AI Could Send the World’s Capital Flowing Uphill to Its Richest Economies

After two centuries of spreading global prosperity, the economic forces that helped poorer countries develop could be thrown into reverse.

In the first half of 2026, the United States received an estimated 78% of the world’s newly arriving AI computing power, according to Epoch AI. That extraordinary lead matters because AI threatens to remove one of the great advantages that helped poorer countries develop: the ability to attract foreign investment by offering cheaper workers. The countries that most need capital to catch up could see AI send capital back uphill to the richest economies.

For generations, investment helped spread prosperity from richer countries to poorer ones. In the nineteenth century, Britain exported capital to finance railways, infrastructure and resource development overseas, where capital was scarce and investment opportunities were plentiful. Much later, manufacturers moved production to countries such as China, Vietnam and Bangladesh for a different reason: cheaper workers. In both cases, investors could earn attractive returns abroad, bringing productive capacity, technology and income to economies outside the established industrial powers.

AI and robotics could change that calculation. One of the most important advantages poorer countries offer investors is inexpensive labor. But when businesses can manufacture their workers instead of hiring them, it may become cheaper to automate at home than move production abroad. Their costs will increasingly depend on the price and productivity of machines, electricity, financing and infrastructure rather than local wages.

Consider a garment maker choosing between Germany and Bangladesh. Lower Bangladeshi wages can compensate for weaker infrastructure or more expensive financing. Once both factories become fully automated, however, the difference in wages ceases to matter. If Germany can operate more productive robots with better infrastructure and cheaper capital, it could become the lower-cost producer. Bangladesh would lose the advantage that attracted the factory in the first place.

Digital services face this threat even sooner. The Philippines has built a major export industry around business services. Overseas customers could replace call center workers and administrative staff with AI without installing a single robot in the Philippines. The country could lose export income before its own industries have had the opportunity to adapt.

Why capital could flow uphill

Artificial labor is unusual because it can become cheaper and more productive at the same time. Wright’s Law describes how manufacturing costs tend to fall as more units are produced and manufacturers gain experience. Meanwhile, improvements in AI allow the same equipment to perform more work. Recursive self-improvement, in which increasingly capable AI helps develop better AI, could accelerate this process. Cheaper, more productive artificial workers make more tasks worth automating, increasing demand for the next generation of machines and intelligence.

This is where America’s enormous starting advantage becomes particularly important. More capable AI makes automated production more profitable, attracting further investment in the technology and infrastructure that improve it. Capital that once moved abroad in search of cheap workers may instead find better returns in economies that already possess advanced AI, industrial infrastructure and deep financial markets.

For poorer countries, this could create a vicious circle. Competition from automated production overseas reduces their export income, leaving less money to buy the machinery and build the infrastructure they need to remain competitive. At the same time, international investors may prefer the returns available in advanced economies. The countries that need investment to catch up could find it increasingly difficult to attract.

What will still be worth paying for?

As AI and robotics make goods and services increasingly abundant, countries may find that some of their most valuable assets are those that cannot be reproduced. The world can build and operate hotels ever more cheaply, but it cannot create more locations overlooking the Swiss Alps. It can grow and process grapes at ever lower cost, but Champagne grapes can only come from Champagne. Robots can stitch handbags with extraordinary speed and precision, but they can never make a handcrafted Italian bag. AI and robots may deliver flawless service, but they cannot recreate the experience of genuine Thai hospitality.

As manufactured goods become cheaper, the income earned from these scarce assets could buy progressively more. A Swiss hotel might charge the same price for its rooms, but if the price of manufactured goods halves, its income can buy twice as many cars, televisions or household appliances. Switzerland could enjoy rising purchasing power without producing anything more. Countries with assets that cannot be reproduced could therefore have an increasingly valuable source of national wealth.

Who controls the new economy?

Much will depend on who owns the businesses exploiting AI. Imagine a British accounting firm buying American AI to serve its customers more efficiently. Much of the resulting profit remains in Britain. Now imagine the American AI provider selling accounting services directly to those customers. Accounting becomes cheaper, but the income that once supported a British business increasingly flows overseas. Foreign-owned automated factories could similarly generate large amounts of output while leaving little income in the country where they operate.

Even countries that successfully use American AI could face a deeper problem. If their factories, banks and service industries depend on American intelligence to remain competitive, much of their economy depends on technology controlled abroad. American model owners would have considerable influence over the price and terms of access, while the US government could restrict its availability. Other governments could exclude American AI, but without equally capable alternatives, they risk making their own economies increasingly uncompetitive.

America would not need to own the world’s factories to exercise enormous influence over them. It might be sufficient to control the intelligence they need to operate. Other countries therefore have a powerful incentive to develop their own AI or secure alternative sources from China or elsewhere. But doing so requires capital just as governments are being called upon to support people displaced by automation.

A world of winner-take-all economies

As employment and export earnings decline, governments will need new sources of revenue to support their populations. Taxing productive capital may provide an answer, but only if its owners and income remain within reach of national governments. If much of the world’s productive wealth becomes concentrated in America, there is little reason to expect American taxpayers to support populations elsewhere that have lost their means of earning a living.

The IMF has already identified the danger that increasingly productive robots could draw capital towards richer countries and widen international economic disparities. The wider risk is that AI could reverse the economic forces that have helped spread prosperity around the world for two centuries. Instead of technology and investment creating new opportunities for poorer countries, they could increasingly reinforce the advantages of a small number of already wealthy economies.

The result could be a world of winner-take-all economies, with a handful of countries controlling much of the world’s productive technology and capital, while many others become uncompetitive across a growing range of industries. Governments will fight to avoid this outcome, but countries that fall behind may find it increasingly difficult to finance their way back into competition.

This is the great paradox of the AI revolution. The world could become capable of producing more wealth than at any point in human history, while entire countries struggle to find a way to earn a living. The danger is not a world of scarcity, but a world of abundance in which many countries can no longer find a seat at the table.

Based on ideas from the #1 book: https://amazon.com/dp/B0HFKSP2WF

For over 400 articles and reports based on the ideas in the book: https://abundanceeconomicsresearch.substack.com

More from Abundance Economics: https://abundance-economics.com

Previous infrastructure booms left behind valuable networks and ruined investors. This one has two important protections: customers are already committed, and the biggest builders have enormous businesses financing them.

The United States could invest nearly $9 trillion in AI infrastructure between 2025 and 2032, an average of 3.2% of GDP each year. That is greater, as a sustained share of the economy, than the railroad and telecommunications booms. Financing is expanding almost as dramatically. Goldman Sachs counted nearly $500 billion of AI-related debt issuance by early August 2026, while CoreWeave’s unsecured bonds reached yields of roughly 15–16% in October. Investors have reason to remember what happened the last time capital poured into an essential new network.

Canals helped transform commerce but contributed to state debt defaults. Railroads attracted enormous amounts of capital and suffered repeated bankruptcies. Telecommunications companies installed the fiber networks the internet needed, only for overcapacity and debt to destroy some of their builders. Electrification ultimately created durable utilities, although speculative holding companies also failed. The common lesson is that useful infrastructure can be a disastrous investment when revenue arrives too late to meet the financing bills.

The construction has customers before it starts

The first difference is the demand behind the new capacity. Real estate adviser JLL reported that 95% of the 66 gigawatts of data-center capacity under construction in North America in August 2026 was already committed to a user. That includes facilities companies are building for themselves, so it does not mean 95% has been sold to outside customers. It does mean the present construction pipeline is largely being built for identified use, rather than on a speculative assumption that tenants will eventually appear.

The economics of those commitments are visible in IREN’s contract with Microsoft. IREN agreed to deliver about 200 megawatts of computing capacity under a five-year contract worth $9.7 billion. It committed to buying roughly $5.8 billion of GPUs from Dell; Microsoft will pay 20% of each delivery tranche in advance. IREN subsequently reported that customer advances and investment-grade financing covered 96% of the associated GPU purchases. The contract gives lenders an identifiable stream of payments against which to finance equipment before it earns revenue.

The financial protection provided by these contracts depends on who has promised to pay, and the majority of contracts have very strong financial guarantors. Microsoft, Alphabet, Amazon and Meta have substantial cash flows from established businesses, allowing them to honour infrastructure commitments even if demand for new AI services grows more slowly than expected. OpenAI and Anthropic are the other largest customers and also have paying customers, but their ability to meet long-term commitments depends more directly on continued AI revenue growth and access to finance. For developers and lenders, a signed order from an established technology group therefore provides stronger financial protection.

The builders have balance sheets previous booms lacked

The second difference is who is paying for the buildout. Boston Consulting Group estimates that Microsoft, Alphabet, Amazon, Meta and Oracle will invest about $4.6 trillion in AI capital expenditure during 2022–2030, accounting for 60–70% of the investment covered by its forecast. That share applies to BCG’s forecast, rather than to the separate $9 trillion estimate of US spending during 2025–2032. These five companies are established businesses with substantial sources of income. Google Search and Meta’s advertising businesses, Microsoft’s enterprise software and Amazon’s retail, marketplace and existing cloud operations were generating large amounts of cash before the current spending surge. Oracle has an established software business too, although its debt burden is considerably heavier.

The contrast with ordinary capital-intensive industries is striking. Taken together, the Big Five have funded debt equivalent to approximately 0.27 times book equity, and their reported operating income covers recognized interest expense about 23 times. Telecommunications companies average 1.56 times debt to equity and just 3.6 times interest coverage. The giants can absorb poor returns on some projects without necessarily losing the cash flows that support their debt.

Cash flow provides a second test of resilience. The Big Five generated around $724 billion of operating cash flow over their latest reported twelve months, while cash capital expenditure absorbed about 95% of it. Their filings also show roughly $3.16 trillion of future contractual commitments for leases, construction, equipment and other purchases. These multi-year commitments already cover part of the planned investment rather than adding another $3.16 trillion to it; they also include non-AI obligations and potentially overlapping categories.

Of BCG’s $4.6 trillion forecast, roughly $3.9 trillion falls in 2026–2030, averaging $780 billion a year. That is more than the group’s current annual operating cash flow. Amazon and Oracle are already investing more cash in capital assets than their operations produce. Sustaining this pace will require continued earnings growth, cash reserves or additional borrowing. The five companies’ operating cash flow grew at a combined annual rate of 18.8% between fiscal 2020 and 2025. Continued growth would help fund the construction, but they are almost able to fully fund it out of current cashflow.

The real test is the return on investment

The point of all this investment is to capture the growth in customer demand. Financial strength and booked demand make a bust less likely, but they do not ensure profitable investment. Nvidia’s Jensen Huang has suggested that a one-gigawatt AI data center could cost $50 billion and generate $40 billion in annual revenue. His number is an illustration, not the cash result of a signed contract. IREN’s Microsoft agreement provides a more grounded case: about $2 billion in annual revenue at full capacity against estimated initial spending of around $9 billion. Gross revenue would equal the initial investment after roughly four and a half years. Electricity, maintenance, interest and equipment replacement push actual cash payback further out.

The potential returns are improving in some contracts. IREN reported newer agreements priced above $20 million per megawatt per year, more than twice the Microsoft contract’s annual revenue per unit of capacity. At similar equipment and construction costs, revenue would match the initial investment in less than 2.3 years, before operating and financing expenses. The Big Five also use AI themselves, giving them another route to returns through better advertising, higher-value software and lower costs in existing businesses. Those gains may prove more profitable than renting computing capacity, but the incremental profits have not yet been measured separately.

The danger is that better chips and software make each unit of computing work cheaper. Demand may grow enough to fill expanding facilities, but competition can push down the prices earned by older equipment while loans and leases remain fixed. Specialist providers, dependent on a few contracts and frequent refinancing, face that risk more directly than Microsoft or Alphabet. Bond markets are already distinguishing them: CoreWeave pays far more to borrow than the strongest hyperscalers.

The AI boom can therefore be different without every participant prospering. Much of today’s infrastructure has a committed user, and much of the spending is backed by companies whose existing businesses can withstand disappointing returns. The financial failures, if they come, are more likely to be concentrated among projects and borrowers unable to recover their costs before the technology moves on.

The networks may transform the economy, as canals, railroads and fiber did. This time, more of the companies paying to build them may survive to collect the benefits.

For the book: https://amazon.com/dp/B0HFKSP2WF

For more than 400 case studies: https://abundanceeconomicsresearch.substack.com

More from Abundance Economics: https://abundance-economics.com

Tesla Can Build Optimus Before It Can Make It Useful. That Is Where the Race Really Starts

Tesla is reaching an unusual point in the development of its Optimus humanoid robot. It can begin manufacturing the machines before it has finished developing the capability that will make them valuable. On Tesla’s latest earnings call, Elon Musk said the company may need roughly 10,000 to 20,000 Optimus robots practising tasks so their experience can be used to improve the system. At the same time, Tesla is still developing the robot’s design, production process and artificial intelligence. The production line can therefore start producing robots while much of the work needed to turn those robots into dependable workers is still taking place.

That is very different from the normal relationship between product development and manufacturing. Once a washing machine has been designed to wash clothes properly, making another one produces another machine that can wash clothes. A car is useful when it leaves the factory because the driver supplies the judgment needed to operate it. With a humanoid robot, the mechanical product and the useful worker can arrive at different times. The motors, hands, sensors and batteries can all function as designed while the machine still cannot perform enough valuable tasks reliably enough to justify its cost.

Musk’s remarks therefore reveal both a problem and a possible advantage. Tesla needs to bridge the gap between making robots and making useful workers, but the process of crossing that gap could itself create a competitive lead. The company is trying to start two reinforcing cycles. One makes the robots more capable as the deployed fleet learns. The other makes future robots cheaper as manufacturing volume increases. If Tesla can get both cycles working together, each could accelerate the other.

The first cycle makes the fleet more useful

Tesla’s experience with cars helps explain the difference. Customers bought Tesla vehicles because they were already useful as cars. Those customers financed the physical fleet, paid to operate it and created opportunities for Tesla to collect experience that could improve its automated driving systems. Tesla still had to fund the software, computing and engineering, but it did not have to own millions of vehicles simply to create the environment in which its driving technology could learn.

Optimus begins from a harder position. If a robot cannot yet perform useful work, someone still has to pay to manufacture it, maintain it, supervise it and give it tasks to practise. That makes the first stage expensive. But the economics begin to change as soon as the robot learns something valuable enough to justify deployment.

Suppose Optimus becomes reliable at moving a particular type of material inside a Tesla factory. Those robots can continue doing that job while other machines work on the next task. The resulting experience can improve the shared software used across the fleet. If a new capability transfers successfully, every compatible robot does not need to learn it independently from the beginning.

This creates a reinforcing process. More deployed robots can produce more useful real-world experience. That experience can improve the shared system. Better capability allows the robots to perform a wider range of tasks, making more deployments worthwhile. Those additional deployments then create still more opportunities to learn.

Fleet size matters because learning can happen in parallel. Ten robots encountering problems will normally discover less in a week than thousands operating across different situations. A larger fleet costs more to build and support, but it may shorten the time required to reach useful performance. That matters because development has substantial time-based costs. Engineers continue drawing salaries, facilities remain open, capital stays tied up and commercial revenue is delayed. Faster learning can therefore reduce part of the economic cost of reaching the next useful capability.

The second cycle makes the robots cheaper

At the same time, a separate process should operate on the physical machines. Manufactured products usually become cheaper as cumulative production increases. This relationship is often called Wright’s Law: the more of something an industry makes, the more opportunities it gets to learn how to make the next unit more efficiently.

The effect comes from many small improvements rather than one breakthrough. Factories become better organised. Designs are simplified. Suppliers improve their own processes. Assembly becomes more automated. Engineers remove unnecessary components, materials and production steps. Humanoid robotics is still too young to know how quickly its costs will decline, but the same manufacturing forces should apply.

Lower robot costs matter because a technically possible task is not necessarily an economically sensible one. A robot might be perfectly capable of moving goods around a small warehouse but still cost too much to justify replacing a human worker. Reduce the cost of the machine far enough and the same capability suddenly becomes commercially useful.

That creates the second reinforcing cycle. More robots produced create more manufacturing experience. More experience can reduce the cost of future robots. Lower costs make deployment worthwhile for a wider range of tasks, increasing demand and production again.

The two cycles become much more powerful when they connect.

A more capable robot can enter jobs that were previously impossible. A cheaper robot can enter jobs that were previously uneconomic. Both changes increase deployment. More deployment creates more real-world experience for the learning system and more production volume for the manufacturing system.

The market can therefore expand from two directions at once. Software increases what the robot can do, while manufacturing reduces how much economic value a task has to create before buying the robot makes sense.

Why Tesla may be unusually well placed to start the cycles

Tesla already owns many of the pieces required to make this work. Its vehicle business has built an artificial intelligence operation around machines interpreting and acting in the physical world. The company has experience with computer vision, prediction, planning, model training and sending software improvements back to deployed machines. Manipulating objects with robotic hands introduces problems that driving does not, but Tesla is not beginning with a conventional industrial robotics team learning artificial intelligence from scratch.

It also has substantial computing capacity for training AI models and, just as importantly, places to deploy immature machines. Tesla controls factories, warehouses, service operations and other environments containing repetitive physical work. That allows it to choose relatively simple early tasks, modify the workplace when necessary and tolerate failures that an external customer might reject.

Internal deployment also creates a possible bridge between research and commercial use. Optimus does not need to become a general-purpose worker before creating value. It only needs to perform one task where the benefit of the work exceeds the cost of operating and supporting the machine. Once that happens, useful robots can contribute economically while the wider system continues learning additional tasks.

Tesla also understands mass manufacturing and has the financial resources to support the expensive period before the cycle becomes self-funding. Those capabilities matter because the first rotation is likely to be the hardest. Early robots may create much more information than economic output. The economics improve only when useful work begins contributing enough value to support further deployment and development.

But Tesla is not alone

The important competitive question is therefore no longer simply who can build a convincing humanoid. Several companies can already do that. The harder question is who can connect learning, manufacturing and deployment strongly enough for the two cycles to reinforce each other.

Xpeng has many of the same ingredients as Tesla: an electric-vehicle business, autonomous-driving AI, manufacturing capability and internal locations where its Iron humanoid can initially be deployed. Hyundai and Boston Dynamics have assembled another version of the system, combining Boston Dynamics’ robotics expertise with Hyundai’s manufacturing base and real factories, alongside an artificial intelligence partnership with Google DeepMind. Figure is attempting to build the same structure through partnerships, using a growing robot fleet to produce real-world experience while obtaining manufacturing environments and computing capacity from other companies.

Tesla’s possible advantage is not that nobody else understands the process. It is that more of the required pieces sit inside one organisation, potentially reducing the friction between robot design, AI development, factory deployment and manufacturing.

That advantage is not guaranteed to last. A partnership can outperform an integrated company if its individual components are better. Shared AI models could also reduce the value of proprietary learning, allowing competitors to catch up without reproducing the entire development programme.

The real race starts after the robot is built

There are many ways the process could stall. More robots do not automatically produce better learning. Thousands repeating easy tasks may generate less useful information than a smaller fleet encountering difficult new situations. Skills may not transfer easily between jobs. Hardware may continue changing so quickly that production learning is repeatedly interrupted. A robot might also become technically useful while still requiring so much supervision and maintenance that its work does not save money.

Competition creates another limit. Even if both cycles work perfectly, Tesla may not capture all the economic benefit. If Xpeng, Figure, Hyundai and Chinese manufacturers follow similar cost and capability curves, falling robot costs could be passed through to customers rather than becoming permanently higher margins for producers.

But Musk’s comments reveal why the next stage of humanoid robotics could look very different from the first. Manufacturing a convincing robot is only the beginning. The larger prize comes from creating a system in which every additional machine helps make the fleet more capable and every increase in production helps make the next machine cheaper.

If that process starts working, the slow and expensive beginning of humanoid robotics could give way to much faster expansion.

The real race is not to build the robot first. It is to get both cycles spinning before everyone else does.

AI companies are turning their subscriptions into competitors for software and services that people currently buy separately. OpenAI has added health and personal finance services to its ChatGPT assistant, while Meta promises more business services. Their ambition is to make the assistant the place customers go to get things done. A specialist business, selling a particular service separately, could find itself competing against something already included in its customer’s subscription.

This is how AI could “Amazonise” services: bring more purchases into one customer relationship. Some functions would be supplied directly; others would come from outside businesses. An AI company could gain considerable control over customer spending without owning every supplier involved.

Why the bundle keeps growing

Falling costs give AI companies a reason to expand. Their models, the software providing their assistants’ capabilities, generate income through subscriptions and usage charges. Customers need not keep paying for the newest model once cheaper alternatives can do their work. Stanford’s AI Index recorded a decline of more than 280-fold between November 2022 and October 2024 in the price of using models roughly matching early ChatGPT’s performance on standard tests. It measures the price of using AI at that standard. Models that other companies can download and run widen the choice further. Additional services give customers another reason to keep subscribing.

Expansion also becomes cheaper when each function can share an existing business. A specialist invoicing application must earn enough to support the whole company behind it. An AI provider starts with the necessary models and a customer relationship. Promoting an additional function does not require finding those customers again. Existing billing and administrative systems can support the new function, although development and support still cost money. The addition only needs to justify its extra expenses through the broader subscription.

Specialist applications have traditionally provided valuable guidance through their design. Their menus and forms translate expert knowledge into choices ordinary customers can make. A common assistant could take over much of that guidance across different activities. Customers could describe an outcome and answer questions without learning another application.

Consider someone asking for a restaurant at a particular date and time. With permission, the assistant could match that request to a calendar appointment for dinner with Alex. Earlier emails reveal that Alex is vegan and uses a wheelchair. The assistant confirms whether this is the dinner being planned, then checks menus and accessibility information. Photographs on Google Maps could help, with confirmation from the restaurant where access remains unclear. The customer never had to supply those requirements, and a restaurant application might never have received the information needed to discover them.

Which businesses become features?

A service is especially exposed when the AI company can reproduce its result using information it can readily obtain. Customers buying an ordinary translation may care little about which software produced it, provided the result meets their needs. Klarna reported cutting spending on external marketing suppliers by 25% while increasing campaign activity using several AI tools. The work continued while external suppliers received less money.

Businesses built around another company’s AI face this threat from their own supplier. The supplier already provides the central capability; it can offer the finished service directly once it can reproduce the surrounding instructions and checks.

The replacement only has to meet the customer’s required standard. A phone camera can satisfy everyday needs while remaining inferior to a professional camera. Avoiding another device to buy and carry can outweigh the quality difference. Software customers face a similar choice when a separate subscription offers improvements they rarely need. Professional users may remain loyal while the specialist loses its broader market.

Low additional delivery costs make this pressure stronger. Scheduling can fit into an existing assistant much as a calculator fits into a phone. Services requiring extensive individual intervention are harder to include cheaply. Reliability belongs in that calculation: repeated checking and correction can eliminate an apparent saving. Even sophisticated products become exposed when customers see too little benefit in buying them separately.

What remains worth buying?

A specialist has stronger protection when it supplies something the assistant still needs. Exclusive legal information can retain value when a general system prepares documents. Thomson Reuters’ partnership with Anthropic illustrates this: the Claude assistant gains specialist legal sources and citation checking. Similarly, booking a hotel does not reproduce the room. Such suppliers may remain essential while losing direct access to customers because the assistant handles selection.

Size also creates management problems. General Electric argued that splitting its businesses would provide clearer accountability and better investment decisions. Shared models and software could make related digital services easier to coordinate than unrelated factories. Nevertheless, managing exceptions and taking responsibility still require resources. A focused supplier retains an advantage where those costs outweigh the savings from combining services.

Regulation imposes a separate constraint. Microsoft’s European competition commitments require cheaper packages without its Teams communication application and improved access for competitors. Such measures can preserve alternative suppliers without banning broad subscriptions. Genuine savings and convenience would still leave specialists needing a reason for customers to buy separately.

Where the customers could move

These economics favor companies whose existing assets make additional services useful immediately. Google could combine travel emails in Gmail with places researched on Maps to improve bookings. Access depends on permissions and which connections are technically available. Its advertising business also reveals where suppliers pay to attract customers, suggesting attractive services to add. Advertising spending is a clue to valuable demand.

Meta has a related opportunity through knowledge of people’s interests and relationships. Its Muse assistant could help with purchases suited to particular occasions, while the Meta One subscription offers business services to sellers. That knowledge cannot simply be pooled without limits: Meta excludes private Muse conversations from its advertising systems.

Microsoft’s advantage lies inside organizations, where documents and communications provide material for services it could add to existing business subscriptions. Separate reporting and meeting applications face an included alternative. Microsoft can buy models from competing developers while retaining that customer relationship.

OpenAI and Anthropic are building customer relationships through work performed inside their assistants. Information gathered during one project can help with the next, making broader subscriptions more useful. Their dependence on outside information and services also leaves room for competing packages. Another company could buy their intelligence and use it to build the broader customer relationship.

xAI’s Grok assistant and the Cursor coding service share an included allowance for AI use. Outside models cost extra, encouraging customers to stay within the package when its AI is adequate. xAI also plans the SpaceX engineering data providing a route to providing bundled engineering services.

An independent assistant could likewise combine inexpensive services while carrying the customer’s preferences between suppliers. Claude already offers experimental memory transfers, although these do not move every permission or ongoing task. Competition between broader providers could also turn displaced spending into savings for customers, leaving the companies with more activity without proportionately more revenue.

Product announcements show services entering these subscriptions; evidence of customers canceling separate purchases would establish the commercial impact. The deciding question is how much value a separate provider adds after the assistant can understand the request and complete enough of the work.

For an exposed business, its product could become one more reason to renew somebody else’s subscription.

For the full report with sources: https://abundanceeconomicsresearch.substack.com/p/will-the-ai-companies-amazonise-the

AI and robotics are likely eventually to perform almost every task better and more cheaply than humans. This raises a question more fundamental than which jobs will survive: when machines can do nearly all the work, what will make a human valuable?

The question goes beyond what work remains for people. It asks which human qualities will remain valuable, why they will matter, and what we should develop in ourselves and our children. The answer has implications for how we live, how societies function and how we prepare for the future.

The answer lies in three places: what remains scarce when production becomes cheaper, the need for societies to make decisions together, and experiences that matter precisely because another human is involved.

Together, they suggest a different purpose for education and personal development.

When abundance meets a limit

Suppose robots make building beautiful homes inexpensive. More people can afford one, but the coastline has not grown. As housing gets cheaper, competition for the best sites may intensify because more buyers can enter.

Abundance does not make every resource unlimited. The scarce thing may be land, seats at popular events or access to a limited opportunity. As most goods and services become cheaper, what remains scarce can become more valuable.

AI could be exceptionally good at measuring competing claims. It could predict the consequences of different choices and apply any agreed rules. But what rules should we agree on? Should a waterfront site go to whoever pays most, a family needing housing or people who have lived nearby for generations? A better calculation cannot settle what a community considers fair.

This creates a place for judgment, wisdom and integrity. Judgment weighs competing interests. Wisdom considers wider consequences, including harms that are difficult to undo. Integrity gives people reason to believe their interests will be treated fairly, even when they dislike the outcome.

Why intelligence cannot settle every disagreement

Scarcity is only one reason people need to agree. Consider the balance between privacy and security. AI might accurately predict how much crime a surveillance system could prevent and what personal information it would expose. But it cannot establish how much privacy a community ought to sacrifice. Different people can understand the same evidence and still want different outcomes.

Similar disagreements arise between personal freedom and the interests of a wider community. Cultures and individuals place different weights on these values. Technology changes what is possible; it does not automatically create agreement about what is desirable. As intelligence becomes cheaper, the ability to generate analysis may matter less than the ability to bring people to decisions they can live with.

That requires trust. A leader may have excellent evidence but still fail if people doubt their motives or feel ignored. Trust grows through conduct: listening, acting fairly, admitting mistakes and accepting responsibility. Such qualities can make someone worth following even when everyone has access to equally good advice.

Societies also need people prepared to challenge existing agreements. Courage allows someone to act when success appears unlikely and the personal risks are real. Resilience keeps that effort alive through resistance and setbacks, giving others time and confidence to join.

Imagine a cause that appears hopeless. Many privately support it, but each assumes resistance is futile. One person takes a public stand despite personal risk. Others discover they are not alone and begin to join. What first seemed impossible becomes more achievable because that initial act changes the chances of success.

Courage matters because another human has chosen to face real consequences. Resilience can keep an effort alive through setbacks, although it also requires knowing when to change course. AI might offer better advice about the risks and chances of success, but the willingness of a person to act can inspire others to act too.

When the person is part of the experience

Some things are valuable not because a human solves a problem, but because a particular human is involved. A robot might one day serve a tennis ball faster and more accurately than any player. Yet a human final could remain compelling. Spectators care about the competitors, their histories, their nerves and their possibility of failure. They are watching people confront one another, not measuring the speed of a ball.

Live music works similarly. Recordings have long offered excellent performances at little cost, yet people pay to see particular musicians. Their choices, character and response to the audience are part of the experience. A flawless imitation does not automatically replace an encounter with the person the audience came to see.

Even personal support can follow this pattern. In experiments on empathy, people rated responses from AI more highly but still preferred receiving empathy from another human. The quality of the words was not the only thing that mattered. Who offered them mattered too.

A chef’s individual taste, a performer’s expression or a care worker’s attentiveness may therefore carry value beyond technical skill. Being human alone is not enough, however. Poor service does not become attractive simply because a person provided it, and preferences may change as AI improves.

Cheaper technology could make venues, travel and supporting services more affordable, bringing distinctive human experiences to more people. Yet a musician can perform only so many concerts, and a person can offer only so much individual attention. As other services become abundant, encounters with particular people may become relatively more valuable.

What should we teach the next generation?

For generations, much of education has prepared people to perform useful work. Qualifications demonstrated knowledge and skills that employers would pay for. If machines eventually perform nearly every task better and more cheaply, that advantage becomes less secure.

Knowledge remains important. Good judgment requires understanding, and musicians still need to master their instruments. But education must develop more than the ability to produce correct answers. It should help people become worth trusting, cooperating with, following and seeking out.

These qualities require practice, not just instruction. Children develop judgment when they make choices affecting others and explain them. They learn integrity when fairness costs them something, courage when they speak up despite discomfort, and resilience when they work through setbacks. Sport, music, drama and shared responsibilities also develop expression and character.

AI can help people learn, but completing a task is not the same as developing the person. A student who submits an excellent AI-written argument may have learned little about weighing evidence. One who must explain a decision, answer objections and accept its consequences is developing something different.

The coming of abundance means that a person’s value will no longer be defined by what they can do, but by the qualities they possess and the value those qualities bring to their fellow humans.

For a more complete version of this article with sources: https://abundanceeconomicsresearch.substack.com/p/when-ai-can-do-everything-better

For most of human history, scientific progress has been limited by the amount of human attention available.

A few decades ago, a scientist searching the literature could spend hours in a library scanning journal titles before deciding which papers justified the time needed to read them. Electronic databases and the internet transformed that process by making titles, keywords and abstracts searchable. Finding papers became much easier.

But the deeper constraint remained. A human still had to read the paper, understand it, decide whether its findings were reliable and work out whether any of it was relevant to the problem at hand.

AI changes that constraint.

The important shift is not simply that AI can search faster. It is that large amounts of scientific attention can now be created in software. That attention can be used to discover research, assess it, connect findings across different fields, generate hypotheses and, increasingly, run experiments.

This matters even if AI never produces anything resembling an Einstein.

Science has always thrown information away

Scientific knowledge has long grown faster than any individual could absorb it.

That forced researchers to filter aggressively. A paper with an unpromising title might never be read, even if one result buried inside it would have been useful. A finding from chemistry might help solve a problem in materials science, but only if someone working in one field happened to encounter the work in the other.

Electronic search reduced the cost of finding papers, but it did not remove the attention bottleneck. Search engines could return thousands of results. Someone still had to choose which few were worth reading.

AI changes the economics of that choice because it can inspect far more material before deciding what matters.

OpenScholar, for example, searches a collection of 45 million open-access scientific papers. Systems such as Kosmos have combined literature search, data analysis and hypothesis generation, with its developers reporting that a single research run could involve reading around 1,500 papers.

Current systems are imperfect. They can misunderstand findings and miss relevant work. But they already show how the scale of scientific attention can expand without requiring the same increase in human researchers.

The result is that more of the existing scientific record can become usable.

The biggest gains may come from connections

A large part of scientific discovery comes from connecting things that were previously separate.

Human researchers specialize partly because specialization is useful, but also because attention is limited. No individual can continuously read deeply across medicine, chemistry, physics, biology and engineering.

AI does not face the same limit in the same way.

A machine can compare findings across thousands of papers and fields, including papers that never cite each other. A measurement made for one purpose may turn out to answer a question somewhere else. A method developed in one discipline may solve a problem in another.

This is important because science may already contain more useful knowledge than we currently know how to use.

A 2019 materials-science study showed that machine learning applied to earlier research could identify promising materials before later scientists investigated those applications. The software was not acting as a full scientist, but the result demonstrated something important: useful relationships can already exist inside the scientific record without researchers having noticed them.

AI can increase the chance that those relationships are found.

Scientific knowledge can also become less static

The same machine attention can be used to assess old research continuously.

Traditional peer review concentrates attention around the moment a paper is published. After that, new studies may support, weaken or contradict its findings, but there is no universal system continually updating the standing of every claim.

Human attention is too scarce for that.

AI makes a different model possible. An old finding could be repeatedly checked against new evidence. A successful replication, where another team repeats the work and gets the same result, could strengthen it. A failed replication or newly discovered mistake could weaken it.

That would make the scientific record more dynamic.

Existing tools already point in this direction. Scite analyzes how later papers cite earlier work, including whether they support or contradict it. Other AI systems have been used to find mistakes in published calculations, formulas, figures and tables.

These tools are not capable of deciding whether every scientific claim is true. They do show that some of the work of checking and reassessing published research can be automated.

The quantity of usable scientific knowledge could therefore rise at the same time as its quality becomes easier to monitor.

AI can also produce new evidence

The next step is more significant.

AI does not have to stop at reading and analysis. Some forms of research can be carried out entirely in software. Mathematical proofs, software testing, simulations and analysis of existing datasets can all be expanded by adding computing capacity.

Other fields require physical experiments, but even there human attention may not remain the main constraint.

Chemistry already provides examples of AI systems connected to laboratory equipment that can choose experiments, run reactions, measure the results and decide what conditions to test next.

The same pattern is appearing elsewhere.

In protein engineering, researchers have connected AI systems to robotic production and testing. In mechanical engineering, automated systems have designed structures, printed them, tested them and selected the next experiment. Similar systems are being developed for batteries, solar cells, cell biology, microbiology, plant science, microscopy and quantum physics.

The important distinction is between a robot that follows instructions and a system that can use the result of one experiment to choose the next one.

Once that loop is automated, laboratories become capable of producing scientific attention as well as carrying out physical work.

The remaining bottlenecks move

Not every part of science can be accelerated without limit.

Some experiments depend on biological growth, long periods of observation or events that simply take time. A tissue experiment may be fully automated but still require weeks for cells to develop. More machines can run more experiments in parallel, but they cannot make every biological process happen instantly.

Missing evidence remains another hard limit. AI cannot recover the result of an experiment that was never performed or data that was never recorded.

Reliability also matters. A machine that reads a thousand papers and misunderstands one important result can spread an error very efficiently. Abundant attention therefore makes checking more important, not less.

There is also a risk that large numbers of AI systems could reproduce the same assumptions. If they are trained on similar data and use similar methods, apparent agreement might simply reflect the same blind spots repeated at scale.

These constraints matter, but they do not restore the old system. They shift the bottlenecks away from the number of hours available to human scientists.

What humans do when scientific attention becomes abundant

As AI takes over more of the technical and operational work of science, the human contribution shifts rather than disappears.

Researchers can spend less time searching literature, checking routine evidence, running standard analyses and repeating familiar experiments. More of their attention can move toward deciding what the scientific system should try to achieve.

That includes judging which questions are useful to people, identifying problems that matter even when the existing literature is thin, and setting goals that emerge from human needs rather than from what has already been studied.

Humans also provide the boundaries. Some experiments may be scientifically useful while still being too dangerous or socially unacceptable to pursue. The ability to perform an experiment does not answer whether it should be performed.

The remaining human role is valuable precisely because it is human. Science does not exist simply to generate more knowledge. It exists within human priorities, risks and values.

AI can make scientific attention abundant. People still decide what that attention is for.

For a detailed report including sources: https://abundanceeconomicsresearch.substack.com/p/ai-will-revolutionize-science-without

Subscription businesses have spent the past two decades making revenue more predictable. Instead of selling a product once and persuading the customer to buy again, they persuade the customer to keep paying every month or every year. That creates an important advantage: once the subscription has started, continuing to pay usually requires no decision at all. Stopping does.

Research published in the American Economic Review shows how valuable that inertia can be. Across ten large subscription services, monthly retention fell four times faster when customers were forced to make an active renewal decision. The researchers estimated that consumer inertia increased seller revenue by 87% to 120% on average, depending on the model used.

AI agents could change this advantage by giving continuous attention to every recurring payment, checking whether the service is still being used, whether the price has risen and whether a better alternative exists. Recurring billing starts to become recurring justification.

The subscriptions most exposed are those we stop thinking about

Some subscriptions pay for products customers use constantly and consciously. A Spotify user who listens every day does not need an agent to explain why they are paying. Others, including gyms, dating apps, creative software and some media products, are useful intermittently, so the customer’s need can fade before the payment stops.

A third group is more exposed. PDF tools, résumé services, travel apps, photo utilities or mod-download services may solve a problem that lasts minutes, days or weeks while billing continues for months.

RevenueCat’s 2026 analysis of more than 115,000 subscription apps and over $16 billion of revenue shows how different these businesses already are. Productivity apps had a median first annual renewal rate of only 23%, while business applications were substantially stickier. AI apps also retained monthly subscribers worse after 12 months, at 6.1% compared with 9.5% for non-AI apps. The key distinction is whether the customer’s need itself genuinely recurs.

Agents make that easier to see by combining the recurring charge with renewal terms, actual use and available alternatives. Instead of requiring the customer to assemble the information, the agent can reduce it to one decision: keep, downgrade, switch, renegotiate or cancel.

Losing customers can make the remaining customers more expensive

Subscription businesses often benefit from customers who continue paying while using very little of what they bought.

Gyms provide an easy example. Research on US health-club members found that people on monthly contracts attended only 4.3 times a month on average. They were effectively paying more than $17 per visit when a $10 pay-per-visit option was available. If an agent repeatedly compares the membership fee with actual attendance, those low-use members should be more likely to question whether the subscription still makes sense.

The members who remain are more likely to be the people who actually use the gym, so revenue can fall much faster than visits. The gym still needs enough equipment, floor space and staff to serve them, worsening the economics from both directions.

The same effect can appear in software. A company might buy 1,000 licences even though only 600 employees use the product seriously. If an agent cuts the contract to 650 seats, the supplier loses 35% of its paid seats but may lose very little actual usage. For software running on the customer’s computer, that is mainly a revenue problem. For cloud services and AI tools, where use creates real compute costs, it can become a margin problem as well.

Price increases create another value test

A company facing that squeeze may raise prices, but AI agents make that harder too. A customer might ignore a service increasing from $10 to $13 a month. An agent can immediately compare the higher price with recent usage and competing products.

That creates a feedback loop. Low-use customers leave, average usage among the remaining customers rises, margins come under pressure, and the supplier raises prices. Those increases trigger another round of comparison, cancellation and switching. A company could therefore maintain headline revenue while the economics underneath it deteriorate.

Some products may stop making sense as subscriptions

This eventually raises a bigger question: should some products be subscriptions at all?

Software companies have spent years moving from packaged products toward subscriptions because recurring revenue is more attractive than repeatedly asking customers to buy a new version. Agents could push some products back in the opposite direction. If a desktop utility costs almost nothing to provide after installation and customers only need it occasionally, charging $60 for a permanent licence may become more attractive than repeatedly asking an agent to approve another $10 monthly payment.

Cloud products face a different problem because heavy usage has a real cost. They may move toward credits, metered consumption or subscriptions combined with usage charges. NVIDIA’s GeForce NOW already limits Performance and Ultimate customers to 100 hours of premium playtime each month, with additional hours available separately. The company says the limit helps maintain quality, speed and shorter queue times.

Gaming illustrates the difference clearly. A $70 game running mainly on the customer’s console creates relatively little continuing computing cost for the publisher. A cloud gaming service must keep paying for processors, bandwidth and infrastructure every time the customer plays. If low-use subscribers disappear while power users remain, outright purchases, expansions, sequels and usage-based pricing become relatively more attractive.

Businesses face the same mechanism at a much larger scale

The same logic applies to corporate spending, where hundreds or thousands of subscriptions can accumulate. xAI gave its Grok procurement agent access to contracts, spending and usage information. It found 43 paid software seats with no activity for 90 days and identified more than $100,000 of direct savings across the spending it reviewed.

This turns occasional procurement reviews into a continuous process. An agent can check unused seats, overlapping products, alternative suppliers, renewal prices and whether the company is still buying what employees actually use. The relationship between employee numbers and software revenue consequently becomes less secure.

Gartner calls a broader version of this “agentic arbitrage” and estimates that up to $234 billion of enterprise application spending could be exposed by 2030, around 20% of projected enterprise application software-as-a-service spending. That does not mean $234 billion disappears. It means a very large pool of recurring expenditure becomes easier to question, optimise and renegotiate.

The subscription economy is larger than it looks

US consumers report spending roughly $1,300 to $1,900 a year on subscriptions and recurring services, depending on the survey and definition used. US software-as-a-service revenue alone was estimated at more than $160 billion in 2025. The affected spending extends beyond obvious subscriptions. An agent may not cancel electricity or car insurance, but it can compare tariffs, shop an insurance renewal or find a cheaper mobile plan.

That does not mean subscriptions disappear. Services providing strong continuing value should survive repeated scrutiny, while bundles may become more valuable because they give customers several reasons to keep paying.

What becomes harder to defend is recurring revenue supported by inattention, underuse and the effort required to reconsider a purchase. The subscription economy was built around recurring billing. AI agents may turn it into an economy of recurring justification.

For a more detailed version of this article with sources: https://abundanceeconomicsresearch.substack.com/p/ai-agents-could-rewrite-the-subscription

Banks perform a useful trick. They let customers withdraw money from checking and savings accounts at short notice, while using much of that funding to support mortgages and business loans that may not be repaid for years.

The system works because customers do not all ask for their money back at the same time. Most balances are remarkably stable. They move slowly enough that banks can treat a large part of their deposits as dependable funding, even though customers are legally free to withdraw them.

AI agents could make that assumption much less reliable.

The obvious concern is a faster bank run. Apollo economist Torsten Slok recently pointed to an average annual interest rate of 0.1% on US checking-account balances while selected alternatives advertised yields of 3.3% to 5.0%. An AI agent that continually searched for a better return could move money far faster than a human customer normally would.

But a bank run may not be the most important effect. The bigger change could be that banks have to start competing much harder for money they currently receive cheaply and keep for long periods.

The customer may stay while the money leaves

People do not choose bank accounts on interest rates alone.

A checking account may provide payments, cash access, customer service, deposit insurance and the comfort of dealing with a familiar institution. For some customers, those benefits justify earning less interest.

That does not mean the customer must keep all their money there.

An AI agent could understand that someone wants to keep $5,000 in a familiar bank for bills, cash withdrawals and everyday payments, but does not need another $50,000 sitting in the same account earning very little. The agent could leave the service relationship intact while moving the surplus money somewhere that pays more.

This is important because the bank could keep the customer but lose much of the funding that made the customer profitable.

Large banks benefit from bundling many services together. Yet several parts of that bundle can already be obtained elsewhere. Smaller insured banks can provide the same standard US deposit-insurance protection within the relevant limits. Shared ATM networks can provide cash access without every bank building its own network. Customer service is increasingly automated.

As AI gets cheaper, agents could make it easier to compare the whole package rather than simply compare interest rates. They could ask: How much cash access does this customer actually need? Do they care about branches? What fees would moving money trigger? Is the alternative account insured? How much money should remain instantly available?

Automatic cash management is not new. Wealthfront introduced Autopilot in 2020, allowing customers to move excess cash according to rules they had already set. The difference with a more capable agent is that it could potentially choose between providers as well as execute transfers. The open question is whether that increase in capability is large enough to change consumer behaviour at scale.

There is already evidence that greater attention changes bank pricing. Research from the Bank for International Settlements finds that US digital banks tend to pay higher deposit rates and adjust them more strongly when central-bank rates change.

Agents could make many more customers behave like highly attentive customers.

Banks would have to start paying for stability

That creates two separate problems for banks.

First, deposits become more expensive. A bank may have to raise its savings rates to stop money leaving.

Second, even after matching a competitor’s rate, the bank has no guarantee that the money will stay. If another bank improves its offer tomorrow, the customer’s agent can move again.

For banks that need reliable funding over longer periods, fixed-term deposits therefore become more valuable. Instead of relying on customers simply leaving money alone, the bank pays them for a contractual commitment to keep it there.

Interestingly, agents could help on this side too. An agent might calculate that a household will not need $20,000 for the next year and place it into a fixed-term account. So the same technology that makes instant-access deposits less stable could increase the supply of deposits committed for longer periods.

The important change is that banks would increasingly have to pay explicitly for stability rather than receive part of it through customer inertia.

There is also a more extreme risk. Silicon Valley Bank lost $42 billion in a single business day during its 2023 crisis. That was a confidence-driven run, not an AI event, but it showed how quickly deposits can already move.

If millions of agents use similar decision rules, withdrawals could become more synchronized. A small pricing problem could become a liquidity problem if a bank suddenly had to replace large amounts of funding at much higher cost.

Who actually pays when deposit rates rise?

Initially, the bank does.

If a bank raises the interest paid to depositors, its costs rise. It then has four broad choices: charge borrowers more, reduce its operating costs, accept lower profits, or make fewer loans.

In practice, all four could happen at once.

This is why higher returns for savers do not automatically mean equally higher mortgage and business-loan rates.

Banks spend a great deal of money operating branches, call centres, compliance teams, administrative processes and large technology systems. If AI reduces those costs, banks could pay depositors more while spending less elsewhere.

Borrowers may also become more price-sensitive. Their agents could compare loan offers, prepare applications and identify when refinancing makes financial sense. That would make it harder for a bank simply to pass every increase in its funding costs to borrowers.

So the extra return paid to savers could come from several places. Borrowers may pay somewhat more. Bank profits may fall. Operating costs may fall. Some loans that no longer make economic sense may disappear.

The result depends heavily on competition.

A lean bank might pay depositors more, charge borrowers no more than today, and remain profitable because its operating costs are much lower. A bank carrying thousands of employees, expensive branches and legacy systems would have a much harder time doing the same.

This does not create specialist finance. It changes the economics of it

Specialist lenders and specialist funding providers already exist at scale.

US finance companies owned or managed about $1.95 trillion of loans and leases in July 2026. In the UK, Hampshire Trust Bank already takes deposits while financing independent lending companies. Nonbank mortgage companies commonly borrow from banks, originate mortgages and then sell those loans to investors.

AI agents would not invent this structure.

What they could change is its cost and competitive reach.

A specialist lender could automate more of its administration, reach customers more cheaply and operate profitably in smaller niches. At the same time, customers’ agents could make those specialists easier to find.

But the large banks would not simply disappear. Specialist lenders often depend on banks for funding. If bank funding becomes more expensive, some of that cost flows through to them too. Large banks also retain advantages in capital, regulation, liquidity and access to emergency funding.

So the likely result is not a clean transfer of business from large banks to thousands of tiny replacements.

It is more likely to be a squeeze on the margins throughout the system.

Savers may capture more of the available return. Borrowers may gain from better comparison. Banks may retain less income between the two. Some of the value may move to specialist lenders, funding providers or the agent platforms themselves.

Fractional-reserve banking probably survives, but in a different form

AI agents do not automatically end fractional-reserve banking.

Banks can respond by paying more for deposits, attracting more fixed-term money, holding more liquid assets and reducing their own operating costs.

The deeper change would occur only if long-term lending became much less dependent on deposits that customers can withdraw immediately and much more dependent on money that savers and investors have committed for longer periods.

That would be more than a change in which company provides the mortgage or savings account.

It would change the type of money supporting the lending.

The first test is much simpler. Once reliable agents can move money easily between banks, do large banks continue to hold huge amounts of cheap, stable deposits?

If they do, the existing model is more durable than this thesis suggests.

If they do not, one of the oldest economic advantages in banking begins to disappear.

Amazon’s decision to block Meta’s Muse shopping agent looks, at first, like a technical dispute over access, security and authorization. In practice, it points to a much larger commercial problem. If customers increasingly hand purchasing decisions to AI agents, retailers may still process the sale while losing control over how the customer chooses what to buy.

That matters because modern retail is built around much more than checkout. Amazon, Walmart and Target do not simply stock products and deliver them. They also control discovery, comparison, advertising, recommendations and the customer’s habit of starting the next purchase in the same place. A capable shopping agent begins to separate those functions.

The customer no longer has to browse a retailer’s site to compare televisions, check whether a product is compatible, read reviews or decide whether a slightly more expensive option is worth it. The agent can do that work across many retailers at once. By the time it arrives at Amazon, the purchasing decision may already have been made.

That creates an obvious conflict. Amazon reported $68.6 billion in advertising-services revenue in 2025. Not all of that depends directly on people browsing shopping pages, but the scale shows why control of the purchase journey matters. Amazon could still get the order while losing some of the surrounding economics.

Why Amazon can resist when others cannot

Amazon is in a stronger position than most retailers because excluding it can make an agent less useful. Its combination of product range, Prime benefits, delivery speed and customer familiarity is difficult to replace completely. If an agent cannot access Amazon, some customers may simply abandon the agent and shop directly.

Many smaller retailers do not have that option. If a mid-sized general retailer blocks an agent, the agent may simply move to another supplier offering a similar product. This creates a collective-action problem. Retailers may dislike the idea of handing customer access to AI platforms, but individually many will still want the orders those platforms can send them.

The likely outcome is therefore not a simple divide between retailers that accept agents and those that do not. It is more likely to be a continuing negotiation over access, customer data, liability, payments, returns and commercial terms. The more distinctive the retailer or its products, the more leverage it has.

AI makes the seller smaller too

The same technology that helps customers shop also reduces the cost of running a retail business. Product descriptions, routine customer support, order administration, stock questions and parts of merchandising can increasingly be handled by software rather than people. That lowers the level of sales a retailer needs to support its organization.

At the same time, shopping agents can reduce the cost of being discovered. Those two changes reinforce each other.

Consider a small retailer specializing in replacement parts for a narrow range of older appliances. Today, such a business may struggle because customers do not know it exists and because staff have to answer endless compatibility questions. A shopping agent could identify the exact appliance, find the correct part, compare suppliers and decide whether the specialist is credible. The retailer’s own AI systems could then handle most routine questions and process the transaction.

The important point is that the specialist does not need to become another Amazon. It only needs to be good enough at one small part of Amazon’s market.

If that happens across thousands of categories, competition starts to look very different. Instead of one giant challenger trying to reproduce the whole Amazon model, many small businesses can attack different parts of it.

This does not necessarily make small retailers more profitable. Lower costs bring more entrants, and better comparison makes price competition harder to avoid. AI can reduce the minimum size of a viable retailer while also reducing the margin each retailer can retain.

Manufacturers get a new route to the customer

The same logic applies to manufacturers. Direct-to-consumer selling is not new, but historically it has required a company to build a retail function around the product. It needs marketing, support, payments, fulfillment, returns and a way to persuade customers to visit its website in the first place.

AI reduces some of those costs, while shopping agents reduce another: discovery.

A manufacturer no longer has to persuade the customer to start on its website. It only has to make its products easy for the agent to understand, compare and buy. That could make direct sales viable at lower volumes than before.

For small manufacturers, that opens a clear opportunity. For large manufacturers, the calculation is more difficult because they depend on powerful retail partners. Procter & Gamble, for example, reported that Walmart and its affiliates represented about 16% of total sales in fiscal 2026. A supplier with that level of dependence cannot simply bypass the retailer without considering the consequences.

But it does not have to choose between direct sales and retail sales in an all-or-nothing way. It can sell different versions, bundles, sizes or colors through different channels. Retailers already use the same strategy.

Exclusivity becomes more valuable, but harder to control

Large retailers often carry slightly different versions of the same product. Televisions are a familiar example. Two retailers may sell essentially the same Samsung television under different model numbers, with small variations in speakers, warranty or accessories.

This can make direct comparison harder and can prevent formal price matching when the retailer requires an identical model number. A sufficiently capable agent weakens that protection because it does not need the product codes to match. It can compare the products on what they actually do.

That makes superficial differentiation less useful, but genuine exclusivity more valuable.

If one retailer has a product the customer genuinely wants and cannot buy elsewhere, an AI agent cannot compare that advantage away. The same applies to useful bundles, better warranties, installation or service. Retailers therefore have a stronger incentive to secure real exclusivity rather than rely on customer confusion.

Manufacturers, however, gain more reasons to resist granting it. If direct sales become cheaper and easier to generate, giving one retailer exclusive access to a product carries a larger opportunity cost.

The balance of power between retailer and manufacturer can therefore move in both directions at once. Retailers want more exclusivity because comparison is becoming easier. Manufacturers may demand more compensation for providing it because their alternatives are improving.

Winning the product may no longer mean winning the customer

Retailers have traditionally benefited from getting the shopper through the door, whether physical or digital. A customer who arrives for one product may buy several others. An exclusive item can therefore be worth more than the margin on that item alone.

Shopping agents can weaken that effect.

An agent can buy one exclusive item from Walmart, another product directly from the manufacturer and a third from a specialist retailer. The customer does not have to manage three separate shopping experiences because the agent handles the coordination.

Physical costs still matter. Three suppliers can mean three deliveries, higher shipping costs and more complicated returns. Large retailers retain real advantages where consolidated delivery and service matter. But the connection between winning one product and winning the entire shopping basket becomes weaker.

That is a significant change because it reduces the value of controlling the shopping session itself.

Amazon can lose control without losing the business

This is why the likely outcome is not simply that agents replace Amazon. Amazon owns warehouses, delivery systems, returns infrastructure and customer trust. Those assets remain valuable even if another company controls product discovery.

Amazon already provides fulfillment for sales that originate elsewhere. In a more fragmented retail system, it could continue to earn money from logistics even when the customer’s AI agent decides what to buy.

The economics, however, are different. Owning the customer relationship, influencing product choice and selling advertising around that decision may be more valuable than simply moving the parcel.

The long-term contest is therefore over which parts of retail still need to sit inside the same company.

AI shopping agents could make the supply side of retail more fragmented while making customer access more concentrated. There may be more specialist sellers, more direct manufacturer channels and more small retailers serving narrow niches, even as a few large AI platforms become important gateways to those businesses.

That concentration is not inevitable. AI also lowers the cost of building competing agents, and customers may use several. The more important question is which functions continue to require scale when the work of buying and selling becomes much cheaper.

Amazon may remain extremely powerful. Walmart and other large retailers may use exclusivity, service and delivery to defend their position. Manufacturers may keep more sales for themselves. Small specialists may become viable in markets that were previously too narrow.

Retail is unlikely to disappear. But the activities that today sit together inside a retailer may no longer need to stay there.

For the book: https://amazon.com/dp/B0HFKSP2WF

For research reports: https://abundanceeconomicsresearch.substack.com

More from Abundance Economics: https://abundance-economics.com

Google May Already Be Showing Us Its Route to AGI

In 2015, Demis Hassabis looked back at a decision he had made more than a decade earlier. By 2004 he was already thinking about what would eventually become DeepMind, but concluded that the technology was not ready. As he later put it, “we still didn’t have enough of the components.” Deep learning had not yet taken off, computing power was inadequate and many of the systems needed to build something approaching general intelligence simply did not exist.

By 2026, Hassabis was describing a fuller set of ingredients: Gemini as a world model, AlphaGo-style search and planning, specialist tools, memory, reasoning and systems capable of acting in the world. The possibility is that Google has not recently changed its theory of artificial general intelligence. It may finally have enough of the pieces to execute a theory DeepMind has pursued for years.

That matters beyond the contest among AI labs. Falling AI costs make individual tasks cheaper to automate, but cheap task execution does not necessarily replace a worker. Jobs are bundles of tasks. Planning, memory, simulation, tool use and action let an AI system move across that bundle. If Google can combine them reliably, the shift from AI as a tool to AI as a worker could happen before anyone agrees that AGI has arrived.

DeepMind was never really building a chatbot

DeepMind’s early history looks very different from modern large language models. Its first major breakthrough was not language generation but an agent learning how to act. In 2013, DeepMind showed that a neural network could look at raw pixels from Atari games, decide what action to take, observe the result and improve from the reward. Hassabis later described the system as moving from “pixels to action”.

Planning became another strand of the same programme. AlphaGo combined neural networks with search, examining possible future positions before deciding what to do. In 2016 Hassabis said one purpose was to see how neural networks could be “meshed with planning approaches”. A year later DeepMind was experimenting with imagination-augmented agents that simulated possible futures before acting.

DeepMind was also working on memory and continual learning. The projects differed technically, but together they reveal a consistent view of intelligence: a capable system would need to perceive, remember, plan, imagine, act and learn over time. What these systems lacked was generality. AlphaGo could defeat the world’s best Go players while being useless at almost everything else.

Then the LLM arrived

Modern large language models changed that part of the problem. Google Brain’s Transformer architecture, published in 2017, became the basis for systems trained across enormous quantities of language, code, images and other information. As they scaled, these models developed broad competence across subjects and tasks that earlier DeepMind agents could not approach.

Former DeepMind researcher Misha Laskin has described the distinction neatly: language models provide “generality”, while reinforcement learning provides “capability”. DeepMind had systems that could act effectively but only inside narrow worlds. LLMs were almost the reverse: remarkably broad, but unreliable when required to pursue objectives autonomously across long sequences of actions.

In 2023 Google combined the two research traditions institutionally when it merged Google Brain and DeepMind into Google DeepMind. Gemini appeared soon afterwards. From the beginning, Google described it as multimodal, designed to work with tools and intended eventually to support memory and planning. Hassabis went further, describing Gemini as combining some strengths of AlphaGo-type systems with the language capabilities of large models.

That makes Gemini easier to understand as more than Google’s answer to GPT. The large language model may instead be the general-purpose cognitive core around which DeepMind’s older work on search, planning, memory and action can finally be assembled.

The old architecture is reappearing

Google’s recent projects increasingly resemble components DeepMind once had to investigate separately. Hassabis has said the ambition is a “world model, not just a language model”. Google’s Genie project explores that direction by generating interactive environments in which actions alter subsequent events. SIMA supplies the agent side: Google has combined Gemini with SIMA so an AI can interpret objectives, reason about them and act inside 3D environments.

Deep Think attacks another weakness of conventional language models by exploring several possible reasoning paths before committing to an answer. Robotics extends the same structure into the physical world. Gemini Robotics ER 2 uses Gemini 3.5 Flash as a high-level reasoning system that can interpret a situation, plan multiple steps and coordinate lower-level models that produce physical actions.

That modularity may be crucial. Human workers do not personally contain every capability they use. They understand an objective, choose a tool, interpret its output and decide what to do next. An artificial worker could operate similarly. This also explains Hassabis’s emphasis on specialist tools: the general model need not contain the best protein-folding system or robotics controller. It needs to know when to call one and what to do with the result.

The economic threshold is below AGI

The important economic threshold is much lower than full human-level intelligence. Consider an analyst. The job may require understanding an objective, finding information, deciding what matters, operating software, creating an output, checking it, recognizing mistakes, revising the work and deciding when it is finished. An AI system that can draft one paragraph cheaply automates a task. One that can perform most of that sequence reliably begins to compete with the role.

That is why falling cost and expanding capability matter together. Inference has been getting cheaper. At the same time, search, planning, world models, tools and memory expand the number of adjacent tasks AI can perform. More of the job therefore becomes technically automatable just as the price of automating it falls.

Better reasoning often requires more computation: several candidate solutions, verification and repeated tool calls. If inference remains expensive, those techniques can make autonomous AI uneconomic. As inference prices fall, more computation can be spent making the system reliable while the total cost of completing the work still declines. Different occupations can therefore cross the threshold at different times, before any universal AGI moment.

Google may still fail

None of this means Google has solved the problem. World models can be wrong. Search can amplify errors. Reinforcement learning works best where success can be measured clearly, whereas many real jobs involve ambiguous objectives and human judgement. Long autonomous workflows accumulate errors, and persistent memory can preserve bad information as easily as useful experience.

Google itself is also showing signs of execution difficulty. Gemini 3.5 Pro missed its expected release window. In August Hassabis handed day-to-day leadership of Google DeepMind to Koray Kavukcuoglu and became Alphabet’s Chief Scientist. Google says the change allows Hassabis to concentrate more heavily on long-term AGI research. A less favourable interpretation is that the long-horizon research programme and commercial Gemini roadmap are becoming more separate.

The architecture may never converge in the way the historical pattern suggests. Genie, SIMA, Deep Think, robotics and memory could remain distinct research programmes while Gemini itself improves mainly through better training and more compute. If AI becomes cheaper and better at individual tasks but remains unreliable when asked to complete long workflows without supervision, the thesis will have failed its clearest test.

But the history makes Google’s current direction more credible than recent announcements alone would suggest. DeepMind was pursuing reinforcement learning, memory, planning, simulated futures and action before modern LLMs existed. Those systems lacked broad generality. Transformer-based models supplied much of that missing capability. Google then merged the two research traditions and began building world models, agents, reasoning systems and robotics around Gemini.

That does not prove Google has discovered the route to AGI. It does suggest that its current work is not merely a collection of fashionable additions to a chatbot. Google may be assembling the machine DeepMind wanted to build before the components existed. If it works, the economically important date will probably not be the day Google announces AGI. It will be the quieter point when an AI system becomes cheap, reliable and broadly capable enough to do the whole job.