Prologue. Before AI had a name: 1936–1955

Artificial intelligence became a named research field in 1956, but its essential ingredients were already converging. The preceding twenty years established four propositions: a general machine could execute any formally describable computation; neurons could be represented as logical units; information and feedback could be measured; and electronic switching could make computation economically scalable. Wartime operations research added a fifth: mathematical models could improve consequential decisions inside a real organization.

Alan Turing’s 1936 work defined a general model of computation and also clarified that some problems are not computable. The inheritance for AI is double-edged. A programmable machine can implement an enormous range of algorithms, but computation has formal limits; scale does not turn every question into a solvable one. Turing’s computability paper

During the Second World War, interdisciplinary operations-research teams used probability, statistics, optimization and field evidence to improve radar deployment, logistics, search and resource allocation. This decision-science lineage later supplied forecasting, queuing, simulation, dynamic programming and mathematical optimization—the hidden machinery beneath many systems now labeled AI. It remains crucial because an LLM may interpret a goal, while a solver or controller calculates the allowable action. INFORMS history

In 1943, Warren McCulloch and Walter Pitts described simplified neurons as logical elements. Their abstraction was biologically limited, but historically decisive: networks of small computational units could collectively implement complex functions. Claude Shannon’s 1948 information theory then formalized information, entropy and channel capacity; Norbert Wiener’s cybernetics connected communication with feedback and control across animals and machines. These traditions supplied concepts that reappear in loss functions, representation learning, compression, reinforcement learning, robotics and system monitoring. McCulloch and Pitts, Shannon, Wiener

The physical prerequisite arrived at Bell Labs in 1947. The transistor offered a smaller, more reliable and more energy-efficient switching element than the vacuum tube. It would take integrated circuits, manufacturing scale and decades of design progress to reach modern processors, but the industrial direction was established: intelligence implemented in software ultimately depends on devices that switch and move information. Bell Labs history

Turing’s 1950 paper Computing Machinery and Intelligence reframed an abstract argument about whether a machine can think into an observable interaction. The imitation game was not a modern benchmark or a definition of intelligence; it was a practical way to discuss machine behavior while anticipating learning, language and common objections. That product insight—capability becomes socially consequential when people can interact with it—would recur with ChatGPT seventy years later. Turing’s 1950 paper

This prologue does not move AI’s formal birth before Dartmouth. It explains why Dartmouth was possible. Earlier intellectual ancestors—probability, formal logic, Babbage and Lovelace’s programmable-machine ideas—belong in the ancestry, but a detailed history of mathematics would obscure this article’s focus: when ideas became a reproducible industrial capability.

0. 1956–2009: research foundations and the hidden computing platform

The economic history in this article begins in 2010 because that is when cloud computing, big-data platforms, digital transactions and mature enterprise analytics began converging at industrial scale. The intellectual history begins much earlier. Modern AI is the result of several research traditions that were repeatedly declared triumphant, repeatedly exposed as incomplete and repeatedly revived when algorithms met better data, compute and software.

The field received its name at the 1956 Dartmouth Summer Research Project. Its organizers proposed that aspects of learning and intelligence could be described precisely enough for a machine to simulate them. Early work focused on symbolic reasoning: represent facts and goals, search through possible actions and apply formal rules. This approach produced theorem provers, planning systems, game-playing programs and, later, commercial expert systems. It established ideas that remain essential—knowledge representation, search, decomposition, explicit constraints and explanations—but it also exposed a production problem. Human expertise is frequently tacit, contextual and difficult to convert into stable rules. Rule bases become brittle as exceptions multiply, while maintenance cost rises with the number of interactions among rules. Dartmouth, AAAI proposal archive

The expert-system booms and AI winters matter because they offer an economic warning for the current cycle. Capability demonstrations can attract capital before reliability, integration and maintenance are understood. A system may work in a narrow laboratory environment yet fail when the domain changes, when its knowledge becomes stale or when responsibility cannot be delegated to software. The lesson is not that earlier AI failed. Scheduling, diagnostic, configuration and decision-support systems created real value. The lesson is that the surrounding knowledge, workflow and governance system usually costs more than the elegant core algorithm.

From encoded rules to learned representations

Neural networks proposed a different answer: instead of writing every decision rule, adjust a network’s internal parameters from examples. The 1986 backpropagation paper by David Rumelhart, Geoffrey Hinton and Ronald Williams showed how error signals could train hidden representations in multilayer networks. The method was not sufficient by itself; datasets were small, processors were slow and deep networks were hard to optimize. Its importance became visible later, when the rest of the stack caught up. Backpropagation paper

During the 1990s, statistical learning became the practical center of gravity. Logistic regression, decision trees, support-vector machines, hidden Markov models, Bayesian networks and ensembles learned correlations from data. Speech recognition, credit scoring, fraud detection, direct marketing, recommendation and forecasting improved without claiming general intelligence. The dominant production recipe was supervised learning: define a target, collect labeled examples, engineer features, train a model, measure error and embed the score in a rule or human workflow.

Neural architectures continued advancing inside that statistical era. Yann LeCun and collaborators demonstrated convolutional networks for handwritten-document recognition, connecting learned visual features with an end-to-end application. Sepp Hochreiter and Jürgen Schmidhuber introduced long short-term memory to address the problem of learning across long delays in recurrent networks. LSTM’s gated state later became foundational in speech recognition, translation, handwriting and sequence modeling. Both examples illustrate a recurring pattern: an algorithm can precede its largest commercial impact by many years because adoption waits for data, compute, tooling and a market with sufficient scale. LeCun and collaborators, LSTM

Data and programmable compute become the hidden platform

The internet then created datasets and workloads unlike the traditional enterprise warehouse. Search queries, links, clicks, advertisements, messages, images and transactions arrived continuously. Google’s 2004 MapReduce paper described a programming model that automatically distributed data processing across large clusters of commodity machines while handling scheduling, communication and failures. Its commercial significance extended beyond one implementation: it helped normalize the idea that data systems should scale horizontally and that software could hide much of the complexity of distributed computation. Google Research

Hardware changed in parallel. Graphics processors contained many arithmetic units designed to execute similar operations across pixels. NVIDIA’s 2006 CUDA architecture exposed that parallel capacity to general computational workloads. This did not guarantee a deep-learning revolution; it created a programmable substrate and a software ecosystem on which researchers could express matrix-heavy algorithms efficiently. CUDA’s libraries, compilers, documentation and backward compatibility later became as strategically important as the silicon. NVIDIA CUDA guide

Data collection supplied the third precondition. ImageNet’s 2009 paper described a large hierarchical image database assembled around WordNet categories. The contribution was not a new learning algorithm. It was a shared training and measurement resource large enough to expose differences among methods and focus a global research community on the same task. This pattern—dataset, benchmark, competitive iteration, reusable software and adequate compute—would recur in language models. ImageNet

Algorithm–hardware co-evolution

AI progress is not a sequence of algorithms floating above the physical world. Every major regime combined a computational idea with a machine, dataset or software system capable of executing it economically. Solving one constraint usually moved the bottleneck elsewhere.

The pre-2010 system therefore contributed more than a list of inventions. It established four durable truths. First, explicit rules remain valuable when policy must be deterministic, even when learning supplies the prediction. Second, data quality and task definition can matter more than algorithm novelty. Third, a hardware advantage becomes durable only when software makes it programmable. Fourth, benchmark success is an intermediate result; production value appears only after a model is integrated into a repeated decision and monitored through change.

1. 2010–November 2022: analytics, deep learning and foundation models become industrial

In 2010, enterprise AI usually meant analytics, forecasting, optimization or decision support. Companies built warehouses, ETL pipelines, reports and specialist teams so a model could improve a repeated decision. Outputs entered dashboards, campaign lists, planning tools or rules; people and deterministic software usually retained action authority.

2010–2015: big data industrializes analytics

Hadoop, Spark, NoSQL systems and cloud storage made web logs, transactions and sensor data more usable. Production models remained narrow—one target, dataset, workflow and owner—but created material value.

The expensive element was usually not model training. It was data integration, software, validation, specialist labor and process change. IDC estimated narrow big-data technology and services revenue growing from less than $5B in 2010 to roughly $15–20B in 2015. IBM reported $18B of analytics revenue in 2015; GE reported roughly $500M of annual internal digital productivity savings; UPS expected ORION route optimization to save $300–400M annually. These company cases established industrial value before generative AI. European Commission/IDC, IBM, GE, UPS

Realization was uneven. McKinsey estimated that by 2016 U.S. retail had captured about 30–40% of previously identified analytics value, manufacturing 20–30%, and healthcare and government only 10–20%. Data quality, workflow ownership and management—not algorithm availability alone—separated leaders from laggards. McKinsey

2012–2022: learned representations become reusable

AlexNet’s 2012 ImageNet result joined a large dataset, benchmark, GPU and trainable convolutional network. Deep learning then spread through vision, speech, translation, search, advertising and recommendation. LSTMs handled sequences; embeddings learned reusable representations; GANs expanded generation; and reinforcement learning plus search produced systems such as AlphaGo. Digital platforms advanced fastest because they owned large interaction streams and rapid experimental feedback. AlexNet, GANs, AlphaGo

The 2017 transformer made sequence learning more parallel and scalable. Self-supervised training extracted supervision from raw text, images, audio and code; BERT and GPT shifted development from one model per task toward reusable pretraining. Scaling laws made larger compute programs legible to laboratories and capital markets, while Chinchilla, mixture-of-experts and retrieval showed that data allocation, sparse computation and external memory could improve the return on compute. Instruction tuning and human preference learning then converted a next-token predictor into a more cooperative assistant. Transformer, Scaling laws, Chinchilla, RAG, InstructGPT

Researchers and competing hypotheses

The point is not to choose one famous researcher as the winner. These programs solve different problems: representation, memory, action, causal reasoning, evaluation and control. The future system is likely to combine several of them.

China before ChatGPT

Alibaba, Baidu, Tencent, ByteDance and financial platforms already operated recommendation, search, advertising, logistics, risk and computer-vision systems at immense scale. Baidu’s ERNIE, the 2020 CPM paper and the 2021 PanGu-α paper established Chinese-language foundation-model research before the global consumer breakthrough. China had capability, investment and engineering depth; it had not yet produced one global product moment that compressed public understanding and commercial urgency.

Traditional analytics versus the new value pool

These ranges are transparent article reconstructions, not audited global profit. The direction matters: traditional analytics improved a limited set of structured decisions; generative and agentic systems add unstructured knowledge, software and coordination. Traditional ML remains inside the new stack for forecasting, recommendation, tabular risk and optimization, while LLMs interpret intent and coordinate tools.

ChatGPT’s discontinuity came from composition: a scaled model, post-training, conversation, low-friction access and immediate feedback. Once executives perceived a platform transition, capital flowed into accelerators, clouds, laboratories, data centers and power. The central question moved from whether models could generate useful output to whether the industrial system could supply and organizations could absorb reliable intelligence.

2. December 2022–2026: mass distribution, multimodality and agents

ChatGPT joined a capable model to conversation, freemium access and rapid feedback. It sharply reduced the cost of trying AI and turned a technical frontier into a mass-market habit. The resulting demand signal pulled capital into models, chips, cloud and applications.

From product launch to global habit

These figures are not market shares. Weekly users, monthly users, embedded assistants, subscribers, developers and API calls have different denominators. The defensible conclusion is that ChatGPT established the leading direct AI destination while Google, Meta and Chinese platforms converted existing distribution into enormous reach.

Distribution became part of capability. Usage supplied product feedback; traffic pulled forward infrastructure; APIs allowed startups and business units to experiment; and default placement inside search, work, devices and messaging became a strategic advantage. Assistants may increasingly sit upstream of websites, merchants and publishers, redirecting advertising and transaction value even while creating consumer surplus.

Coding agents: the first mass market for agentic work

Coding moved first because software provides machine-readable context, tools, tests, version control and rollback. The product evolved from autocomplete to chat, then to repository agents that plan, edit, run commands, inspect errors and return reviewable changes. The innovation is the harness around the model—not code generation alone.

The evidence is deliberately mixed. Anthropic reports Rakuten reducing one feature cycle from 24 working days to five; Google reports a major internal refactor compressing a two-year schedule to three months. Yet METR found experienced maintainers 19% slower on a different set of familiar-repository tasks, and 46% of Stack Overflow respondents distrusted AI-output accuracy. The relevant metric is time and total cost per accepted, verified change, including inference, tests, review, failures and rework. Rakuten, METR, Stack Overflow

Three shifts after ChatGPT

1. Models became multimodal, post-trained systems. Text expanded to images, audio, video and interfaces; preference training and safety operations shaped product behavior.
2. Compute moved into inference. Reasoning, search, tool calls and verifiers made the cost of one task variable, while batching, caching, quantization and routing improved utilization.
3. Answers became actions. Agents connected models to data and tools, but errors, prompt injection and excess authority made identity, permissions, tests and rollback essential.

Open weights from Llama, Mistral, Qwen, DeepSeek, GLM and Kimi made capable models cheaper and more portable, although deployment still requires compute, security, evaluation and support. By 2031, one user task may route across several closed, open, specialist and traditional models. The durable control point is therefore likely to be the system holding consented context, identity, tools, evaluation and the customer relationship.

The post-ChatGPT race is no longer only to train the strongest model. It is to assemble the most reliable and economical system for completing useful work.

3. The AI factory: chips, data centers, power and water

AI appears as software but is produced by a physical chain: semiconductor tools, leading-edge logic, HBM, packaging, networks, servers, cooling, power, water and cloud operations. A shortage or delay at one qualified link can strand investment everywhere else.

From cloud expansion to an industrial buildout

A concentrated global supply chain

No country needs to reproduce every layer. Resilience comes from competitive model and cloud access, enough governable compute for critical work, reliable networks and power, technical talent and the ability to switch suppliers.

The bottleneck moves

Three facts organize the factory.

1. Complete systems matter more than chip counts. Accelerators can wait on HBM, packaging, optical networks, switchgear, cooling or grid connections. Revenue flows first to whichever qualified component cannot expand quickly.
2. Training and inference need different machines. Training emphasizes synchronized cluster throughput; interactive inference emphasizes latency, memory, batching and reliability; edge inference emphasizes privacy and low power.
3. Software determines usable capacity. Compilers, kernels, distributed systems, caching, quantization and routing convert theoretical performance into cost per completed task.

The production stack remains heterogeneous: CPUs orchestrate; GPUs and custom accelerators perform parallel work; HBM feeds them; packaging and networks connect them; serving software schedules them; identity, retrieval and verification determine whether output becomes an authorized action. The economically useful unit is not a transistor, GPU or token. It is a reliable task completed by the whole system.

Power, water and materials

Power is permission to operate. A campus can be built faster than new generation or transmission, and the IEA estimates that 20% of planned data-center projects could face delays without grid action. Efficiency lowers energy per fixed task, but reasoning, video and agent loops expand demand. Operators should report verified tasks per dollar, watt and liter alongside absolute resource use.

Annual renewable matching does not prove carbon-free supply in every hour. Credible claims identify local physical supply, firming, additional generation and attributable grid cost. Large loads should fund the infrastructure they require and use flexible workloads where service and data rules permit.

Water is local. LBNL estimates about 66B liters of direct U.S. data-center consumption in 2023 and 800B liters indirectly associated with electricity, with direct use potentially reaching 150–280B liters in 2028. Withdrawal, consumption, indirect use and replenishment are different measures; disclosure should be facility- and basin-specific. LBNL

The environmental ledger also includes steel, concrete, copper, specialty chemicals, semiconductor water and frequent server replacement. Buildings, substations and fiber can serve several accelerator generations, so adaptable sites and equipment reuse reduce both capital and embodied impact.

The 2031 physical architecture

The likely system combines frontier clusters, custom silicon for stable workloads, smaller edge models and schedulers that route by quality, latency, price, residency and resource conditions. Batch work can sometimes shift in time or geography; real-time services and synchronized training cannot.

The IEA’s electricity outlook, Morgan Stanley’s investment estimate and Synergy’s capacity forecast answer different questions and must not be added. Together they show that advantage belongs to the operator that converts chips, energy, water and networks into utilized, reliable and socially permitted intelligence.

4. The model and cloud economy: laboratories, training and inference

Model laboratories convert compute into reusable capability; clouds meter and distribute capacity; applications combine models with data, tools and customers. These layers have different costs and should not be evaluated with one revenue or margin model.

From data to a production service

A final training run is only one cost. Epoch AI estimates roughly $4.5M for GPT-3, $12.4M for PaLM and about $390M for the largest public 2024 estimates. It also describes a path toward $100B-plus frontier clusters by 2030. A laboratory can therefore report a training run in the hundreds of millions while relying on a many-billion-dollar reusable industrial platform. Epoch AI

Capital structures now mix equity, cloud credits, compute reservations and commercial agreements. The key question is who owns costly capacity and who owns the customer. Asset-light laboratories avoid construction and depreciation but pay a provider margin; integrated clouds capture more of the stack but must keep short-lived chips and long-lived sites utilized.

Inference and cloud become the recurring economy

Inference is continuous. Cost depends on model size, context, reasoning steps, modalities, caching, memory bandwidth, retries and service guarantees. The price of GPT-3.5-level capability fell from about $20 to $0.07 per million tokens between late 2022 and late 2024, but longer context, video, search and agents increased consumption. The useful metric is successful, verified tasks per dollar, joule and minute, not tokens alone.

Cloud infrastructure revenue grew from $129B in 2020 to $227B in 2022 and about $500B for the trailing twelve months through Q2 2026. Hyperscalers combine diversified demand, proprietary chips and product distribution; neoclouds specialize in accelerator operations; sovereign providers sell jurisdiction and continuity; enterprise or edge deployments serve private, high-volume or latency-sensitive work. Their common risk is paying for capacity before utilization arrives.

Routing is likely to become a durable control point. It can assign routine work to small or local models, difficult planning to frontier systems and exact calculations to deterministic tools, while preserving evidence and budget. If routing and data remain portable, buyers retain leverage; if hidden inside one platform, lock-in deepens.

Closed models, open weights and China’s alternative stack

Public evaluations show meaningful convergence, but benchmark proximity is not production parity. Buyers must also test uptime, tools, security, capacity, license, support and total cost per successful task.

Chinese open models matter because they widen supplier choice, compress prices and show how algorithmic efficiency can offset some hardware constraints. They also shift competition toward deployment engineering, local language, evaluation and integration. “Open weight” remains the accurate term when code, data, training recipe or license rights are incomplete. OSI, OECD

There may be no single most-used model in 2031. One family can lead conversation, another embedded search, another enterprise tokens and many open models can run without central measurement. The likely winner is the organization that can route among them without losing its data, permissions, evaluations or customer relationship.

5. What current AI cannot reliably do—and the research attempting to fix it

Current systems are useful but probabilistic. They can write, code, search and call tools, yet still invent facts, miss constraints, lose state and execute the wrong action. Longer context and more test-time compute improve many tasks; neither guarantees truth, memory or safe autonomy.

The practical question is not whether every limitation disappears. It is whether better algorithms, memory systems, chips and controls can lower the cost of a verified outcome enough for wider deployment.

Five limitations that shape the 2031 market

Three research and deployment priorities

1. Verify the outcome, not the explanation. Use tests, retrieved evidence, formal rules and sampled human review that are independent of the generating model.
2. Give agents limited authority. Identity, short-lived credentials, action budgets, checkpoints and rollback should expand only after measured reliability improves.
3. Optimize the complete system. Route routine work to smaller models, preserve structured state and measure dollars, joules, latency and recovery cost per successful task.

The largest open risk is a verification bottleneck: generation becomes abundant while trustworthy review remains scarce. The largest opportunity is the same market in reverse. Organizations that combine strong models with evidence, secure tools and recoverable workflows can turn probabilistic capability into dependable service.

6. Adoption becomes application: consumers, enterprises, industries and government

Mass adoption is not the same as economic absorption. Consumers can obtain value from a useful answer immediately; companies and governments must connect models to data, permissions, workflows and accountability. The gap between those two speeds explains why user growth is spectacular while enterprise return remains uneven.

Breadth rose faster than depth

Three findings survive across the evidence. First, bounded work can improve materially: customer-support agents produced 13.8% more resolutions per hour, and three developer experiments found about 26% more completed tasks. Second, task boundaries matter: consultants improved inside the model’s capability frontier but performed worse outside it, while experienced maintainers in one METR study were 19% slower with early coding tools. Third, firmwide return requires complementary software, data, training and process redesign; survey evidence shows many pilots still fail to scale. NBER, Harvard Business School, Management Science, IBM

Applications and investment economics

The 2031 base case is one waterfall: $6.42T annual value enabled → $4.10T captured before cost → $700B program cost → $3.40T net direct benefit. The figures are an article scenario, not vendor revenue or GDP. They extend the economic architecture developed in Agentic AI: Transforming Industries, From Better Answers to New Industry Economics and Agentic AI: $16Tn Enabled.

Three application and investment rules

1. Fund an outcome, not a feature. The strongest application owns a measurable workflow, regulated permission, system of record or customer relationship. A thin interface is vulnerable to model and platform bundling.
2. Price the full operating system. Include inference, integration, review, failure recovery, security and organizational transition. Time saved becomes profit only when output, capacity, price or loss changes.
3. Scale in stages. Begin with high-volume, reversible and checkable work; expand authority only when cost, quality and intervention rates remain acceptable.

Coding agents are the clearest reference market because software provides repositories, tests, version control and rollback. Claude Code’s company-reported $1B run rate shows willingness to pay, while Codex, Gemini, Grok, GitHub and open-model tools keep the category competitive. The appropriate metric is not generated code but time and total cost per accepted, verified change. Anthropic, OpenAI Developers

Government has three priority application groups: administration and revenue such as tax, procurement and benefits; human services such as health, education and workforce support; and infrastructure and security such as grids, emergency response and cyber defense. In every group, the return is reliable public capacity per dollar, while legal authority, adverse decisions and appeals remain human and institutional responsibilities.

The transition is therefore from seats and prompts to verified outcomes. Leaders should track completion, cycle time, quality, intervention, loss and resource use. Adoption is an input; retained value is the result.

7. The money flow: capex, venture capital, revenues, valuations and returns

AI finance follows a sequence: funding → equipment orders → installed capacity → usage → productivity → cash return. Suppliers are paid before buyers prove value, and the same dollar can appear as customer capex, supplier revenue and cloud cost. Those objects must not be added.

The capex reset

Definitions and fiscal periods differ, and Amazon includes logistics and other assets. The direction is nevertheless clear: post-ChatGPT capital formation entered a different regime. Company filings and Goldman Sachs

Where returns appear

Three conclusions from venture research

Andreessen Horowitz’s payment data, Sequoia’s return questions, General Catalyst’s service thesis, Coatue’s cross-stack view and Matrix’s infrastructure portfolio reinforce those conclusions from different vantage points. They are useful directional evidence, not neutral market measurement. a16z, Sequoia, General Catalyst, Coatue, Matrix

Three investment rules follow. First, match financing to asset life: accelerators, buildings and power infrastructure have different depreciation and refinancing risks. Second, require evidence of utilization and customer return before treating bookings, capacity announcements or valuation marks as demand. Third, prefer durable control points—energized sites, proprietary data, trusted distribution, regulated permission, systems of action and verified outcomes—over reproducible features.

Private valuations are claims on future cash flow, not present economic value. Paper revaluations, preferred terms and strategic compute agreements can obscure operating performance. The 2031 test is simple: does incremental gross profit and free cash flow justify capex, leases, research, inference and transition cost?

8. National and human consequences: sovereignty, work, security and trust

AI use is global, but ownership of chips, clouds, models and capital is concentrated. Capability can therefore diffuse faster than profit or policy power. The United States leads frontier laboratories, accelerator design, clouds and venture finance; China combines a vast domestic market, engineering depth, manufacturing and influential open-weight models. Both remain dependent on an international supply chain spanning Taiwan, Korea, Japan and Europe.

National strategy and local value capture

China reported 602M GenAI users, more than 6,000 AI enterprises and a core industry above RMB1.2T in 2025. U.S.-based firms received about 75% of reported global AI venture capital in the OECD dataset. These figures use different definitions, but they show two different systems: U.S. private capital and frontier platforms versus Chinese deployment scale, industrial policy and efficient-model competition.

For countries without a frontier laboratory, sovereignty should mean continuity, not prestige. The World Bank’s four priorities—connectivity, compute, context and competency—are a stronger foundation than one national chatbot. Add multi-provider procurement and an ability to evaluate, secure and switch critical systems. World Bank

This allocation is not a GDP forecast. Value created locally can still be captured by a foreign chip, cloud or model provider. Governments should test whether infrastructure projects create reliable power, tax revenue, skills and supplier capacity rather than only exporting subsidized land, water and electricity.

Three public functions

Public procurement should reuse secure components, publish outcome measures and preserve portable data. More use cases are not success: the U.S. agency inventory rose from 571 to 1,110 between 2023 and 2024, while only a fraction had moved beyond early stages. Rights, security and recourse are part of the return, not obstacles outside it. GAO

Work will reorganize around tasks

Evidence supports broad exposure but not a single job-loss forecast. The ILO finds one in four jobs has some GenAI exposure and expects transformation to be more common than full replacement. The World Economic Forum’s employer survey expects large gross job creation and displacement across all modeled trends; it is not a realized AI employment count. Early studies also show uneven pressure on younger and more exposed workers. ILO, WEF

Coding agents make the transition visible first. Engineers spend less time typing every implementation and more time specifying intent, preparing context, designing tests, reviewing changes and accepting production responsibility. Demand may expand because lower software cost makes more products and modernization viable, but routine implementation and labor-billed services face pressure. The immediate risk is apprenticeship: if firms remove junior tasks and hiring, they weaken the future supply of reviewers and architects. Supervised practice, debugging, incident rotations and independent assessment should remain part of the job.

The human supply chain also includes data creators, annotators, safety raters, moderators and domain reviewers. Provenance, consent, fair working conditions and governed access to authentic scientific, operational and local-language data are economic inputs. Synthetic data helps when simulation or verification supplies ground truth; it cannot replace observation indefinitely.

Access, security and trust

Export controls and provider safeguards restrict some chips, models and uses because the same systems can support cyber defense, vulnerability discovery, science or harmful operations. A credible regime combines customer screening, secure weights and compute, least-privilege tools, monitoring, incident reporting and human authorization for consequential actions. It should also avoid turning safety into unreviewable commercial gatekeeping.

Organizations should design for capability continuity: portable data and evaluations, more than one viable model route, locally controlled identity and logs, and fallback procedures for provider or policy changes. High-impact public and commercial decisions also require notice, evidence, human authority and appeal. Trust is economic infrastructure because institutions will not delegate important work to systems they cannot challenge or recover.

The distribution test is therefore concrete: who owns the scarce assets, who receives the productivity gain, whether consumers obtain lower prices or better services, whether workers retain mobility, and whether countries can switch providers. Technology creates the opportunity; institutions distribute it.

9. 2027–2031: four futures and one operating agenda

The next five years depend on several variables moving together: model reliability, inference efficiency, energized capacity, capital cost, workflow conversion, labor response and public legitimacy. The following are conditional futures, not precise predictions.

Four plausible intelligence-age systems

The planning envelope

A simpler production architecture

The likely system is a governed portfolio rather than one universal model.

MCP, A2A and NIST’s agent initiative indicate a move toward interoperable tools, agent communication, identity and evaluation. The standards are early, but the direction is clear: context and authority need explicit control outside the model. MCP, A2A, NIST

Three actions through 2031

1. Prove workflow conversion. Enterprises and governments should select a few measurable processes, track cost and quality per completed task, and expand autonomy only after evidence.
2. Underwrite utilization, not announcements. Investors should connect chips, buildings, power and leases to energized demand, cash yield and appropriate asset lives.
3. Build diffusion and exit rights. Governments, employers and buyers should support skills, apprenticeships, shared access, portable data, plural suppliers, independent evaluation and recourse.

Three indicators reveal which future is emerging: AI cash yield, or incremental gross profit and free cash flow relative to commitments; workflow conversion, or the share of important processes reliably completed; and diffusion quality, or whether smaller firms, workers, public institutions and countries without frontier labs can access and retain value.

The robust strategy does not require certainty about model progress. Modular systems, flexible infrastructure, trained people, independent verification and the ability to switch suppliers remain valuable across every scenario.

10. Beyond 2031: six questions that will define the 2035 intelligence economy

2031 is the article’s decision horizon because today’s fabs, data centers, power contracts, model investments and workforce programs can be assessed against it. Extending precise market-share or valuation forecasts to 2035 would create false confidence. A longer horizon is still useful if it is framed as questions, alternative paths and observable signals, not a single numerical destiny.

The range in energy analysis illustrates the problem. The IEA’s Energy and AI work places 2035 global data-center electricity demand across cases spanning roughly 700–1,700 TWh, with a high-efficiency pathway around 970 TWh. The width is not analytical failure. It reflects uncertainty about model architecture, task demand, chip efficiency, utilization, flexible load and policy. The same uncertainty applies to labor, market structure and national access. IEA executive summary, IEA demand analysis

What these questions imply

First, the 2035 economy may be more heterogeneous than today’s model race suggests. A user may experience one assistant while a router calls a small private model, a specialist predictor, a frontier reasoner, a search system, a formal solver and a human approver. The strategic asset shifts from one model checkpoint toward the institution’s context, permissions, evaluations, action surfaces and evidence of successful outcomes.

Second, physical constraints will not disappear merely because algorithms improve. Efficiency changes which projects are economical, but lower cost expands demand. Power, transmission, HBM, packaging, cooling, water and skilled construction remain important; so do the permissions of communities that host them. The best long-horizon infrastructure is therefore modular and adaptable: reusable power and networks, buildings that can accept new equipment, diversified customers and financing matched to the life of each asset.

Third, the decisive economic uncertainty is distribution, not just capability. A powerful system can create consumer surplus while reducing wages, improve a public service while weakening due process, or raise corporate output while destroying the apprenticeship that produces future experts. By 2035, the important scorecard should join technical performance with cash yield, resource intensity, competition, worker mobility, service access and recourse.

Finally, the longer horizon strengthens rather than weakens the article’s central recommendation. No government, company or investor needs to predict artificial general intelligence to act intelligently now. Build portable data and evaluations; buy plural model routes; preserve human authority for consequential decisions; finance adaptable infrastructure; train people inside redesigned roles; and measure verified outcomes. Those capabilities retain value across every plausible 2035 path.