<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Startup Sapience on Medium]]></title>
        <description><![CDATA[Stories by Startup Sapience on Medium]]></description>
        <link>https://medium.com/@startupsapience?source=rss-75bd4ad3fb37------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*mgOOHNNu07kTqB8AlBG6bw.png</url>
            <title>Stories by Startup Sapience on Medium</title>
            <link>https://medium.com/@startupsapience?source=rss-75bd4ad3fb37------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 18:10:03 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@startupsapience/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Your mirror should pay you]]></title>
            <link>https://startupsapience.medium.com/your-mirror-should-pay-you-dff5325f8602?source=rss-75bd4ad3fb37------2</link>
            <guid isPermaLink="false">https://medium.com/p/dff5325f8602</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[investment]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[advertising]]></category>
            <dc:creator><![CDATA[Startup Sapience]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 15:35:28 GMT</pubDate>
            <atom:updated>2026-10-27T15:35:28.753Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EXUf8QZRpQE7twWk_gm63A.png" /></figure><p>Mirrors have been in our homes for centuries. We use them to shave, get dressed, check our hair before leaving, then walk away. For most of the day, a mirror is a piece of glass at eye level in a place we pass repeatedly.</p><p>A smart mirror can do something with that time. It shows your reflection whenever you want it. When you finish, it can display ads, and the company selling those ads can pay you a share.</p><p>Companies have spent years finding places to put screens in our homes. Here is a screen people would use even if it showed nothing else.</p><p>Take a hallway mirror. You check it on your way out. Later, you pass it while putting down your keys or taking off your coat. A café nearby could advertise a breakfast offer in the morning. A restaurant could show a dinner offer in the evening. A household could choose which sorts of businesses appear, along with the hours the display runs.</p><p>The mirror needs to stay a mirror first. No sound, no video suddenly playing beside your face. If you want the display off for a week, you switch it off for a week.</p><p>Playing ads to an empty house would make a poor business. An advertiser would pay for the short periods when someone is nearby, and perhaps pay more when an offer leads to a sale. A presence sensor could count those periods. A code on the display could let a customer claim an offer without the mirror knowing who they are.</p><p>Those details decide whether the household gets paid anything worthwhile. A mirror in a hallway used by four people has a different audience from one in a spare room. A local business might pay well to reach households a short walk away, while a national brand might want thousands of mirrors across a city. The company operating them would have to sell those placements, cover the hardware and send the household its share.</p><p>One mirror might pay part of a bill or cover the price of the mirror over time. The business becomes much larger when a company can offer advertisers a network of homes. Each household earns a little; the network sells access to a large audience.</p><p>We already have screens in our homes that carry ads. We often paid for the devices ourselves, yet the company controlling the screen collects the ad money. A smart mirror could be sold on different terms. The household provides the wall, the electricity and permission to display the ads. It gets paid.</p><p>That brings up an argument I expect to see if these mirrors spread: who gets the payment in a rented home?</p><p>A landlord could buy and install the mirror. The tenant walks past it every day. The landlord could ask for the advertising income because they paid for the device. The tenant could ask for it because the ads appear in the home they occupy.</p><p>I think the person living in the home should decide whether the ads run and receive the larger share. The landlord can be paid back for installing the device through an agreed fee. Owning the glass should not give someone control over what appears in another person’s hallway.</p><p>People will disagree on the price. Some will welcome a small payment for something they barely notice. Others won’t want a single ad inside their home. That choice should stay with them.</p><p>I can see companies offering this within the next five years: a smart mirror installed cheaply, a clear monthly statement showing what it earned, and controls on the display that anyone in the house can understand. The harder work will be convincing advertisers to pay enough and convincing households the mirror is worth living with.</p><p>If that works, mirrors could be the start. A kitchen panel or a wardrobe door could also display something while idle. At that point, we’ll be discussing who has the right to sell advertising on surfaces inside a home.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dff5325f8602" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Next Workforce May Never Become a Middle Class]]></title>
            <link>https://startupsapience.medium.com/the-next-workforce-may-never-become-a-middle-class-997d83528ef3?source=rss-75bd4ad3fb37------2</link>
            <guid isPermaLink="false">https://medium.com/p/997d83528ef3</guid>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Startup Sapience]]></dc:creator>
            <pubDate>Thu, 10 Sep 2026 20:13:51 GMT</pubDate>
            <atom:updated>2026-09-10T20:13:51.511Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VvIe4hkSSldEC7n9TdG1VA.png" /></figure><p>A country could spend years attracting a new factory, build the roads and electricity connections it needs, celebrate its opening, and eventually discover that relatively few households can afford the products coming out of it. The factory might be profitable. Exports could be rising. The disappointment would sit in the surrounding towns, where people had expected the investment to bring a much larger number of decent jobs.</p><p>I think we need to consider that possibility more seriously as governments compete for automated production. The link between industrial growth and household prosperity has never been guaranteed. But employment has been an important way of connecting them, and some of the machinery being installed could allow output to grow considerably faster than the payroll supporting it.</p><p>I would call this <strong>Synthetic Demography</strong>: the growing separation between a country’s human population and its capacity to perform economically useful work.</p><p>The term describes a change in how we should assess a country’s economic prospects. Population forecasts tell us how many people are likely to enter work or retire. They tell us much less about the work that machines could perform, who will own those machines, and which households will receive the resulting income. As automation expands, those questions deserve a place beside the population figures.</p><p>Machines have increased output per worker for centuries. It would be difficult to explain modern living standards without them. What interests me now is the possibility that more physical tasks and parts of professional work can be reproduced through systems whose expansion depends on capital, engineering and electricity. Their capacity can grow on a different timetable from the population that once would have supplied the work.</p><p>For a country with an aging workforce, that could be welcome. For a country expecting millions of young people to find their first formal jobs, it creates a different calculation. Both may adopt the same technology and experience very different consequences.</p><p>Industrial employment has often allowed households to move beyond subsistence or irregular earnings. The work could be exhausting, and the gains depended heavily on wages, labour rights and public investment. Still, a dependable salary made it possible to plan. A family could pay for education or begin saving towards a home. Local businesses gained customers whose income arrived regularly enough to support spending beyond immediate necessities.</p><p>Over time, that income could become ownership. Savings entered banks and pension schemes. Some employees started businesses. A wider middle class emerged through a process in which people first earned enough from work to accumulate something that could later earn money for them. The sequence was imperfect and uneven, but it gave a large part of the population a route into the gains from industrial growth. Automation can alter the size of that first opening, before a generation has earned enough to acquire assets of its own.</p><p>The International Federation of Robotics reported that China installed 295,000 industrial robots in 2024, accounting for 54% of global installations. Asia accounted for 74%. The worldwide operating stock reached about 4.66 million. Those figures cover industrial robots performing defined tasks. They cannot be translated into an equivalent number of displaced workers, but they show the scale and geographical concentration of investment in automated production.</p><p>A robot still needs integration into a working process. Factories require maintenance, materials and people who can deal with situations the equipment cannot handle. The economic outcome depends on the whole operation. A manufacturer that becomes more competitive might expand enough to employ more people, including in jobs created by the new production methods.</p><p>Yet governments cannot assume that every increase in output will bring a comparable increase in employment. A factory could double production while adding relatively few employees. That might be an excellent result for the business. Whether it produces the local prosperity anticipated when public money was committed is a separate question that needs answering.</p><p>Consider a hypothetical plant owned by an overseas company. It imports much of its machinery and uses software licensed from abroad. Its products are sold internationally. Local employees receive wages, domestic suppliers win some contracts, and the government collects whatever taxes apply. The remaining income may belong largely to investors elsewhere.</p><p>Nothing about that arrangement makes the investment inherently undesirable. Without it, the country might have had less production and fewer jobs. Foreign investment can bring valuable technical knowledge and help domestic suppliers develop. But the benefits need to be traced through the project. The construction budget and expected export revenue cannot tell us how much income will eventually reach local households.</p><p>Gross domestic product measures value added within a country. It includes profits generated there even when those profits accrue to foreign owners. National income adjusts for income flows across borders, although that still leaves open how income is distributed among residents. A government could report strong production figures while ordinary families see only modest improvement in their finances.</p><p>Suppose the plant’s next expansion requires more imported equipment and a relatively small increase in local staffing. The country gains additional output, but the connection to household income remains narrow. Across enough industries, it could become increasingly productive without giving enough people a dependable route into the middle class. The project announcements would still look encouraging. You would have to follow the income for several years to see how much of the promised prosperity had reached the surrounding population.</p><p>I suspect this will also complicate the development strategy built around exporting office work. A business serving overseas clients can earn foreign currency without shipping a physical product. Its wage bill then supports spending at home. If software takes over a growing share of the work, export earnings might continue rising while fewer additional employees are needed to deliver the service.</p><p>Some of the revenue could also go back abroad to pay the companies supplying that software. A country might retain the client relationship and report a growing services industry while capturing less income from each contract. This is worth watching in businesses whose attraction has been access to skilled employees at competitive wages. Their future position may depend increasingly on whether they own useful technology or depend on a foreign supplier that can also sell directly to their customers.</p><p>The evidence does not support treating every exposed job as a disappearing job. The ILO’s 2025 assessment found that roughly 1 in 4 workers worldwide were in occupations with some exposure to generative AI. It judged changes within jobs to be the more likely outcome because most occupations still contain tasks requiring human input.</p><p>That leaves considerable room for people to benefit. A small firm may serve customers it previously lacked the staff to reach. Someone with limited capital could start a business using software that performs work once requiring a much larger team. These possibilities belong in any serious account of the transition. They also make predictions of permanent mass unemployment difficult to defend.</p><p>My concern is how widely those opportunities will spread, and how long it will take. A profession can remain busy while becoming harder to enter. A successful company can increase salaries for its existing staff while hiring fewer graduates than it would have hired before. Employment may eventually adjust elsewhere, but households still have bills during the intervening years.</p><p>There is also a question about customers. A machine performing paid work does not independently develop the spending needs of a household. It has operating costs, and the people who own it can spend its earnings. The proceeds nevertheless reach the economy through a different route from a salary distributed among many employees.</p><p>That does not mean demand must collapse. Profits can finance new investment, lower prices can expand consumption, and governments can redistribute income. It does mean that the ownership of productive systems becomes more important to understanding who can afford what is being produced. The additional output alone does not settle the question.</p><p>Cheap software would be useful to a household whose earnings were under pressure. It would not necessarily compensate for higher rent or the cost of childcare. Different parts of the household budget respond to productivity improvements at different speeds. The price of a digital service could fall sharply while the cost of living near a good school remains difficult to meet.</p><p>Aging makes the ownership question more urgent. A country with fewer workers will usually find it harder to support a growing retired population. Automation can help by increasing the output available from the remaining workforce. But the ability to produce more and the ability of retirees to pay for it depend on arrangements that population ratios do not describe. The same automated factory could help finance somebody’s retirement thousands of kilometres away while doing relatively little for pensioners living nearby.</p><p>A retired person with a pension that owns shares in profitable businesses has some claim on the income those businesses generate. If automation improves their earnings, part of the benefit can reach the pension, subject to investment costs, market prices and the fund’s obligations. A retiree relying on a poorly funded public system occupies a different position, even if highly productive factories operate nearby.</p><p>Norway offers an existing example of a country establishing broad public claims on income generated abroad. Its Government Pension Fund Global invests overseas, and withdrawals help finance the government budget. It is not an individual pension account for each citizen. The relevant point is that national spending can be supported partly by ownership of assets beyond the domestic workforce.</p><p>Norway accumulated that position with oil revenue and decades of investment. Countries cannot recreate it simply by announcing a fund. Nor can every country rely on being a net recipient of investment income from elsewhere. Still, the example helps explain why an aging population with substantial shared assets may face a more manageable adjustment than its population profile alone suggests.</p><p>Financial claims also have physical limits. A pension portfolio cannot ensure that a nurse is available in a town where nobody wants to work. Some services remain dependent on local skills and human attention. Automation may make goods cheaper and free people for other work, but access to care will continue to require planning beyond the purchase of securities.</p><p>The ownership problem becomes harder for households arriving late. People with established savings can participate in productivity gains through assets they already hold. Younger workers generally need earnings before they can build that position. If stable employment becomes harder to obtain, the route into ownership narrows for the people being told they need to become investors.</p><p>Employer pension contributions illustrate the difficulty. They connect work today with an ownership claim on future production. If fewer people obtain jobs offering those contributions, the machinery of retirement saving reaches fewer households. Advising everyone to buy shares does little for someone who cannot consistently cover rent. The ability to own capital still depends, for much of the population, on receiving enough income from labour first.</p><p>Governments face a related issue. Public finances collect substantial revenue through employment, while the income generated by automation may arrive as business profits or payments to suppliers in another jurisdiction. The IMF has warned that a declining labour share could weaken tax bases and has questioned incentives that favour rapid worker replacement. It has also acknowledged that AI could ultimately increase employment and wages.</p><p>The outcome will depend on where income arises and how effectively it can be taxed. A country could gain enough new business activity to improve its finances. Another could provide substantial public support for investment and later discover that the taxable benefits are smaller than expected. The cost of supporting households during the transition would remain a domestic responsibility.</p><p>I would therefore expect negotiations over major industrial projects to become more concerned with participation in the income they generate. When a government provides valuable land or finances infrastructure, the public contribution has an economic value. There is a case for examining whether part of the return should reach residents through shared ownership or other enforceable claims, alongside wages and taxes.</p><p>Public ownership brings risks of its own. A badly governed fund can lose money or direct investment towards political favourites. Requiring a state stake in every project could deter useful investment. Employee ownership can also leave workers exposed to the same company for both their earnings and savings. Any attempt to broaden participation needs diversification, competent management and a clear account of who bears losses.</p><p>The purpose would be to give more households access to productive returns as the economy changes. It would not remove the need for good jobs. Education would remain essential, including the ability to use new tools and move between kinds of work. But education policy cannot carry the entire burden of deciding how productivity gains are distributed.</p><p>At the company level, employing fewer people to deliver the same output may represent progress. At the national level, we need to follow what happens next. Do lower costs create enough additional demand? Can domestic suppliers develop profitable capabilities? Are households accumulating assets, or using debt to maintain spending while income falls behind? Those outcomes will determine whether the transition strengthens the middle class.</p><p>I would want demographic analysis to examine household ownership of productive assets alongside the age structure of the population. Industrial policy should track the wages and domestic income generated after construction has finished. Investment announcements would be more useful if they explained how local participation could grow through later expansions, when the first round of hiring is already complete.</p><p>We may eventually see countries with similar population profiles taking sharply different economic paths. In one, automation could support higher wages and widely held savings. In another, output could grow while much of the income accrues to a small group or leaves the country. The machines would not explain the difference by themselves. Ownership and the institutions distributing their returns would account for much of it.</p><p>The next workforce may never become a middle class because a growing part of the work could be performed by systems that do not become households. The people around those systems can still become more prosperous. Achieving that will require dependable ways for them to participate in the income, including through work that becomes more valuable as technology improves.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=997d83528ef3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Higher for Eternity]]></title>
            <link>https://startupsapience.medium.com/higher-for-eternity-09e281e25bbc?source=rss-75bd4ad3fb37------2</link>
            <guid isPermaLink="false">https://medium.com/p/09e281e25bbc</guid>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[stock-market]]></category>
            <dc:creator><![CDATA[Startup Sapience]]></dc:creator>
            <pubDate>Mon, 07 Sep 2026 18:01:00 GMT</pubDate>
            <atom:updated>2026-09-07T18:01:00.971Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h4F3nJwxlcQ5OY6BBMu8Hw.png" /></figure><p>The phrase “higher for longer” has always seemed to contain a promise. Borrowing is expensive, but eventually we will get through this and resume whatever we put on hold. I understand the appeal, especially for someone postponing a house purchase while trying to make the monthly payments work on a spreadsheet. I am less convinced that patience will deliver the rate they have in mind.</p><p>Suppose a company borrowed $100 million at 2%. When its debt approaches maturity, the available refinancing rate is 6%. Management waits, conditions improve and the company eventually secures 4%. The finance director has saved $2 million a year against the earlier offer. The annual interest bill has also doubled from $2 million to $4 million. If sales have barely grown, the planned expansion may be the first expense the directors postpone.</p><p>I suspect that experience will be familiar across much of the economy. We can receive encouraging news about interest rates while discovering that our own finances have become harder to manage. The rate in the announcement is being compared with the previous announcement. The rate on the new loan has to be compared with the one it replaces.</p><p>“Higher for Eternity” is how I would describe a borrowing environment that remains expensive relative to the years on which many financial plans were built. Rates would still rise and fall. There could be recessions severe enough to produce substantial cuts. My concern is that the conditions people are waiting for may remain unavailable through the years when they need to buy a home or build a business.</p><p>We need to be careful about which rate we are discussing. The central bank sets the price of very short borrowing. A lender offering money for 5 or 20 years considers what short rates might average over that period, along with compensation for uncertainty. Lending to a company adds the possibility that it will fail to repay. Mortgage pricing also reflects lenders’ funding, capital and operating costs, and competition for customers.</p><p>A cut can therefore help without producing the mortgage or business loan somebody had hoped for. If the economy weakens, expectations for policy rates may fall while lenders become more concerned about borrowers’ ability to pay. The borrower experiences the combined result. Being correct about the next central bank decision is only part of the calculation.</p><p>The delay in refinancing is already visible. In its May 2026 Financial Stability Review, the European Central Bank reported that rates on outstanding floating loans had declined since its first cut in 2024, while average rates on outstanding fixed loans had continued rising. Existing fixed contracts do not change whenever the central bank meets. The average changes as old lending is repaid or replaced and new loans enter the stock.</p><p>That leaves households and companies living through different stages of the same interest rate cycle. A borrower whose expensive loan adjusts immediately might already be receiving relief. Someone protected by an old fixed deal could be approaching their first substantial increase. Aggregate figures can conceal how far apart those experiences are.</p><p>For many people, the period after the financial crisis lasted long enough to become their understanding of normality. Someone beginning a career in 2010 could spend years learning how property, business valuations and debt worked under unusually low interest rates. Decisions made under those conditions were reinforced by the next deal that worked. Eventually the assumption of cheap refinancing required very little discussion.</p><p>Low rates had economic causes. Research by Bank of England economists examined how increased demand for saving, weaker investment and slower growth helped explain the decline in global interest rates over several decades. Central bank purchases added support to bond markets, but the earlier environment cannot be explained entirely by decisions taken inside central banks. Some of those underlying forces may persist.</p><p>The question is what now sits alongside them. Governments face substantial spending commitments, while private businesses are financing projects that require years of construction before earning a return. An investor deciding whether to lend has to consider competing uses for the money. There is no obligation to supply it at the price a borrower remembers paying.</p><p>A government trying to expand domestic defence production may accept financing costs that would persuade a private business to abandon a factory. The political commitment survives the higher quote. Similar choices arise when a country needs to replace ageing electricity infrastructure. Some investment can be deferred, but doing so has consequences beyond the financial return on the project. I would expect this to make parts of the demand for capital less responsive to its price. Financing can expand through bank lending and foreign investment; whether it stays cheap also depends on how quickly the economy can supply the capacity being financed.</p><p>The OECD projected that governments and companies would borrow $29 trillion from bond markets in 2026, 17% above 2024. That includes refinancing, so it should not be read as $29 trillion of additional spending. It does indicate the volume of funding that has to be arranged.</p><p>The buyers have been changing too. In the government bond markets covered by the OECD’s analysis, domestic central bank holdings fell from 30% of the total in 2021 to 20% in 2025. Investors more sensitive to price have taken a greater role as issuance has remained elevated. A pension fund or household buying bonds has different objectives from a central bank implementing monetary policy.</p><p>Higher borrowing does not automatically produce higher yields. Saving can increase, capital can arrive from abroad and investment plans can be cancelled. Nevertheless, it seems unwise to assume that the demand for financing can expand indefinitely while the terms remain as generous as before. Some adjustment has to take place, whether through the price of money or through the projects that receive it.</p><p>This is where my earlier idea of <strong>Computational Crowding Out</strong> fits. The construction of AI infrastructure brings another substantial claim on financing, alongside the physical capacity needed to deliver electricity and remove heat. A developer trying to fund housing may be competing for investors’ money with projects backed by much larger companies.</p><p>There is also no simple conclusion that better technology must bring interest rates down. Lower production costs can reduce inflationary pressure. Profitable new investment opportunities can increase the demand for capital at the same time. The timing of those effects will influence who benefits, especially for a borrower whose refinancing date arrives well before the expected productivity gains.</p><p>Public finances make the timing problem larger. The IMF’s April 2026 Fiscal Monitor put global public debt at just under 94% of GDP in 2025 and projected 100% by 2029. It identified spending pressures and rising interest burdens among the strains on budgets. These are projections, with outcomes dependent on growth and policy, but they hardly support treating the demand for government financing as a temporary inconvenience.</p><p>A government that issued debt cheaply can take years to feel the full cost of replacing it. During that period, even falling market yields may leave the average interest rate on its debt moving upwards. The budget then has to accommodate additional interest through taxation, reduced spending elsewhere or more borrowing. A household may encounter the consequence through a public service long before making any connection with the bond market.</p><p>I would not assume that this becomes an endless spiral. Governments can change their budgets, and stronger growth can improve their finances. But the political willingness to make those changes forms part of the risk a lender is taking. The BIS has warned that doubts about fiscal sustainability can raise the extra return investors demand, complicating the effect of monetary policy on borrowing conditions.</p><p>The household version is clear. On an illustrative $400,000 repayment mortgage with 25 years remaining, a 2% rate gives a monthly payment of about $1,695. At 5%, the payment is about $2,338, excluding fees. That is roughly $7,715 a year that must come from income which already has other demands on it. The calculation holds the balance and remaining term constant to isolate the effect of the rate.</p><p>The same calculation also tells us something about buying power. At 5%, a $1,695 monthly budget supports a loan of about $290,000 over 25 years. A buyer would need roughly $110,000 more equity to purchase the same property with the same payment, assuming the price was unchanged. A relative able to provide that equity can alter the buyer’s circumstances more than a modest improvement in salary.</p><p>Property prices can fall, of course, and rising incomes can restore some affordability. Yet the adjustment has to work through existing owners, lenders and a supply of homes that cannot change quickly. An owner who can afford to stay may withdraw a property from sale. A buyer may remain in unsuitable housing. The market can spend a long time adjusting through decisions that never become transactions.</p><p>There is evidence of the uneven burden in Britain. The Bank of England’s July 2026 Financial Stability Report projected that nearly 750,000 households paying less than 3% would leave fixed deals during 2026, with an average monthly repayment increase of £170. It also judged household finances resilient in aggregate and expected smaller increases for many other borrowers. The pressure is concentrated in individual circumstances, including when the existing deal was signed.</p><p>I find the effect on opportunity more troubling per se. A person with little debt has greater freedom to take a lower salary while changing career. Someone whose mortgage has just absorbed most of their spare income may have to remain in a job they intended to leave. A few additional years of expensive borrowing can influence which people are able to start businesses, move cities or accept an uncertain opportunity.</p><p>Those consequences can outlast the rate itself. If borrowing eventually becomes cheap after someone has abandoned a business plan or spent years delaying a move, the later improvement does not automatically restore what was possible earlier.</p><p>Businesses face their own version of this constraint. A company can remain profitable and still struggle to fund the debt it took on at an earlier valuation. A buyer calculating what to pay for that business has to allow for current financing costs. Owners who expect the previous valuation to return after a few cuts may spend years postponing a sale, while the business continues to need investment.</p><p>The return available to shareholders also has to be examined after the refinancing. A company reporting stronger operating profits may send much of the improvement to its lenders. I would want to know when its debt matures, the cost of replacing it and how much cash remains after necessary investment. An earnings forecast built around the old interest bill could give a flattering picture of what shareholders will eventually receive.</p><p>Some companies would benefit from an environment like this. A business able to finance its expansion from cash generation has options that a heavily indebted competitor lacks. It can pursue a project while others wait for approval from lenders. It may also acquire assets from owners who need to reduce debt. Growth would continue, although access to it could become more dependent on the balance sheet a company brings into the period.</p><p>Savers can benefit too, provided the return they keep after tax compensates adequately for inflation. Higher yields can improve the income available from bonds bought at those yields, even as existing bonds suffer when rates rise. The effects reach both sides of household finances. Someone with cash and limited borrowing may experience a period of higher rates quite differently from a younger household trying to acquire the same assets.</p><p>There are reasons this argument could prove too pessimistic. A prolonged downturn could weaken investment demand and bring very low rates back. Strong demand for saving could continue to support bonds. Technological progress could improve public finances as well as household incomes. I would be cautious about attaching a single direction to ageing populations, since longer retirements encourage saving while pension and healthcare commitments increase demands on government budgets.</p><p>Central banks could also return to substantial bond purchases. Their ability to influence borrowing costs has not disappeared. The difficulty would be sustaining that support if it conflicted with inflation control or damaged confidence in the currency. Low nominal yields imposed under those conditions might leave savers with poor returns after inflation. The distribution of the cost would change; the underlying claims on future income would still have to be met.</p><p>For me, the more plausible concern is an extended period in which relief arrives in portions that disappoint the people waiting for it. Market rates fall, but refinancing remains expensive against an old contract. Inflation eases, but the accumulated increase in living costs leaves less income available for debt payments. A borrower gets a better offer than last year and still cannot make the original plan work.</p><p>That would gradually change the assumptions beneath buying and investment decisions. A property would need to be affordable at the rate available when it is bought. An acquisition would need to earn enough to support its financing without depending on an early refinancing. People would still take risks, but the expectation that cheaper debt will eventually repair a stretched budget would deserve much closer examination.</p><p>How much of the future we expect depends on a return to the financing conditions of the 2010s? I would still welcome lower rates, and there may be good reasons for them to fall substantially. But for someone who has already spent 4 years waiting to buy a home, another 4 years changes a considerable part of their life.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=09e281e25bbc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Last Bank Account: Balance Sheet Hypervisor]]></title>
            <link>https://startupsapience.medium.com/the-last-bank-account-balance-sheet-hypervisor-bc9338e618a3?source=rss-75bd4ad3fb37------2</link>
            <guid isPermaLink="false">https://medium.com/p/bc9338e618a3</guid>
            <category><![CDATA[banking]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[investing]]></category>
            <dc:creator><![CDATA[Startup Sapience]]></dc:creator>
            <pubDate>Sun, 30 Aug 2026 12:24:07 GMT</pubDate>
            <atom:updated>2026-08-30T12:24:07.717Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X7sXZvfbv7EsHKKHq04vKQ.png" /></figure><p>Every financial company that promised to kill the banks eventually needed one.</p><p>The early neobanks rented licences and placed customer deposits with regulated partners. Crypto exchanges needed bank accounts so customers could move between national currencies and digital assets. Stablecoin issuers built a new form of dollar, then backed much of it with Treasury bills, cash and deposits held inside the system they were expected to replace. Some challengers later obtained licences of their own and became banks in the legal sense. The revolution kept returning to the institution it was supposed to make obsolete.</p><p>They kept coming back because what the customer experiences as banking is an app, a card and a balance on a screen. Underneath sits a licensed balance sheet connected to deposit insurance, payment systems, courts, regulators and central bank money. The institution can turn deposits into loans, absorb losses with its own capital and provide somebody for the government to hold responsible when the numbers go wrong. An attractive interface still depends on all of that machinery.</p><p>I think whoever finally weakens the banks will accept that much of the old plumbing has to remain. Too many parts of the economy depend on it. The opening lies in taking control of the valves through a layer I call the <strong>Balance Sheet Hypervisor</strong>.</p><p>In computing, a hypervisor separates software from the physical machine underneath. Several virtual systems can run on the same hardware, and a workload can move from one machine to another without the user rebuilding everything. The hardware remains expensive and necessary, although the software above it can treat each machine as a replaceable source of capacity.</p><p>A Balance Sheet Hypervisor would apply the same idea to finance. A person or company would keep one financial identity, one permanent address and one set of instructions. Several regulated institutions could supply the deposits, credit, settlement and investment products underneath. The customer would see one financial life even though no single bank controlled all of it.</p><p>Suppose your salary arrives on the final working day of the month. Today it lands in the account named on your employment records. Your mortgage, subscriptions, card and household bills are attached to that account. Moving the relationship requires a formal switch, and even when the process works well, most people need a reason strong enough to bother.</p><p>Under a hypervisor, the salary would arrive at an address that belongs to you. The first $2,000 could remain available for bills. The next $8,000 of savings might be placed with whichever insured institutions currently need deposits and are willing to pay for them. Money beyond your chosen cash reserve could move into Treasury bills or a money market fund. If you travelled, payments could be made from the cheapest available currency balance. None of this would require your employer, landlord or electricity supplier to learn a new account number.</p><p>Your mortgage could sit inside the same system. Lenders would be able to bid for it when the fixed period ended, using a verified record of income, expenditure, assets and repayments. The paperwork would already be present. If another lender offered a meaningful saving after fees, the system could prepare the transfer and ask for permission. For a business, the process could extend to invoices, working capital, foreign exchange and the movement of spare cash between countries.</p><p>The customer would no longer switch banks. The banks would switch beneath the customer.</p><p>Britain has already proved that some of the administrative friction can be removed. The Current Account Switch Service moves a current account within 7 working days and redirects payments sent to the old account for at least 3 years. It processed more than 1 million switches in 2025. That is useful infrastructure, although the customer still leaves one bank and joins another. The account number changes and one institution remains at the centre of the relationship. A Balance Sheet Hypervisor would have no main bank to leave. The permanent address sits above all of them.</p><p>Open banking has gone further. By December 2025, the UK had reached 16.5 million active open banking connections, 36% more than a year earlier. Customers can allow outside services to read accounts and initiate certain payments. Yet most of these permissions still treat the bank account as the fixed object around which everything else is arranged. They let another company look into the account or reach inside it. They do not make the underlying bank interchangeable.</p><p>There is a more obscure precedent in deposit networks. In the US, services already divide large balances among participating banks so each portion can qualify for deposit insurance, while the customer deals through a single relationship. A deposit shown as one total can already be assembled from claims on several banks. The process is usually designed around large cash holders and participating institutions. A hypervisor would extend the principle into the rest of a person’s finances.</p><p>Digital money adds the ability to move these claims continuously. The Bank of England now describes a possible retail system in which traditional deposits, tokenised bank deposits, stablecoins and a digital pound operate together. The BIS has been working with 7 central banks and 43 regulated financial institutions through Project Agora, which explores using tokenised commercial bank deposits and central bank reserves for cross border settlement. These projects are trying to modernise money without discarding the legal protections and credit creation of banking.</p><p>Seen side by side, these developments expose what is still missing. Data has become portable. Payments can be redirected and cash can be divided among institutions. New forms of money can move outside normal banking hours. The customer’s identity and legal relationship remain tied to providers that each want the person to enter their system, repeat parts of the same verification and accept the products they have chosen to sell. Banks have little reason to remove that dependence because their economics are helped by people who stay.</p><p>The FCA found that 9 of Britain’s largest savings providers passed through only 28% of the increase in the Bank of England’s base rate to easy access accounts between January 2022 and May 2023. Fixed and notice accounts received more. Smaller providers generally offered better rates because they had to compete harder for funding. Later FCA work found that the largest banks had the lowest funding costs, helped by their current account bases and the tendency of customers to leave cash where it was. Easy access accounts produced the highest savings margins despite allowing withdrawal at any time, partly because banks treated much of the money as behaviourally stable over several years.</p><p>A true hypervisor would auction that asset continually. Deposits would go to institutions offering the best combination of yield, insurance, access and financial strength. Credit would go to lenders willing to accept the lowest return for a given risk. Payments would travel through whichever regulated route offered the right balance of speed, cost and recourse. The customer could set rules against constant movement, exclude certain banks or require a proportion of money to remain instantly available.</p><p>I suspect this would change banks without making them disappear. They would become <strong>Balance Sheet Foundries</strong>.</p><p>A semiconductor foundry provides the factories, specialised processes and capital equipment needed to manufacture a chip. It does not have to own the final relationship with the person buying the phone. A Balance Sheet Foundry would provide regulated financial capacity. One bank might become unusually good at funding mortgages. Another could specialise in small business credit, foreign exchange or insured cash. Community banks could compete for deposits beyond their local branch network. Large banks would still benefit from scale, risk systems and access to markets, but their brands would carry less power if the hypervisor could replace them without disturbing the customer.</p><p>Banking would begin to resemble a wholesale market. Banks currently compete for whole relationships and then try to sell several products to the same person. A foundry competes for each piece of the balance sheet. It may hold somebody’s savings without providing their card, fund the mortgage without seeing their salary, or process a payment without owning the account shown in the app.</p><p>The company controlling the hypervisor would occupy the valuable position. It would know the customer’s cash flows, obligations, preferences and risk limits. It would decide which provider receives each piece of business. The largest financial company of that period might not have the largest balance sheet. It may own the address and the permission to route what passes through it.</p><p>The obvious candidates are not limited to banks. A payroll company sees income before it reaches a current account. Accounting software sees the working life of a small business, including invoices, taxes and unpaid bills. A brokerage already holds investments and spare cash. A large wallet controls payments and identity. An insurer understands household risks that a current account cannot see. Any of them could begin assembling the relationship from the edge.</p><p>There is also a dangerous version of this future, and we have already seen an early warning.</p><p>Synapse operated as a middle layer connecting financial apps to regulated banks in the US. When it collapsed in 2024, customers lost access to funds while banks and fintech companies struggled to agree on the balances recorded in different ledgers. The banks themselves had not failed, so ordinary deposit insurance did not simply repay everybody. The FDIC responded by proposing stronger record keeping requirements for custodial accounts, including daily reconciliation and the ability to identify the beneficial owner and balance of each customer.</p><p>Synapse leaves a fairly brutal lesson for anyone trying to build a Balance Sheet Hypervisor. A clever interface with vague responsibility for the ledger is dangerous. The system needs an authoritative record showing where every unit of money sits and which institution owes it. Customers must be able to see the legal owner, the insurance position and the route taken by each transfer. If the control layer fails, another operator or regulator should be able to reconstruct everything without waiting for several companies to argue over incompatible records.</p><p>I would assume the operator needs its own form of licence. Calling it a technology company becomes absurd once it starts moving deposits, selecting lenders and deciding how much liquidity a household should retain. It would be acting as a fiduciary even if the assets remained at other institutions. Somebody has to carry legal responsibility when its instructions cause a loss.</p><p>Identity is just as awkward. Banks cannot simply outsource every judgement about who a customer is. Money laundering rules require institutions to understand their customers and monitor transactions. A reusable financial passport would need verifiable credentials that can be checked and updated, with each receiving institution able to satisfy its own obligations. The consumer should not have to upload a passport 12 times, although the existence of 1 earlier check cannot absolve everybody else forever.</p><p>Credit will be harder to make portable than deposits. Cash is a claim whose value is supposed to remain stable. A loan depends on underwriting, collateral, servicing and the possibility that the borrower stops paying. Moving a mortgage every week would create costs and legal problems. The more plausible model is continuous observation combined with occasional refinancing when the saving exceeds a threshold. Standardised credit records and documents could make lenders compete without turning every household debt into a day trading instrument.</p><p>The system becomes dangerous once people trust it enough to automate movement. Bank deposits are valuable partly because they are sticky. If millions of hypervisors move cash whenever another institution offers 0.10% more or shows a small increase in risk, funding could leave a bank within minutes. A rumour that once developed into a queue outside a branch could become an instruction executed by software before breakfast.</p><p>Regulators may place limits on how quickly the system can move, particularly during stress. Deposits might be divided into an instantly portable layer and a committed layer carrying a higher rate. Banks could issue notice deposits that cannot disappear automatically. The hypervisor would then optimise within the promises the customer had already made instead of pretending all money can be liquid, safe and high yielding at the same time.</p><p>Borders create another problem. Governments are used to supervising banks within their territory. A hypervisor serving customers in several countries could route savings towards foreign institutions, currencies or government bills whenever domestic returns became unattractive. That would improve individual choice while weakening the captive deposit base supporting local lending. Smaller countries could find that household money leaves through allocation settings instead of a visible bank transfer.</p><p>Perhaps the eventual system remains national at the legal level and international at the allocation level. Each country could require a locally regulated node responsible for identity, tax, consumer protection and emergency controls. The underlying assets could still include claims issued elsewhere. This would allow financial life to travel without pretending that law has become borderless.</p><p>Investors should follow the profit pool. Deposit spreads should narrow if cash is automatically sent towards better rates. Payment fees face pressure when routes compete transaction by transaction. Mortgage distributors gain influence because they control access to borrowers. Banks with strong underwriting, cheap operations or specialised knowledge could do well as foundries even after losing the main customer interface. Banks relying on sleepy deposits and habitual cross selling would have more to lose.</p><p>The hypervisor itself could earn a subscription, a disclosed share of the savings it produces or routing fees paid by providers. The way it gets paid could corrupt the model. If a bank can secretly buy preferential placement, the system becomes another sales channel. I would prefer a fiduciary arrangement in which the customer pays and the operator is required to show why each allocation was chosen. A service trusted with an entire financial life needs incentives that point towards the person whose money it controls.</p><p>I do not expect banks to volunteer for this future. An institution will happily use software to improve its own deposit gathering, underwriting and payments. It is less likely to build a machine that moves profitable customers to a competitor. The Balance Sheet Hypervisor probably has to begin outside the large banks, gain permission to see several accounts and gradually become the place from which the customer manages them.</p><p>At first, it may only sweep idle cash. Then it can move subscriptions and bills onto a permanent address. Credit offers follow once the income history becomes reliable. Investments, insurance and tax records can be added later. The old account remains underneath during the transition, doing the regulated work while losing a little more of the relationship each year.</p><p>That is how I think banks could finally be displaced. There may be no dramatic moment when the branches close and deposits flee into a new currency. The banks continue operating, their licences remain valuable and their balance sheets still carry credit into the economy. What changes is our attachment to them.</p><p>The last bank account will be opened when a person can keep the same financial address for life while the institutions underneath compete to serve it. After that, asking which bank someone uses may sound as dated as asking which physical server holds their email. The banks may still be there, working somewhere below the surface.</p><p>We simply will not belong to them anymore.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bc9338e618a3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Jobcasting: The Expertise Harvest]]></title>
            <link>https://startupsapience.medium.com/jobcasting-the-expertise-harvest-df5e4b8c92c4?source=rss-75bd4ad3fb37------2</link>
            <guid isPermaLink="false">https://medium.com/p/df5e4b8c92c4</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[jobs]]></category>
            <dc:creator><![CDATA[Startup Sapience]]></dc:creator>
            <pubDate>Sat, 29 Aug 2026 13:08:36 GMT</pubDate>
            <atom:updated>2026-08-29T13:08:36.557Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TeVqyGaSGl4ntLHqftWpsA.png" /></figure><p>Most career advice has a timing problem. It tells someone aged 17 what an accountant earns today, how many nurses were hired last year or which degree has the highest starting salary. The person then spends 3, 5 or 7 years preparing for a working life that may last until 2070. If an investor valued a 40 year asset from a page of current prices, we would call the analysis useless. Students are expected to stake their working lives on something close to it.</p><p>Job descriptions add very little. They describe what an employer needs done now, often in language copied from the advert used for the last person. By the time the successful applicant has learned the work, parts of it may be cheap, automated or pushed onto the customer. Other parts may have become harder. The title remains on the office door and gives everyone the impression that the job itself survived unchanged.</p><p>I have been trying to find a word for the forecast we actually need. <strong>Jobcasting</strong> is the best I have come up with.</p><p>A jobcast would follow a profession through time: the tasks inside it, the number of beginners entering, how long it takes before anyone trusts them, the law around the work and whatever is creating demand. Technology belongs in the forecast, obviously, though it can be a surprisingly small piece of the answer. An electrician’s future may depend on grid investment and apprenticeship places. The future of teaching can turn on birth rates, public budgets and class sizes. Insurance losses can change the claims profession before any new software does. I doubt a forecast could tell us the salary to expect in 2036, but it could give us a reasonable picture of the job that will be waiting.</p><p>Consider a call centre. A present day description might mention answering customers, resolving billing problems and meeting a target for call time. A jobcast would notice that the simple calls are increasingly handled before a person picks up the phone. Password resets, delivery dates and balance enquiries go first. The calls left for staff involve bereavement, suspected fraud, a broken payment, an angry customer or a problem that has already defeated the automated system. The company may report that software reduced call volume by 40%. Nobody on the phones experiences a 40% easier day. A greater share of it is spent with people who are frightened, furious or caught in a problem for which the script has no answer. Average stress can rise while total calls fall. Training should improve because the easy conversations once gave new staff time to learn the systems and listen to how experienced colleagues dealt with customers. Whether employers will spend the savings that way is another question.</p><p>Perhaps capable staff become scarce enough for pay to rise, although employers may keep wages low and accept worse service. Regulation could require a person to make certain decisions. Better systems could make the role much more tolerable. That range of outcomes is why lists of jobs that are supposedly safe or doomed have always seemed shallow to me. A profession does not move as 1 solid block. Work leaves from one side and arrives at another, often with a different level of difficulty attached. A school gives another example. If software cuts lesson preparation, the headteacher could give teachers more time with struggling pupils. The finance department could see the same saving as a chance to enlarge classes. Anyone forecasting the teacher’s job needs to know which choice the institution is likely to make. Knowing what the software can do tells you almost nothing about that.</p><p>What concerns me is the person entering a profession after the easier work has gone. They may be asked to exercise judgement before the job has given them much chance to acquire it.</p><p>Ask the best person you know in any profession what they were doing at 23 or 24. The answer will probably sound unimpressive. A senior accountant once reconciled invoices and chased missing receipts. A lawyer read contracts that did not deserve a partner’s time. An engineer checked drawings, sat through site visits and made corrections that felt embarrassingly obvious afterwards. A nurse learned which small change in a patient could be ignored and which one meant trouble. Those tasks produced work for the employer, although that was only part of their value. They were also the means by which the profession made its next accountant, lawyer, engineer or nurse. Beginners see normal cases in large numbers. They make cheap mistakes under supervision and discover that 2 situations which look similar on paper can carry very different risks. Being wrong under supervision never appears in a university prospectus, yet it is a large part of how people learn work in which mistakes are eventually expensive. Nobody remembers every small job that built this knowledge. Years later, the person simply looks experienced, as though judgement arrived with age.</p><p>Employers are finding that software can prepare the first draft, match accounts and search a large set of documents before a new employee has finished opening the files. Removing some junior positions makes sense when the decision is viewed through this quarter’s budget, especially since the senior staff are still there to check the result. Clients may see no change. Output might even improve for several years. The trouble begins when the firm needs another senior person and discovers that fewer people spent the intervening years becoming one.</p><p>I call this the <strong>Expertise Harvest.</strong></p><p>The business uses judgement grown through 10 or 20 years of human practice while cutting away some of the work that would grow the next supply. An orchard can keep producing after someone cuts the planting budget; there is no sudden warning in the next harvest. I suppose that is why the temptation is so strong. The saving belongs to the present manager and the shortage lands in the accounts much later, perhaps under somebody else’s name.</p><p>This can happen without a deliberate plan to eliminate a profession. A manager is asked to lower costs by 8% and discovers that 20 experienced employees using new software can handle work that previously required 20 experienced employees and 8 graduates. Dropping the graduate intake reaches the target, and another firm soon reaches the same conclusion. When both need experienced people 5 years later, each tries to recruit from the other. This works well enough for an individual firm and fails when an industry tries it collectively. There is no outside employer from which the whole profession can recruit. The Expertise Harvest is therefore a supply problem with an unusually long delay. Companies can spend years drawing down a stock of judgement without recording any liability for its replacement. By the time salaries reveal the shortage, rebuilding the pipeline may take a decade.</p><p>There are signs of the squeeze already, although the weak economy deserves some of the blame. A 2026 survey for Lancaster University’s Work Foundation found that 36% of UK employers had reduced entry level opportunities during the previous year. Nearly half of medium and large employers said they had fewer openings. A separate government and LinkedIn review found entry hiring declining in 30 of the 38 occupations it tracked. The falls included accountants at 29%, graphic designers at 28% and software engineers at 27%. The authors were careful to say this did not prove that automation caused the decline. Hiring was weak across the market. Still, the occupations under the greatest pressure are worth watching.</p><p>The better finding came from PwC. Its 2026 jobs study examined 2.4 million entry level roles in the US and found that jobs with high exposure to AI were 7 times more likely than in 2019 to ask for skills previously associated with senior staff, including leadership, team building and judgement. Employers have started asking the beginner to arrive with qualities that used to emerge after years inside the profession.</p><p>I am not sure where employers expect that judgement to come from. A university can teach accounting standards, anatomy, contract law or programming. It cannot reproduce 5 years of being responsible for what happens when those ideas meet a bad client, an unusual patient, an old building or a system nobody documented properly. Simulations will help, and training will get better. Even so, a simulated mistake carries a different weight from one made on a live job with a more experienced person beside you.</p><p>We could end up with large numbers of graduates fighting over a smaller set of poorly paid openings while firms bid against each other for anyone with 10 years of credible practice. Employers will complain that applicants lack experience, which will be true, though it will leave out the uncomfortable fact that many of them stopped providing it. The shortage would sit in the middle of the profession. Plenty of people would want to enter; too few would have been allowed to grow up inside it.</p><p>A career ladder with fewer lower rungs becomes easier to climb for people who already have access to the building. A graduate whose parent owns a firm can acquire judgement informally. Someone who can afford a low paid internship gets another route. Everybody else is told to demonstrate experience before being allowed to earn it.</p><p>Jobcasting could expose the problem earlier by paying attention to the route into a job. For any profession, I would want to know how many trainees enter, how many remain after 3 years, which tasks teach them, how much supervision they receive and the age of the people currently carrying the difficult decisions. Strong pay tells a very different story when the trainee intake underneath it is collapsing. It may be evidence of a profession becoming valuable, or the first price signal from a shortage nobody can repair quickly.</p><p>The issue reaches well beyond office work. An electrician may have strong demand for decades because grids, housing, factories and data centres all need physical installation. That does not guarantee enough qualified electricians. If apprenticeship places fail to keep up, we would see rising prices, longer waits and a growing premium for people allowed to sign off work. Better tools can improve productivity at the same time as the licence and training bottleneck becomes more expensive.</p><p>Healthcare has its own version. Software can help interpret routine scans or prepare clinical notes, but an ageing population brings more patients with several conditions at once. The remaining human cases may be the ambiguous ones. If junior staff see fewer routine cases, hospitals will have to create another way for them to build pattern recognition. Otherwise the system ends up with excellent tools and too few people capable of noticing when the tool has produced a plausible answer for the wrong patient.</p><p>Teachers, claims handlers, mechanics, architects, public servants and financial analysts will each move differently. Birth rates affect teachers. Climate losses alter the work reaching claims staff. The age of the vehicle fleet changes demand for mechanics. Public budgets decide whether caseworkers use saved time with difficult families or simply carry twice the caseload. A useful jobcast has to follow the money and the training arrangements because capability alone cannot tell us which version an employer will choose.</p><p>It should also ask who benefits when productivity rises. If 1 employee can do the work of 2, the gain could appear as higher pay, a shorter week, lower prices, better service or a larger profit margin. Those outcomes create very different careers. Anyone choosing a field should care which outcome its institutions have delivered in the past.</p><p>Employers need jobcasts too, even if some would dislike what they reveal. A company may report a 15% reduction in staff costs as an efficiency while becoming dependent on a shrinking group of senior employees. It looks fine until 3 of them retire, leave or refuse to supervise software they no longer trust. The headcount figure improved and the concentration of judgement quietly became dangerous.</p><p>Investors should ask about this. We spend a great deal of time examining customer concentration, supplier concentration and debt maturity. Expertise concentration can damage a company just as easily. If 6 people understand a critical process and nobody has been trained beneath them, the business has a liability that rarely appears in an annual report.</p><p>I would like to see large employers disclose the number of trainee roles they create, the share of managers promoted from within and the time required for a new employee to handle important work without supervision. A company that keeps hiring senior people from outside may look dynamic. It may also be living off the training budgets of its competitors.</p><p>Governments have a reason to care because an expertise shortage eventually reaches public capacity. Regulators need investigators. Courts need experienced lawyers. Hospitals need consultants. The power grid needs engineers who have already handled difficult failures. You cannot announce a training programme after the shortage becomes visible and expect the people to appear next year.</p><p>For workers, the first useful jobcast can be done without a government database. Ask what the most respected people in your field did when they were young. Find out whether those tasks still exist and whether beginners are still trusted with them. Look at the difficult work flowing towards the profession, not only the easy work leaving it. Notice whether responsibility is arriving earlier and whether pay reflects it. Most of all, look for an employer that treats training as production instead of charity.</p><p>There will be professions where removing routine work makes careers better. People may reach interesting decisions sooner, learn from a wider range of cases and avoid years spent copying numbers between files. I believe that is possible. It requires employers to redesign training with the same seriousness they bring to cutting costs. Giving a graduate a powerful tool and a larger workload is not a training system.</p><p>The phrase “future of work” has become so broad that it can mean almost anything. Jobcasting brings the question down to the size of a career. What will this job ask of me in 5 years? What will it pay? Who will teach me? Which work will disappear, and what gets harder once the easy cases are gone? If I enter now, is there still a route from beginner to trusted judgement?</p><p>That final question is where Jobcasting meets the Expertise Harvest. A society can keep using the knowledge created by its existing experts for quite a while. We can put some of it into software, procedures and databases, then ask a smaller number of experienced people to supervise more output. Productivity may rise during this period and, for all I know, some professions will become much better. The risk is easy to miss because nothing breaks on the day a graduate place is removed.</p><p>Eventually the world changes. New laws appear. Machines fail in unfamiliar ways. Diseases behave differently. Markets produce a fraud that was not in the training data. Someone has to recognise the new case, and recognition usually belongs to a person who spent years seeing the old ones.</p><p>Whenever somebody tells me a job will still exist in 10 years, I now find myself asking a second question: will the profession still be producing people who know how to do it?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df5e4b8c92c4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Computational Crowding Out: Your Mortgage Is Now Competing With AI]]></title>
            <link>https://startupsapience.medium.com/computational-crowding-out-your-mortgage-is-now-competing-with-ai-331754d432ae?source=rss-75bd4ad3fb37------2</link>
            <guid isPermaLink="false">https://medium.com/p/331754d432ae</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[investing]]></category>
            <dc:creator><![CDATA[Startup Sapience]]></dc:creator>
            <pubDate>Sat, 29 Aug 2026 11:53:57 GMT</pubDate>
            <atom:updated>2026-08-29T11:53:57.340Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RekBGPxUngIq0HA9vc_7LA.png" /></figure><p>A household applying for a mortgage and a technology company building an AI data centre do not appear to have much in common. The household might need $400,000 to buy a home. The technology company may need $40B to construct buildings, secure power, install cooling systems and fill the site with processors.</p><p>They will never meet. They may still end up asking the same pension fund, insurer or asset manager for money.</p><p>I think this connection has been largely missed because the AI boom is usually described through physical shortages. We talk about processors, memory, copper, electricity, transformers, water and land. Finance is treated as something that will arrive automatically once the equipment and power become available.</p><p>AI infrastructure has grown so expensive that the companies building it are moving beyond their own cash flows and into bond markets, private credit, project finance and asset backed lending. They are beginning to compete with governments, banks, companies and households for long term capital.</p><p>I call this <strong>Computational Crowding Out</strong>.</p><p>It begins when the amount of money required to build machine intelligence becomes large enough to affect the price of money elsewhere in the economy. A data centre does not need to take money directly from a homebuyer for the effect to appear. It only needs to offer investors a return high enough to change what they demand from Treasury bonds, mortgage backed securities and other forms of debt.</p><p>The machine consumes electricity when it runs. Before it can run, it consumes years of someone else’s balance sheet.</p><p>In finance, those years are called duration. Pension funds and insurers buy long dated assets because their own obligations extend far into the future. A data centre bond, a government bond and a mortgage backed security can all compete for that demand.</p><p>At the old price, there may not be enough willing buyers for everything. The market solves the problem through higher yields.</p><p>The household eventually receives the bill.</p><h3>Capital Has Become the Next AI Bottleneck</h3><p>The largest technology companies entered the AI cycle in an unusually strong position. Their advertising, cloud, software and commerce businesses produced enough cash to fund enormous amounts of investment internally. For a while, it appeared that the AI buildout could continue without relying heavily on outside financing.</p><p>The numbers have now grown beyond that assumption.</p><p>On August 28, 2026, Reuters reported that AI related debt issuance had exceeded $220B during the year, around 2x the amount issued in 2025. JPMorgan estimated that AI debt could reach 50% of US Treasury coupon issuance by the end of 2026. Total US corporate bond issuance had already reached $1.68T.</p><p>The Bank of England reached a similar conclusion in its July 2026 Financial Stability Report. The 5 largest AI focused technology companies represented only 3% of outstanding US investment grade debt at the end of 2025. By early May 2026, however, they accounted for more than 15% of new issuance during the year.</p><p>I believe the difference between those 2 figures deserves attention. The 3% figure describes debt accumulated over many years. The 15% figure shows who is arriving in the market now and asking investors to provide new money.</p><p>Bond prices are set by the next buyer. A sector can therefore influence borrowing costs long before it becomes the largest part of the market.</p><p>The scale keeps being revised upwards. In December 2025, market estimates placed hyperscaler capital spending for 2028 below $600B. By July 2026, the estimate had moved above $1T. Barclays expects around $240B of hyperscaler investment to be financed through investment grade bonds in 2026. For comparison, the 6 largest US banks normally issue between $150B and $170B of senior debt each year.</p><p>Technology companies are beginning to borrow at a scale once associated mainly with banks and governments.</p><p>So far, investors have absorbed the issuance. The largest technology companies remain profitable, highly rated and lightly indebted compared with many traditional borrowers. Credit spreads have not shown clear signs of stress. The Bank of England found little evidence that AI investment had already prevented other companies or governments from accessing debt markets.</p><p>That does not make Computational Crowding Out irrelevant. It tells us that the market has accepted the new supply at current prices. The next $500B may require a different price, especially if government borrowing remains high and every other industry continues refinancing debt at higher rates.</p><p>Capital can expand, but it is not unlimited at a given yield.</p><p>If technology companies offer more interest, some investors will move money towards technology bonds. Treasury yields may need to rise to retain buyers. Mortgage backed securities may also need to offer a better return. Other companies could postpone issuance or pay more to complete it.</p><p>Nobody needs to decide that AI should receive money instead of housing. Thousands of ordinary portfolio decisions can create the same outcome.</p><h3>How It Reaches a Mortgage</h3><p>Most American mortgages are eventually packaged into securities and sold to investors. Those investors compare the return with Treasury bonds and other long term assets.</p><p>The 10 year Treasury yield is often used as a reference for a 30 year mortgage because the expected lives of the 2 assets are closer than their stated maturities suggest. Homeowners sell properties, repay loans and refinance when rates fall. A mortgage written for 30 years may therefore remain outstanding for a much shorter period.</p><p>In May 2026, the Federal Reserve Bank of Boston noted that the average 30 year mortgage rate was around 6.5%, while the 10 year Treasury yield was around 4.5%. The difference covered mortgage risks, lender costs and the value of the borrower’s right to repay early. Treasury yields still provided the base on which the rest of the mortgage rate was built.</p><p>By August 27, 2026, Freddie Mac placed the average 30 year fixed mortgage rate at 6.66%.</p><p>AI borrowing did not create those rates. Inflation, Federal Reserve policy, government deficits, energy prices, economic growth and foreign demand for US debt have much greater influence today. I would not look at a mortgage quote and claim that a specific number of basis points came from data centres.</p><p>Computational Crowding Out works through the marginal pressure added to an already crowded market.</p><p>Governments need trillions of dollars. Companies must refinance debts issued when interest rates were lower. Power producers need money for generation and transmission. Defence spending is rising. Housing still needs financing. AI has joined the queue with capital requirements large enough to be measured beside Treasury issuance.</p><p>An investor can now compare a mortgage backed security with a bond issued by Microsoft, Amazon, Meta, Oracle or a data centre project supported by one of them. If the technology bond offers a better return for an acceptable level of risk, money moves towards it. The mortgage security may then need a higher yield to remain attractive.</p><p>The cost filters through to the borrower.</p><p>Consider a $400,000 mortgage lasting 30 years. At 6%, the monthly principal and interest payment is around $2,398. At 6.5%, it rises to roughly $2,528. The additional 50 basis points cost about $130 each month and almost $47,000 across the life of the loan.</p><p>I am not suggesting AI will add 50 basis points to mortgage rates. The calculation shows how a small movement in the bond market becomes a large expense when carried through 360 monthly payments.</p><p>The people building the data centre can see the cost of capital in a financial model. The household experiences it as a bedroom it can no longer afford.</p><p>The parties being crowded out may never know who they competed against.</p><h3>The Debt Can Outlive the Machine</h3><p>The financing also contains a timing problem that I find increasingly difficult to ignore.</p><p>The debt can survive much longer than the technology purchased with it.</p><p>According to the Bank of England, much of the debt issued for AI infrastructure has a duration above 10 years. Most of the money currently finances buildings, electrical systems, cooling equipment and the wider facility instead of the processors themselves.</p><p>A building can stand for decades. A power connection can remain valuable for even longer. It might therefore seem sensible to finance the site with long term debt.</p><p>The earning power of a frontier data centre still depends on whether it can house and operate the machines customers want to use.</p><p>AI facilities are designed around certain power densities, cooling requirements, network architectures and processor generations. If new hardware requires far more electricity at the rack, a different cooling system or a new internal power standard, the old building may need expensive modifications.</p><p>The Bank of England warned that a facility could lose value if it became unable to support the latest processors. It also cited JPMorgan’s estimate that more than $2T could be required to finance AI chips during the next 5 years.</p><p>Processors have uncertain economic lives. Some will remain useful for a long time, especially if supply remains tight or older hardware can run smaller models. Others may lose value quickly when a new generation produces more computation using less electricity.</p><p>Debt follows a legal clock. Technology follows an engineering clock. They do not move at the same speed.</p><p>A 12 year financing agreement can depend on processors replaced after 4 years, tenants whose own products have existed for 2 years and revenue forecasts based on a market that barely existed when the facility was designed.</p><p>Perhaps the equipment will generate enough money during its short life to repay the longer debt. That is possible. The risk sits in the assumptions required to make the numbers work.</p><p>The lender is relying on continued demand for compute, high utilisation, financially healthy tenants, manageable electricity costs, regular hardware upgrades and a facility capable of accepting those upgrades. A weakness in any of them reduces the cash available to service the debt.</p><p>The machine does not need to stop functioning. It only needs to become too expensive relative to the next machine.</p><p>This is common in technology. A phone from 8 years ago may still switch on, yet few people would value it according to its original price. AI processors could experience the same decline on an industrial scale, except the equipment may have been financed through structures extending well beyond its competitive life.</p><h3>The Supplier Is Becoming the Financier</h3><p>Nvidia’s recent actions show how far the financing problem has moved into the AI supply chain.</p><p>In August 2026, Nvidia partnered with 6 major financial institutions to create platforms intended to raise more than $500B for AI infrastructure. Nvidia retained the option to guarantee up to $125B of the potential financing.</p><p>It later agreed to provide a guarantee of up to $105B connected with OpenAI’s lease of a large data centre in Ohio. The facility is being developed by SB Energy, which is owned by SoftBank. Nvidia also agreed to invest $1.5B in the developer and supply the processors used inside the site.</p><p>There is a logical commercial reason for doing this. Nvidia cannot sell $500B of processors unless its customers can finance the buildings, power and equipment required to operate them. Helping those customers gain access to capital expands the market for Nvidia’s products.</p><p>Industrial companies have used vendor finance for decades. An aircraft manufacturer may help an airline secure funding. A heavy equipment company may provide credit to a buyer. Nvidia is applying a similar model to compute.</p><p>The chip supplier invests in the developer. It guarantees some of the financing. The developer builds a site for an AI company. The AI company signs a long term lease. The lease supports the debt used to construct the facility. The facility buys processors from the company providing part of the guarantee.</p><p>Each contract can be sensible. When combined, they depend on the same future demand appearing in several places.</p><p>The developer assumes the tenant will keep paying. The lender assumes the guarantee remains strong. Nvidia assumes the customer will need increasing quantities of compute. The customer assumes its future revenue will justify the lease. Investors assume the entire arrangement will retain value long enough for the debt to be repaid.</p><p>I do not think every one of these structures should be dismissed as circular finance. The infrastructure is physical, customer demand is growing and Nvidia produces enormous cash flow. Yet the connections make it harder to identify where the final risk sits.</p><p>A failed software product can disappear quickly. A failed data centre project leaves behind debt, buildings, power contracts, equipment leases and guarantees spread across many institutions.</p><p>Computational Crowding Out therefore comes with a 2nd effect. AI does not only compete for money. It carries financial risk into institutions that may appear far removed from technology.</p><h3>When Future Productivity Meets Present Debt</h3><p>The optimistic answer is that AI investment will pay for itself through higher productivity.</p><p>If machine intelligence reduces the cost of research, software, healthcare, logistics, administration and manufacturing, the economy could become more productive. Company profits could rise. Governments might collect more tax. New industries could appear around capabilities that do not yet exist.</p><p>Under that outcome, today’s debt finances a larger economy tomorrow.</p><p>I believe some version of this will happen. The uncertainty lies in the timing and who captures the benefit.</p><p>The debt exists now. The productivity gain may arrive over 5, 10 or 20 years. During the gap, interest must be paid, facilities must remain occupied and processors must earn enough money to justify replacement.</p><p>The economy is therefore borrowing against an intelligence dividend before the dividend has been produced.</p><p>If AI revenue grows quickly, the financing becomes easier to carry. If revenue grows slowly, the pressure may appear through refinancing long before companies default. Investors could demand higher yields, stronger guarantees and shorter maturities. Projects with weaker tenants may struggle to raise money. Suppliers may need to provide more support to keep customers buying equipment.</p><p>Higher financing costs increase the price of compute. More expensive compute slows adoption. Slower adoption weakens revenue forecasts. Weaker forecasts make lenders more cautious. The industry then needs additional guarantees from the same companies already financing large parts of the expansion.</p><p>None of this requires an AI crash.</p><p>A moderate reduction in expected returns could be enough to change how creditors value a 15 year claim on a facility filled with short lived technology.</p><p>At the same time, investors may continue finding AI bonds attractive compared with mortgages, municipal debt or ordinary corporate lending. Capital then keeps moving towards the machine while other borrowers pay a little more.</p><p>Computational Crowding Out is a story about money being repriced as machine intelligence becomes one of the largest investment projects in history.</p><h3>The Cost Will Appear Somewhere Else</h3><p>People living far from an AI data centre may assume the buildout has little connection to them. They do not own Nvidia. They do not use advanced models. Their town may never host a large computing facility.</p><p>Their pension fund may finance one.</p><p>Their insurer may own its bonds.</p><p>Their bank may lend to the developer.</p><p>Their electricity provider may borrow to build the power station supplying it.</p><p>Their mortgage may be packaged into a security competing against it.</p><p>The effects do not need to arrive through a single dramatic event. A mortgage rate moves by 0.1%. A factory pays an extra spread on its debt. A local authority postpones a project. A pension fund changes its allocation. A power company directs investment towards a data centre instead of another customer.</p><p>Each decision is small. Together they show capital settling around the machinery required to produce intelligence.</p><p>I suspect this may become one of the defining financial questions of the AI cycle. Investors have spent years calculating how much revenue AI must generate to justify the capital spending. We should also ask what the rest of the economy must pay while that capital is being gathered.</p><p>The cloud was never weightless. It was built from land, metals, electricity and water. It now has a balance sheet large enough to influence other balance sheets.</p><p>Every prompt begins somewhere inside a machine worth millions of dollars. Behind that machine sits a building, a power contract, a lease, a lender and an investor expecting to be paid.</p><p>The AI economy is therefore entering ordinary life through a route few people will notice. Its cost may appear in the interest rate on a home, the return demanded from a public project or the loan a small company cannot justify.</p><p>That is Computational Crowding Out.</p><p>The machine may never take your job. It may still compete with you for the money needed to buy your house.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=331754d432ae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Forked Dollar]]></title>
            <link>https://startupsapience.medium.com/the-forked-dollar-ad1eae01a3b5?source=rss-75bd4ad3fb37------2</link>
            <guid isPermaLink="false">https://medium.com/p/ad1eae01a3b5</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[cryptocurrency-investment]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Startup Sapience]]></dc:creator>
            <pubDate>Tue, 25 Aug 2026 16:49:06 GMT</pubDate>
            <atom:updated>2026-08-25T16:49:06.316Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VgU3adj88TpPqtQ2D7eOgQ.png" /></figure><p>For most of my life, I assumed a dollar was a dollar. It could sit in a bank account, pass through a payment card or appear as a number on a trading screen, but the substance seemed unchanged. One currency, one monetary system and one set of assumptions about where the money came from, who stood behind it and what happened when something went wrong.</p><p>I no longer think that picture holds.</p><p>The dollar is beginning to divide into two separate financial lives. One remains inside the familiar world of banks, insured deposits, domestic lending and central-bank supervision. The other is spreading through digital wallets, trading platforms and payment networks, backed largely by short-term US government debt and circulating across borders at all hours. Both trade under the same currency symbol. Both aim to hold the same value. Yet they are taking different paths through the financial system, funding different activities and handing influence to different institutions.</p><p>I call this the <strong>Forked Dollar.</strong></p><p>My view is that the consequences will extend far beyond the argument over whether stablecoins are useful. This split could change who funds the US government, who finances American businesses, how foreign countries protect their currencies and how software begins paying other software. It may strengthen the dollar’s reach around the world even as it changes the relationship between American banks, public debt and the people whose savings once supported both.</p><p>To see why, we need to begin with a simple question that most investors rarely ask: who owes you the dollar you think you own?</p><p>When you hold cash, the money is a direct liability of the Federal Reserve. When you keep money in a commercial bank, you hold a claim on that bank. The bank takes deposits, makes loans, buys securities and relies on its own capital, regulation and access to emergency support to remain solvent. In the US, standard deposit insurance generally covers up to $250,000 per depositor, per insured bank, for each ownership category. That promise is one reason people can treat balances held at different banks as the same thing.</p><p>Now consider a dollar stablecoin. It is normally a claim on a private issuer that promises redemption at $1. The issuer holds reserves intended to support that promise, often using Treasury bills, cash deposits and short-term agreements backed by government securities. The token moves across a digital network, but its safety depends on the issuer, the quality of its reserves, the terms of redemption and the institutions holding the assets.</p><p>There is another detail people often miss. Many ordinary holders cannot redeem a token directly with its issuer. They rely on exchanges, dealers or approved institutions to move between the token and bank money. The $1 price survives because those intermediaries are willing and able to keep the two connected. When that link weakens, two claims that were treated as equal can suddenly trade at different prices.</p><p>The domestic banking dollar has spent decades growing through lending. When a bank extends a mortgage, funds a business or provides working capital, its balance sheet expands and new bank money enters the economy. A reserve-backed stablecoin works differently. An issuer usually receives existing dollars and places a large share of those funds in short-term government assets. The customer receives a digital token. The US Treasury gains another buyer for its debt. The chain ends in a different place.</p><p>This is where I believe the story becomes much larger than payments.</p><p>The dollar has lived abroad before. The offshore dollar market developed because companies, governments and banks outside the US needed dollar funding without holding accounts directly inside the American banking system. Over time, those offshore balances helped finance trade, corporate borrowing and cross-border investment. By March 2026, the Bank for International Settlements estimated that dollar-denominated credit to borrowers outside the US had reached $14.7 trillion, with roughly 30% linked to emerging and developing economies.</p><p>Stablecoins could become the next stage of that history, although their financial purpose follows a different route. Offshore bank dollars helped create private credit across the world. A Treasury-backed digital dollar can turn demand for American currency abroad into direct demand for American government debt. A shopkeeper in Lagos, a trading firm in Dubai and an AI agent in Singapore may all end up funding the same US Treasury market without ever deciding to buy a bond.</p><p>The numbers have already grown beyond the point where this can be dismissed as a side experiment. The Federal Reserve put the stablecoin market near $320 billion in its May 2026 financial stability report. An IMF official said in August that the market was around $300 billion and that nearly 99% of stablecoins were denominated in dollars. These totals are still small beside the wider banking system, but the direction of travel is clear. Almost the entire emerging market for digital money has chosen the dollar as its reference point.</p><p>Circle reported $73.3 billion of USDC in circulation at the end of the second quarter of 2026. Tether reported that its own tokens in circulation had passed $186 billion by the end of 2025, alongside more than $141 billion in direct and indirect exposure to US Treasuries. Those are not simply payment companies collecting transaction fees. They are becoming large holders of American government debt because people across the world want a dollar they can carry outside a bank account.</p><p>I suppose Washington has every reason to find this attractive. The Congressional Budget Office projected a $1.9 trillion federal deficit for 2026 and more than $1 trillion in net interest costs. In August, the Treasury said it expected to borrow another $739 billion during the Q3 quarter and $628 billion during the following quarter. Its borrowing committee also highlighted a projected $1.45 trillion funding shortfall across fiscal years 2027 and 2028 under existing issuance assumptions. Every new source of steady demand for government debt will get attention.</p><p>Imagine a merchant in Argentina converting the equivalent of $10,000 into a dollar stablecoin to protect working capital. If that demand eventually leads to new token issuance, the issuer receives dollars and invests much of the reserve in Treasury bills. The merchant gains a portable claim linked to the dollar. The US government gains financing. The issuer collects interest. The local banking system may lose a deposit, and the local currency may face another source of selling pressure.</p><p>One decision creates several consequences, spread across several countries, and none of them appear on the merchant’s phone.</p><p>The US has now started giving this arrangement a formal legal home. The GENIUS Act was signed in July 2025 and requires permitted payment stablecoin issuers to maintain at least 1:1 backing through approved assets, including certain short-term Treasuries, qualifying bank deposits and government-backed funding agreements. Treasury said in August 2026 that the law is expected to take effect on January 18, 2027, with federal or state licensing required for issuers operating in the US.</p><p>I would assume many people see this as ordinary financial regulation. To me, it looks closer to the legal construction of a new export channel for the dollar. The government is allowing private companies to distribute dollar claims across the world while directing the reserves toward assets that support the American financial system.</p><p>The outcome could be powerful, although the benefit will fall unevenly across the Treasury market.</p><p>Stablecoin reserves favour short-dated government paper. Under the new rules, qualifying Treasury securities are generally limited to maturities of 93 days or less. That creates a steady buyer for the front end of the government’s borrowing curve. It does far less for the 10-year and 30-year bonds that fund longer commitments and shape mortgage costs, corporate borrowing rates and public investment.</p><p>An IMF working paper found that a stablecoin demand shock linked to a 1% rise in the combined value of USDC and USDT reduced one-month Treasury yields by about 1.9 basis points at its strongest point over the following months. The effects were much weaker at longer maturities. I think that offers an early glimpse of a growing imbalance: digital demand can support short-term government borrowing while the long end remains exposed to deficits, inflation worries and the need for a wider pool of buyers.</p><p>The same study found that the broader dollar index weakened by roughly 0.09% at its lowest point after the demand shock. That may sound strange. More people were using dollar tokens, yet the exchange value of the dollar softened as short-term yields fell and investors adjusted their portfolios. A currency can expand its global reach without rising against other currencies. The two outcomes answer different questions.</p><p>This is already visible in the questions Treasury is trying to manage. On August 19, 2026, it announced plans to at least double the size of certain longer-dated bond buyback operations from $2 billion to $4 billion. That move reflects liquidity conditions in long-maturity government debt. It should not be presented as proof that stablecoins caused the pressure. Still, the picture is worth watching: private digital dollars gather around short-term bills while the government looks for ways to support trading further out on the curve.</p><p>In my opinion, this could produce a government borrowing market with two different sources of support. Short-term bills attract global demand through digital dollar reserves. Long-term bonds must continue to compete for pension money, foreign reserves, household savings and other investment capital. The stronger the pull toward bills, the greater the temptation for governments to rely on short-term borrowing, even though that can leave debt costs more exposed when interest rates rise.</p><p>There is another question sitting beneath all of this. Who gets paid for the dollar?</p><p>When someone holds a stablecoin, the reserves behind it can earn interest on Treasury bills. The holder generally receives none of that income. Circle reported $668 million of reserve income in the second quarter of 2026 and $412 million of distribution, transaction and other costs. Tether reported more than $10 billion in profit for 2025. The economics are easy to follow. Users supply the demand for digital dollars, the reserves earn a return and the issuer shares part of that income with the platforms that help distribute the token.</p><p>The GENIUS Act prohibits issuers from paying interest directly to payment stablecoin holders, though questions remain around rewards offered through affiliates and other partners. That rule helps explain why the future battle may centre on ownership of the wallet, the customer relationship and the flow of reserve income. Whoever controls access to the user may claim a growing share of the interest generated by the money sitting behind the token.</p><p>Interest rates add another twist. Higher Treasury yields can raise the income available to issuers even when token holders receive nothing. Lower rates can shrink the pool of money used to pay distributors and attract customers. The Federal Reserve may therefore influence the growth of digital dollars through company profits and payment partnerships as well as through the usual channels of bank lending and borrowing.</p><p>Perhaps this is where the Forked Dollar becomes politically sensitive. A person in an emerging market may hold a digital dollar because their local currency is unstable. The US government receives financing. A private issuer earns interest. A payment company earns distribution fees. Yet the person carrying the currency receives no share of the income produced by the reserves, even though their savings helped create it.</p><p>Banks have their own reason to be uneasy. A deposit inside a local bank can help support mortgages, business loans and other forms of private credit. A dollar moved into a stablecoin backed by Treasury bills changes the route that funding takes. If enough savings follow that path, governments could gain a new source of demand for their debt while some lenders face higher funding costs.</p><p>I would be careful about claiming that every stablecoin dollar removes a dollar from the banking system. Money often returns to another bank when an issuer buys a Treasury bill from an existing holder. Some reserves remain in bank deposits, and banks can also provide custody, payments and other services around the tokens. The question is where the deposits end up, which banks retain them and whether the same money still supports private lending in the same places.</p><p>Even the White House has pushed back against the largest forecasts of bank disruption. An April 2026 analysis estimated that banning stablecoin yield would increase bank lending by just $2.1 billion under its central assumptions, equal to 0.02% of lending. Under an extreme combination of assumptions, the estimate rose to $531 billion, or 4.4% of loans. The spread between those figures tells me the outcome will depend on how the products are designed, where the reserves are held and whether users treat tokens as everyday money or as another place to store cash.</p><p>There is a further complication. Community banks and regional lenders do not occupy the same position as the largest custodians. If stablecoin reserves gather at a small number of major financial institutions, deposits can move around the banking system without remaining evenly spread. A bank that once funded local businesses could lose access to cheap customer deposits, even while total deposits across the country stay broadly intact.</p><p>The first branch of the dollar has traditionally supported local lending through thousands of banks. The second may concentrate cash, custody and government debt in a much smaller group of institutions. That could reshape who gets funded, even if the total quantity of dollars barely changes.</p><p>I think the international effects could be even more dramatic. An IMF report on Nigeria found that the country accounted for roughly 60% of stablecoin inflows into sub-Saharan Africa since 2019. The same report noted that sending $200 to the region through conventional channels can cost around 9%, compared with a global average near 6%. For a family receiving money from abroad or a business paying an overseas supplier, a digital dollar can solve an immediate problem.</p><p>Yet each movement into a dollar token can also reduce demand for the local currency. IMF research covering four dollar stablecoins and 27 national currencies found that a 1% rise in stablecoin inflows was linked to a 40 basis point increase in the gap between digital dollar prices and traditional foreign exchange prices, alongside pressure on local currencies. The effects became stronger when financial intermediaries were under stress.</p><p>This creates a difficult position for emerging market central banks. Raising domestic interest rates may do little to stop households from moving into dollar tokens if their main concern is protecting savings. Restricting bank access to crypto platforms can push trading into informal channels. Blocking wallets may encourage users to move activity abroad. Every defence becomes harder when the competing currency lives on a phone and can circulate outside the local banking system.</p><p>Dollar use abroad once required physical notes, a foreign bank account or access to a regulated financial institution. The new version needs an internet connection and a willing counterparty. I suspect that lowers the barrier to currency substitution more sharply than many policymakers have accepted.</p><p>The challenge extends well beyond countries with unstable currencies. In February 2026, a European Central Bank board member warned that dollar stablecoins could spread through European cross border payments, online commerce and machine-to-machine transactions. He noted that international card networks already account for roughly two-thirds of card transactions in the euro area, and that 13 of its 21 member countries lack a domestic card network. The concern is that Europe could lose deposits, payment income and control over the way digital transactions are settled.</p><p>I would expect governments to respond by encouraging local currency stablecoins, digital central bank money and bank-issued digital deposits. Some will try to restrict foreign dollar tokens. Others may allow them but require reserves, licensing or reporting inside their own borders. The result could be a struggle over who controls the currency used by the internet, with the dollar starting from a large advantage.</p><p>That advantage comes with a complication for Washington. A private digital dollar can extend American financial influence across borders, but it can also carry American sanctions and legal authority deeper into foreign markets. In April 2026, Tether said it had helped freeze more than $344 million across two wallets after coordination with US authorities. Treasury’s proposed rules also call for foreign issuers serving US markets to be able to comply with lawful orders and relevant cross-border arrangements.</p><p>For the US, that can look like an extension of financial oversight. For another country, it can look like foreign control embedded in its payment system. I believe this tension will shape the next round of global monetary competition. Governments may accept the convenience of digital dollars until they start asking who can freeze them, who can monitor them and whose courts decide when access should be blocked.</p><p>We should also be honest about how far adoption has progressed. Circle reported $14.8 trillion of USDC activity during the second quarter of 2026, a striking figure for a token with $73.3 billion in circulation. Yet gross activity includes automated trading, repeated transfers and transactions that have little to do with buying goods or sending money to family. The IMF has cited an estimate of only $390 billion in payment-related stablecoin flows during all of 2025, within a global cross-border payments market measured near $1 quadrillion annually. Large trading numbers should never be confused with equally large consumer adoption.</p><p>Visa’s own experience gives a better sense of current scale. In April 2026, the company said its stablecoin settlement programme had reached a $7 billion annualised run rate across nine digital networks. That is meaningful growth, but still small beside global card and bank-payment volumes. The opportunity is visible. It has yet to reach the size suggested by some of the louder claims around crypto.</p><p>Even so, there is one future customer for digital dollars that may grow faster than human adoption. Software agents.</p><p>A machine cannot walk into a bank branch, produce a passport or wait for a payment to clear on Monday morning. It can hold a digital wallet, follow spending rules and pay for a service within seconds. As AI systems begin hiring computing power, purchasing data and carrying out transactions for businesses, they will need money that can move at the same pace.</p><p>Circle launched payment tools for AI agents in May 2026 and said by the following quarter that more than 900 paid services were using its agent platform. It also reported that 99.3% of payments using one agent-payment standard were settling in USDC. Mastercard has separately introduced services designed to let machines pay one another across cards, bank accounts and stablecoins. These are early products, and their future size is uncertain, but they point toward a form of commerce built around software that can earn, spend and settle without keeping office hours.</p><p>I find that possibility more interesting than many of the current arguments about crypto. The digital branch of the dollar may become the currency used by machines to buy services from other machines. Each payment could be tiny. The number of transactions could be enormous. A system built for monthly invoices and business-day settlement may struggle to serve an economy where software buys a fraction of a second of computing time or pays for a single piece of data.</p><p>That connects the Forked Dollar to the wider changes taking place in financial markets. A share trading through a global network on a Sunday needs money available on Sunday. A hedge fund using automated agents across several countries needs a settlement asset that can move between markets after local banks close. An AI system buying electricity, cloud capacity or financial information needs a payment method it can control through software.</p><p>The dollar that serves those activities may still be worth $1. It may simply operate under a different set of rules from the dollar sitting in a household checking account.</p><p>There are risks in treating the two as interchangeable. When Silicon Valley Bank failed in March 2023, Circle had $3.3 billion of reserves trapped at the bank, equal to about 8% of its backing at the time. USDC traded below its intended value until US authorities intervened to protect depositors. A token designed to move continuously had run into the same problem that has troubled banking for centuries: users began questioning whether the promise behind their money would hold.</p><p>The danger becomes larger as these tokens move deeper into government debt markets and payment systems. A rush to redeem stablecoins could force issuers to sell Treasury bills or draw heavily on bank deposits. A problem at a reserve bank could spread back into the token. Stress could pass between government funding, private payment networks and traditional lenders much faster than regulators expect.</p><p>My own assumption is that banks will answer by issuing digital versions of their deposits. These would keep the protections and lending links of ordinary bank money while gaining some of the speed and flexibility offered by stablecoins. Central banks may also provide digital settlement tools for banks and financial institutions. Over time, the dollar could develop several digital forms, with each serving a different purpose and carrying a different kind of risk.</p><p>Still, the central split would remain. One branch of the dollar would be created and circulated through the banking system, supporting private lending and domestic economic activity. The other would spread through global digital networks, drawing strength from government securities, private distribution and demand from users who may have no connection to the US beyond the currency they hold.</p><p>I think the deepest question is whether those two branches continue to support each other. If they do, the US could gain a wider global market for its currency, stronger demand for government debt and faster channels for international commerce. If they begin pulling in different directions, the country could end up with a dollar that grows stronger abroad while parts of its domestic credit system become more expensive to fund.</p><p>The next phase of dollar dominance may therefore arrive through a change in form. People across the world could hold American money without entering an American bank. Machines could transact in dollars without opening accounts. Foreign savings could reach the Treasury market through a digital wallet. Governments could find themselves competing with a currency that enters their economies through software.</p><p>The dollar would still carry one name. Its financial life would be divided between two systems, each with its own institutions, risks and beneficiaries.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ad1eae01a3b5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Computational Attrition Wall]]></title>
            <link>https://startupsapience.medium.com/the-computational-attrition-wall-0835c9d0890f?source=rss-75bd4ad3fb37------2</link>
            <guid isPermaLink="false">https://medium.com/p/0835c9d0890f</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[quantum-computing]]></category>
            <dc:creator><![CDATA[Startup Sapience]]></dc:creator>
            <pubDate>Mon, 24 Aug 2026 18:55:24 GMT</pubDate>
            <atom:updated>2026-08-24T18:55:24.063Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6udv7NiHvbNpe_VuaI44-w.png" /></figure><p>The AI trade has behaved like a pressure wave moving through a machine.</p><p>It began with the chip because the chip was the easiest part to see. Demand for intelligence suddenly became demand for GPUs, and the market placed an enormous price on the companies that could design, manufacture and package them. NVIDIA’s full-year revenue reached $215.9 billion in fiscal 2026, up 65%, while its data centre business produced $62.3 billion in the final quarter alone. The first great fortune of the AI cycle came from owning the silicon through which the models had to pass.</p><p>Then the market discovered that a chip without electricity is expensive furniture. Utilities, gas turbines, transformers, nuclear plants, switchgear, copper and grid connections became part of the AI trade. The International Energy Agency expects global data-centre electricity use to more than double to around 945 terawatt-hours by 2030, slightly above Japan’s total electricity consumption today. In the US, data centres could account for almost half of electricity demand growth through the end of the decade.</p><p>The money now committed to the buildout is almost difficult to absorb. Alphabet has guided to $195 billion to $205 billion in 2026 capital expenditure, Meta to $130 billion to $145 billion, Microsoft to roughly $190 billion and Amazon to about $200 billion. Those figures cover spending across their businesses, so they cannot all be assigned to AI. Together, however, they amount to roughly $715 billion to $740 billion in one year, and a large share is tied to the physical infrastructure supporting cloud and machine intelligence. Alphabet alone reported negative free cash flow of $5.9 billion in the second quarter after spending $44.9 billion on capital projects. Once that much money has been installed, every hour it sits idle becomes a financial problem.</p><p>Memory came next. A processor can calculate only after the data arrives, so scarce high-bandwidth memory began to decide how much of the GPU could actually be used. Micron’s HBM4 for NVIDIA’s Rubin platform now moves more than 2.8 terabytes a second, 2.3 times the bandwidth of its HBM3E product, with more than a 20% improvement in power efficiency under the company’s test conditions. The memory surrounding the processor became nearly as important as the processor itself.</p><p>Then the machine grew too wide for electricity and memory alone. Thousands of GPUs needed to exchange data as one system, and copper links began to run into distance, speed and heat limits. Photonics moved towards the centre. Lasers, optical transceivers, switches, fibre and co-packaged optics became the nervous system of the AI factory. NVIDIA says its Spectrum-X photonic switches for Rubin can deliver five times better power efficiency, ten times greater reliability and five times longer uptime than conventional designs. The language is interesting. Even in an announcement about faster networking, the selling point is already beginning to shift towards survival.</p><p>After that came heat. Every watt entering a data centre eventually has to leave as heat, and the old method of blowing cold air through a room could not keep up with racks drawing 100 kilowatts or more. Liquid cooling, coolant distribution units, pumps, chillers, heat exchangers, pipes, valves, filtration and leak detection became investable parts of the machine. A GB300 NVL72 rack can draw more than 120 kilowatts. Vertiv cites Uptime Institute data showing that power distribution caused 45% of serious data-centre outages in its 2025 survey, while cooling caused 14%.</p><p>Investors have been trained to ask what comes after cooling. Perhaps it is water. Perhaps it is copper, natural gas, nuclear fuel, power semiconductors, storage or some new optical material. All of those will have periods of scarcity. I have been thinking about a deeper answer, because the next constraint may no longer sit inside one type of component.</p><p>I think the next great AI infrastructure wave will be the fight against attrition.</p><p>By attrition, I mean the steady loss of useful computation as a giant machine is exposed to faults, delays, voltage changes, failed links, corrupted memory, thermal stress, ageing components, software errors and slow recovery. The industry has spent years measuring how much computation it can install. The coming years will force it to measure how much trustworthy computation survives long enough to produce an answer.</p><p>I call this the <strong>Computational Attrition Wall</strong>.</p><p>The law behind it is simple. As an AI cluster grows, the number of possible failure points grows with it, while the chance that every part remains healthy and synchronised falls. Beyond a certain size, adding more processors can create surprisingly little additional output unless reliability, fault isolation and recovery improve at the same time. A machine made from reliable pieces can still become fragile as a whole.</p><p>The mathematics turns ugly very quickly. Imagine 100,000 GPUs, each with a hypothetical 99.999% chance of surviving a full day without failure. That sounds close to perfection. Across the whole cluster, it still produces an average of one GPU failure every day. The probability that all 100,000 survive the day together is only about 36.8%. At one million GPUs, the same component reliability gives roughly ten failures a day and only a 0.0045% chance of an interruption-free day. Networks, memory, storage, power supplies and cooling equipment would add their own failure probabilities.</p><p>This is already happening. Meta trained Llama 3.1 405B using more than 16,000 H100 GPUs. During one 54-day period, the training run suffered 419 unexpected interruptions, roughly one every three hours. NVIDIA now uses that example to explain why large training jobs save checkpoints every 15 to 30 minutes. Meta has separately said that hardware faults across SRAM, HBM, processing grids and network switches caused more than 66% of training interruptions during its work on the Llama 3 family.</p><p>A stopped training run tells the operator something has gone wrong. The machine can isolate the damaged node, restore an earlier checkpoint and continue. Silent data corruption is more dangerous. A faulty memory cell or processor can produce a believable number that happens to be wrong. That error then enters the shared training process, spreads through the gradients and contaminates the model while the system appears to be working.</p><p>Meta warns that these errors can take weeks or months to detect. Some cause obvious numerical failure. Others allow training to continue without useful improvement, consuming electricity and GPU time while the engineers search for a fault they cannot see. At that point the machine is technically online, the utilisation dashboard may look healthy and the output is still worthless. I suspect this becomes one of the largest hidden costs in AI because accounting systems record electricity consumed and equipment depreciated. They do not record intelligence that was quietly damaged before it was born.</p><p>There is a deeper consequence for the semiconductor supply chain. A chip can pass its initial tests and still reveal a weakness after months of heat, vibration or electrical stress. As processors become more tightly packed and memory stacks grow taller, the burden shifts towards longer testing, better inspection, error detection and continuous health checks after installation. Meta estimates that silent corruption can occur in roughly one out of every 1,000 devices across modern accelerator systems. At the scale of a million processors, even a small defect population becomes a permanent operating condition.</p><p>This changes the meaning of capacity. A company may own 100,000 GPUs, yet the financially useful number depends on how many are available, correctly connected, properly cooled, supplied with clean power and producing calculations that can be trusted. Installed compute is a physical count. Useful compute is an outcome.</p><p>I would express it this way:</p><p><strong>Useful Compute = Installed Compute × Utilisation × Availability × Synchronisation × Computational Integrity</strong></p><p>Each term multiplies the others. If a large system achieves 90% on all four operating factors after installation, it delivers only about 65.6% of its theoretical capacity. Moving each factor to 95% raises the figure to roughly 81.5%. The improvement sounds small when each line is viewed separately. Together, it creates almost 24% more useful output from the same installed hardware.</p><p>That is why I do not see reliability as a dull maintenance category sitting at the end of the AI boom. It may become one of the highest return forms of capital spending in the system. If a $10 billion AI facility loses 10% of its effective capacity to stalls, failures and recovery, improving that loss by half creates the equivalent of $500 million of additional working equipment before counting power savings or faster model releases. Five points of utilisation can be worth more than another building full of chips.</p><p>The first part of this wave is already arriving through power delivery. For the past two years, investors have asked where the next gigawatt will come from. The newer question is how that gigawatt enters a dense rack without wasting too much of it during conversion or destabilising the equipment it supplies.</p><p>In August 2026, NVIDIA described an 800-volt direct-current design developed with Google and Microsoft through the Open Compute Project. More than 80 equipment and infrastructure companies are building products around the standard. The first hybrid power racks are due in the second half of 2026, followed by row-level systems supporting as much as two megawatts per row in 2027.</p><p>The reason becomes clear through basic electrical physics. Delivering one megawatt at 48 volts requires more than 20,000 amps. At 800 volts it requires 1,250 amps. Because resistive loss rises with the square of current, higher voltage can sharply reduce the copper, heat and conversion burden, although the final design still depends on conductor size and the number of conversion steps. The AI power trade is moving inside the building, from the grid connection towards busways, converters, solid-state transformers, capacitors, magnetics, circuit protection and the final centimetres before the processor.</p><p>Power quality will become part of the product too. An AI cluster does not simply need a large annual supply of electricity. It needs voltage and frequency held inside narrow limits while a highly dynamic workload changes demand. A short disturbance can interrupt thousands of processors at once. The equipment that detects a transient, carries the load for a few seconds, disconnects a fault or restarts a row cleanly may protect more value than its own purchase price suggests.</p><p>The pressure also moves in the other direction. In May 2026, the US Department of Energy warned that synchronised AI training can create repeating electrical oscillations that travel beyond the data centre and interfere with equipment at nearby power plants. Existing grid-monitoring instruments may miss some of the faster movements. A large AI cluster can therefore destabilise the same grid it depends on, creating demand for better sensors, local storage, power conditioning and systems that smooth electrical demand before it reaches the wider network.</p><p>This is where batteries, uninterruptible power supplies, fast switches, relays, fuses, sensors and local energy storage take on a different financial role. They stop being support equipment and begin to determine how much of the expensive compute can be sold. The same will happen to electrical monitoring. A sensor capable of finding insulation damage, harmonic distortion or a failing connection before it trips a rack can create an enormous return from a very small component.</p><p>Cooling will follow the same path. The first liquid-cooling trade was about installing pipes and coolant distribution units. The attrition trade begins after those systems have been running for several years. Pumps wear. Filters fill. Seals weaken. Connectors leak. Coolant chemistry changes. Corrosion and small particles move through channels sitting millimetres away from high-value silicon. Thermal cycling expands and contracts materials thousands of times.</p><p>The cooling system will therefore need redundancy, continuous flow monitoring, leak detection, water treatment, filtration, chemical testing, spare parts and trained technicians. Vertiv’s current cooling units already include redundant pumps, dual power feeds, 50-micron filters and controls designed to prevent thermal shock. Those features offer a preview of where spending moves once liquid cooling becomes common. The sale of the equipment is the beginning of the revenue stream. Its useful life will be supported by inspection, replacement and service contracts.</p><p>Photonics will have its own attrition curve. A faster optical network does not remove dirty connectors, ageing lasers, temperature drift, misalignment, link failures or the difficulty of servicing optical components inside a dense rack. As co-packaged optics moves lasers and optical links closer to the switch, efficiency improves and repair becomes harder. Test equipment, optical monitoring, replaceable laser sources and automated fault location may earn a growing share of the spending. The market first valued the ability to move data faster. It will begin to value the ability to keep moving it after years of heat and vibration.</p><p>Storage may be one of the biggest beneficiaries. Every time a large training run fails, the system has to reload a saved state. The checkpoint is a financial memory of all the electricity and computation spent before the interruption. Slow checkpointing forces the operator to save less often, increasing the amount of work lost when something breaks. Fast checkpointing consumes bandwidth and creates sudden storage demand across thousands of machines.</p><p>Meta says storage and interconnect performance have grown more slowly than AI compute and remain a major cause of GPU stalls. Its engineers also explain how one slow GPU can hold back an entire synchronised training step. This gives high-performance flash, local caches, checkpoint compression and low-latency storage a second AI story. Memory feeds the processor during the work. Storage allows the work to survive a failure.</p><p>Software then has to decide how failure spreads. Most large training jobs still depend on thousands of processors reaching the same point together. One failed node can stop everyone. Researchers working on training across 100,000 GPUs have described frequent faults and long recovery times as a direct cause of low efficiency. Their proposed systems divide the job into smaller groups, allowing one section to recover while the others continue.</p><p>I believe that idea will move from research into the design of the whole AI factory. Power rows, cooling loops, storage pools, optical links and software jobs will be divided into fault zones. A damaged component will be isolated while the rest of the facility continues producing tokens. The best system may be the one that bends under failure without stopping, in much the same way that the internet routes around a damaged link.</p><p>NVIDIA said on 24 August 2026 that AI-factory economics are defined by delivered output, including tokens per second, tokens per watt, utilisation and uptime. It also placed continuous health monitoring and component-level service while the factory keeps running alongside network performance as a design requirement. That choice of language says a great deal. The factory’s value comes from the useful output reaching the customer after every failure, delay and repair along the way.</p><p>The same attrition appears again when the model becomes an autonomous agent. Suppose each step in a 100-step task has a hypothetical 99.9% chance of success and every step depends on the one before it. The chance of completing the full chain falls to roughly 90.5%. Extend the job to 1,000 steps and the success rate drops to about 36.8%, before accounting for broken data feeds, payment failures, changing permissions or inaccurate information.</p><p>An agent managing a trading book, running a warehouse or moving cash across several banks therefore needs its own version of the infrastructure being built for the data centre. It needs checkpoints, error detection, approved access, audit trails, recovery procedures and a way to stop one failed action from damaging the whole process. The hardware suffers attrition across components. The digital worker suffers attrition across decisions. Both need an operating system that can survive mistakes.</p><p>This should change how the assets are financed. An investor lending against a data centre may eventually ask for effective compute hours, fault rates, recovery time and useful tokens per installed dollar. Two facilities with the same chip count could deserve different borrowing costs if one loses twice as much output to interruption. Insurers could price coverage from component histories. Equipment leases could include uptime guarantees. Used GPUs may carry a health record showing thermal exposure, memory errors and remaining life.</p><p>Depreciation makes the pressure harder to escape. A hypothetical $5 billion computing fleet written off over four years consumes around $1.25 billion of capital a year, or about $143,000 an hour, before electricity, debt, staff or cooling are included. The clock keeps running when the cluster stops. A business that cuts recovery time from an hour to five minutes is protecting an asset that loses value even while it waits.</p><p>Failure data itself could become a moat. The company operating hundreds of thousands of accelerators sees error patterns that a component maker may never observe in a laboratory. It learns which batch of memory fails under a certain temperature, which optical link degrades after repeated cycling and which firmware update creates a rare interaction between otherwise healthy parts. Meta says diagnosing silent corruption across its systems can involve hundreds of engineers. The operator with the longest history of failures may eventually run the most reliable fleet.</p><p>I can imagine a new group of companies built around this data. They would monitor the AI factory as an aircraft engine is monitored, looking for small changes in vibration, temperature, current, flow, optical power and error rates. They would predict which component is moving towards failure and schedule replacement before a training run begins. AI would then be used to preserve the machines running AI.</p><p>The opportunity reaches far beyond software. It includes power conversion, solid-state transformers, high-voltage DC equipment, switchgear, protection devices, busbars, connectors, sensors, metrology, optical test, error-correcting memory, flash storage, coolant filtration, pumps, seals, valves, water chemistry, fire suppression, robotics and field service. Many of the companies involved will look like ordinary industrial suppliers. Their component may cost a few hundred dollars and protect a rack worth several million.</p><p>Uptime Institute’s 2026 survey gives some sense of the economics. It found that 57% of respondents said their most recent major outage cost more than $100,000, while one in five reported a cost above $1 million. Power remained the leading cause of serious outages, with UPS systems, transfer switches and generators among the main sources. As AI racks become denser, a single event can strand more capital in the same room.</p><p>Countries will be affected too. In the Electrostates framework I developed, national power is converted into computation and exported as intelligence. The Computational Attrition Wall adds another requirement. An Electrostate needs reliable grids, clean power, trained technicians, spare parts, secure networks and enough engineering depth to keep the machines operating. Cheap electricity can attract the first data centre. A strong maintenance and reliability base may decide where the next ten are built.</p><p>Perhaps compute will eventually be quoted in a way that resembles electricity. A power plant has nameplate capacity and a capacity factor showing how much energy it actually produces. An AI factory may develop its own compute capacity factor, measuring useful tokens against theoretical tokens over time. Nations and companies would then compete through dependable output, not the number of processors announced at an opening ceremony.</p><p>This does not mean chips, power, memory, photonics or cooling disappear as investment themes. They remain inside the machine, and many are still scarce. The deeper point is that every solved bottleneck adds another working part whose failure can stop everything around it. Complexity accumulates. The old constraints do not leave. They become sources of attrition.</p><p>The AI cycle began by rewarding whoever could make the machine larger. I think its next phase will reward whoever can keep the machine alive.</p><p>That is the Computational Attrition Wall. It appears when the theoretical size of an AI system grows faster than its ability to remain healthy, synchronised and trustworthy. Crossing it requires a new layer of investment, spread across the quiet components that prevent expensive machines from damaging themselves.</p><p>There is something almost biological about where the cycle has arrived. The chip became the brain. Memory fed it. Photonics connected its nerves. Electricity became its metabolism. Cooling became its circulation. Now the machine needs an immune system, a way to detect injury, isolate failure, repair itself and remember where it was before the damage occurred.</p><p>The next great AI trade may be built around the mortality of the machine.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0835c9d0890f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Sovereign Exchange Mesh]]></title>
            <link>https://startupsapience.medium.com/the-sovereign-exchange-mesh-593f766505a2?source=rss-75bd4ad3fb37------2</link>
            <guid isPermaLink="false">https://medium.com/p/593f766505a2</guid>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Startup Sapience]]></dc:creator>
            <pubDate>Sun, 23 Aug 2026 10:41:52 GMT</pubDate>
            <atom:updated>2026-08-23T10:41:52.309Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fmQVAVVWyw1UUcAw1jJloA.png" /></figure><p>For most of financial history, the stock exchange was a place. Traders gathered in Amsterdam, London or New York because buying and selling shares required people to stand close enough to find one another. The buildings changed, the shouting disappeared and the trading floors became server rooms, but the geography survived. A company still lists somewhere. A share still belongs to a national market. An exchange still opens and closes according to the working hours of the country where it happens to sit.</p><p>I believe that arrangement is approaching the end of its natural life.</p><p>The world already prices companies as though capital has no borders. A fund in Singapore follows the same American earnings call as a trader in London and a pension manager in Toronto. News travels everywhere at once. Money does not. Shares do not. Ownership records do not. The price can cross the planet in a second, while the legal and financial machinery beneath it remains divided between countries, currencies, custodians, brokers and settlement systems.</p><p>According to the World Federation of Exchanges, global listed equity markets reached roughly $148.4 trillion in value, with 60,633 listed companies and more than $20 trillion of share trading during the period covered by its December 2025 dashboard. That is an enormous pool of capital. It is also broken into thousands of smaller pools, each with its own rules, opening hours and gatekeepers. I think the next stage of financial history will be shaped by the attempt to connect those pools without asking governments to give up control over their own markets.</p><p>To understand where this could lead, we first need to look at what a stock exchange actually does. Most people see a price on a screen and assume the exchange is the system that makes the transaction happen. That captures only a small part of the story. The exchange admits companies to the market, sets listing standards, runs auctions, publishes prices, matches buyers with sellers and watches for manipulation. It also creates the public reference price that investors, index funds and lenders use to value assets across the financial system.</p><p>Once a trade is agreed, another world takes over. A clearinghouse steps between the buyer and seller and guarantees that both sides will honour the transaction. It collects collateral and combines large numbers of trades so that firms settle only what they owe after their purchases and sales have been offset. A securities depository records where the shares are held. Custodians keep accounts for investors. A transfer agent maintains the company’s shareholder records and helps process dividends, voting and other changes in ownership. Banks move the money. Regulators stand behind the entire chain.</p><p>When someone says they bought a share, what they usually mean is that a long chain of institutions agreed that the trade happened and will eventually make the ownership and payment line up. The exchange is the front door. The building behind it is much larger.</p><p>Even inside one country, that building is full of separate rooms. In the US, shares trade across exchanges, private trading venues and firms that execute customer orders internally. An SEC filing put off-exchange trading at 50.4% of US equity volume in the first quarter of 2025. Nasdaq later reported that off-exchange trading moved above 50% during 2025, while openly displayed exchange trades accounted for less than a third of continuous intraday volume. Much of the market now depends on public exchange prices even though the trades happen somewhere else.</p><p>I find this fascinating because markets are pulled in opposite directions. Buyers and sellers benefit when they can meet in the deepest possible pool of trading. At the same time, exchanges, brokers and trading firms earn money by controlling their own piece of that pool. Some investors want their orders displayed publicly. Others need privacy because announcing a large purchase would push the price against them. Regulation encourages competition between venues, even though each new venue can make the market harder to see as a whole.</p><p>The result is a financial system that behaves as one market in theory and many markets in practice. The same stock may have dozens of available prices, depending on where the order travels, who receives it and how quickly the buyer can reach the seller. Entire businesses exist to stitch these prices together, route orders and sell access to the fastest information. Fragmentation has become somebody’s business model.</p><p>Intercontinental Exchange, the owner of the New York Stock Exchange, reported $9.9 billion in net revenue for 2025. Its exchange division produced $5.4 billion, including about $1.03 billion from data and connectivity and $495 million from listings. Another $2.4 billion came from its fixed income and data services division. These figures show why the future of exchanges cannot be understood by looking only at trading commissions. Exchanges sell information, access, trust, market position and the ability to stand at the centre of financial activity.</p><p>Crypto offered an early glimpse of another way. It showed that markets can remain open 24 hours a day, that buyers and sellers can participate across borders and that ownership can move through a shared digital record. Centralised crypto exchanges made global trading feel familiar, while decentralised exchanges showed that buyers and sellers could trade through software without handing the same level of control to one operator.</p><p>I do think crypto changed the conversation for good. It forced traditional finance to ask why a share should stop trading at 4 pm. when the company behind it keeps operating, when overseas investors are awake and when important news can arrive on a Sunday. It also made people question why an asset and its payment cannot move together, why ownership must pass through several separate databases and why moving money between markets still feels so slow.</p><p>But crypto did not settle the harder questions. A token that follows the price of a stock is not automatically a share in the company. Holding a digital claim issued by an offshore platform does not guarantee voting rights, dividends, protection in bankruptcy or a recognized place on the shareholder register. If the platform fails, the token holder may discover that they owned an unsecured promise from the platform instead of a direct stake in the business.</p><p>Decentralized markets have their own problems. Trading is split between different networks and pools. Prices can depend on outside data feeds. Public transactions can expose buyers to traders who jump ahead of them. Bridges between systems have been hacked. A pension fund cannot simply explain to its trustees that the shares disappeared because a piece of code failed. Once public companies, retirement savings and national financial systems are involved, someone has to answer for ownership, custody and mistakes.</p><p>In my opinion, the future will borrow from both sides. It will take the continuous access and shared records that crypto introduced, while keeping the legal protections, supervision and financial safeguards that stock markets spent centuries building. The first signs of that combination are already appearing.</p><p>In January 2026, the New York Stock Exchange announced a proposed platform for tokenized securities with 24/7 trading, fractional ownership, immediate settlement and funding through stablecoins. The plan would combine its existing order-matching technology with digital post-trade records, subject to regulatory approval. In March, the NYSE named Securitize as its first proposed digital transfer agent partner, with a focus on shares issued directly under the authority of the company. That second announcement tells us more about the future than the first. Trading around the clock is useful. Knowing who legally owns the share is the foundation of the whole market.</p><p>Nasdaq has also proposed extending US equities trading to 23 hours a day, five days a week. The SEC announced in July 2026 that it would hold a public roundtable on the move toward 24-hour equity trading. Across the industry, the question has already shifted from whether markets will stay open longer to how the rest of the financial system will support them once they do.</p><p>There is an obvious problem hiding inside that ambition. A share trading on Sunday needs money that can settle on Sunday. It needs a custodian that can update ownership on Sunday, a clearinghouse that can monitor risk on Sunday and a bank that can supply collateral on Sunday. If a Japanese investor buys an American stock using yen, the foreign exchange transaction also needs to work. Otherwise, a 24-hour market becomes a screen showing trades that the rest of the system will have to repair on Monday morning.</p><p>This is where the idea of instant settlement deserves a closer look. It sounds like an improvement because we have learned to associate speed with progress. But settlement exists for a financial reason, and moving every transaction through the system immediately can increase the amount of money needed to keep markets functioning.</p><p>DTCC reported that, on an average day in 2024, it reduced $2.2 trillion of trading activity to just $33.5 billion of final settlement obligations through netting. That is a 98% reduction. The shift to next-day settlement also cut the clearing fund by an average of $3 billion, or 23%. Both figures point to the same lesson. Faster settlement can lower some risks, while settling every trade one by one would remove a large part of the savings created when purchases and sales are combined.</p><p>Imagine a market-making firm buying and selling the same stock thousands of times in one afternoon. Under the current system, many of those positions cancel each other out before cash and shares finally move. If every trade had to settle immediately, the firm could need far more cash, more shares and more collateral throughout the day. That expense would flow back into wider trading spreads and higher costs for investors. A system designed to make markets faster could leave them less liquid.</p><p>I believe the next exchange will offer different settlement speeds for different situations. A cross-border transaction between unfamiliar parties may need immediate final payment. A busy trading firm may benefit from having hundreds of transactions combined every few minutes. A pension fund may choose same-day delivery, while a government bond pledged as collateral could move within seconds. Settlement speed should follow the transaction’s risk, funding needs and effect on market liquidity.</p><p>DTCC itself is moving in that direction. In 2026, it announced plans for tokenized securities that could operate across more than one digital network, and by July it reported completed US securities transactions involving tokenized assets. Its role is especially important because ownership in the US stock market already passes through institutions connected to its existing systems. Any lasting change will need to carry those existing legal rights forward.</p><p>The same pattern is appearing in corporate ownership records. In May 2026, digital asset company Bullish announced a $4.2 billion agreement to acquire Equiniti, a transfer agent business serving nearly 3,000 issuers and 20 million shareholders while processing around $500 billion of annual payments. The deal is expected to close in 2027, subject to approval. I suspect this will come to be seen as an early sign of where the power is moving. If you want to build the next stock exchange, owning the system that records the shareholder may be just as valuable as owning the system that matches the trade.</p><p>This brings me to what I call the <strong>Sovereign Exchange Mesh</strong>.</p><p>I use the term to describe a connected global market where trading, ownership, payments and settlement can move between authorized local nodes, while every country keeps control over its own laws, currency, investor protections and supervision. The market becomes global at the level of liquidity and access. It remains local at the level of legal authority. Its governing rule is simple: liquidity follows the asset, while legal authority stays with the node.</p><p>I do not think the world ends up with a single company owning every exchange. Governments are unlikely to hand control of domestic capital markets to a foreign operator, and central banks will continue to decide how their own currencies settle. Sanctions, capital controls, tax rules and shareholder rights will remain national concerns. The more believable outcome is a shared market that works through many connected institutions. One market, many legal homes.</p><p>Under this model, New York, London, Singapore, Tokyo, Dubai and Sao Paulo could each operate as a regulated node in the same wider trading network. A local exchange would continue to supervise companies and trading activity inside its jurisdiction. A local custodian would hold assets. A local bank or central bank system would handle the currency. A shareholder register would confirm who owns the security. The nodes would share enough information and common standards to allow a trade in one country to connect with capital in another.</p><p>Suppose a pension fund in Toronto wants to sell shares in a Brazilian company to a buyer in Tokyo. Today, that transaction can involve several brokers, custodians, currency conversions, settlement instructions and layers of communication. In a Sovereign Exchange Mesh, the seller could enter through a Canadian node, the buyer through a Japanese node and the share through its Brazilian ownership register. The payment would move through approved currency channels, while the Brazilian market would retain authority over the company and its shareholders. Each country would keep its own gate. The gates would open into the same market.</p><p>I think Europe is already giving us an early preview. In July 2026, the European Securities and Markets Authority authorized EuroCTP to combine trading data for shares and exchange traded funds into a single stream. EuroCTP was created by a group of 15 European exchange operators. The exchanges remain separate, but investors can begin to see their activity through one common view. Once markets share a view of prices, the next question is whether they can share access to the buyers and sellers behind those prices.</p><p>The foreign exchange market also offers a clue. The Bank for International Settlements estimated that currency trading reached $9.6 trillion a day in April 2025, with the US dollar appearing on one side of 89% of transactions. Yet currencies do not trade on one giant global exchange. They move through a connected network of banks, dealers and trading centres. The market is global because its participants can find one another across institutions, even though the institutions remain separate.</p><p>Of course, shares are harder than currencies. Every share carries a relationship with a company. Someone must receive the dividend, vote at the annual meeting, pay withholding tax and respond if the company merges, splits its stock or goes bankrupt. These obligations cannot be replaced by a price feed. They have to travel with ownership as the share moves across the network.</p><p>This is why I think the shareholder register becomes one of the most valuable pieces of financial property in the coming decade. Once the official ownership record can be updated across approved markets, trading venues no longer need to create their own separate versions of the same share. A company could issue one recognized security and allow it to trade through many authorized entry points. Liquidity would gather around the asset itself, even though investors arrive through different countries and institutions.</p><p>The same idea extends to money. Stablecoins may help investors move funds outside banking hours, but a national stock market cannot depend entirely on the promise of a private token issuer. The BIS has proposed a system built around tokenized central bank reserves, commercial bank deposits and government bonds. Its Project Agora has also tested cross-border payments using tokenized central bank and commercial bank money. I believe a future exchange network will support several forms of approved money, with each country deciding which forms can settle transactions inside its own borders.</p><p>The design would also change how companies raise capital. Today, a business often chooses between listing locally, moving to a larger exchange or maintaining listings in several countries. A shared market could allow a company to keep its home listing while reaching investors across approved foreign nodes. That could help smaller markets hold on to domestic companies that currently leave in search of deeper capital pools. A business in Nairobi or Jakarta would still answer to its home regulator while gaining access to buyers who previously found the market too difficult or expensive to enter.</p><p>I can also imagine the largest companies becoming almost permanent global listings. An American company could trade through Asian nodes during Asian hours, European nodes during the European morning and US nodes during the American session, with ownership tied back to the same recognized register. The shares would follow the sun while the company stayed rooted in one legal home.</p><p>This does not mean every moment would produce an equally useful price. A stock trading at 3 am in its home market may have far fewer buyers and sellers than it does during the main session. A small overnight order could push the price around, especially when company executives, analysts and institutional investors are asleep. I do think the opening and closing auctions will survive because they bring large numbers of participants together and produce prices used by index funds, pensions and exchange traded funds. A market can stay open all day while still having shared moments when everyone agrees to gather.</p><p>Privacy would survive too. A pension fund trying to buy $500 million of a stock cannot display its full intention to the entire planet without inviting other traders to push up the price. Some orders will need to remain hidden or appear only to approved counterparties. The future system will have to make prices visible enough to keep markets honest while giving large investors ways to trade without exposing their entire position.</p><p>There is also a physical limit to the idea of one global order book. A trading signal still takes time to travel between New York and Tokyo. No software can remove the distance between continents. Local nodes will continue to match orders close to where they are placed, while market-making firms and automated trading systems keep prices connected across regions. The market can become economically connected even when its computers sit on opposite sides of the world.</p><p>AI will accelerate this shift because machines have no reason to observe the opening hours of New York or London. A trading agent can scan hundreds of markets, watch currency prices, calculate settlement costs and decide which local node offers the best route. Another agent can monitor a company’s shareholder register, adjust collateral and hedge the position before a human trader has finished reading the news.</p><p>I believe this will change the customer of the exchange itself. For centuries, exchanges were built around human brokers and human schedules. The next generation may be built around software agents acting for funds, banks, companies and individuals. An investor might simply tell an agent to buy exposure to a business at the lowest total cost. The agent would choose the venue, the currency, the settlement window and the amount of collateral to use. The exchange becomes a set of connected services that financial software can call whenever it needs to trade.</p><p>That opens another possibility. Stocks, government bonds, company debt, private shares and fund interests could eventually travel through parts of the same network. A trader might sell a stock, move the proceeds into a government bond and post that bond as collateral for another position without waiting for three different systems to catch up. The promise here is capital that can move between uses with less delay, while each asset keeps the legal rules that belong to it.</p><p>Still, putting an asset on a digital network does not guarantee that buyers will appear. A private company share can trade 24 hours a day and remain hard to sell if investors do not trust the valuation or lack information about the business. A thinly traded corporate bond does not become liquid simply because its ownership record moves faster. Liquidity comes from confidence, information and people willing to take the other side of a trade. Technology can bring those people together. It cannot manufacture their appetite for risk.</p><p>The business winners may also be different from what many investors expect. Companies that control trusted ownership records, clearing systems, approved payment channels and market data could gain more influence than a new trading app with a polished interface. Exchanges may compete to become the preferred local nodes for global capital. Transfer agents and custodians may become gatekeepers to recognized ownership. Banks may earn new fees by making cash and collateral available across borders and outside normal working hours.</p><p>I believe countries would compete differently too. A jurisdiction with reliable courts, sensible taxes, dependable custody and cash available around the clock could attract global trading without winning every company’s listing. Capital would gather around the nodes offering the strongest combination of trust, access and funding. The exchange would become an extension of national economic policy, and a country’s legal system could shape its cost of capital as directly as its interest rates.</p><p>At the same time, some old revenue streams could come under pressure. If investors can compare prices across a shared network, it becomes harder to charge high fees simply for showing them where the market is. If one ownership record can support trading across many venues, duplicate wrappers and separate listings become less attractive. The struggle will be over who controls the standards, the data and the connection between a trade and the official record of ownership.</p><p>There are dangers here too. A more connected market can spread panic faster. A failure in one settlement node could affect investors across several countries. A cyberattack on a widely used ownership system could create confusion about who holds what. Governments would need ways to isolate local problems, pause trading and protect their own financial systems without shutting down the wider network. A global market can only earn trust if the failure of one part does not bring down the rest.</p><p>I suspect the transition will happen quietly. First, longer trading hours. Then common price feeds, tokenized versions of existing shares, digital shareholder registers and bank money that can move outside normal business hours. After that, approved links between exchanges, custodians and central bank systems. Investors will keep entering through familiar local institutions, even as those institutions become connected to a much larger pool of global capital.</p><p>Perhaps that is how the next stock exchange arrives, without a new building or a bell ringing on its first morning. Old exchanges, banks, clearinghouses and ownership records would gradually recognize one another across borders. The national exchange would still exist, serving as one trusted entrance into a market that never fully closes.</p><p>I believe the Sovereign Exchange Mesh is where all of these trends are heading. Global capital wants access to the deepest possible pool of buyers and sellers. Governments want control over their own financial systems. Companies want wider access to investors without losing their legal home. Investors want their ownership protected, their cash available and their trades completed without unnecessary delay.</p><p>The stock exchange of the future may satisfy all of those demands at once. It may be one market in the economic sense, spread across many local nodes, with ownership, money and regulation still anchored to the places that give them legal force. The world would keep its national exchanges, while the market itself gradually becomes global.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=593f766505a2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Electrostates: The countries that turn electricity into intelligence will inherit the next economy]]></title>
            <link>https://startupsapience.medium.com/electrostates-the-countries-that-turn-electricity-into-intelligence-will-inherit-the-next-economy-edf18f3cf510?source=rss-75bd4ad3fb37------2</link>
            <guid isPermaLink="false">https://medium.com/p/edf18f3cf510</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[investing]]></category>
            <dc:creator><![CDATA[Startup Sapience]]></dc:creator>
            <pubDate>Fri, 21 Aug 2026 16:10:55 GMT</pubDate>
            <atom:updated>2026-08-21T16:14:06.269Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*inYKcnFPgLf6mNwFKZ0SvA.png" /></figure><p>For most of the past century, a country’s importance could often be understood by looking at what lay beneath its soil. Oil built fortunes, funded governments, shaped currencies and decided which places the rest of the world could afford to ignore. The countries sitting above large reserves discovered that geology could become foreign policy. Tankers carried that advantage across oceans, and an entire global order grew around the movement of fuel.</p><p>I think another kind of state is beginning to take shape. Its strength will come from the ability to produce electricity, move it through a reliable grid, convert it into computation and sell the resulting intelligence across borders. Its most valuable export may never appear on a ship. It may leave through a fibre cable as a medical diagnosis, a financial forecast, a drug design, a piece of software or the decision of an autonomous agent.</p><p>The word electrostate already appears in discussions about electrification and clean energy manufacturing. China is often described as the first major example because of its position in solar panels, batteries, electric vehicles and power equipment. That definition describes an important stage of the transition. I believe the next stage deserves its own meaning. A <strong>computational electrostate</strong> is a country that turns domestic electricity into globally tradable intelligence and keeps enough control over the process to capture the wealth it creates.</p><p>Electricity is hard to store, expensive to transport over long distances and usually trapped inside regional grids. A country with excess hydroelectricity cannot easily send that power across an ocean. A geothermal field in East Africa cannot supply a factory in Europe without a vast chain of transmission and interconnection. AI gives electricity another way to travel. The energy stays where it is generated, runs a data centre and leaves the country as intelligence. A stranded electron becomes an export.</p><p>I find this idea more important than most of the debate around the latest AI model. For generations, countries exported what nature gave them or what factories could physically manufacture. AI allows them to export the work performed by energy itself. The output might be legal research, financial analysis, protein modelling, autonomous coding or thousands of digital workers serving clients somewhere else. The customer never needs to see the power plant. The power plant becomes part of the customer’s invisible workforce.</p><p>The scale is already moving beyond ordinary technology infrastructure. The International Energy Agency estimates that global data-centre electricity consumption reached around 485 terawatt-hours in 2025 and could rise to roughly 950 terawatt-hours by 2030. Power use by AI-focused facilities is expected to triple over the same period. By 2027, a single advanced server rack may draw as much peak power as 65 households. This is what happens when thought becomes an industrial activity.</p><p>The old petrostate lived from a stock. Oil sat underground until someone extracted it, and each barrel removed part of the reserve. The electrostate lives from a flow. Electricity has to be produced continuously, balanced through the grid and delivered when the machines need it. Its wealth depends on uptime, coordination and the ability to keep the entire system working. A country can inherit an oilfield. It has to build an electrostate.</p><p>That difference makes government competence more valuable. A functioning computational electrostate needs dependable generation, transmission, cooling, fibre, chip access, legal stability, capital and customers. Weakness in any one of those areas can reduce the value of everything else. Cheap electricity is of little use if transformers take years to arrive. A beautiful data centre cannot run without chips. The strongest model cannot serve foreign customers if the host country has poor connectivity, unpredictable rules or sanctions risk.</p><p>The number that counts is deliverable power. A gigawatt of solar capacity cannot support a gigawatt of continuous computing without storage, backup generation or a wider grid able to fill the gaps. Wind, hydro, nuclear, geothermal and gas each bring a different pattern of reliability and cost. Countries will compete on their ability to shape those resources into uninterrupted electricity at a predictable price. An attractive headline tariff means little if the machines go dark every evening or the cooling system fails during a heatwave.</p><p>I think the best way to understand the system is as a chain. Energy becomes electricity. Electricity becomes compute. Compute becomes intelligence. Intelligence becomes decisions, services and income. The country that owns only the first link earns the smallest share. The country that controls the later links begins capturing the larger economic surplus. This is where the difference between an energy exporter and a computational electrostate becomes visible.</p><p>An oil refinery earns a margin between the crude it buys and the products it sells. The electrostate will develop its own version of that margin. I would call it the <strong>electron-to-intelligence spread</strong>: the value of the intelligence produced from a unit of electricity after deducting generation, cooling, hardware depreciation, financing and payments to the owners of the models and chips.</p><p>Two countries can have the same power price and very different spreads. One may own its data centres, train domestic models, serve global customers and keep the profits at home. The other may provide subsidised electricity to a foreign operator that imports the hardware, owns the models and books the revenue elsewhere. Both appear successful when measured by megawatts. Only one is building an enduring national industry.</p><p>That brings us to a second question: how much of the value remains inside the country? I think of this as the sovereign capture rate. It includes local ownership, tax receipts, domestic model development, skilled employment, affordable access for local companies and the profits retained by national institutions. A country can have a large computational footprint and a very low capture rate. It can host the machines while someone else owns the minds.</p><p>The United Arab Emirates offers a first glimpse of what a more ambitious model looks like. G42, OpenAI, Oracle, Nvidia, Cisco and SoftBank are developing Stargate UAE, a one-gigawatt compute cluster inside a planned five-gigawatt AI campus in Abu Dhabi. The first 200 megawatts were targeted for 2026. The country is combining sovereign capital, energy, international partnerships and a desire to build domestic AI capability. Abu Dhabi’s digital strategy also aims to deploy more than 200 AI solutions across public services, supported by AED 13 billion of investment.</p><p>Saudi Arabia is following a similar path. HUMAIN, owned by the Public Investment Fund, has announced plans with Nvidia for AI factories reaching up to 500 megawatts, using hundreds of thousands of advanced processors over five years. Qualcomm is separately working with HUMAIN on 200 megawatts of infrastructure aimed at global AI inference. The ambition is clear enough. The kingdom wants to turn its energy, capital and geopolitical relationships into a position inside the next global production system.</p><p>There is a certain historical symmetry here. The countries enriched by oil are trying to use their remaining hydrocarbon wealth to acquire the infrastructure of an economy that may depend less on oil. Their first inheritance came from geology. The second has to be assembled from power plants, chips, data centres, talent and access to foreign technology. A successful transition would allow them to export intelligence long after the last barrel has lost its strategic importance.</p><p>Yet the Gulf also exposes the weaknesses of the model. Heat increases the cooling burden. Water is scarce. Advanced chips remain subject to American export controls. The best models may be owned elsewhere. Sovereign money can purchase buildings and electricity, but access to the rest of the system depends on diplomacy. A country that controls its power supply can still discover that another government controls the processor required to convert that power into intelligence.</p><p>China approaches the problem from the opposite direction. It has spent years building renewable generation, batteries, power equipment, transmission lines and a domestic industrial base. Its East Data, West Computing programme links demand in wealthy eastern regions with computing hubs in the interior, where land and energy are more available. By mid-2024, the government reported more than 200 billion yuan of investment connected to eight national computing hubs and over 1.95 million data centre racks. China is attempting to control the chain from electricity and equipment to compute and industrial use.</p><p>The challenge for China sits at the semiconductor frontier. Its system is strong in power equipment, manufacturing and state coordination, while the most advanced foreign chips face export restrictions. The United States has the opposite advantage in leading AI models, chip design and financial markets, while struggling with grid congestion, slow permitting and rising electricity costs. Each country owns part of the future the other needs. The contest between them may be decided by which side repairs its missing link first.</p><p>Texas shows that an electrostate does not have to be an independent country. It has abundant gas, large wind and solar resources, industrial experience, capital markets, land and a power system managed largely within its own borders. The Dallas-Fort Worth, Houston and Austin-San Antonio corridor, together with the westward extension towards Abilene, can connect energy production to data centres, finance, semiconductors and enterprise demand. It behaves like a subnational electrostate forming inside the wider American economy.</p><p>The pressure is already changing how its grid is managed. In June 2026, Texas approved a new process for assessing large electricity users in groups, covering projects of 75 megawatts and above. ERCOT described it as the first such system among American grid operators. The point is to understand the combined effect of data centres before they connect and to identify the transmission upgrades the system will need. Computational ambition has reached the stage where it must negotiate directly with the physical limits of the grid.</p><p>Canada and the Nordic countries enter the picture through a different set of advantages. Hydro, nuclear, cool climates, water and political stability make them attractive locations for machine industry. In Sweden, Fortum has agreed to supply an EcoDataCenter facility hosting CoreWeave with electricity connected to its hydroelectric plants along the Dalälven river. The chain is easy to picture: river water turns turbines, electricity runs chips, and the resulting intelligence is sold into a global market.</p><p>Quebec is already asking how much of that value should belong to its citizens. Hydro-Québec expects demand from data centres to rise roughly sevenfold to more than 1,000 megawatts by 2035. It has proposed a new rate of around 13 Canadian cents per kilowatt-hour for large facilities, roughly double the rate paid by existing large-power customers. Projects above five megawatts remain subject to a selection process. The province is deciding that cheap electricity is a public asset, and that companies consuming it should share more of the benefit with the people who own the resource.</p><p>This may become one of the central political questions of the electrostate era. Who owns the surplus created when public energy becomes private intelligence? A government can attract foreign investment by selling electricity cheaply. It can also end up subsidising machines that serve overseas customers while households face higher bills. The first serious political crisis of an electrostate may arrive through a family electricity statement.</p><p>There is a broader danger I would call electrical Dutch disease. In the familiar version, a resource boom pushes up the currency and makes other exports less competitive. In the electrical version, foreign demand for compute bids up power prices, absorbs transmission capacity and makes life harder for domestic manufacturers, small businesses and households. A country can become richer on paper while weakening the industries that gave it a balanced economy. The more valuable its electricity becomes to foreign machines, the more carefully it must decide who receives the next available megawatt.</p><p>Finland offers a more constructive version of the social contract. Fortum has started operating heat pump plants beside planned Microsoft data centres near Helsinki. Once the data centres are connected, excess heat is expected to cover roughly 40% of the district heating needs of around 250,000 people. A machine producing intelligence for distant customers also helps warm homes nearby. The energy creates value twice, and the local community participates in the result.</p><p>I think the places most likely to surprise us will be countries whose energy was previously difficult to export. Kenya has reached around 1,000 megawatts of geothermal capacity, making it a leading producer in Africa. Geothermal provides steady power without depending on sunshine or wind. If Kenya can build fibre connectivity, reliable grids, data centres and trusted legal frameworks, it could sell computational services far beyond the region where that electricity is generated. Its geothermal fields would become a source of digital exports.</p><p>India has a different opportunity because it combines a huge domestic market, technical talent, growing renewable capacity and a government interested in sovereign digital infrastructure. Adani has announced a $100 billion plan to develop renewable-powered AI data centres by 2035, building on a two-gigawatt platform and seeking to attract another $150 billion across the wider ecosystem. The importance lies in the attempt to connect generation, grid infrastructure, data centres and domestic users. India does not need to rely entirely on foreign demand if its own businesses and public institutions become major consumers of locally produced intelligence.</p><p>This introduces a contrast between the export electrostate and the domestic electrostate. One earns foreign income by selling intelligence abroad. The other gains by reducing its dependence on imported models and applying local compute across healthcare, education, public administration and industry. The strongest countries may do both. They will sell intelligence into foreign markets while giving domestic businesses access to cheaper and more secure computational capacity.</p><p>There is a compounding effect here that oil never offered. Each new customer can generate domain knowledge, evaluations and operational feedback that improve later systems, provided the data can be used lawfully. A country serving hospitals develops stronger medical models. A country serving banks builds a deeper understanding of financial behaviour. The electricity is consumed once, yet the knowledge created around its use can survive and improve. The electrostate with the strongest feedback loop may keep widening its advantage even after other countries match its power price.</p><p>The weaker version is the computational colony. It has electricity, land and a welcoming government, yet foreign companies own the chips, models, cloud platform and customer relationship. The host keeps some construction jobs, property taxes and utility payments. The greater share of the profit leaves. A country may celebrate the opening of a billion-dollar campus without noticing that it has become the low-cost energy base for another country’s intelligence industry.</p><p>This is where ownership becomes more important than the number of megawatts announced. A five-gigawatt campus sounds impressive, but its national value depends on who owns the hardware, who controls the models, where the income is booked and whether local companies can use the capacity. A small country with fewer data centres and a stronger position in software, models and financial services may capture more value than a larger country that provides cheap power alone.</p><p>The financing will also change. Power plants, data centres and chips have different useful lives, yet investors are being asked to fund them as one connected project. A reactor may operate for decades. A data-centre building may last twenty years. A leading processor can lose much of its competitive value in a few product cycles. A project that looks safe when all three are bundled together can carry an awkward mismatch underneath.</p><p>Morgan Stanley expects global AI-related debt issuance to approach $570 billion in 2026. Nvidia is working with major financial institutions on a proposed $500 billion compute-financing effort. These numbers suggest that national AI ambition will soon depend on the same bond markets used to finance governments, housing and infrastructure. A country with cheap electricity and an expensive cost of capital may struggle against a country with slightly higher power costs and easier access to long-term financing.</p><p>Perhaps sovereign wealth funds will begin underwriting entire intelligence corridors, from generation to transmission to compute. Governments may issue bonds backed by long-term data-centre power contracts. Banks may finance future inference revenue as they once financed oil cargoes. Energy traders could follow the electron-to-intelligence spread across countries, comparing the value of local power with the income earned by the models running on it. A nation with a widening spread would be moving up the chain. A nation with a shrinking one might be giving away its strategic advantage.</p><p>There is also a currency question. Oil exporters earned dollars because oil was priced internationally. Countries exporting AI services may earn whichever currency governs the models, cloud contracts and customer relationships. If American companies own the dominant platforms, the dollar may remain embedded in the trade even when the electricity comes from another continent. China will try to build a competing system around its own technology and financial relationships. The struggle over the currency of intelligence could become as important as the struggle over the currency of oil.</p><p>The security risks are different too. An oil exporter worries about pipelines, shipping lanes and ports. An electrostate worries about substations, fibre cables, cooling systems, semiconductor deliveries and the software controlling its grid. A damaged transformer can interrupt a much larger quantity of economic activity than its purchase price suggests. A chip export restriction can leave a country with abundant electricity and no efficient way to convert it into high-value intelligence. The new border may run through the power system.</p><p>That also means an electrostate can become vulnerable to its own success. Large data centres create concentrated demand. When too much capacity is connected before generation and transmission are ready, electricity prices rise and reliability falls. PJM, the largest American grid operator, has already introduced emergency procedures involving backup generation at data centres and other large users. The more a country depends on machine intelligence, the more expensive an interruption becomes.</p><p>The environmental question will not disappear. The IEA estimates that renewables will meet a large share of new data centre demand, yet natural gas, coal and nuclear will also remain important. A country can become an electrostate while increasing emissions if it relies heavily on fossil generation. Another can build a cleaner system around hydro, geothermal, nuclear and storage. I suspect customers and governments will eventually distinguish between the carbon content of intelligence produced in different places, much as they already examine the origin of physical goods.</p><p>The labour question may prove harder. A vast AI campus can absorb billions of dollars and employ only a modest permanent workforce. Its machines can then perform work previously carried out by people elsewhere. The country hosting the campus gains investment and tax revenue, while the wider labour market absorbs the disruption. If ownership is concentrated, the benefits collect around a small number of companies and state institutions. The old resource curse could return in electrical form.</p><p>The successful electrostate will need a way to circulate the surplus. That might mean public ownership of generation, a national investment fund, discounted compute for local businesses, technical training, lower household bills or a direct citizen dividend. Quebec’s electricity pricing and Finland’s heat-recovery projects offer early hints. The deeper principle is that a national resource should create a national benefit even when the immediate customer sits elsewhere.</p><p>I believe the countries that do this well will begin to look very different from the countries that merely attract data centres. They will own a larger share of their generation, protect grid reliability, negotiate access to advanced chips, develop domestic models, keep profits within their tax base and use the resulting intelligence across their own economies. Their strength will come from control over the whole chain.</p><p>By the middle of the 2030s, I expect governments to measure intelligence exports, contracted compute capacity and the share of domestic electricity devoted to machine work. Rating agencies may begin asking whether a country’s AI infrastructure is supported by firm generation or exposed to volatile power markets. Trade negotiations may include chip access, model access and cross-border compute rights. Sovereign wealth funds may own data centres as strategic national assets. The language of energy policy and the language of industrial policy will start sounding like the same conversation.</p><p>Some countries will rise because they have abundant power and know how to convert it into valuable services. Others will discover that cheap electricity is only the beginning, and that the largest profits belong to whoever owns the models, customers and decisions at the end of the chain. The map of power will change with them.</p><p>For a century, nations became important because the world needed what lay beneath their land. In the coming decades, some will become important because the world needs what their electricity can think. The next great resource may be the intelligence a country can produce before its energy disappears into the air.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=edf18f3cf510" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>