According to the Stanford AI Index Report, US private AI investment reached $285.9 billion in 2025, while China recorded just $12.4 billion, a gap of roughly 23 times.1 US private investment was also significantly greater than Europe and China combined. With such a stark difference, how is China coming so close to American frontier models? Capital is supposed to be one of the defining barriers to AI development, yet China is showing that spending alone does not determine technological performance.

China is surprisingly close given the investment disparity. DeepSeek-R1’s strong performance in early 2025 demonstrated that powerful models can be built with far fewer high-end chips than many competitors. After US restrictions limited China’s access to Nvidia’s most advanced chips, DeepSeek trained its model using around 2,000 specialised chips alongside thousands of lower-grade chips, reducing the computing resources available to it.2 The restrictions created challenges, but they also strengthened the incentive for Chinese firms to improve efficiency and technological independence.

Other Chinese models point in the same direction. Alibaba’s Qwen model has demonstrated the ability to work autonomously for extended periods on coding tasks, writing code, fixing errors and refining its work with limited human intervention.3 This type of continuous optimisation could reduce the time and human input required for certain AI tasks.

The performance gap is narrowing

The chart below shows how the US-China AI model performance gap has narrowed.

Performance of top United States and Chinese models on the Arena. Source: Stanford HAI, 2026 AI Index Report.

The red line tracks the highest-performing US model at each point in time, while the black line tracks the highest-performing Chinese model using Arena’s Elo-like performance score.4

DeepSeek-R1 nearly matched the top US model in February 2025, trailing only by 5 points. By March 2026, the gap between the leading US and Chinese models was only around 40 points. The US has continued moving the frontier outward, but Chinese models have repeatedly caught up despite vastly lower measured private AI investment.

The striking feature is not simply that China is improving. It is the speed of convergence.

However, headline private investment figures do not capture the entirety of resources being committed to AI. The US and China have fundamentally different economic systems. While US AI financing is dominated by private companies and capital markets, China makes greater use of the state to direct resources towards strategic industries. As a result, the headline private-investment gap is distorted by three major factors:

1) The decline of traditional venture capital

Historically, Chinese AI start-ups relied heavily on Western venture capital. Geopolitical tensions, US investment restrictions and Beijing’s own regulatory crackdowns have reduced foreign private capital. Chinese AI firms have therefore had to move away from the traditional Silicon Valley venture-capital model.

2) State capital masked as market capital

Funding for Chinese AI start-ups can resemble private-market transactions on paper while ultimately being supported by government guidance funds, local governments and state-owned enterprises. Among AI firms receiving both government and private venture-capital funding, 71% received government investment first.5 This suggests that private investors may sometimes follow signals created by state capital rather than taking the initial risk independently.

3) Subsidies instead of private investment

Some operating costs for Chinese AI companies can also be supported directly by government programmes. These resources do not necessarily appear in measures of private AI investment.

The headline gap: US private AI investment in 2025: $285.9bn · China: $12.4bn · US investment: roughly 23× higher.

The investment gap therefore partly reflects different financing models. Even after accounting for this, however, China still appears to be developing competitive models with substantially fewer financial resources.

Scarcity may be forcing greater efficiency

Restrictions on advanced chips have forced Chinese laboratories to make more intensive use of the hardware they can access. Certain high-performance computing chips remain restricted, creating a stronger incentive to optimise models around less-advanced hardware.

Chinese labs also make use of techniques such as distillation, where smaller “student” models learn from the outputs of larger frontier systems. This can extract strong performance from more limited computing budgets. By building lower-cost models and optimising the use of less-advanced chips, Chinese developers can reduce the performance disadvantage created by hardware constraints.

Model distillation: A technique in which a smaller model learns to reproduce aspects of a larger model’s behaviour, allowing strong performance with lower computing and deployment costs.

AI research also crosses borders through technical diffusion. Once an architectural innovation, research paper, or technique becomes publicly known, competitors can build upon it. This means being the technological leader can be considerably more expensive than being a fast follower.

It is therefore tempting to look at the enormous US-China investment gap and conclude that American spending is inefficient. But model performance is only one return on AI investment. US models often push the performance frontier higher before Chinese competitors subsequently close part of the gap. American companies bear much of the cost of frontier research, while fast followers can concentrate resources on improving efficiency, lowering costs and adapting techniques that have already been demonstrated.

America is investing in future capacity, not just today’s models

A large share of US AI spending is not intended to improve today’s chatbot benchmarks. It is being invested in GPUs, data centres, energy capacity and cloud infrastructure. Economically, this is investment in future productive capacity and potential output.

That makes a direct comparison between 2025 investment and 2025 model performance potentially misleading. Major technology companies are spending hundreds of billions of dollars on infrastructure that may take years to reach full utilisation. UBS has projected roughly $4.1 trillion of hyperscaler capital expenditure between 2026 and 2028.6

China’s current ability to compete therefore does not necessarily tell us whether it can match the future productive capacity the US is creating. The two countries may be optimising for different constraints: China for efficiency under scarcity, and the US for scale and future capacity.

The economics of diminishing returns

The relationship between investment and AI performance is unlikely to be linear. Large gains may be achieved initially, while each additional dollar of investment generates progressively smaller improvements in benchmark performance.

The law of diminishing marginal returns offers one explanation for how the US can spend dramatically more without possessing dramatically better models. Once models reach a high level of capability, pushing the technological frontier further may become increasingly expensive.

The law of diminishing marginal returns: indicates that continuously adding more of one type of input, while keeping all other inputs constant, eventually leads to a decrease in the efficiency of additional input use.

China’s rise therefore does not prove that America’s AI spending is excessive. But it does challenge a more fundamental assumption underlying the investment boom that spending significantly more on AI will produce a proportionately greater technological advantage.

China’s progress emphasises that financial investment alone is not a sufficient measure of AI competitiveness. If Chinese firms can remain near the frontier at much lower measured cost, investors may need to reconsider the assumption that ever-increasing AI capital expenditure automatically translates into proportionately greater technological advantage.

The question is no longer simply who is spending more? It is what return will investors ultimately receive on the next dollar of AI capital?

Footnotes

  1. Stanford HAI, The 2026 AI Index Report: Economy (opens in a new tab), 13th April 2026.

  2. BBC News, DeepSeek: How China's 'AI heroes' overcame US curbs to stun Silicon Valley (opens in a new tab), 28th January 2025.

  3. CNBC, Alibaba shares rally after unveiling its 'most powerful' AI model as U.S.-China competition heats up (opens in a new tab), 3rd August 2026.

  4. Stanford HAI, The 2026 AI Index Report: Technical Performance (opens in a new tab), 13th April 2026.

  5. Stanford Center on China's Economy and Institutions, Government Venture Capital and AI Development in China (opens in a new tab), 1st December 2024.

  6. CNBC, How the AI debt binge shattered hyperscalers’ ‘unspoken contract’ with investors (opens in a new tab), 23rd February 2026.