The $650B Zero-ROI Disconnect: AI's Biggest Bet vs the Data
Four companies will spend $650 billion on AI infrastructure this year. That exceeds Sweden's GDP. Goldman Sachs says the contribution to US economic growth has been basically zero. An MIT study found 95% of enterprise AI pilots delivered no measurable P&L impact. An NBER survey of 6,000 executives across four countries found roughly 90% of firms reported zero productivity gains from AI over the past three years. These findings come from the most credible institutions in economics and technology research, and they all point the same direction.
The gap between what is being spent and what is being produced is the defining financial story of 2026. As a CFA charterholder who spent 13 years reading financial statements before I ever touched an AI model, this is the kind of disconnect I was trained to notice. The numbers on both sides are real. The question is what they mean for the people actually building with this technology.
The Spending: $650 Billion in One Year
The scale is hard to internalize. Here's how it breaks down:
| Company | 2026 Capex (Projected) |
|---|---|
| Amazon | ~$200B |
| Alphabet/Google | $175B-$185B |
| Microsoft | ~$150B |
| Meta | $115B-$135B |
| Combined | ~$650B |
J.P. Morgan projects $5 trillion in total AI infrastructure spending through 2030. Goldman Sachs projects hyperscaler capex at $1.15 trillion from 2025 through 2027, more than double the $477 billion spent from 2022 through 2024. A TIME analysis noted this spending exceeds the Apollo space program, the US interstate highway system, and the railroads as a percentage of GDP.
These are real dollars leaving real balance sheets. And someone has to earn them back.
Goldman's Verdict: Basically Zero
In February 2026, Goldman Sachs Chief Economist Jan Hatzius told the Atlantic Council that AI investment spending had contributed basically zero to US GDP growth in 2025. His explanation was blunt:
"A lot of the AI investment that we're seeing in the U.S. adds to Taiwanese GDP, and it adds to Korean GDP but not really that much to U.S. GDP."
Roughly 75% of the cost of a data center comes from imported parts. Chips manufactured in Taiwan and South Korea. When US companies spend billions on Nvidia GPUs, that spending registers as imports, which subtract from GDP calculations. The money flows out. The silicon flows in. The net contribution to the US economy is close to nothing.
Hatzius added: "We don't actually view AI investment as strongly growth positive... there's been a lot of misreporting of the impact that AI investment had on GDP growth."
This matters because the popular narrative throughout 2024 and 2025 was that AI capex was driving GDP growth. Goldman's data says those figures conflated gross investment with net contribution. Big difference.
The Revenue Math Doesn't Close
The investment-to-revenue ratio tells the story more simply. According to Fortune's analysis, Microsoft, Meta, Tesla, Amazon, and Google invested approximately $560 billion in AI infrastructure over two years and generated $35 billion in combined AI-related revenue. That's 16:1.
Goldman Sachs analyst Ben Snider calculated that maintaining acceptable returns on capital from current spending would require an annual profit run-rate exceeding $1 trillion. More than double the consensus estimate of $450 billion for 2026. His conclusion: there is a diminishing probability that all of today's market leaders generate enough long-term profits to justify today's investors.
OpenAI exemplifies the tension. The company hit $25 billion in annualized revenue by February 2026 but projects $14 billion in losses for the year. It spent $1.69 for every dollar of revenue in 2025. Cumulative cash burn through 2029 is expected to reach $115 billion. Revenue is growing fast, but losses are growing faster.
95% Failure Rate: What the Studies Found
The academic research paints a consistent picture.
The MIT study ("The GenAI Divide: State of AI in Business 2025") examined 52 executive interviews, 153 leader surveys, and 300 public AI deployments. The headline finding: 95% of AI pilots delivered no measurable P&L impact. Only 5% of integrated systems created significant value. US businesses had collectively invested $35-40 billion in enterprise AI by that point.
The NBER survey ("Firm Data on AI," February 2026) covered 6,000 executives across the US, UK, Germany, and Australia. Key findings: 69% of firms actively use AI. Roughly 90% reported no impact on employment or productivity over three years. Average executive AI usage: 1.5 hours per week. Despite seeing nothing yet, executives forecast AI will increase productivity by 1.4% and output by 0.8% over the next three years.
RAND Corporation (2025): 80.3% overall AI project failure rate. Broken down: 33.8% abandoned, 28.4% delivered no value, 18.1% couldn't justify costs.
S&P Global: 42% of companies scrapped most of their AI initiatives in 2025, up from 17% the year before.
The root causes across these studies are structural: poor data quality, unclear ownership, weak cross-functional coordination, and attempting to automate workflows that were already broken. The technology works. The organizations deploying it mostly don't.
Where the Money Actually Goes
Most of the $650 billion flows to infrastructure: chips, data centers, power, cooling. Very little reaches the applications that would actually generate productivity gains. This creates a timing mismatch. Nvidia, TSMC, and construction firms capture the capex dollars today. The productivity returns, if they arrive, show up at the companies deploying the models. Years later.
Bridgewater Associates quantified this: AI capex is projected to contribute roughly 140 basis points to US GDP growth in 2026 and 150 basis points in 2027, driven purely by the spending itself. On par with the contribution of business investment during the tech bubble. The spending boosts GDP mechanically. It tells you nothing about whether the investment produces returns.
The crowding-out effects are starting to bite. Private data center construction runs at $41 billion annualized, matching state and local government spending on transportation construction dollar-for-dollar. Chatham Financial estimates total AI infrastructure capex at $830 billion in 2026 when including adjacent spending. That represents roughly 50% of last year's entire US investment-grade bond volume.
Bridgewater's comparison is useful here. Meta's $1.5 billion Texas data center creates 100 permanent jobs. A comparably priced battery manufacturing plant supports 1,620 positions. That's 16x fewer jobs per dollar invested. AI capex generates fewer jobs per million dollars than virtually any other form of capital expenditure. High interest rates driven by competitive capital demand are creating headwinds for more labor-intensive sectors. Non-AI corporate borrowing sits at low levels relative to the past 15 years.
The Human Cost of GPU Shopping
Meta provides the clearest case study of how companies are funding this buildout.
The company budgeted $115-135 billion for 2026 capex, roughly double its 2025 spending. Simultaneously, it laid off several hundred employees across five divisions in March 2026 and is weighing a larger reduction that could eliminate up to 20% of its global workforce. Roughly 15,000-16,000 of its 79,000 employees.
Meta's stock climbed nearly 3% on the layoff reports. The market rewarded the trade: humans out, GPUs in. Across the broader tech sector, layoffs surged to 59,000 in 2026 as companies shifted resources from headcount to compute.
When a company doubles its infrastructure budget while cutting a fifth of its workforce, the capital allocation thesis is explicit. The bet: GPU hours are more valuable than human hours. At current ROI numbers, that bet remains unproven.
The Solow Paradox, Round Two
In 1987, economist Robert Solow wrote: "You can see the computer age everywhere but in the productivity statistics." Productivity growth had dropped from 2.9% annually (1948-1973) to 1.1% during the IT boom. The pattern repeats almost exactly with AI.
Companies are adopting AI everywhere. 374 S&P 500 companies mentioned AI in earnings calls, most claiming positive implementations. Usage keeps growing. Confidence in the technology's long-term value remains high. And the macro productivity numbers refuse to move.
The resolution of the original Solow paradox took over a decade. Productivity growth surged 1.5% from 1995 to 2005, vindicating the IT investments of the 1970s and 1980s. But only after organizational structures, business processes, and worker skills caught up with the technology. The infrastructure came first. The returns came 15 years later. Goldman Sachs forecasts AI will begin measurably impacting US GDP and labor productivity around 2027.
The dot-com parallel offers a different angle. Telecom companies laid more than 80 million miles of fiber optic cable in the 1990s, driven by WorldCom's fraudulent claim that internet traffic was doubling every 100 days. Four years after the bubble burst, 85-95% of that fiber remained dark. Entirely unused. The AI capex wave dwarfs the dot-com era by 17x in inflation-adjusted terms.
But the dot-com story had two acts. The infrastructure that survived the bust became the foundation for the next generation. The companies that won the next decade (Google, Amazon, Facebook) built on top of that overinvestment. They used infrastructure someone else paid for and lost money building.
The 5% That Works
The 95% failure rate has a flip side. The MIT study found that successful deployments share specific characteristics.
The companies seeing real ROI from AI focus on back-office automation: administrative functions, repetitive workflows, internal tooling. More than half of failed AI spending went to sales and marketing, areas researchers say still need heavy human involvement. The wins are unglamorous. Automated data entry. Faster document processing. Standardized internal reporting. Tasks where the inputs are consistent, the outputs are predictable, and the cost of errors is manageable.
The 5% also tend to be well-scoped. Focused deployments targeting a specific workflow with clear success criteria. The enterprise failures share a different profile: company-wide AI transformation programs without defined outcomes, driven by board pressure and vendor pitches rather than a specific problem that needed solving.
Some economists, including Erik Brynjolfsson, have argued that productivity data shows early positive signals. Individual workers using AI tools complete tasks faster. Coding assistants measurably reduce development time. Customer support agents handle more tickets per shift. The gains are real at the task level. They just don't show up at the macro level because the organizations deploying AI absorb the gains in coordination costs, integration overhead, and process redesign. The question has never been whether AI generates value. It does. The question is whether the value justifies the spend at the scale these companies are investing.
The Systemic Risk Question
A TIME analysis identifies two scenarios for how this plays out.
Scenario one: dot-com replay. A correction hits. Some companies fail. Stock prices drop. But the infrastructure remains and becomes the foundation for the next generation. Brief recession, lasting benefit. This is the optimistic case.
Scenario two: 2008-style systemic event. The AI buildout is increasingly financed through circular financing structures, special purpose vehicles, private credit loans, and asset-backed securities. These instruments distribute risk into retirement accounts and pension plans. If the underlying assets (data centers, GPU clusters, long-term compute contracts) lose value simultaneously, the losses don't stay concentrated among cash-rich hyperscalers. They propagate through the financial system.
The critical variable is where the debt sits. If it stays on the balance sheets of companies that can absorb a correction (Apple, Google, Microsoft), the downside is manageable. If it's been securitized and sold to pension funds, the downside becomes structural.
I don't know which scenario is more likely. Nobody does. But the distinction matters because it determines whether the correction is healthy (overbuilt infrastructure gets repurposed, as it did after the dot-com bust) or destructive (retirement portfolios absorb losses from assets they never understood).
What This Means If You Build Things
The macro picture is one thing. The builder's calculus is different.
If 88% of enterprise AI pilots never reach production and the average executive uses AI 1.5 hours per week, the organizational friction killing these projects is extraordinary. A solo builder or small team faces none of it. No committee reviews. No six-month pilot programs. No data governance boards. No vendor selection processes.
The $650 billion infrastructure buildout is, in practical terms, a subsidy for anyone who knows how to use the outputs. API access to frontier models costs $200-2,000/month. Not $200 billion. The hyperscalers are building the plumbing. The returns accrue to the people who build the products that run on top of it.
Target back-office automation first. The MIT data is clear: this is where AI consistently delivers ROI. Sales and marketing AI is where most money gets wasted. If you're building something, start with the boring stuff. Data entry, document processing, internal workflows. The failures are concentrated in customer-facing AI. The wins are in the back office.
Small bets beat big bets. The 5% that succeed tend to be focused, well-scoped deployments. Company-wide AI transformation is where budgets go to die. A single workflow, a defined success metric, a 90-day timeline. That's the shape of the projects that work.
Model depreciation works in your favor. GPT-4 reportedly cost over $100 million to train two years ago and is now outperformed by open-source models that cost a fraction to run. As trained models lose value, API prices fall. The infrastructure overinvestment subsidizes your access. Every quarter, the same capability costs less. I wrote about this dynamic in The Frontier Model Tax: fine-tuned small models are already replacing frontier API calls in production at companies like Airbnb and Cursor.
The asymmetry is real. The companies spending $650 billion may or may not earn adequate returns. That's their problem. The builders using the resulting infrastructure at commodity prices face a completely different equation: ship something useful for $200/month on top of infrastructure someone else spent $200 billion to construct.
Whether the macro bet resolves well for the investors is something we'll know in five to ten years. The plumbing is getting built regardless. For builders, the bet is simpler: learn to use it before the window closes.
Research compiled April 2026. All figures sourced inline. This reflects publicly available data and analysis from Goldman Sachs, MIT, NBER, Bridgewater, RAND, and S&P Global.