Imagine a world where every AI model lives up to its hype, every agent works flawlessly, and every demo ships on schedule. Even under those optimistic conditions, the economics of today’s AI spending spree look questionable. The core issue isn’t whether the technology can deliver; it’s whether anyone will actually pay for the astronomical volume of tokens being generated. Analysts estimate that to justify the current capital outlay, buyers would need to purchase roughly $1.2 trillion worth of AI tokens annually, a figure that could swell to $2.5 trillion by 2031. That scale dwarfs most existing software markets and raises a fundamental question: where will the money come from, and will those buyers ever see a return on their investment?
Let’s start with the spending side. Goldman Sachs projects global AI capex at about $765 billion this year, climbing to $1.6 trillion annually by 2031, which sums to roughly $7.6 trillion over the period. The largest tech players are already allocating massive fractions of their revenue to this effort. Microsoft guides toward $175 billion of reported capex, Meta toward $130‑145 billion, Amazon near $220 billion, and Alphabet between $195‑205 billion. Oracle, despite its different fiscal calendar, signals another $95 billion gross. Even when we adjust for reporting differences, the combined outlay lands in the $830‑840 billion range for the coming year—levels that would have seemed absurd just five years ago.
Can operating cash flow cover these outlays? Not comfortably. Alphabet, the strongest cash generator in the group, produced only $53 billion of free cash flow after spending $132 billion on capex over the past year. Oracle ran a $24 billion free‑cash‑flow deficit. Sector‑wide, the Bank of England estimates that the AI buildout will require about $1.5 trillion of external financing, with roughly $800 billion expected to flow from private credit markets. The remainder will come from traditional debt, leases, and various capacity‑booking arrangements, meaning a lot of the money is borrowed rather than earned.
Now consider the payback math. AI hardware does not earn forever; chips lose their premium edge long before accounting depreciation ends, as newer generations undercut them. If we treat the spend as an annual recurring cost, the question becomes: what revenue must the entire AI stack generate each year just to break even on hardware? Assuming sellers retain about 65 cents of each revenue dollar after direct serving costs, the $765 billion of yearly capex demands roughly $1.18 trillion in annual revenue to cover the chips alone. That figure excludes salaries, buildings, interest, and any profit margin. By 2031, the required annual revenue climbs to about $2.5 trillion—a sum that must be sourced from somewhere outside the vendors’ own balance sheets.
A significant portion of today’s apparent demand is actually the supply side buying from itself. Microsoft’s funds flow to OpenAI, whose compute runs on Microsoft’s cloud, and regulatory reviews have shown these partnerships often include clauses requiring recipients to reinvest a large share of the money back into the provider’s infrastructure. Similar patterns appear with Amazon’s investment in Anthropic, Google’s backing of the same firm, and Nvidia’s stakes in CoreWeave. While these transactions generate real revenue, they represent money moving within an ecosystem rather than fresh spending by independent end‑users. This dynamic echoes the vendor‑financing tricks seen during the late‑1990s telecom boom, where companies like Lucent and Nortel lent money to their own customers to inflate order books.
End‑user consumption remains modest. AI‑powered mobile apps pulled in just over $4 billion of in‑app purchases during the first half of 2026—a figure that excludes web subscriptions and API deals but is growing quickly. Even if that pace doubled, it would still be only a few percent of the trillion‑dollar annual token demand. Consumers are real participants, but their wallets are far too small to shoulder the bulk of the compute bill. The market therefore needs a far larger, more systematic source of demand to make the spending add up.
That source turns out to be payroll. In a competitive market, the price a token can command is anchored to the cost of the human worker it would replace. If a company can avoid paying a $100,000 salary by deploying an AI agent that performs the same task, the token’s value is roughly tied to that avoided wage cost. The total pool of wages for work that can be performed remotely—roughly 37 % of U.S. jobs, representing about 46 % of wages because desk jobs tend to pay more—amounts to roughly $6‑7 trillion domestically, and perhaps $18‑25 trillion worldwide after weighting for international differences. A $1 trillion token bill therefore equals about five cents of every addressable wage dollar, a proportion that is mathematically payable but still represents only a third of the total addressable wage pool, leaving a large gap that bulls must fill with entirely new economic activity.
To understand what might happen next, we can look at three distinct scenarios that reveal themselves in hiring and employment data. The first is a wave of layoffs: firms simply substitute tokens for people, cutting headcount and saving wages in a visible, measurable way. The second is a hiring freeze, where the need for new roles never materializes because AI handles the scalable, volume‑driven tasks, leaving only exception‑based work for humans. The third scenario is genuine expansion, where companies use AI to create fresh revenue streams, hire additional staff, and grow faster than before. Each of these pathways leaves a different imprint on the labor market, and early indicators already hint at which direction we are leaning.
Economists often frame the impact of automation through three gates, a concept pioneered by Daron Acemoglu and Pascual Restrepo. Gate one asks whether the technology delivers a big dividend—meaning substantial cost savings—or only a small, “so‑so” improvement. Small savings, like those from self‑checkout kiosks, replace workers without lowering prices, leaving consumers to pick up the slack. Big savings, such as those from the assembly line or containerization, slash prices, expand markets, and often create more jobs than they destroy. Gate two examines demand elasticity: even with huge cost cuts, will people actually want more of the now‑cheaper good or service? If demand is inelastic, the gains flow elsewhere, as happened with agricultural mechanization. Gate three concerns reinstatement: do new kinds of work emerge to absorb displaced labor? Historically, about 60 % of today’s U.S. occupations did not exist in 1940, but this process is slow, and the lag between job loss and job creation is where social pain concentrates.
When we look for concrete signs that the AI dividend is materializing, the evidence is thin. Price surveys show no meaningful decline in the cost of software, legal services, consulting, or other support‑heavy offerings that would be expected if tokens were delivering large‑scale savings. Corporate disclosures reveal only modest efficiency gains—JPMorgan’s cited $600 million of improvements amount to about a third of one percent of its 2025 revenue, a figure more consistent with so‑so automation than with transformative change. Surveys indicate that only a small fraction of large firms credit AI with any measurable revenue growth, and volumes of AI‑related activity (such as surging app store submissions) are not matched by comparable consumer spending.
Behavioral clues also point toward the so‑so outcome. Companies like Uber have had to cap spending on internal AI coding tools after teams burned through monthly budgets in weeks, a reaction inconsistent with accessing a genuine gusher of value. Klarna’s experience—initially boasting that its AI assistant replaced 700 support agents, then re‑hiring humans after quality suffered—suggests that early deployments often over‑promise and under‑deliver. Preliminary internal studies find that the vast majority of corporate AI pilots never appear in profit‑and‑loss statements, reinforcing the idea that many projects remain experimental rather than profit‑driving.
On the macro front, the San Francisco Federal Reserve notes that while labor productivity has been strong since 2023, total factor productivity—the metric that should capture a true technological leap—has barely moved, with only a 21 % chance of shifting into a high‑growth regime based on current data. An NBER survey of nearly six thousand executives echoes this: 89 % reported no productivity impact from AI over the last three years, and over 90 % saw no effect on headcount. Of course, every general‑purpose technology experiences a statistical lag (electricity took decades to show up), but the absence of any clear signal so far warrants caution.
Two narrow areas do show clearer signs of impact. Machine translation has demonstrably reduced growth in translator employment, with estimates suggesting a one‑point rise in adoption cuts translator job growth by 0.7 points across U.S. cities, amounting to tens of thousands of positions that never materialized. Many translators report income losses and lost work to generative AI. Coding is the other bright spot: developers undeniably enjoy velocity gains from AI assistants, though the ability to monetize that extra speed remains unproven, as evidenced by the disconnect between surging app submissions and flat consumer spend. The critical question is whether these isolated successes are the first hints of a broad wave or the only places where AI will ever generate a meaningful dividend.
The seller side of the equation, meanwhile, is flourishing. Google Cloud’s 82 % quarterly growth to $24.8 billion, paired with an $8.8 billion operating income at a 36 % margin, shows that someone is definitely getting paid for providing the raw compute. This thick smoke at the shovel‑sellers contrasts sharply with the faint signals coming from the diggers—the end‑users who are supposed to benefit. Such a pattern is not evidence against the thesis; it is precisely what we would expect if most of today’s spending is defensive or insurance‑like rather than driven by clear, profit‑generating use cases.
Layoffs make headlines, but hiring freezes are silent killers. The so‑so world does not arrive with a bang; it creeps in as a series of unopened requisitions, stalled promotions, and stagnant wage growth that never gets announced as a policy shift. That is why labor‑market economists watch entry‑level hiring in AI‑exposed roles as a canary: early‑career workers in those occupations have already seen roughly a 16 % relative decline in employment while their more experienced peers have held steady. The absence of new postings speaks louder than any press release about AI transformation.
Token spending itself splits into two behavioural categories with very different durability. The first is cost‑of‑goods spending: tokens baked into a product that is sold, such as the compute power that runs an AI‑enhanced bike‑sharing platform. Because the expenditure is directly tied to revenue, firms are unlikely to cut it as long as the product sells—much like airlines continue to pay for fuel despite historically poor returns on capital. The second category is insurance‑like spending: defensive purchases made simply to avoid the perception of falling behind, with no direct link to a sold product. This sort of expenditure is highly elastic; when budgets tighten or growth stalls, it is the first line item to be cut. Current evidence suggests that a large share of today’s AI outlay falls into this fragile insurance bucket.
From these observations we can outline three possible endings. In the first, the race continues: cost‑of‑goods spending compounds, fear keeps the insurance renewed, and the compute layer gets paid out of adopters’ margins. Hiring freezes deepen gradually, with no dramatic layoffs but a steady erosion of the labor market’s ability to absorb new workers. In the second, the insurance gets cancelled en masse after a macro shock—perhaps a rate spike or earnings downturn—prompts CFOs to demand proof of return. The cost‑of‑goods component survives, but the insurance‑style spending evaporates, potentially stranding trillions of dollars in capex that never earned back. In the third, the so‑so world prevails: modest, real savings appear, but they are insufficient to lower prices or expand the economic pie. Wages disappear without being replaced by new, better‑paid work, leaving workers worse off and the customer base that fuels AI demand slowly eroding.
All three outcomes share a common feature: the firms buying the tokens do not come out ahead in the long run. In a competitive market, any excess returns tend to get competed away, leaving the question of how much value the compute layer can retain before its own prices are driven down. Exceptions exist—firms with proprietary data, locked‑in distribution channels, or regulatory moats can capture lasting gains—but it remains doubtful whether such niches are large enough to absorb a $7.6‑trillion tab.
For investors and executives navigating this landscape, the practical takeaway is to scrutinize the nature of AI spending. Distinguish between expenditures that are directly tied to a sellable product or service (cost‑of‑goods) and those that are purely defensive or exploratory (insurance‑like). Track leading indicators such as hiring rates in AI‑exposed occupations, changes in contractor spend, and any measurable impact on pricing or margins within your supply chain. Pay close attention to wage‑level data in roles most susceptible to automation, as early‑career trends often foreshadow broader shifts. Finally, diversify bets: while the infrastructure layer may continue to enjoy strong near‑term returns, allocate capital to companies that demonstrate clear, monetizable use cases where AI drives genuine revenue growth rather than merely substituting labor at flat or declining prices.