Morgan Stanley: AI Investment Contributes Approximately 0.42 Percentage Points to U.S. GDP
nashnova research
Morgan Stanley estimates that after stripping out routine activity, new investment driven by the AI boom adds roughly 0.42 percentage points to U.S. real GDP — a material growth source, but far from the full economic picture of AI.
0.61 vs. 0.42 — what's the difference between the two numbers?
Morgan Stanley splits AI investment into broad and narrow measures. The broad measure covers data centers, power, computers, telecom gear, software and more — averaging about 0.61 pp of GDP contribution from 2025 through H1 2026.
The narrow measure benchmarks against 2010–2019 trends, strips out investment that would have happened without the AI boom, and counts only the excess. That drops the figure to about 0.42 pp.
This means → roughly one-third of "AI investment" is routine economic activity. Related to AI ≠ caused by AI.
The contribution is rising — but not because more is being built?
Morgan Stanley's breakdown shows that across each half-year since 2025, the gross pull from infrastructure spending — before subtracting imports — has hovered around 0.5 pp, largely unchanged.
What actually lifted the net number is weakening import offset: the import drag was nearly 0.4 pp in H1 2025 but fell below 0.2 pp by H1 2026.
In plain terms = the U.S. didn't build dramatically more; imported equipment simply "subtracted less from the ledger" — if imports normalize, the GDP contribution could shrink even as firms keep spending.
Why did imports suddenly drop?
Morgan Stanley acknowledges the weaker import offset may reflect statistical recording, corporate stockpiling, and front-loaded purchasing ahead of trade restrictions.
Current evidence is insufficient to prove the U.S. has achieved large-scale domestic substitution of the relevant equipment.
This means → the "improvement" in net contribution may be temporary. If import patterns revert, the GDP number could worsen even with sustained corporate capex.
Where is the money going — infrastructure far outweighs software?
In H1 2026, infrastructure investment contributed about 0.36 pp and software about 0.18 pp — infrastructure roughly double software.
On an annualized nominal basis, AI-specific investment totals about $450 billion: infrastructure ~$390 billion, software ~$60 billion. Computers and peripherals account for 60%–70% of AI infrastructure spending.
This reflects an AI economy still in the "roads and bridges" phase — very few applications are actually running on the road.
Software spending is up — does that mean AI is being adopted?
Morgan Stanley notes that software is only a proxy for the application layer. The statistics bundle corporate AI deployments together with model developers' own software R&D, and cannot be reliably separated.
In Q2 2026, major cloud providers' R&D spending equaled roughly half of their capex, but earnings reports lack the detail to split software development, model training, and other R&D.
In plain terms = rising software spending may just mean "the people building AI spent more," not "the people using AI grew" — proving real adoption requires paid-usage rates, renewal data, workflow changes, and unit-cost evidence.
What's missing from this ledger?
The 0.42 pp counts only direct investment. Employment, consumption, and spillovers into other industries are all excluded — it cannot be read as a precise net value of AI's total economic impact.
Morgan Stanley sees two verification checkpoints: first, whether equipment orders accelerate further, so the net contribution no longer depends on a temporary import dip; second, whether U.S. domestic value-added genuinely rises, so the same dollar of investment leaves more output at home.
This means → a better number ≠ a confirmed trend. The two wild cards — import structure and domestic capacity — have not yet been revealed.
市场有风险，内容仅供研究参考，不构成投资建议。