China’s Cheaper AI Models Are Reshaping the US-China Race

The United States still leads the global AI race on frontier capability and chips, but J.P. Morgan now frames the US-China contest as three different strengths, with China winning on price-per-task and India building the readiness and capital-market breadth to absorb AI at scale.

The bank’s report, titled “Semiquincententacles: The US grip on global markets at 250,” argues the next phase of AI competition won’t be decided at the frontier alone. Token pricing, benchmark scores, API traffic, enterprise procurement, and chip supply all carry weight, and each country now leads in at least one of those layers.

The US Still Owns the Frontier

J.P. Morgan frames the US as the most vibrant and prepared country for AI, with China close behind on some readiness measures, citing Stanford’s Vibrancy Index and other indicators. That lead rests on four reinforcing pieces: frontier AI laboratories, the dominant chipmaker, hyperscale cloud providers, and deep enterprise customers all sit in the United States.

Nvidia alone accounts for the bulk of global AI accelerator revenues, the report notes, while Google, Amazon, Microsoft, and Meta are deploying custom-designed chips to lower costs and expand capacity. The custom-chip build-out at hyperscalers is starting to compress Nvidia’s margins on commodity inference. Big Tech’s 2026 AI infrastructure and hiring pivot underscores the capital intensity behind the US position.

Productivity data back the position. The US leads the G10 in labour productivity and total factor productivity, and growth in information and data-processing sectors accelerated after generative AI tools reached production. That link between AI and measurable output is the hardest advantage for any rival to replicate quickly.

Where the US Lead Actually Sits

The US position is most concrete in semiconductors and cloud. Nvidia continues to dominate AI accelerator revenues, the report notes, but competition from custom hyperscaler chips is closing the gap on margin and supply.

Frontier capability also remains concentrated in US firms. J.P. Morgan says advanced uses such as cybersecurity, scientific discovery, agentic systems, and large-scale reasoning still depend on closed frontier models, where capability, reliability, and integration can outweigh price. In those workloads, “good enough” rarely is.

Stanford’s Vibrancy Index, which feeds J.P. Morgan’s framing, shows the US lead is widest in research and development, responsible AI, and infrastructure. China consistently ranks second across the same pillars. India, the new third-place country, is climbing on talent and economy but lags on infrastructure and the deepest research layers.

The next battle sits one layer down. As AI usage shifts from experiments to embedded business workflows, per-token cost starts to dominate procurement. The bank notes that some enterprises have already begun shifting workloads from expensive frontier models to lower-cost alternatives, a pattern that fits the Chinese cost story directly.

China’s Play Is Cheaper, Not Better

J.P. Morgan says China dominates the efficient frontier in intelligence-per-dollar, naming a small set of firms whose open-weight models are pulling away on price without falling far behind on capability.

• DeepSeek
• MiniMax
• Xiaomi
• Alibaba

The framing matters. China isn’t trying to out-frontier OpenAI or Anthropic on every benchmark. The strategy is to make “good enough” AI much cheaper, which is enough to capture the high-volume commercial workloads that frontier labs treat as secondary. In the report’s own scoring, Chinese models cluster tightly around closed frontier systems, while US models appear further out on cost relative to capability.

The Real Cost Gap on Tokens

The hard numbers come from the J.P. Morgan report. By April 2026, leading Chinese open-weight models scored within a few dozen Elo points of closed frontier models and cost 10x-50x less per token, according to the bank’s analysis.

That gap is showing up in real traffic. The report cites a surge in API calls to Chinese models on OpenRouter. OpenRouter’s December 2025 study analysed over 100 trillion tokens of real-world usage, and its leaderboard by early March 2026 showed the three most popular models on the platform were all Chinese: MiniMax M2.5, Moonshot AI’s Kimi K2.5, and Zhipu GLM-5. An independent read of OpenRouter’s leaderboard data traced the surge to developer and agentic workloads, not enterprise core systems.

• 78.6: US score on Stanford’s 2024 AI Vibrancy Tool
• 21.59: India’s score, third place
• 10x-50x less: per-token cost of leading Chinese open-weight models vs closed frontier, April 2026
• 17% to ~40%: S&P 500 market cap held by the 10 largest US stocks, 2015 to now
• All 3: top models on OpenRouter by early March 2026 were Chinese

The funding backdrop inside China mirrors the cost story. Moonshot AI’s recent funding round shows the country’s leading model companies are drawing record capital as they compete on price. For the long tail of enterprise work where price compounds by the billion tokens, the J.P. Morgan framing is clear: the next dollar of AI spend may not go to the highest-scoring system.

India Is an Adoption Story, Not a Frontier Story

India doesn’t yet rank among frontier AI powers, but it sits firmly on the readiness map. Stanford HAI’s Global AI Vibrancy ranking tool, updated in December 2025 with data through 2024, ranks India third globally with a score of 21.59, up from seventh place in 2023. The United States scored 78.6 and China 36.95, a lead the index attributes to frontier capability, infrastructure, and research depth.

The index covers seven pillars: research and development, talent and education, economy and investment, policy and governance, infrastructure, responsible AI, and public opinion. India’s gain reflects expansion in AI-skilled professionals and rising activity in its software industry, alongside growth in scientific publications, patents, and global collaborations.

The opportunity is deployment, not invention. India’s technology services base, large digital economy, and enterprise sector make it a natural absorber of AI systems across industries. J.P. Morgan notes that India remains “far behind” the US and China in overall AI capability and semiconductor depth, but its readiness footprint is harder to dismiss, especially for the long-tail work that AI cost curves are now opening up. Infosys chairman Nandan Nilekani’s framing of AI as an amplifier of Indian IT, not a replacement for it, captures the same deployment-side logic. The country is positioning to absorb and distribute, not to out-frontier anyone.

The Capital-Market Lens on India

The same J.P. Morgan report puts a second number behind the India story. The full J.P. Morgan AI readiness findings for India note that the ten largest US stocks represented 17% of the S&P 500’s market capitalisation in 2015. That figure has since risen to around ~40%, the report says.

India sits on the other end of that spectrum. The bank calls the 40% concentration “among the three lowest equity concentration figures in the world; only Japan and India have less.” For investors looking at AI exposure with broader market breadth, that profile gives India a structural appeal that frontier-lab returns cannot match on their own.

What a Split AI Market Looks Like

The clearest read of the report is that AI leadership is no longer a single ranking. The US leads on frontier capability, chip dominance, cloud infrastructure, investment, and enterprise distribution. China is gaining on cost-efficient models that could absorb high-volume commercial workloads. India is building readiness, but still has to close gaps in core AI capability and semiconductor development.

For procurement, the choice gets more complicated. The best model isn’t always the right model. Enterprises are weighing performance against cost, data security, compliance, reliability, governance, and vendor dependence. A cheaper model suits routine work; a frontier model suits sensitive and mission-critical uses. A frontier model may be too costly for every task.

J.P. Morgan warns that policy restrictions and supply-chain vulnerabilities could affect the future US lead, putting semiconductors, export controls, cloud capacity, and domestic AI investment at the centre of the next phase. The bank tracks the AI race through token pricing, benchmark performance, API usage, enterprise procurement, chip supply, cloud spending, and national AI policies. Three open questions sit inside that frame: whether China’s cost advantage persists, whether US firms respond with cheaper and more efficient offerings, and whether India can turn readiness into deeper capability.

Frequently Asked Questions

What does the Stanford Global AI Vibrancy Tool measure?

It ranks 36 countries across seven pillars: research and development, talent and education, economy and investment, policy and governance, infrastructure, responsible AI, and public opinion. The December 2025 release draws on data through 2024, with the United States scoring 78.6, China 36.95, and India 21.59.

Why are Chinese AI models so much cheaper than US frontier models?

J.P. Morgan attributes the gap to a combination of cheaper electricity, algorithmic advances such as Mixture-of-Experts architectures that activate only a subset of parameters per query, and a domestic price war that has pushed token prices sharply lower. The bank names DeepSeek, MiniMax, Xiaomi, and Alibaba as the firms dominating the “efficient frontier” in intelligence-per-dollar.

Does this mean the United States is losing the AI race?

The bank doesn’t say so. The US still leads on frontier capability, chip supply, cloud infrastructure, and productivity-linked AI adoption, and frontier US models remain essential for cybersecurity, scientific discovery, agentic systems, and large-scale reasoning. The shift J.P. Morgan describes is in the high-volume, cost-sensitive layer, not the frontier.

What does India’s third-place ranking actually mean?

India’s 21.59 score reflects the conditions for absorbing and deploying AI at scale, including R&D, talent, economy, infrastructure, policy, governance, and public opinion. India moved from seventh in 2023 to third in 2024. The ranking is readiness, not frontier capability, and India remains behind the US and China on semiconductors and core AI development.

What should companies do with a split AI market?

Match the model to the workload. Frontier US systems still anchor sensitive, mission-critical, and high-stakes tasks. Lower-cost Chinese open-weight models can absorb routine, high-volume work such as coding, summarisation, document processing, and customer service where per-token cost compounds at scale. Procurement is becoming multi-axis rather than single-rank.