The most expensive technology bet in corporate history has a GDP contribution of approximately zero. The math is simple and brutal.
Goldman Sachs Chief Economist Jan Hatzius told clients this week that massive AI investment spending contributed "basically zero" to U.S. economic growth in 2025. Morgan Stanley reached the same conclusion independently. Two of Wall Street's most influential economics desks agree: the AI boom is not showing up in the numbers that matter.
The explanation is not that AI doesn't work. It's that America doesn't manufacture the hardware AI requires.
The Import Problem
Roughly three-quarters of AI data center costs go toward chips and equipment manufactured in Asia — primarily TSMC in Taiwan, SK Hynix and Samsung in South Korea, and assembly operations across Southeast Asia. When an American company spends $50 billion on Nvidia GPUs, the majority of that money flows to foreign manufacturers. In GDP accounting, domestic spending on imported goods gets subtracted from output. The investment shows up on corporate balance sheets. It does not show up in GDP.
This is not a rounding error. Meta committed $65 billion to AI infrastructure in 2026. Microsoft committed $80 billion. Amazon, Google, and Oracle each committed tens of billions more. Goldman estimates AI investment will contribute approximately 1.5 percentage points to U.S. capital expenditure growth in 2026 — but the net GDP impact will be 0.1 to 0.2 percentage points. Almost all of it leaks offshore.
The companies spending the money know this. They are not investing to grow GDP. They are investing to capture markets. The distinction matters because the political narrative around AI — that it drives American economic growth — rests on a category error. What drives growth is domestic production. What AI drives is demand for Taiwanese semiconductors.
The Productivity Question
Goldman expects meaningful productivity gains from AI to begin appearing in 2027 and to compound through the late 2030s. That is the actual economic case for AI: not the spending itself, but the eventual output gains from workers using AI tools. The problem is that those gains remain theoretical at current adoption rates.
McKinsey's most recent survey found that 72% of companies have adopted AI in some form, up from 55% a year earlier. But adoption does not equal productivity. MIT's Project NANDA found that 95% of custom enterprise AI deployments deliver zero measurable ROI. Anthropic's own research shows developers using AI code assistants scored 17% lower on code comprehension tests. The tools are everywhere. The output gains are not.
Hatzius draws an explicit parallel to the late 1990s internet investment cycle. Companies spent aggressively on fiber optic cable, web servers, and e-commerce platforms for years before productivity statistics moved. The dot-com crash intervened. The productivity gains eventually materialized — between 2003 and 2007 — but they arrived for different companies than the ones that spent the money.
What the Math Says
SK Hynix announced this week that its entire 2026 production of HBM4 memory chips — the next-generation component required for frontier AI training — is already sold out. The company plans to increase advanced DRAM capacity eightfold, from 20,000 to 190,000 wafers per month. Prices will keep rising. "We cannot meet the needs of all customers this year," the company stated.
Every dollar of that price increase flows to South Korea, not to the American economy. Every wafer fabricated at TSMC's Arizona plant — which handles a fraction of total demand — competes with higher-volume, lower-cost production in Taipei. The CHIPS Act allocated $52 billion to reshoring semiconductor manufacturing. Total AI chip demand in 2026 alone will exceed $200 billion.
The arithmetic is not complicated. America is financing the most expensive infrastructure buildout in history. The infrastructure is being built with foreign components. The economic returns, if they arrive, will arrive years from now. In the meantime, GDP measures what gets produced domestically, and what gets produced domestically is the electricity bill.
Goldman Sachs is not saying AI is a bad investment. It is saying the investment and the economic growth are happening in different countries. The companies spending the money are American. The factories collecting the money are not.
That is not a technology problem. It is a trade balance problem wearing a technology costume.
If you work with AI tools daily, check out my prompt engineering resources on Polar.sh — practical prompt packs for developers who want better outputs from LLMs.
Top comments (0)