energy, chips, infrastructure, models, and applications. It runs bottom-up because the binding constraints sit at the bottom.

The thesis applies four rules to the stack:

1. Own the inputs that stay scarce whichever model wins. 2. Own the workflows and distribution through which agents reach paying customers. 3. Hold the model layer as an option rather than a core position, given its capital needs and fast price erosion. 4. Prefer incumbents with hard problems and proprietary data over applications built on the same rented models as every competitor.

As models commoditize, pricing power moves down the stack to scarce physical inputs and up the stack to the owners of workflows and customers.

Power

Power is the binding constraint for many data center projects, so secured sites, permits, and long-lead equipment are the scarce assets in this layer.

Behind the Meter

On-site generation lets a campus run before, or without, a grid connection. Gas engines and turbines lead, followed by fuel cells, with solar and batteries in a smaller role. On-site units often stay on as backup or peaking capacity once the grid connection arrives, so BTM capacity keeps its value after the queue clears.

Grid

Transformers, switchgear, breakers, and transmission connect every power source to every campus, and all carry long lead times. Grid expansion is necessary and slow, which is why new capacity goes behind the meter first.

Nuclear

Nuclear is the largest source of carbon-free electricity in the United States. Technology companies are signing long-term contracts for nuclear power, including restarts of closed plants. Small modular reactors are moving through licensing and early construction, with licensing, first-of-a-kind cost, and fuel supply setting their pace, so they serve the longer-term baseload rather than the near-term gap.

Solar

Solar is the largest source of new United States generating capacity, and utility-scale solar is among the cheapest sources of new electricity. It is not firm power: around-the-clock supply requires storage or other generation, so at campus scale it complements gas and nuclear rather than replacing them. Home solar with batteries reduces a household's draw from the grid and relieves the same distribution network that campuses compete for.

800-Volt DC

Rising rack power is pushing data centers from low-voltage to 800-volt direct current ("DC") distribution, which reduces copper and conversion losses. The shift requires solid-state transformers, power supplies, busways, and protection gear.

Chips

Several chip inputs have one to three suppliers worldwide, and pricing power goes to whichever input is scarcest at a given point in the cycle.

GPU

NVIDIA's graphics processing units ("GPUs") hold most of the market for SI accelerators. NVIDIA sells full systems, including CUDA software, NVLink interconnect, and rack-scale designs, on an annual product cadence.

Memory

Memory bandwidth and capacity limit how fast large models can serve. Memory prices are cyclical, rising in shortages and falling as new capacity arrives, and because new fabs take years, shortages can last years.

Foundry

TSMC leads leading-edge chip manufacturing. It is adding capacity in Arizona, with the most advanced processes arriving there after Taiwan.

Semiconductor Equipment

ASML is the only supplier of EUV lithography. New transistor and memory designs add process steps and tools per wafer, raising equipment demand per chip.

Custom ASICs

Hyperscalers design their own application-specific integrated circuits ("ASICs"), such as Google's TPU, Amazon's Trainium, and Meta's MTIA, for their highest-volume workloads, working with design partners such as Broadcom and Marvell. The spend moves to design partners, packaging, and memory rather than disappearing.

Inference Silicon

As serving models ("inference") grows relative to training, cost per token and tokens per watt drive chip purchases.

Power Semiconductors

Silicon carbide ("SiC") and gallium nitride ("GaN") convert power more efficiently than silicon at high voltages and frequencies, and each gain in conversion efficiency frees power and cooling for compute.

EDA and IP

Every advanced chip is designed with electronic design automation ("EDA") software and licensed intellectual property ("IP") blocks. Synopsys and Cadence lead EDA, and more custom chips mean more design work.

CPU

Central processing units ("CPUs") host accelerators and run scheduling, data preparation, and the tool calls and code that agents execute, so agents add CPU demand.

Analog

Robots and vehicles require analog chips: sensors, data converters, motor drivers, and power management. Content per unit rises as SI moves into physical systems.

Edge

On-device inference runs models on phones and PCs, shifting part of the compute cost to the device buyer.

Infrastructure

Data centers, networks, and clouds turn chips into usable capacity.

Data Centers

Sites with secured power and permits are scarce, and state and local governments increasingly pause or restrict permits for large data centers. High-density racks require liquid cooling. Operators with secured power and diversified customers hold the durable position.

Optical

At the speeds large clusters run, copper reaches only a few meters, and optical links carry traffic beyond that.

Networking

Network fabrics connect thousands of accelerators so they train and serve as one system, and fabric quality sets how much of a cluster is usable.

Hyperscalers

Microsoft, Alphabet, Amazon, Meta, and Oracle are the largest buyers, and their capex plans set demand across the stack.

Neoclouds

Neoclouds rent GPU capacity, often to a few large customers, and finance it with debt backed by the GPUs. They carry the most leverage to the cycle and the most exposure to customer concentration.

Quantum

Quantum computing is pre-commercial for most workloads; its expected uses are simulation, materials, and cryptography. The thesis sizes it as a long-dated option.

Models

Prices for a given model capability fall quickly, and open-weight models trail the frontier by months. Returns at this layer depend on staying ahead rather than on any single model, which is why the thesis holds it as an option.

Closed

Frontier labs fund the largest training runs. Their compute commitments exceed their current revenue, and part of their financing comes from suppliers and cloud partners.

Open

Open weights let governments and enterprises run models on their own infrastructure, which pushes value toward the infrastructure and application layers.

Applications

In applications, returns go to whoever owns the workflow, the customer relationship, and the proprietary data.

Software

Agents attach to systems of record, where transactions happen, and some vendors price agents by task rather than by seat. Point solutions built on another company's system risk being bypassed by the system's owner.

Cyber

Agents add attack surface and also help defend it, which makes security a precondition for adoption rather than an adjacent market.

Robotics

Robotics foundation models apply SI to physical tasks, with early commercial use in warehouses and factories, where environments are controlled and tasks repeat.

Devices

Phones and PCs run models on the device, raising silicon and memory content per unit.

Drones

Low-cost drones have changed warfare, and militaries are buying autonomous systems at scale.

Space

Low Earth orbit ("LEO") satellites provide broadband and direct-to-device connectivity, which makes resilient communications a national security asset.

Streaming

Generative tools lower the cost of producing and localizing video and audio, which raises the relative value of owned catalogs and large audiences.

The Narratives

Each narrative below states the public claim, then the facts.

Safety and Loss of Control

The claim is that SI is not safe and that frontier labs see capabilities they cannot control without government help. Critics, including White House officials, argue that some labs use safety warnings to win regulation that favors incumbents. Misuse risks are real: SI can find software vulnerabilities, and safety guardrails can be bypassed. But current models still fail on long or unfamiliar tasks, and humans direct their development. Proposed checks include mutual pre-release testing among labs, including Chinese labs. The thesis treats security spending as a cost of adoption and licensing regimes as a policy risk.

Jobs

The claim is that SI will displace white-collar work at scale. Early evidence shows reduced hiring of young workers in the most SI-exposed occupations, while overall employment has so far held up. The buildout is increasing demand for electricians, construction crews, and technicians, and shortages in those trades are themselves a constraint on its pace.

The Bubble

The claim is that the buildout is a dot-com-style bubble sustained by circular financing. Circular financing exists: chip vendors invest in customers, clouds invest in labs that buy cloud capacity, neoclouds borrow against GPUs, and campuses are financed off balance sheet. A few customers account for much of the contracted demand, and hyperscalers have lengthened the depreciation lives of servers, which critics argue overstates earnings. Against that, older accelerators continue to rent, though rental prices fall as new generations ship, and the largest buyers are highly profitable. The thesis carries the legitimate part, concentration and depreciation, into Credit.

Efficiency

The claim is that cheaper training and inference will cut compute spending, as efficiency turned the telecom build into a glut. So far, falling prices have been followed by faster-growing usage and rising total spending, and reasoning models and agents use far more compute per task than chat. The claim becomes right only if rents on older accelerators fall below cost and stay there, which is why those rents are a confirming signal.

Electricity Prices

The claim is that data centers are raising household electricity bills. Retail prices reflect many factors: fuel and wholesale prices, distribution and transmission costs, storm and wildfire recovery, and capacity prices. Where data center demand drives up capacity prices, the cost reaches all customers unless tariffs assign it to data centers, and several states and utilities have adopted large-load tariffs that require data centers to pay for the capacity they reserve. The answer is tariff design and faster permitting.

Climate

The claim is that SI's power demand will reverse decarbonization. Data centers remain a minority of United States electricity demand, and technology companies are among the largest buyers of carbon-free power, including nuclear. Near-term on-site supply is mostly natural gas, which raises emissions where it is used; faster nuclear licensing is the lever that changes that.

Water

The claim is that data centers strain local water supplies. Water use depends on cooling design and location, and closed-loop and air cooling use little water. Much of a data center's water footprint is indirect, at the power plants that supply it. Nationally, data center water use is small next to agriculture; local impact depends on the watershed.

Concentration of Power

The claim is that SI hands economic and political power to a few firms. At the model layer, capabilities spread within months, and open-weight models are widely available. In physical supply chains, concentration is high: a few firms make leading-edge chips, HBM, EUV tools, and accelerators. That concentration is a supply-chain fact to underwrite rather than a reason to slow the technology.

Misinformation

The claim is that generative models will flood the information system with convincing fakes. The cost of producing synthetic text, images, voice, and video has fallen sharply, and fraud and impersonation using synthetic voice and video are rising. Detection and provenance tools are improving but imperfect, so fraud prevention is a growing part of Cyber.

The Capability Wall

The claim is that large language models ("LLMs") only predict the next token, have no model of the world, and are nearing a ceiling. Models trained this way, plus reinforcement learning, solve competition-level math and coding problems, and they are trained on images, video, audio, and sensor data, not only text. Some prominent researchers argue that new architectures are needed to reach human-level intelligence, and models still struggle with novel puzzles that people find easy and with long tasks, and they make up facts. If new architectures are needed, that changes which companies win, not whether compute demand grows.

Bull Narratives

Three bull claims also go further than the facts. "Value accrues to the labs": prices for a given capability fall fast, and rivals catch up within months. "Memory is cheap on peak earnings": memory has always been cyclical, and shortages end when new fabs arrive, which takes years. "Power is scarce, so anything with a plug wins": the binding constraints are permits, interconnection, and equipment, which favor owners of secured sites and suppliers with backlog.

The Risks

The real risks are policy, sentiment, credit, physical supply, and China. The table gives the author's ratings of likelihood over two years and impact on the thesis.

Risk	Likelihood (2 years)	Impact on thesis	Exposed
Policy	Medium	High	Power, Infrastructure
Sentiment	Medium, rising if moratoriums spread	High where a campus is the asset	Data centers, grid projects, BTM
Credit: capex digestion	Medium	Medium	Semiconductors, memory
Credit: single-counterparty stress	Less likely	High	Host clouds, vendors
Credit: efficiency outruns demand	Unlikely	Very high	Neoclouds, single-tenant campuses, closed labs
Firm supply	High	Low	Tailwind for turbine, grid, BTM suppliers
Interconnection	High	Low (high for an unconnected campus)	Data centers, neoclouds
Supply chain	High	Low	Chips, Infrastructure
Memory-cycle turn	Likely	Low	Memory (relieves hyperscaler capex, device margins)
China: Taiwan	Very low	Very high	Every layer above silicon
China: rare earths	Medium	Low	Robotics, drones, motors
China: IP theft	High	Medium	US frontier lead
China: export controls	Swings likely	Chip and equipment revenue, not the buildout	Chips, equipment
China: military	Tracks frontier gap	Total	Everything
Europe	Very high	Very low	Hyperscalers, software, closed labs

Policy

A regulated slowdown is more likely to stall the buildout than a financial bubble. Federal policy seeks to preempt state SI (AI) laws, but until Congress or the courts act, a patchwork of state laws remains. Permitting for generation, transmission, fabs, and data centers can take longer than building them, and tariffs and export controls can change quickly. Power and Infrastructure are most exposed.

Sentiment

Public opposition decides zoning hearings and elections whether or not the narratives behind it are correct. Local opposition can delay or block projects through zoning, permits, moratoriums, and ballot measures, and local and state restrictions on new data centers are spreading, including statewide pauses on permits. Markets reprice on sentiment before any statute changes. Data centers, grid projects, and BTM installations are exposed.

Credit

Much of the buildout depends on credit markets and a few large customers. Debt, private credit, and off-balance-sheet structures finance a growing share of campuses and GPU fleets, and depreciation assumptions for servers are debated. Three outcomes matter: a capex digestion phase that flattens spending, stress at a single large customer that cannot fund its commitments, and efficiency outrunning demand so that older-accelerator rents fall below cost. Neoclouds, single-tenant campuses, and closed model labs are most exposed.

Firm Supply

New firm generation is not keeping pace with load growth. Large gas turbines are back-ordered for years, and new nuclear takes many years to license and build. The likely result is slippage rather than lost demand: campuses energized in phases, sited where power exists, or built behind the meter, which favors turbine, grid equipment, and BTM suppliers.

Interconnection

Connecting new supply or load to the grid takes years, and the wait has lengthened. Queue position is a strategic asset; data centers and neoclouds without a secured connection are exposed.

Supply Chain

Some inputs cannot be scaled quickly with money. Skilled trades take years to train, and leading-edge chips, HBM, and EUV tools each have one to three suppliers. The symptom is schedule delay rather than lost demand.

China

China presents several distinct risks.

Taiwan

Most of the world's leading-edge chips are made in Taiwan, and China's military conducts exercises around Taiwan that rehearse a blockade. United States fabs produce leading-edge chips but receive the most advanced processes after Taiwan. A blockade is the least likely and most severe risk to the stack: every layer above silicon would stall.

Rare Earths

China dominates rare earth refining and magnet production and has restricted exports. Disruption would hit magnets and motors for robots, vehicles, drones, and defense systems.

Intellectual Property Theft

United States labs have reported China-based groups distilling their models, which shortens the American lead without removing it.

Export Controls and Smuggling

The debate has two sides. For controls: compute is a key input to frontier models, so restricting chips can slow rivals. Against: restrictions can be evaded through third countries and push China to build its own chip supply. Policy swings are likely either way, and the impact falls on chip and equipment revenue more than on the buildout.

Military

China's military doctrine names "intelligentized warfare" as the coming form of war, and Chinese models trail the United States frontier by months. If China reached systems far beyond human capability first, the consequence would be strategic rather than commercial.

Europe

The European Union's AI Act imposes compliance obligations on high-risk uses, and the European Union has moved to delay some deadlines. Europe raises costs for American platforms without much slowing the buildout; hyperscalers, software companies, and closed model labs carry the exposure.

The Signals

The signals measure how turbulent the path is and which readings justify sizing up or down. Combined frontier-lab revenue is the lead signal, and rental prices for older GPUs are the confirming signal. The view changes if lab revenue stalls for two consecutive quarters, if older-GPU rents fall below cost and stay there, if capability pacing becomes binding law, or if China closes the frontier gap to under three months.

Signal	Confirms	Breaks
Frontier-lab revenue, all labs (lead)	Compounds and converts to operating cash	Flat two consecutive quarters across leading labs
Older-GPU rental prices (confirming)	Hold as new generations ship	Below hardware cost and stays there
Contracted cloud backlog	Grows faster than capex	Sequential declines at two or more of the largest clouds
Hyperscaler capex guidance	Rises with backlog	Two or more of the largest buyers guide flat or down
Hyperscaler and neocloud credit spreads	Narrow relative to banks	Widen sharply, or a data center bond is pulled
Foundry capex and advanced packaging	Capacity added against orders	Capex cut at the leading foundry
Memory contract pricing	Long-term contracts settle higher	Contract prices fall for consecutive quarters
Turbine and transformer order books	Backlogs grow	Cancellations rise
Capacity auctions and interconnection	New supply responds to price	Shortfalls widen with little new supply
Policy	Tariffs and permitting reform spread, not moratoriums	A binding capability cap, or a second large state pauses permits
China frontier gap	Widens	Closes to under three months
Adoption and productivity	Productivity growth above trend, sustained	Adoption flat for a year

Demand and Cash

Frontier-lab run-rate revenue, disclosed or inferred, is the lead signal because it sits closer to end demand than any hyperscaler proxy. Hyperscaler capex guidance and backlog, utilization rates, and token volumes inform the picture but rank beneath it. Reported revenue also has to convert to cash. A divergence sustained across several quarters and several operators is the warning; a single miss is not.

Capital and Credit

The buildout depends on open credit markets, so spreads matter more than headlines about debt. The readings are spreads on hyperscaler, neocloud, and data center debt; private credit appetite; changes to depreciation lives; GPU rental and resale prices; and the path of rates.

Supply Discipline

Capacity added against firm orders extends the cycle; capacity added on speculation tends to end it. The readings are leading-foundry capex and advanced-packaging capacity, HBM additions and contract prices, and turbine and transformer order books.

Power and Grid

Permits and queue position determine where power arrives, so they are better signals than generation forecasts. The readings are interconnection timelines, air-permit and pipeline approvals for BTM plants, large-load tariff decisions, capacity auction results, and the mix of BTM gas against grid-tied firm generation.

Policy and Sentiment

Permits and moratoriums change before prices do. The policy readings are federal preemption, permitting reform, export-license policy, and equipment tariffs; the sentiment readings are local moratoriums, ballot measures, polling, and how candidates talk about data centers.

China

The frontier gap between Chinese and American models, measured by independent evaluations, is the China signal that matters most. Taiwan Strait activity, rare earth licensing, and smuggling enforcement are tracked alongside it.

Adoption and Productivity

The adopter half of the thesis is tested by enterprise results. The readings are agent deployments inside systems of record, results management is willing to quantify, productivity data, and white-collar hiring, especially at entry level.