Ghost GDP

McKinsey says AI will add $13–22 trillion in cumulative value by 2040. PwC says $15.7 trillion by 2030. Goldman Sachs said +7% to global GDP, then quietly noted in February 2026 that AI had contributed "basically zero" to US GDP through 2025.

I keep staring at these numbers trying to figure out where the money comes from. I mean the actual mechanics. Which line items in GDP go up, and why?

Because the more I look at the mechanisms, the less they look like GDP growth.

I went and read the methodology sections. Here's what's inside the headline numbers.

Three of the four assume displaced workers find new jobs. McKinsey doesn't even model what happens if they don't. Goldman's model is built on historical patterns from electricity and personal computers—technologies that automated physical tasks and left cognitive work for humans. AI automates cognitive work. The thing humans moved to last time.

1. Productivity is not GDP

When a company replaces 4,000 customer service workers with AI and handles the same volume of calls, output per worker goes up. That's a productivity gain. But total output (the number of calls handled, the value of the service) hasn't changed. GDP measures total output. It stayed flat.

What changed: costs went down and margins went up. The company is more profitable. Its stock price probably jumped. But the economy didn't grow. Income shifted from wages to profits.

The GDP forecasts are mostly measuring expected productivity gains and calling them GDP growth. McKinsey's $13–22 trillion is a supply-side, technical-potential estimate. It measures what AI could do to output per worker. It doesn't measure whether total output value increases. If prices fall, it might not.

2. The digital economy is 10% of GDP. AI is making it cheaper.

The US digital economy is about 10% of GDP ($2.6 trillion as of 2022, the last year BEA measured it). Software and IT investment was historically a rounding error in GDP growth, contributing about 0.3 percentage points per year. Then in the first half of 2025, it accounted for all of GDP growth. A 28% annualized surge, entirely from the AI investment boom. That's not organic growth. That's a capex cycle.

But what happens when the capex cycle ends and the AI starts producing? When the marginal cost of producing software approaches the marginal cost of compute, and compute costs keep falling, software prices collapse. Right now the AI boom is growing GDP through investment spending. The question is what happens to GDP when that investment starts working. When AI agents can build a SaaS product in hours and the value of software as a product approaches zero.

GDP is measured in dollars. More output at lower prices can mean less GDP, not more. RAND modeled this directly: "Nominal GDP is substantially lower if AI systems lead to substantive deflation, even though real GDP growth rate remains similar." The economy produces more. The number gets smaller.

Sam Altman gets this: "GDP is going to be a terrible metric because AI is so deflationary." He's right about the measurement problem. But deflation only helps if you still have income. If the mechanism that creates the deflation (replacing workers with AI) also destroys the paychecks, cheaper goods don't fix much.

3. Consumption is 68% of GDP. Workers are consumers.

GDP = C + I + G + NX. Consumption (C) is over two-thirds of the number. Consumption comes from income. Most income comes from wages. If AI displaces workers at scale, wages fall, consumption falls, GDP falls.

Citrini Research modeled this as a spiral: each company's decision to cut staff and deploy AI is individually rational. Collectively, the cuts shrink the customer base for every other company's products. Declining consumer spending creates margin pressure, which forces more AI investment, which causes more layoffs. Citrini calls it "no natural brake."

Citadel Securities published a full rebuttal. Their argument: productivity shocks are historically growth-enhancing. Falling costs expand the consumption frontier. Keynes underestimated the elasticity of human wants, and Citrini is making the same mistake. That's the standard economic counter, and historically it's been roughly true. But it assumes the income exists to consume with. Workers spend 70–90 cents of every dollar. Capital owners don't. The velocity changes when the composition of income changes.

Acemoglu's framework gives this some structure: automation has a displacement effect (AI takes tasks, labor demand drops) and a reinstatement effect (new tasks emerge for humans). Historically these balanced. But Acemoglu's own data shows reinstatement has been weakening since the 1980s. And that was before AI—before the tool doing the displacing could also do the new tasks.

An NBER survey of nearly 6,000 executives across the US, UK, Germany, and Australia found 89% reported no impact of AI on their labor productivity over the previous three years. Robert Solow noticed the same thing about computers in 1987: "You can see the computer age everywhere but in the productivity statistics."

But the layoffs are real. Challenger, Gray & Christmas counted 55,000 AI-attributed job cuts in 2025. That's up from near zero in 2023. An HBR analysis from January 2026 found 60% of AI layoffs are anticipatory—companies cutting staff because they expect AI to replace those roles, not because it already has. Only 2% came from actual AI implementation making workers redundant. And Forrester found 55% of employers who made AI-related layoffs already regret the decision.

Oxford Economics calls some of this "AI-washing"—companies using AI as cover for ordinary cost-cutting because the market rewards the narrative. The 55,000 is 4.5% of 1.17 million total layoffs in 2025.

So the scoreboard right now is: GDP impact near zero, but layoffs real and accelerating. Dario Amodei predicts half of entry-level white-collar jobs gone within one to five years. VCs surveyed by TechCrunch expect AI to "aggressively impact" the employment rate in 2026. The productivity gains haven't arrived. The displacement has started anyway.

The optimist response to all of this is: give it time. Electricity was invented in the 1880s and didn't show up in productivity statistics until the 1920s. Factories had to be physically redesigned around electric motors. Forty years between invention and impact.

I don't think this analogy holds anymore.

Electricity automated physical power. Humans moved to cognitive work: management, design, analysis, communication. That's where the new jobs came from. AI automates cognitive work. It automates the thing humans moved to last time. And unlike electricity, it can learn new tasks. When a new role emerges, AI can do that too.

The ATM example gets cited a lot: ATMs didn't kill bank tellers. Tellers shifted to advisory and relationship work. True—but ATMs could only dispense cash. They couldn't learn to give financial advice. AI can. Research from 2025 shows AI alone outperforming AI-plus-human teams on legal research and medical diagnosis. The "centaur" model (humans working alongside AI) is already losing to the pure AI version in measurable domains.

Goldman's model assumes displaced workers find "slightly less productive" new jobs, following patterns from electricity and PCs. But those patterns held because the new jobs required capabilities the automation didn't have. If AI has those capabilities, or develops them in months rather than decades, there's nowhere to move to.

Adoption is different too. Electricity required rewiring factories. You had to physically rebuild infrastructure. AI requires a subscription. The bottleneck isn't hardware installation. It's a decision to cancel headcount. Companies are making that decision before the AI even works—60% of AI layoffs are anticipatory.

All of these assume AI works. The capability is real. The question is what happens to the numbers and to the people.

These aren't mutually exclusive. The Measurement Collapse and the Ownership Economy could happen simultaneously—real GDP stagnant while corporate profits and quality of life for asset owners both surge. The Demand Spiral could trigger the policy interventions (UBI, fiscal stimulus, job guarantees) that prevent the worst version of the Ownership Economy. Or not.

What all three share: the $15 trillion number doesn't tell you anything useful about what happens to people. The Measurement Collapse means life might actually improve while the headline number stagnates. The Demand Spiral means the number was always fiction. The Ownership Economy means the number is real but describes someone else's life. Different flavors of the same problem.

None of this is landing in a vacuum. For fifty years, productivity gains have not become wage gains.

Daniel Citrini calls the AI-accelerated version of this "Ghost GDP." Output that shows up in national accounts but never circulates as income people actually spend.

The institutions that used to connect productivity to pay (unions, tight labor markets, progressive taxation) are weaker than at any point since the 1920s. AI doesn't create the disconnect. It arrives into a system that's been disconnected for fifty years and applies pressure in exactly the direction the gap has been widening.

I don't know which scenario we end up in. I think the Measurement Collapse is underrated—GDP probably does become a bad metric, and some version of Altman's argument is right. But I also think the Ownership Economy is the default, because it's just the last fifty years with the dial turned up. Nothing about AI changes who captures productivity gains. It just makes the gains bigger and the question harder to ignore.

The forecasts might land. $15 trillion, $26 trillion, whatever. Or GDP might shrink while life gets better for some people and worse for others, and the number captures none of it. I genuinely don't know. But every time someone says "AI will grow the economy," I want to ask them to show their work. Which line items? Through what mechanism? Who ends up holding the money?

I haven't found anyone with a good answer to that yet. I'm still looking.