The $7.6 Trillion Question: Can AI Companies Ever Actually Make Money?

Estimated reading time: 18 minutes

In the spring of 2026, something remarkable happened that almost nobody noticed. Goldman Sachs — one of the most influential financial institutions on earth — quietly published a research note confirming what a growing number of economists had suspected for months: $700 billion in AI investment during 2025 contributed essentially zero to US GDP growth. The tech industry had spent the equivalent of Sweden’s entire annual economic output on artificial intelligence, and the macroeconomic needle did not move.

This is not a story about whether AI works. It clearly does, in narrow and often impressive ways. This is a story about money — who is spending it, where it goes, what it returns, and what happens to the global economy if the answer to that last question remains, as it does today, “almost nothing.”

The numbers are staggering in their contradictions. OpenAI reports $20 billion in annualised revenue while simultaneously projecting $14 billion in losses for the same year. Anthropic reaches a $380 billion valuation while its gross margins hover around 40% — in an industry where software companies typically run at 70–80%. Nvidia briefly exceeded a $5 trillion market cap, surpassing the GDP of every country on earth except the United States and China, while the applications its chips power struggle to demonstrate economic value at scale.

“FOMO has proven a stronger incentive than poor stock performance. Hyperscalers have prioritised being involved in the AI arms race over their current shareholders.”

Goldman Sachs analyst Peter Covello, May 2026

The question is no longer whether some AI companies will fail. They will — in large numbers. The question is whether the foundational premise of the AI economy, that vast infrastructure investment now produces transformative productivity returns later, survives contact with reality. And if it does not, what happens next.

The Profitability Paradox: Revenue Without Profit

Start with the unit economics. OpenAI is perhaps the most successful product launch in the history of consumer technology. ChatGPT reached 53% population adoption faster than the PC, faster than the internet, faster than the smartphone. The company now serves 900 million weekly users. And yet, only 5.5% of those users pay for a subscription. The other 94.5% access the service for free — while OpenAI absorbs the compute cost of every single query across that enormous user base.

The result is a business structure that resembles, in the words of former Fidelity asset manager George Noble, “every warning sign I’ve ever seen.” OpenAI posted a $13.5 billion net loss in just the first half of 2025. Internal projections show $14 billion in losses for 2026, against roughly $13 billion in revenue — meaning the company spends approximately $2.20 for every $1 it earns. Total projected losses could balloon to $44 billion before the company reaches profitability in 2029 — if everything goes according to plan.

Sources: The Wall Street Journal / The Information / ainvest.com · OpenAI internal projections, Q1 2026

The AI Math

Sam Altman acknowledges the math. He has said publicly that training a single GPT-5-class model now exceeds one billion dollars before anyone is paid, leases settled, or electricity bills cleared. The compute cost problem, analysts emphasize, is structural rather than temporary. Inference costs quadrupled in a single year. Data center power demand grows at a pace the electrical grid struggles to match. And DeepSeek’s January 2025 release — demonstrating that a Chinese startup could match frontier model performance at a fraction of the cost — permanently undermined the assumption that massive spending creates an unassailable competitive moat.

Anthropic’s story looks better, but only relatively. The company hit $19 billion in annualized revenue in March 2026, doubled its valuation to $380 billion, and projects positive cash flow by 2027. Its enterprise-focused strategy — building deep integrations with Microsoft 365, Salesforce, Deloitte, and Cognizant — produces more predictable and higher-margin revenue than OpenAI’s consumer approach. But Anthropic’s gross margin still stands at roughly 40%, against the 70–80% margins that investors expect from software companies. Its path to genuine profitability remains steep.

$700 Billion Spent. Zero Added to GDP.

The profitability problem of individual companies might be forgivable if AI were clearly transforming the broader economy. Technology investments frequently run at a loss for years before delivering systemic returns. The internet required a decade of infrastructure investment before the productivity dividend arrived. The question is whether AI follows the same arc — or whether something more troubling is happening.

The data, as of mid-2026, should alarm anyone who has been sold the productivity revolution narrative.

Sources: Goldman Sachs Research, March 2026 · Stanford HAI AI Index 2026 · Futurism analysis of Bureau of Economic Analysis data

Goldman Sachs found no meaningful relationship between AI and economy-wide productivity as of March 2026, though the bank noted isolated gains of around 30% on specific tasks where companies actively measured outcomes. A separate Goldman analysis confirmed that $700 billion in AI investment during 2025 contributed essentially zero to US GDP growth.

No meaningful relationship between AI and economy-wide productivity

Dario Perkins, head of macroeconomics at TS Lombard, was blunt in a Financial Times interview: “There is no evidence that AI deployment is either boosting productivity or damaging US employment.” His analysis concluded that while US productivity had been strong and hiring weak, cyclical forces — not automation — were responsible.

Johns Hopkins economist Steve Hanke put it without diplomatic softening in April 2026: “AI didn’t deliver. Welcome to the real world. Forget the AI bubble — you know, it didn’t deliver.” A survey commissioned for the same period found that 95% of organisations were getting zero return on their AI pilots, despite $30–40 billion in enterprise investment in generative AI.

The structural explanation for the investment-GDP gap has two components. First, geographic: when US companies buy chips from Taiwan, that money boosts Taiwan’s economy, not America’s. The supply chain for AI hardware runs almost entirely through TSMC’s foundry in Taiwan, a single point of failure that also means American AI investment exports significant economic benefit abroad. Second, the investment itself is primarily capital expenditure — building infrastructure that theoretically becomes productive later. As one economist summarised it: economic growth during a bubble phase depends on continually building infrastructure, not using it.

The Arms Race Nobody Can Win — Or Afford to Lose

Perhaps the most revealing indicator of where the AI industry stands comes not from the companies losing money, but from those that should know better: the hyperscalers.

Microsoft, Amazon, Google, and Meta have dramatically increased their spending on AI infrastructure even as their stocks have lagged the S&P 500. These are companies with mature, profitable businesses and sophisticated financial teams. They have burned through all their free cash flow from operations and are now issuing debt to fund the build-out. Data center debt issuance doubled to $182 billion in 2025 alone.

Is AI a bubble?

Goldman’s Peter Covello described the dynamic precisely: hyperscalers have prioritised being involved in the AI arms race over their current shareholders. The phrase “arms race” carries a specific implication. Arms races, by definition, are not about winning — they are about not losing. No participant can afford to stop, regardless of whether the expenditure makes economic sense, because the perceived cost of falling behind exceeds the very real cost of continuing.

“A lot of capital has been deployed that doesn’t deliver returns. It’s kind of an industrial bubble.”

Jeff Bezos, Amazon Executive Chairman, October 2025

The financial structure that has emerged to sustain this race deserves particular scrutiny. OpenAI raises billions from Microsoft — and then spends much of it at Microsoft Azure. SoftBank invests $40 billion — and proceeds flow partly back to SoftBank’s own corporate partners. Investors are, in a meaningful sense, funding their own future revenues. The circularity creates the illusion of a functioning market while obscuring the fact that money circulates within a closed loop rather than generating genuine external economic value.

Sources: Wikipedia AI Bubble entry (May 2026) · Yale Insights · MacroStrategy Partnership · Shiller P/E data · Goldman Sachs Research

The Shiller price-to-earnings ratio for the US market exceeded 40 for the first time since the dot-com crash. In late 2025, 30% of the S&P 500 and 20% of the MSCI World index was held up by just five companies — the greatest concentration in half a century. Over the year 2025, AI-related enterprises accounted for roughly 80% of all gains in the American stock market. When a single sector drives that much of the market’s total return, the risk of contagion in a correction becomes existential rather than merely painful.

The Warning Indicators: Reading the Gauges

Financial bubbles rarely announce themselves. But they leave fingerprints — patterns that, viewed in aggregate, define the structure of speculative excess. The current AI market displays an uncomfortable number of them.

Bridgewater Associates co-chief investment officer Ray Dalio said early in 2025 that current levels of investment in AI are “very similar” to the dot-com bubble. Jamie Dimon of JP Morgan acknowledged in October 2025 that “the level of uncertainty should be higher in most people’s minds,” warning of a higher chance of a meaningful stock drop over the following two years than markets were then pricing. And economist Ruchir Sharma identified rising US interest rates as the specific trigger that could convert a slow deflation into an acute crisis.

The Bull Case: Why the Sceptics Might Still Be Wrong

Intellectual honesty demands that the counter-argument receive its full hearing. Several credible voices reject the bubble thesis, and their reasons are not without substance.

The productivity argument also deserves more nuance than its critics allow. Stanford HAI’s 2026 AI Index reports that on a key coding benchmark — SWE-bench Verified — AI performance rose from 60% to near 100% in a single year. AI agents made a leap from 12% to approximately 66% task success on OSWorld, which tests agents on real computer tasks. These are not trivial improvements. The question is whether they translate into economy-wide productivity gains, and if so, on what timeline.

Increasingly becoming critical to how businesses work

PwC’s April 2026 AI Performance Study found that 74% of AI’s economic value is captured by just 20% of organisations — companies actively using AI as a growth catalyst rather than a cost-cutting tool. This suggests the productivity dividend exists; it simply has not yet distributed itself across the economy. The lag between technology adoption and productivity gains is well-documented in economic history. The steam engine, electricity, and the internet all produced delayed productivity curves. Economists who dismiss AI on the basis of current productivity data may be committing the same mistake as those who dismissed the internet in 1997.

JPMorgan’s December 2025 analysis applied a five-factor diagnostic framework to the AI rally and concluded that investment is linked to actual enterprise revenue rather than speculation alone. Anthropic CFO Krishna Rao’s characterisation that Claude is “increasingly becoming critical to how businesses work” finds support in adoption data: organisational AI use reached 88% globally, with the technology embedded in workflows across banking, law, healthcare, and software development.

“AI will pay off — just like cars in total paid off, and TVs in total paid off. But most people involved in them didn’t do well.”

Jamie Dimon, CEO JPMorgan Chase, October 2025

Dimon’s framing is perhaps the most useful of all: the technology itself may be transformative while the investment structure around it remains deeply irrational. Cars and televisions genuinely changed civilisation. The companies that pioneered them, and their early investors, largely did not capture that value. The technology won; the financiers lost. AI may follow precisely the same pattern.

How a Crash Actually Unfolds: Three Scenarios

Financial bubbles do not burst instantaneously. They deflate through stages, and the trajectory depends heavily on the trigger. Analysts at Oliver Wyman, the World Economic Forum, and Yale’s Chief Executive Leadership Institute have each modelled the mechanics of an AI market correction. The scenarios diverge significantly in their severity.

The equity scenario carries echoes of the dot-com era but with a crucial amplifier: debt. The financing of AI capital spending is shifting from free cash flow to credit. If half of the projected $6 trillion in AI capital spending between now and 2030 is debt-financed, the resulting credit buildup would exceed all broadband infrastructure investment since the beginning of the internet. In a serious AI downturn, the dependence on debt financing could trigger a wave of credit defaults. The large scale of AI projects makes that debt concentrated, lumpy, and vulnerable — characteristics that, in the Global Financial Crisis, proved catastrophic.

Sources: Goldman Sachs Global Institute “Tracking Trillions” (May 2026) · Gartner Worldwide AI Spending Forecast · Stanford HAI 2026 AI Index

What a Crash Would Mean for the Global Economy

The geographic concentration of AI investment creates a specific vulnerability. The United States hosts 5,427 data centers — more than 10 times any other country — and almost every leading AI chip flows through a single TSMC foundry in Taiwan. A US-centred AI market correction would not stay in the US.

The World Economic Forum’s January 2026 analysis of a burst AI bubble concluded that the immediate economic fallout would be concentrated but significant. Job losses at speculative AI firms would be acute. Tens of thousands of AI engineers currently earning $200,000–$400,000 annually would enter the job market simultaneously, creating a skills glut that suppresses wages across the tech sector broadly. The geographic and skills concentration of the AI industry means that fear of unemployment would be relatively contained to those directly in the technology sector — but those individuals earn and spend at a rate that amplifies any consumption shock.

AI could wipe out half of all entry-level white-collar jobs

S&P Global’s scenario analysis estimated that a full AI bubble burst could wipe out over 2.5 million US jobs in tech and related sectors. Anthropic CEO Dario Amodei has separately predicted — irrespective of any bubble — that AI could wipe out half of all entry-level white-collar jobs and spike unemployment to 10–20% within the next one to five years. The compound effect of a burst bubble and ongoing AI-driven automation represents perhaps the most severe labour market scenario in the post-war era.

For emerging economies and markets outside the US, the contagion mechanism runs through three channels: capital flight from risk assets globally, reduced demand for commodities used in data center construction, and — most dangerously — the potential seizure of credit markets if AI debt defaults cascade through the financial system the way mortgage defaults did in 2008.

Deutsche Bank warned in September 2025 that the United States could already be in an economic recession without the tech industry’s AI spending spree. Manufacturing had contracted for seven consecutive months. The only meaningful prop under growth was AI-related capital expenditure. When — not if — that expenditure decelerates, the underlying weakness of the non-AI economy becomes visible.

The Startup Graveyard: 99% Are Already Toast

While the macro debate about bubbles and crashes plays out among economists and central bankers, the first wave of casualties is already visible at the startup level.

The AI startup ecosystem of 2023–2025 built itself largely on a single flawed premise: that wrapping OpenAI or Anthropic’s API in a vertical-specific interface constituted a viable business. It does not. As one Reddit post that circulated widely in late 2025 summarised the situation: “Most are just wrappers — zero moat, zero differentiation, and the second the underlying models get cheaper or offer native features, these companies are toast.”

Srinivas Rao predicted in May 2025 that 99% of AI startups would be dead by 2026. Industry data suggests this was not hyperbole. Single-purpose AI agents for tasks like email management or scheduling have been commoditised overnight by foundational model updates from the major labs. Valuations are crashing as users discover that “AI-powered” products frequently fail to deliver the promised return on investment. Enterprise SaaS companies including Salesforce and ServiceNow saw their stocks fall more than 20% from their January 2026 peaks as AI agents began automating workflows that previously required their platforms.

The pattern mirrors the dot-com era almost point for point. Then, appending “.com” to a business name guaranteed venture capital. Now, labelling anything “AI-powered” produces the same effect — and will produce the same eventual outcome. The technology is not the problem. The valuation structure around undifferentiated applications of that technology is.

What Survives the Winter

History suggests that the assets that survive technology bubble collapses are those with genuine infrastructure value, not those built on the hype layer above it. The dot-com crash destroyed thousands of companies but left the internet itself intact. The fibre optic cable laid at enormous expense in the late 1990s eventually supported businesses its original investors never imagined. The question for AI is which layer survives.

The answer almost certainly includes the foundational infrastructure — the chips, the data centers, the electrical grid upgrades, the fibre networks. Nvidia’s hardware will remain valuable regardless of which AI companies succeed or fail. The major cloud providers — AWS, Azure, Google Cloud — will continue to provide compute, even if demand for AI-specific workloads falls.

Among the model companies themselves, survival depends on a factor that most analysts underweight: the structure of revenue. Anthropic’s B2B focus — deeply embedding Claude into enterprise workflows at Deloitte, Cognizant, Microsoft, and Salesforce — creates switching costs that protect revenue even in a market downturn. Organisations that restructure their legal departments, financial operations, or software development pipelines around a specific AI system do not abandon it when headlines turn negative. OpenAI’s consumer-facing model, by contrast, depends on the continued willingness of users to pay for subscriptions to a product that free alternatives increasingly approximate.

The companies that will not survive, in large numbers, are those operating as thin wrappers without genuine differentiation, those with revenue dependent on continued speculative investment rather than real enterprise contracts, and those whose valuations assumed a technology adoption curve that reality has declined to provide.

The Reckoning That Is Already Arriving

The optimists are not wrong that AI will eventually transform the global economy. The historical record of general-purpose technologies — steam, electricity, computing — strongly supports the view that transformative infrastructure investment produces enormous economic returns, just not on the schedule that speculative capital demands.

But the structure of the current moment carries specific risks that distinguish it from prior technology cycles. The scale of debt financing is unprecedented. The circular interdependencies between the major players create opacity that regulators and markets cannot easily price. The gap between executive expectations and measurable outcomes is historically large. And the political economy of the moment — with AI investment propping up headline GDP growth in the United States — creates pressure to sustain the illusion long past the point where prudence would counsel retreat.

AI “didn’t deliver” may prove premature

Steve Hanke’s blunt assessment that AI “didn’t deliver” may prove premature. But the more uncomfortable possibility — that it will deliver, eventually, for a small number of companies with genuine structural advantages, while destroying the capital of the investors, the workers, and the economies that funded the journey — demands to be taken seriously.

Goldman Sachs estimated that AI will drive a 1.5% increase in global labour productivity over the coming decade. If that forecast proves correct, the productivity dividend will arrive — but over ten years, not two. Against cumulative AI capital expenditure projected at $7.6 trillion between 2026 and 2031, a productivity gain that arrives slowly and unevenly across the global economy is an extraordinarily poor return on investment. The math, on almost any conservative assumption, does not close.

The $7.6 trillion question deserves a precise answer. Can AI companies ever actually make money? Yes — some of them, eventually, in a market that consolidates aggressively around a small number of companies with genuine structural advantages, durable enterprise relationships, and the financial patience to survive years of losses. The rest of the ecosystem — the wrappers, the speculative infrastructure plays, the companies built on the assumption that every technology hockey stick curves upward without limit — will not.

The global economy’s exposure to that reckoning is larger than almost anyone outside a small community of sceptical economists currently acknowledges. The time to acknowledge it is before the trigger event, not after. That has always been true of every bubble in history. It remains true now.