China Could Close the U.S. Lead in Top AI Models, Goldman Says
Goldman Sachs research says China could close the gap with the U.S. in building the world's most powerful AI models.
That is the central finding of "The US-China Tech Race," published Dec. 4, 2025. Goldman Sachs Research The report treats leadership in frontier models, the largest and most capable AI systems, as contestable rather than locked in for U.S. labs.
Goldman gave a follow-up update on Oct. 10, 2026: "Competition Among China's AI Models Is Intensifying as Chip Sector Advances." Goldman Sachs In that update, Ronald Keung said "the gap in performance, especially in coding, is narrowing between leading AI models." In plain terms, test scores are bunching at the top, with coding as the current point of difference.
The spending behind that race runs into the hundreds of billions. Goldman's baseline points to $765 billion in annual AI CapEx in 2026, rising to $1.6 trillion in 2031. Goldman Sachs CapEx is capital spending, money for data-center buildings, AI chips called accelerators, plus power and networking. The limit is compute, meaning processing power. The question is how that power gets shared out.
China is aiming its share at self-sufficiency. It is ramping up work on a home-grown AI system that can run without Western technology, the Wall Street Journal reported on July 29, 2025. Wall Street Journal A separate Journal report on July 1, 2025 said China is investing heavily in its own AI supply chain to cut into America's lead. Wall Street Journal
China is also sharing its work more widely. Most Chinese AI labs publish model source code and weights, the settings that shape how a model acts, according to a March 23, 2026 paper. U.S.-China Economic and Security Review Commission This open-source method cuts the cost of fine-tuning, distillation and fitting models into apps. It also lets more outsiders test what the models can do and where safety risks sit.
The rule book is different. In China, companies must get Beijing's approval before launching chatbots built on large language models, the Journal reported on June 20, 2024. Wall Street Journal That approval step sits next to the Journal's Aug. 30, 2025 account of a different Chinese vision for rolling out AI. Wall Street Journal
The broader context here is that one gap has become two races. One is pre-training, teaching the most capable models. The other is diffusion, spreading use through open weights, tools and home-built infrastructure. Coding gains count because they can be sold through software productivity and agent workflows. If test gaps close there first, closed paid models face easier switching by customers.
In my view, the spending path and the open-source path need to be read together. An envelope of $765 billion to $1.6 trillion a year brings a risk of overbuilding if use, inference yields and business adoption lag. For savers and investors, that matters because open sharing can lift use by pulling in more developers, yet it squeezes profits at the model level and pushes value toward compute owners, distributors and systems integrators. The watch list is not model launches alone. It is steady training throughput under supply limits, the pace of domestic substitution, and whether approval frictions cancel out sharing advantages.
Looking at what this means for risk assessment, Goldman is making an odds-based call. Could close the gap is not has closed the gap. The October 2026 update narrows the claim to measured gains, chiefly in coding, amid tougher home competition and chip progress. Whether any catch-up lasts depends on repeat access to fast compute, talent density and capital discipline over a multi-year build.
