NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Three Numbers That All Rise. The One That Falls Is the Difference.

1.8, 2.5, 2.8. Goldman Sachs gives those for US AI investment as a share of GDP in 2026, 2027 and 2028.

2026-11-17 · 4772 words · NeuroAI
Three Numbers That All Rise. The One That Falls Is the Difference.

1.8, 2.5, 2.8. Goldman Sachs gives those for US AI investment as a share of GDP in 2026, 2027 and 2028.

All three rise. The increments are 0.7 and 0.3.

That is the whole content of "growth peaks in the first half of 2027."

What peaks is the difference

A peak in scale would mean AI investment stops rising. Goldman does not say that; it puts the peak in absolute scale in 2028.

A peak in growth means the yearly addition shrinks. That is the claim.

Hatzius's "Macro Outlook 2027" has global nominal AI investment at $1.1 trillion in 2026 and $1.8 trillion in 2027, and the US share of GDP at 1.8%, 2.5% and 2.8%, against 0.9%, 1.3% and 1.4% globally.

That is what "peak" means in sell-side research: the second derivative, not the level. The $1.8 trillion and the 2.8% are still there; only next year's addition is smaller.

Chinese headlines about "the AI feast nearing its end" erase the distinction. The feast has not broken up. Fewer new dishes arrive each day.

Growth peaks in the first half of 2027, scale in 2028, eighteen months apart. That gap is the second-derivative peak: scale still makes new highs while growth has turned, with a year of "still rising, more slowly" between.

How the date was derived

The first half of 2027 is not a forecast; it is where a line crosses a dashboard.

The line is extrapolation: draw the GDP share forward on Goldman's own estimates through 2028 and the year-over-year change turns in that half by construction. Arithmetic, not judgment.

The dashboard is leading indicators. Rather than guess demand, Goldman watched semiconductor equipment imports into Taiwan and Korea, PMI sub-indices, import prices, memory prices and GPU rent, asking whether they lead US AI capex. All sit near highs last seen in 2022.

The ruler is where to push. Share of GDP ties AI investment to GDP, and imported equipment—chips, servers—enters GDP differently from domestic capital formation. The denominator distorts. Goldman concedes it: business investment data understate semiconductor spending, putting AI's 2027 contribution at 0.2 points, or 0.6 after correction.

0.2 and 0.6 differ by a factor of three. That error is larger than whether the peak lands in the first half or the second.

Three percentages

What matters is not $1.8 trillion but three percentages.

The cross-asset team, Trivedi and Wilson: AI contributed 21% of US GDP growth, 62% of S&P 500 earnings growth, and 92% of the market cap gain at the peak.

21 to 62 to 92, climbing.

The economy absorbed a fifth, earnings three fifths, market cap nine tenths.

That is why market impact exceeds economic impact: the same inflection is a footnote at 21% and a valuation event at 92%. Their wording: AI's contribution to market gains far exceeds its contribution to growth, so a peak hits markets harder than the economy.

Who is worst off

Three parties: compute sellers, buyers, and the application layer.

Sellers are not worst off. Investment still rises through 2027, and $1.8 trillion is 64% more than $1.1 trillion. Goldman recommends overweighting hyperscalers, whose P/E is 19.7 times, below the 20-year range and below consumer staples. They are buyer and seller at once, buying chips and renting compute; Goldman picks the resale position.

The application layer is not worst off either: it buys tokens, token prices are falling, and cheaper models mean more margin. A peak there means upstream discounting, not vanishing demand.

The worst off is the buyer—not the buyer of tokens, the buyer of AI stocks who pays for growth.

Earnings growth halves almost everywhere: the US from 36% to 11%, Europe 18% to 7%, Japan 24% to 9%, Asia from near 80% to under 30%. 36 to 11 is not profit falling; it is the slope, and valuations are set on it. The S&P 12-month target is 8,700 against a spot near 7,619, a 14% gain, on 11% earnings growth.

14 over 11 is 1.03. The extra 3% is multiple expansion. Goldman, while halving earnings growth, still assumes the P/E drifts up. That is the thinnest layer of the outlook: whether 11% lands decides if the 3% is expansion or contraction.

What is really paying is debt

One number deserves more attention than 92%: $687 billion of AI-related debt issuance in 2026 against $115 billion of equity.

Debt is 6 times equity.

That is why Goldman keeps returning to credit: it forecasts US investment-grade spreads widening from 88 basis points to 105, high yield from 320 to 410, and warns that indigestion in large-scale AI debt financing would widen spreads across the sector.

Translated: data centers are increasingly built with borrowed money. Pickier creditors raise funding costs, and higher costs worsen the crowding out of other investment in 2027. The most uncomfortable party may be neither shareholders nor chipmakers but AI debt holders: upside capped at the coupon, downside a repricing.

The soft spot

Goldman's 10-year Treasury target is 4.6%; when written, the market was 4.99%, later 5.15%. An allocation call built on a rate path was off 55 basis points on rates first.

55 is not small.

A sell-side outlook maps which variables entered the model and at what weight. Goldman put the weight in two places: AI investment growth, which sets timing, and its share of market cap, which sets pain.

The checklist has three items: Q1 2027 AI capex growth actually below 2026; investment-grade spreads at 105; S&P earnings growth at 11% or single digits.

Two of three, and the outlook holds.

Whether the feast is ending is another question. Goldman says the yearly addition is shrinking, not that the table is being cleared.

Those two sentences are a long way apart.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "Three Numbers That All Rise. The One That Falls Is the Difference," 17 November 2026, https://neuroai.site/a/na-capital-gs-ai-peak-1117

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Capital & Markets” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories