AI's Economic Balance Sheet: Trillions Invested, Zero GDP Growth

Goldman Sachs chief economist Jan Hatzius recently said something that made a lot of people uncomfortable:

AI investment contributed “basically zero” to 2025 US GDP growth.

This comes after Microsoft, Google, and Amazon collectively spent hundreds of billions on data centers. After ChatGPT crossed 200 million users. After every AI company claimed they were “disrupting” their industry.

Basically zero.

Why This Isn’t Surprising

First, a sobering historical fact: the internet’s infrastructure was deployed at massive scale from 1999 to 2003. But the internet’s actual impact on productivity didn’t show up until after 2005 — a lag of five to seven years.

Technology economics has a concept called the J-curve: costs come first, returns come later. The curve dips before it rises. We’re probably at the bottom of AI’s J-curve right now.

What’s already happened (cost side): Data center construction and GPU procurement. Power consumption far exceeding projections. AI engineers as one of the most expensive job categories. Months of time to train foundation models.

What hasn’t materialized yet (return side): Workflow automation is mostly still in pilot phase at enterprise scale. How much labor cost has actually been replaced? Hard to measure. What new services has AI created? Too early to say.

GDP measures what was produced today, not what you invested for tomorrow. The accounting doesn’t line up yet.

The Measurement Problem

Even when AI creates genuine value, GDP might not capture it.

Example: if AI saves a developer two hours per day, and they spend that time resting or with family, GDP sees nothing. If they use those two hours to write more code, and the product eventually sells — then some portion enters GDP. The chain from AI productivity to GDP is long, and breaks easily.

Knowledge work productivity improvements have historically been GDP’s blind spot. When AI cuts a code review from three hours to thirty minutes, the project’s output value is unchanged. The ninety minutes saved simply disappears from the statistics.

So Goldman’s number doesn’t necessarily mean AI created no value — it may just mean current GDP measurement frameworks can’t capture knowledge worker efficiency gains.

A Real 10x Case

Ladybird browser developer Andreas Kling published a real case study two days ago.

Using Claude Code and Codex, he ported Ladybird’s JavaScript engine (LibJS) from C++ to Rust. The result: 25,000 lines of Rust code, two weeks of work, zero regressions — all 52,898 test262 tests passed, all 12,461 Ladybird regression tests passed. He estimates doing this by hand would have taken “multiple months.”

Critically: he’s explicit that this was “human-directed, not autonomous code generation.” He decided what to port, in what order, what the Rust code should look like. The AI did translation and first drafts; he did multiple rounds of adversarial review.

This is a real 10x productivity improvement — but with two requirements: an engineer who knows both C++ and Rust, and a specific, well-defined task.

And Ladybird is open source. This productivity gain doesn’t enter GDP in any form.

So Does AI Have Value?

Yes, but not in the form most people imagine.

Genuinely valuable AI applications today:

• Personal productivity tools (the stakeholder pool is small enough that GDP doesn’t register it)
• Vertical domain-specific deployment (AlphaFold’s protein folding breakthrough is a genuine scientific advance)
• Reducing cost for well-defined software development tasks (the Ladybird kind)

Applications that haven’t proven value yet:

• Enterprise-scale deployment (most companies are still in POC)
• White-collar job replacement (evidence is weak; augmentation is more accurate than replacement)
• Autonomous AI agents completing complex workflows (technically feasible, but reliability and trust-building take time)

What This Means for Developers

Don’t overestimate near-term returns. The J-curve says you might be at the bottom for another two to three years. Your competitors are waiting too. This isn’t “too late to enter” — it’s “still early.”

Focus on real problems. Genuinely valuable AI products solve specific, measurable problems — not “we integrated GPT.” “Reduce customer service cost by 40%.” “Cut code review from two hours to thirty minutes.” Specific problem, specific answer.

Personal tools before platforms. The most certain AI value right now is personal productivity. Helping yourself work faster has the most definite ROI — you don’t have to convince customers or shareholders.

The Bottom Line

Goldman’s number isn’t pessimism. Goldman themselves describe this as an early stage.

This number is an honest reality check: we’re in infrastructure-building mode, not returns mode. The internet took five to seven years to appear in GDP. AI might be faster (software’s replicability) or slower (reliability and trust take time to establish).

The key: don’t plan as if expected returns are existing returns.

---

Sources: Goldman Sachs Chief Economist Jan Hatzius’s recent public statements; Ladybird case from Andreas Kling’s official blog post at ladybird.org/posts/adopting-rust/ (published 2026-02-23).