The AI Bubble Is Not the AI Economy

The Claim Worth Taking Seriously

A man who publishes under the name Asian Dad Energy posted a video this summer predicting that the current AI investment boom is approaching a crash. He accepts that AI will improve and eventually transform the economy, but sees the replacement of expensive human labor as the only source of value large enough to support the extraordinary amount of capital being committed to chips, model training, data centers, electrical capacity, and the other infrastructure required to produce and operate the most capable systems. Because mass white-collar displacement is not occurring quickly enough, he doubts that the investment can be repaid.[1]

His second argument imposes an even narrower threshold. If AI is not independently producing radically new technologies—something on the order of “new physics”—then he sees no other plausible source of productivity large enough to justify the buildout. AI must either remove millions of people from payrolls or produce some spectacular technological discontinuity. If it does neither, the benefits may be real but economically insufficient.[1] I am not singling Asian Dad Energy out here; his argument is useful because it is a particularly clear example of a broader category of AI-bubble claims built on the same underlying assumptions.

There is a serious investment concern underneath both arguments. The costs arrive before the full benefits do. Data centers have to be built, accelerators purchased, electrical capacity secured, engineers and researchers paid, and financing serviced while many of the applications are still being discovered. Open models and increasingly capable smaller systems may weaken the pricing power of companies spending heavily to remain at the frontier. Falling inference prices could distribute enormous benefits to users while leaving some of the firms that financed the capacity with mediocre or disastrous returns.

My suspicion is that even a severe repricing will look less like the popping of one coherent bubble than like an uneven shakeout. That is not a prediction that AI stocks cannot fall sharply, that data-center construction cannot slow, or that investors will not lose enormous amounts of money. It is a claim about structure. The buildout rests on many different applications, business models, assets, and sources of repayment. The failure of some of them does not tell us what happened to all the others.

To see why, it helps to begin with the accounting.

The Wrong Ledger

The phrase “AI must repay the investment” sounds straightforward until one asks who, exactly, is supposed to repay whom.

There is no single AI corporation that receives all the money spent on models and data centers, deploys the resulting systems throughout the economy, collects every benefit, and then settles one consolidated account. There are chipmakers, utilities, construction firms, data-center operators, cloud platforms, frontier-model developers, application companies, ordinary businesses, public institutions, and individual users. Each incurs different costs, receives different benefits, and has a different ability to charge for the value it helps create.

Three questions are routinely compressed into one. How much economic value does AI create? Who captures that value? And did a particular investment, made at a particular price, earn an adequate return?

A strong answer to the first question does not guarantee a strong answer to the third. A weak answer to the third does not imply a weak answer to the first.

Suppose a company pays $1,000 a year for an AI tool and uses it to produce $10,000 of additional value. The provider captures $1,000. The customer retains the other $9,000. If competition later drives the subscription price down to $500, the provider may have a worse business even while the customer has a better bargain and adoption spreads to many more users. The economic value has not disappeared. Its distribution has changed.

The same separation continues upstream. The application company spends only part of its revenue on model access. The model provider spends only part of that revenue on inference infrastructure. A cloud company may own the data center, lease capacity from someone else, or provide infrastructure to a laboratory in which it also owns an interest. The chip manufacturer may have earned a terrific return before the company that borrowed money to buy the chips discovers that its debt cannot be serviced.

There is no reason to expect the financial result at one layer to summarize the economic result of the entire system.

The distinction becomes even clearer when AI prevents losses. No one records revenue for the ambulance that was not dispatched, the factory that did not shut down, the fraudulent transfer that was blocked, the customer who did not leave, or the month of work that was not lost to an injury. Preventing destruction can actually reduce measured expenditure by eliminating repairs, litigation, emergency response, replacement purchases, and other transactions. The economy is richer because fewer resources were consumed repairing preventable damage, not poorer because the repair spending never happened.

Some benefits will remain with the people buying AI products. Others will pass through lower prices or better products. Others will accrue to people who never entered the transaction at all. A safer vehicle benefits occupants of other vehicles. Better cybersecurity protects customers and counterparties. A better software tool may improve a system subsequently used by millions of people who never know AI contributed to its construction.

None of this means social value somehow pays private bills by magic. A data-center operator needs actual cash flow. A model company needs paying customers, a parent company willing to fund it, or some other source of support. If too little of the value flows back to an investor, that investor loses money no matter how useful the technology becomes elsewhere.

But that is precisely the point. Failure to capture value is not failure to create it.

Nor does enormous downstream value rescue every upstream decision. Two companies can spend billions developing capabilities the market regards as interchangeable. A useful data center can be built at a price its future cash flow cannot support. A technically excellent frontier model can create enormous customer value while its owner discovers that a free model is good enough for most applications.

A large AI economy can coexist with a brutal AI industry.

And that is where the stronger bear case begins.

The Stronger Bear Case

A sophisticated AI bear does not need to deny anything in the previous section.

AI can create enormous economic value and the present capital buildout can still be badly mispriced.

The cash requirements are genuinely immense. In its July 2026 Monetary Policy Report, the Federal Reserve said business fixed investment grew at an annualized 11 percent in the first quarter and that most of the strength appeared connected to infrastructure needed to support AI services. Outside AI-related categories, investment spending was fairly weak on net.[2] A sharp retreat in the buildout would therefore not be merely an event for technology shareholders. It would affect construction, electrical equipment, utilities, semiconductor suppliers, financing markets, and the regions reorganizing around new data-center demand.

The buildout is also putting pressure on cash flow. A July Reuters analysis of consensus estimates concluded that Microsoft, Alphabet, Amazon, Meta and Oracle were on a trajectory to spend more collectively on capital expenditures than they generated in free cash flow by 2027. That is an estimate, not destiny, but it captures the financial question accurately: the infrastructure is consuming cash at a rate that increasingly requires investors to believe future revenue will justify what is being built now.[3]

Some of the demand is also financially entangled. Microsoft is both a major investor and commercial partner of OpenAI, while OpenAI has contracted for an additional $250 billion of Azure services. Nvidia is simultaneously an investor, supplier, and backstop customer of CoreWeave. In addition to a $2 billion equity investment, Nvidia has agreed to purchase up to $6.3 billion of CoreWeave cloud capacity that would otherwise go unsold through April 2032, while CoreWeave itself is built heavily around Nvidia infrastructure and is raising enormous amounts of capital to construct more capacity.[4][5]

There is nothing imaginary about the compute being consumed or the data centers being built. But these relationships mean some flows that look independent are connected through financing, ownership, supply, and customer contracts. A downturn can propagate through relationships that seemed diversified while capital was abundant.

The large technology platforms also share a common factor even though their business models differ. Microsoft, Alphabet, Amazon, Meta, Alibaba, Nvidia, and the frontier laboratories do not all require the same source of revenue. But every one of their AI theses becomes weaker if demand for machine intelligence grows much more slowly than expected, if enterprises sharply ration usage, or if cheap local and open models satisfy most valuable applications without requiring expensive frontier inference.

Diversified business models are not the same thing as uncorrelated cash-flow drivers.

Accelerator economics add another real risk. Amazon shortened the estimated useful life of a subset of its servers and networking equipment from six years to five because of the increased pace of technological development, particularly in AI and machine learning. It also recorded accelerated depreciation associated with equipment it decided to retire early. Meta, in contrast, extended the estimated useful life of most servers and network assets to 5.5 years, while Alphabet generally uses six years for servers and network equipment.[6]

Those different estimates tell us something useful. There is no universal technological expiration date. Companies are continually reevaluating how long particular assets remain economical for the work they perform.

New accelerators can process far more work per watt and per rack than older ones. In a data center limited by available power or floor space, a perfectly functional older GPU may cease to make economic sense because replacing it allows the same megawatt to produce much more valuable output. That is genuine technological depreciation.

But obsolescence is not binary. AWS currently lists P4d and P4de accelerated-computing instances alongside P5 and multiple P6 generations, while still maintaining a separate previous-generation catalog for older equipment. An accelerator can move from frontier training to ordinary training, inference, smaller models, batch workloads, or specialized computing before it finally becomes uneconomic to operate. Meanwhile the building, substation, cooling system, network connection, and power rights around it can have still different useful lives.[7]

A distressed AI data center should therefore be treated neither as dark fiber that will remain valuable for decades regardless of what happens nor as a pile of worthless silicon the moment Nvidia ships another generation. Some of the productive capacity survives a shakeout. Some does not.

Open models pose an even more fundamental challenge to the frontier-model business. If a mid-tier open model combined with good application scaffolding is sufficient for most commercially useful work, AI adoption could accelerate while the economics of frontier development deteriorate. Users get cheaper intelligence. Application companies get lower costs. More projects become viable. Demand for inference may grow.

Meanwhile a laboratory spending billions to move from “very capable” to “slightly more capable than everyone else” discovers that almost nobody is willing to pay enough for the difference.

That outcome would be bullish for the AI economy and disastrous for parts of the AI industry.

The labor story can also be harsher than visible layoffs suggest. Productivity gains can reduce junior hiring, slow replacement hiring, or allow an organization to expand output without expanding headcount. A ten-person team may absorb work that would previously have required twelve people without anyone being fired. Two jobs have still vanished from the counterfactual. The first rung on a career ladder can contract much faster than the installed workforce.

Taken together, these are serious objections. Capital spending is large enough to matter macroeconomically. Free cash flow is under pressure. Some commercial relationships are financially entangled. The largest investors share a common dependence on broad AI demand. Hardware faces real economic obsolescence. Open models can destroy pricing power even while AI use expands. Productivity gains can impose painful labor-market adjustments without appearing first as mass layoffs.

These facts should raise skepticism about particular projects, public-market multiples, financing structures, and the cash conversion of the 2026–2028 investment vintage. They may increase the probability of a difficult year or two for AI-linked assets and investment spending.

What they do not establish is a single imminent event in which the underlying usefulness of AI is suddenly revealed to have been fictitious. They tell us there is plenty of capital to sort. They do not tell us in advance what survives the sorting.

For that we have to look downstream.

Productivity Without Disappearance

The labor-replacement argument assumes something close to a fixed quantity of useful work. If AI makes software development, analysis, design, research, customer support, legal work, and administration cheaper, the gain should appear as fewer people performing the same amount of work.

But organizations do not already perform all the useful work worth doing.

They have enormous backlogs of systems they would build, questions they would investigate, processes they would automate, defects they would repair, tests they would write, documents they would produce, and analyses they would perform if the work were cheaper. Legacy systems survive because replacing them would consume too much engineering capacity. Small companies tolerate manual processes because custom software would cost more than it saves. Researchers abandon lines of inquiry because the preliminary work is too laborious. Managers accept incomplete information because a comprehensive answer would arrive after the decision has to be made.

Reducing the cost of skilled cognitive work changes which work is economical.

A task that cannot justify forty hours of expert attention may become worthwhile when AI reduces the human requirement to ten. A project that would have required a large team may become possible for a small one. A person who could manage three active projects may be able to sustain five.

None of those gains requires anybody to disappear from payroll.

Spreadsheets did not need to eliminate accountants to become economically indispensable. CAD did not need to eliminate engineers. Search engines did not need to eliminate researchers, and compilers did not need to eliminate programmers. Each changed the amount, speed, complexity, or quality of work people could perform.

Human supervision can even become more important as AI reaches more consequential work. A model used to draft boilerplate code can operate with relatively little oversight. A model recommending a security response, interpreting evidence for a legal decision, or contributing to a safety-critical system will be supervised precisely because the output matters.

The workflow can change profoundly while the org chart changes slowly.

The same point applies to the “new physics” threshold. Most technologies do not justify themselves by discovering another law of nature. They justify themselves by applying knowledge more effectively. Electricity became economically transformative through motors, lighting, transmission, factory redesign, controls, appliances, and thousands of applications. Computing became transformative through databases, networks, operating systems, development tools, spreadsheets, logistics systems, and applications no semiconductor engineer had anticipated.

AI does not need to invent another technology before it becomes one.

Bad Futures That Do Not Occur

Tesla’s Full Self-Driving system is useful here because it is already a real AI application attached to an expensive physical product. In the second quarter of 2026, Tesla reported that more than 55 percent of its new North American deliveries included an FSD subscription. That is not fleet penetration, and it tells us nothing by itself about long-term retention. It does show substantial willingness to pay for a concrete AI-mediated capability.[8]

Now impose an intentionally extreme assumption. Suppose Tesla eventually recovers no more than the full cost of developing, training, deploying, and operating FSD. It earns no excess return from the system. Ignore Cybercab and robotaxi entirely. Assume only that FSD eventually produces a meaningful reduction in the frequency or severity of crashes.

NHTSA estimated that motor-vehicle crashes occurring in the United States in 2019 produced roughly $340 billion in tangible economic costs. Those costs included property damage, medical care, lost productivity, congestion, insurance administration, legal costs, emergency services, and workplace losses. People not directly involved in the crashes ultimately bore roughly three-quarters of those costs through insurance premiums, taxes, congestion, and other channels.[9]

The $340 billion still excludes much of what people actually care about. When NHTSA applied standard Department of Transportation valuations to lives lost, pain, disability, and diminished future quality of life, its central estimate of total societal harm approached $1.4 trillion.[9]

Tesla cannot invoice the employer whose worker was not injured, the hospital whose trauma bed remained free, the city whose emergency crews were never dispatched, or the driver who did not spend an hour sitting behind an accident. Tesla’s private economic return could be zero while the system produced a very large recurring social return.

The valuable output is partly a collection of bad futures that fail to occur.

Cybersecurity has the same economic structure. An attacker searches for weaknesses, but a successful breach usually requires some combination of exploitable software, stolen credentials, misconfiguration, poor judgment, weak processes, or systems that fail to behave as intended.

The best breach is no breach.

AI can identify vulnerable code and dangerous configurations, detect phishing and impersonation attempts, recognize unusual account behavior, prioritize weaknesses, and connect warning signs before an intruder gains meaningful access. When prevention fails, it can also help analysts correlate alerts, understand malicious code, reconstruct an attack, and contain it more quickly.

The most defensible empirical evidence is about analyst leverage rather than breaches prevented. Microsoft reported that experienced security analysts using Security Copilot completed assigned tasks 22 percent faster and with 7 percent greater accuracy. That does not tell us how many breaches were prevented. It does show that analysts can use generative AI to process complex security evidence more effectively.[10]

The difficulty of demonstrating prevention is itself part of the economic problem. A system that prevents a ransomware attack produces an uneventful Tuesday. No customer database is copied. Production continues. Nobody spends six months rebuilding systems or repairing identities. No dramatic transaction records the value created.

The buyer still has reason to pay because a single avoided serious incident can justify years of fees. The application captures only part of that expected benefit; customers, suppliers, insurers, employees, and public institutions keep the rest.

That is the beginning of the commercial bridge between social value and private infrastructure.

More Software, and More Things Worth Doing

Software development gives us a shorter path from the new data centers to economic output.

The empirical literature is already telling us that “AI coding productivity” is not one number. Randomized field experiments at Microsoft, Accenture, and an unnamed Fortune 100 company covering 4,867 developers found a combined increase of about 26 percent in completed tasks among developers given coding assistance. METR, working with a much smaller group of highly experienced open-source developers modifying mature repositories they knew intimately, found that early-2025 AI tools made them 19 percent slower. The tasks, users, tools, and quality requirements were very different.[11]

That is not a contradiction requiring one result to be discarded. “Software development” contains many kinds of work.

My own experience is anecdotal, but it illustrates a mechanism that a stopwatch can miss completely: AI can make a project economical enough to exist.

With the help of an AI coding tool, I recently built a data mart in ten weeks. The initial schema was operational within hours. Most of the project was not spent producing the mart itself but validating it.

I selected ten reports that would exercise it heavily and a set of customers representative of the demanding end of our base. I then had the tool construct a harness that ran every report for every selected customer against both the new mart and the legacy aggregation environment.

The harness uncovered undocumented business and data rules in the legacy structures, missing records, records populated in violation of those rules, and mathematical errors in the existing aggregations. Several discoveries forced parts of the mart to be rebuilt. By the end, we had validated the static reporting structures, validated the processes that updated facts and dimensions as source data changed, and accounted for every discrepancy between the old system and the new one.

I estimate that AI saved four to six months of additional engineering effort. That estimate is necessarily counterfactual; I did not build the same data mart twice. But the more important fact is easier to establish.

Without AI, the project would have been too expensive to begin.

AI did not replace a developer. It caused a data mart to exist that otherwise would not have existed.

That mechanism extends beyond programming. A general model can supply language, visual interpretation, synthesis, retrieval, classification, and reasoning, while an application organizes those capabilities around a particular problem.

Be My Eyes provides a useful example. Its platform now connects more than one million blind and low-vision users with more than ten million volunteers and also incorporates AI-enabled visual assistance. The value of identifying a medication, interpreting a document, navigating an unfamiliar object, or obtaining visual information privately cannot be inferred from the marginal cost of the model call.[12]

Khanmigo supplies another version of the same mechanism. Khan Academy has tested ways of giving the tutor structured information about a student’s learning history and prerequisite gaps. Across more than fifteen million tutoring threads, some of those changes improved its immediate next-problem performance while others did nothing. The economic point is not the particular percentage. It is that the application layer—curriculum, history, workflow, evaluation, and pedagogy—changes what the same underlying model can usefully do.[13]

The eventual AI economy need not be dominated by five famous killer applications. It can contain millions of comparatively small systems, each making a particular worker, customer, student, patient, machine, or institution more capable.

Many will fail.

That is not a problem for the argument. Most products have always failed.

The important question is what happens to the successful ones.

A Thin Toll on a Vast Field of Value

Downstream benefits do not automatically pay data-center bills. They have to become somebody’s willingness to pay.

A driver subscribes to FSD. A company licenses a security system because a serious breach could cost far more than years of subscription fees. An engineering organization buys coding tools because the software produced is worth more than the seats. A school buys educational software. An application company pays for inference and incorporates that cost into something else it sells.

The payment does not need to equal the benefit.

The coding tool that helped me complete the data mart did not need to collect the value of the new reporting system or the four to six months of additional effort I believe it saved. It needed to collect enough, along with payments from other users, to support the models, software, operations, and capital required to provide the service.

I kept most of the surplus.

That is normal economics, not a defect in the business model.

Competition can make the relationship more extreme. If inference prices fall, providers may capture less value per unit while applications become economical for many more users. A weaker tollbooth can support a larger road system.

This raises the objection that some model access may currently be priced below its fully loaded cost. The private laboratories disclose too little to determine precisely what a particular token costs after allocating research, unused capacity, infrastructure, and capital expenses. But suppose the critics are right and current prices are subsidized.

A stand-alone model company cannot do that indefinitely. Eventually its revenue has to support its ongoing costs or someone has to keep supplying capital.

But temporary subsidy can also be part of creating the future market.

AWS explicitly gives new customers credits and a free plan intended for experimentation and proof-of-concept development before they make a larger cost commitment. Anthropic likewise offers API credits through startup and venture programs. The point is not that free access somehow becomes profitable. The provider is subsidizing complementary investment: developers learning the platform, experimenting, building integrations, discarding bad ideas, and discovering workloads that eventually become durable paying customers.[14]

A twenty-year-old building the hundredth mobile photo-editing app with AWS credits may never produce a successful photo business. But the failed experiment can still produce a developer who later works for a company that desperately needs somebody who knows how to move large blobs through S3 for a DNA-sequencing pipeline. AWS can benefit whether the original experiment becomes Instagram or tuition.

Model providers are attempting something similar. They are priming the pump from which future inference demand may flow.

Some companies can go much further because selling model access is not their principal business. Mark Zuckerberg has said this explicitly about Meta: selling access to AI models is not Meta’s business model. Meta wants a broad Llama ecosystem because it expects to benefit from better tools, integrations, efficiency improvements, and freedom from dependence on a closed competitor’s platform.[15]

Meta may also have concluded that whatever direct subscription revenue it might extract from Llama is too small to justify the organizational risk and strategic damage of keeping the technology closed.

Alibaba occupies a hybrid position. It distributes Qwen broadly while also selling AI compute, models, and agent services through Alibaba Cloud. In the final quarter of fiscal 2026, Alibaba said AI-related products represented 30 percent of Cloud Intelligence’s external revenue.[16]

Microsoft and Alphabet offer still more routes. Microsoft says its own, imperfectly specified measure of the AI business exceeded a $37 billion annual revenue run rate in its fiscal third quarter of 2026, while its cloud margins and free cash flow were simultaneously being pressured by the infrastructure required to serve that demand. Alphabet reported 82 percent year-over-year Google Cloud growth in the second quarter and a $514 billion backlog driven in large part by AI infrastructure and AI solutions. That backlog is contracted future business, not cash in the bank, and does not tell us the eventual margins on the infrastructure required to fulfill it. But it demonstrates that AI demand is arriving through enterprise cloud relationships, not only consumer chatbot subscriptions.[3][17]

These companies are not making independent bets. They all depend to some degree on continued growth in AI usage. But they possess multiple pathways through which a successful transformation can become revenue or protected profit: model access, cloud consumption, enterprise software, advertising, internal productivity, recommendation, security, and third-party applications.

The upstream stack does not need to capture most of AI’s economic value.

It needs to collect a sufficiently thin toll from a sufficiently vast field of valuable activity.

What a Shakeout Would Actually Mean

That still leaves ample room for spectacular capital-allocation mistakes.

Some data centers will be built at costs their eventual workloads cannot support. Some accelerators will be retired earlier than expected. Others will spend years migrating down the workload hierarchy. Some laboratories will discover that the market treats models costing billions to train as substitutes for models available nearly free. Some application companies will find that users love the product but not enough to pay its cost. Some investors will discover that they bought a strategically sound company at a price that assumed an impossible future.

A severe financial adjustment is therefore entirely compatible with the argument of this essay.

It could also be economically painful outside the technology sector. AI-related investment is now large enough that cancelled projects and falling equipment purchases could reduce aggregate demand.[2] Financing losses could spill into credit markets. Regions that expanded electrical infrastructure and construction around projected data-center demand could discover that some of those projections were fantasy.

There is no need to sanitize the word “shakeout.”

The question is what the shakeout tells us.

A collapse in stock-market valuation is one event. A collapse in capital formation is another. A collapse in the economic usefulness of the technology is a third.

They can occur together, but they do not mean the same thing.

The present evidence gives good reasons to question some projects, some valuations, and some assumptions about the speed at which current investment will turn into cash flow. It does not tell us that the only remaining uncertainty is which week the entire AI complex implodes. Nor does a difficult 2026–2028 investment vintage imply that demand for the underlying capability evaporates.

There are observations that would make me substantially more skeptical of the broader economic case.

If enterprises spend two or three budget cycles experimenting with AI and then systematically reduce rather than expand those budgets when the projects have to survive ordinary P&L scrutiny, that would matter. If inference prices keep falling but usage fails to rise, that would matter. If developers continue using the tools but measurable output and quality improvements remain trivial, that would matter. If companies that claim large internal AI benefits cannot eventually show those benefits in revenue, margins, product quality, cycle times, or some other operating measure, that would matter. If the applications remain perpetually dependent on subsidies because customers will not pay even a fraction of the value their advocates claim, that would matter.

Those are tests of the economic transformation.

A laboratory going bankrupt is not.

Asian Dad Energy is right about several important things. AI can be economically transformative while investors lose fortunes. Infrastructure must be financed before much of the eventual demand is known. Open models and falling prices can make the economics of frontier development much worse than the economics of AI use.[1]

Where his argument becomes too narrow is in treating mass labor replacement as the only value pool large enough to matter, with scientific revolution as the only alternative.

The transformation is likely to be more ordinary than that, and for that reason much broader.

A developer completes a project that otherwise would never have begun. A security system stops an attack before anyone notices. A blind person obtains information that previously required another person’s eyes. A student receives better-targeted help. A vehicle avoids a collision. A business incorporates a little machine intelligence into a process and finds that the process becomes cheaper, faster, safer, or simply possible.

Each event may contribute only a tiny payment to the infrastructure underneath it. The user may keep almost all of the value.

Multiply that pattern across an economy and the thin toll can become substantial.

Some of the people who financed the roads will still have paid too much.

A capital shakeout will sort those investments.

It will not refute the transformation.

Notes

[1] Asian Dad Energy, “AI Is About to Crash. Here’s Why.”, July 23, 2026. The Substack post reproduces the argument made in the accompanying video.

[2] Board of Governors of the Federal Reserve System, Monetary Policy Report, July 2026 — “Capital spending growth has been brisk, reflecting strong demand for investment related to artificial intelligence.”

[3] Patturaja Murugaboopathy and Gaurav Dogra, Reuters, “AI investment boom puts Big Tech’s free cash flow under pressure.”, July 22, 2026.

[4] Microsoft, Fiscal Year 2026 First Quarter Earnings Conference Call. Microsoft stated that OpenAI had contracted for an incremental $250 billion of Azure services.

[5] Reuters, “Nvidia invests $2 billion in CoreWeave, expands partnership.”, January 26, 2026; and “CoreWeave, Nvidia sign $6.3 billion cloud computing capacity order.”, September 15, 2025. The latter agreement obligates Nvidia to purchase qualifying CoreWeave capacity that would otherwise go unsold through April 2032.

[6] Accounting-life disclosures: Amazon, 2025 Form 10-K; Meta, 2025 Form 10-K; Alphabet, 2025 Form 10-K.

[7] Amazon Web Services, “Specifications for Amazon EC2 accelerated computing instances.” The current catalog lists P4d/P4de, P5-family, and P6-family accelerated instances simultaneously, alongside older accelerated families.

[8] Tesla, Q2 2026 Update. Tesla reported that more than 55 percent of new North American deliveries during the quarter included FSD subscriptions.

[9] National Highway Traffic Safety Administration, The Economic and Societal Impact of Motor Vehicle Crashes, 2019, revised technical report. NHTSA estimated $340 billion in tangible economic costs and nearly $1.4 trillion in total societal harm when quality-of-life valuations are included.

[10] Microsoft Security, “Microsoft Copilot for Security is generally available on April 1, 2024 with new capabilities.” The cited randomized study reported that experienced security analysts completed assigned tasks 22 percent faster and with 7 percent greater accuracy when using Copilot.

[11] Zheyuan Cui et al., “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers,” Management Science; and METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.” The studies examine substantially different developer populations and work settings.

[12] Be My Eyes, “Be My Eyes Reaches 1 Million Blind and Low-Vision Users and 10 Million Volunteers.”, March 12, 2026.

[13] Khan Academy, “How Khan Academy Is Building a Better AI Tutor: Our Most Recent Learnings.”, May 2026. The reported product tests covered more than 15 million tutoring threads over six months.

[14] Amazon Web Services, “AWS Free Tier now offers $200 in credits and 6-month free plan to explore AWS at no cost.”, July 16, 2025; and Anthropic, “Claude for Startups.” Both programs explicitly offer credits to support experimentation and product development.

[15] Mark Zuckerberg, Meta, “Open Source AI Is the Path Forward.”, July 23, 2024. Zuckerberg writes that selling access to AI models is not Meta’s business model and argues for an open Llama ecosystem.

[16] Alibaba Group, 2026 Form 20-F. Alibaba reported that AI-related products accounted for 30 percent of Cloud Intelligence Group external revenue in the final quarter of fiscal 2026.

[17] Microsoft, Fiscal Year 2026 Third Quarter Earnings Conference Call, reporting that its company-defined AI business annual revenue run rate surpassed $37 billion; and Alphabet, “Q2 2026 earnings call: Remarks from our CEO.”, reporting 82 percent Google Cloud revenue growth and a $514 billion Cloud backlog.