Goldman has the US-China AI capex gap narrowing to 7.5x in 2026 before reopening wider in 2027

Goldman Sachs Global Investment Research's latest hyperscaler numbers imply a US-to-China capex ratio of 7.8x in 2025, 7.5x in 2026, and 8.3x in 2027, on $764B US versus $102B Chinese cloud spend in 2026. Goldman's November 2025 publication had China's 2026 figure above $70B, so part of that dip is a revision, not a fresh forecast.

For the narrowing to be real, something has to be true about what the money buys. The visibly cheap things are the artifacts themselves, the published weights and the published prices. The cheapest frontier output price undercuts Claude Fable 5 by about 3.3x, with Moonshot's Kimi K3 at $3.00/M input and $15.00/M output against $10 and $50. A remarkable amount of this field's output is simply given away: Alibaba ships Qwen3.5 under Apache 2.0, Tencent's final Hy3 carries Apache 2.0, and Robbyant, an embodied-AI company under Ant Group, open-sourced a robot foundation model called LingBot-VLA 2.0 under Apache 2.0 covering both code and weights, so the artifact itself is close to the cheapest thing anyone here produces. Publishing an artifact costs almost nothing; standing up the capacity to serve it at scale does not.

That is where the narrowing stops looking like convergence. Epoch AI puts Chinese companies at just over 5% of the cumulative compute of leading AI chips as of end-2025, on a share that has been decreasing. Three days after K3 launched, Moonshot paused new paid membership signups because demand pushed its GPUs to capacity. You cannot spend your way out of that inside one fiscal year.

What cuts against it is that the capability is real, and it is also unpriceable from the outside. K3 took #1 on the blind Frontend Code Arena and sits #9 on Text Arena, as of 2026-07-17 since those boards drift. The costs are equally unverifiable: DeepSeek's $5.576M is the paper's own arithmetic at an assumed $2 per GPU-hour and excludes prior research and ablations, and Moonshot's CEO disavowed the $4.6M figure outright, so neither can be compared to a national capex line without a category error, and neither is audited while the scope of what each covers is unclear.

For a shareholder this resolves less than it appears to. The 2026 narrowing is in the model; the capacity to make it durable is not demonstrated. Holding any of these names means underwriting either the reacceleration to 8.3x or the narrowing persisting below 8x, and the model does not tell you which.

Positions: long index funds and US large-cap tech including Nvidia and Microsoft. No shorts, no options.