China Unlikely to Overtake US in AI Race Within Five Years, Experts Say

China will not overtake the United States in artificial intelligence (AI) within the next 3 to 5 years that is the verdict from senior researchers inside Alibaba, Tencent, and Zhipu AI. The 3 main obstacles are a compute deficit 10 to 100 times larger than most people realize, no domestic extreme-ultraviolet (EUV) lithography machines to build advanced chips, and a cultural shortage of risk-taking in foundational research.

These are not outside critics. These are China’s own AI leaders speaking plainly at a major industry conference in Beijing. Their candor matters because it cuts through the hype around models like DeepSeek and Qwen and forces a clear-eyed look at the structural gap between the two countries.

This article breaks down what each researcher said, why the compute gap is so severe, where China still has real strengths, and what needs to change before any Chinese company can challenge OpenAI or Google DeepMind. For more on the current state of AI models, read our guide on which AI model is best in 2026.

3 Numbers That Define the Gap

What Alibaba’s Lin Junyang Actually Said

Lin Junyang is the technical lead of Alibaba’s Qwen AI team the group behind one of China’s best-performing open-source large language models (LLMs). His assessment was blunt: fewer than 1 in 5 Chinese companies will surpass OpenAI or Google DeepMind in 3 to 5 years, and he called that estimate “highly optimistic.”

The reason is raw compute. U.S. companies operate with computing capacity 1 to 2 orders of magnitude that is 10 to 100 times larger than what most Chinese firms can access. Lin explained that OpenAI and others pour enormous computational resources into next-generation research, while Chinese teams are stretched just meeting daily inference demand. There is nothing left over for exploratory training runs or architectural experiments.

This is a structural problem, not a funding problem. Fresh capital cannot buy chips that export controls block. Two IPOs totaling over $1.1 billion by Zhipu AI and MiniMax in January 2026 generated headlines, but industry insiders were quick to note that money alone does not close a hardware gap measured in orders of magnitude.

Zhipu AI’s Tang Jie: ‘The Gap May Actually Be Growing’

Tang Jie, co-founder and chief AI scientist at Zhipu AI, offered the starkest framing. The performance gap between Chinese and U.S. models is not narrowing it may be widening. His explanation points to a strategic blind spot in how China’s AI progress gets reported.

When Chinese open-source models score well on public benchmarks, observers often conclude that China is catching up. Tang rejects this reading. U.S. labs particularly OpenAI, Anthropic, and Google DeepMind do not release their most advanced models publicly. Benchmarks compare published Chinese models against published U.S. models, but the published U.S. models are not the frontier. The frontier is tested and refined behind closed doors, giving American firms a strategic edge that no benchmark can capture.

This is the gap competitors rarely discuss: not just the compute gap, but the closed-frontier gap. U.S. labs quietly build and test models that never appear in public leaderboards. By the time a Chinese model matches a visible U.S. benchmark, the real frontier has already moved further ahead.

Tencent’s Yao Shunyu: The Risk-Taking Culture Problem

Yao Shunyu is Tencent’s newly appointed chief AI scientist and a former OpenAI researcher. His perspective carries weight on both sides of the Pacific. Yao does not dismiss China’s chances but his 4 structural obstacles are the most sobering list from any of the 3 researchers.

The 4 structural obstacles Yao identified:

• No domestic EUV lithography machines China cannot build the advanced semiconductor fabrication tools needed for leading-edge chips at scale.
• Slower enterprise adoption Chinese businesses are not integrating AI into operations at the pace required to generate the feedback loops that improve models.
• Limited foundational research investment China excels at applying and optimizing existing frameworks. Building new ones is a different skill.
• Shortage of the risk-taking spirit Yao’s words, not a paraphrase. China needs researchers willing to define the next paradigm, not just extract maximum performance from the current one.

Yao points to China’s success in electric vehicles (EVs) and advanced manufacturing as evidence that scaling works when the framework already exists. AI leadership requires something harder: creating the framework itself.

The Hardware Reality: Huawei vs. Nvidia by the Numbers

The chip gap is not theoretical. According to analysis from the Council on Foreign Relations (CFR), the best U.S. AI chips are currently about 5 times more powerful than Huawei’s best offerings. By 2027, that gap will widen to 17 times.

More striking: Huawei’s own public roadmap shows its next-generation chip in 2026 will perform worse than its best chip today. SMIC, China’s leading chip foundry, is stuck at 7-nanometer (nm) process technology because the equipment to go further EUV lithography machines from ASML is blocked by export controls from the Netherlands and the U.S.

Even under aggressive production assumptions, Huawei could generate only about 4% of Nvidia’s aggregate AI computing power in 2026, falling to 2% by 2027. A quantity strategy producing more inferior chips is not working because the quality deficit is too large.

Washington briefly approved Nvidia H200 chip sales to China in late 2025, but Beijing subsequently urged some Chinese tech companies to pause those orders, pushing domestic alternatives instead. This creates a policy paradox: Chinese AI developers need more compute to compete, but national self-reliance goals push them away from the hardware that would help most.

Where China Genuinely Leads: Open-Source and Optimization

The picture is not entirely bleak for China. 3 areas show real, measurable progress:

• Open-source LLMs Models like Qwen, DeepSeek, and others have closed much of the benchmark gap with U.S. models and sparked global developer adoption. China’s open-weight strategy has moved faster than anything U.S. labs have released publicly.
• Compute efficiency Chinese researchers are world-class at squeezing maximum performance from limited GPU resources. DeepSeek’s training cost efficiency impressed even U.S. engineers. This skill becomes more valuable as compute remains scarce.
• Enterprise AI applications Chinese firms are rapidly deploying AI tools in manufacturing, logistics, and consumer apps, building practical experience at scale that feeds back into model improvements.

These strengths are real. They explain why China narrowed the model performance gap from roughly 3 years in 2023 to 6 to 12 months in 2025, measured by public benchmarks. The problem is that the frontier U.S. labs actually work on is not the public benchmark frontier. See Tang Jie’s closed-frontier argument above.

For more on China’s technology push across sectors, read our coverage of China’s 100 GHz light-speed chip development and Xiaomi’s fully automated dark factory.

Why the U.S. Advantage Is Structural, Not Just Temporary

5 main reasons the U.S. lead is durable rather than momentary:

• Capital scale Five U.S. companies (Meta, Alphabet, Microsoft, Amazon, Oracle) are expected to spend over $450 billion on AI-specific capital expenditures in 2026 alone. Chinese investment is a fraction of that.
• Chip design dominance Nvidia designs chips with no viable domestic Chinese equivalent. TSMC manufactures the most advanced chips, and Taiwan is effectively aligned with U.S. export control policy.
• Closed-model advantage OpenAI, Anthropic, and Google DeepMind do not publish their most capable models, protecting the real frontier from benchmarking.
• Ecosystem depth U.S. firms control roughly 70% of global AI compute capacity. Chinese firms control about 10%. This gap compounds over time.
• Talent pipelines U.S. universities and labs attract global AI talent. China has strong domestic researchers but still faces emigration of top engineers to U.S. labs.

What China Would Need to Close the Gap

Yao Shunyu’s conditional was specific: a Chinese company could lead global AI in 3 to 5 years only if several major obstacles are addressed simultaneously. Those conditions are:

• Domestic EUV lithography ASML holds a near-monopoly on EUV machines. China building its own would take years and requires material science breakthroughs, not just engineering effort.
• Rapid enterprise adoption China needs its largest companies to integrate AI deeply and fast, creating the feedback data that drives model improvement.
• Paradigm-defining research Not optimization of existing frameworks, but willingness to pursue architectures that might fail spectacularly before they work.
• Sustained foundational investment Basic research with 5 to 10 year payoff horizons, funded at U.S. university and national lab scale.

None of these are impossible. But none will happen within 3 to 5 years. The 3 researchers agree on this even the most optimistic framing puts the probability below 20%.

What This Means for Global AI in 2026 and Beyond

The US–China AI race shapes every major technology decision made today from export control policy to which AI models enterprises adopt. 4 practical implications:

• Open-source will remain China’s primary competitive strategy expect more high-quality open-weight models from Chinese labs through 2026 and 2027.
• Efficiency research gains importance Chinese teams will push compute-efficient training methods further than any U.S. lab because they have to.
• The benchmark gap will mislead as Tang Jie warned, public scores will continue to suggest near-parity while the real gap grows. Read benchmarks as floor measurements, not ceiling comparisons.
• Chip policy is the decisive variable if advanced chip exports to China expand significantly, the compute gap could shrink faster than hardware alone suggests; if restrictions tighten further, the gap widens faster.

For related reading on AI’s broader societal impact, see our article on whether AI could make medical school unnecessary and the latest AI model comparisons for 2026.

Frequently Asked Questions

Will China overtake the U.S. in AI within 5 years?

No the probability is below 20%, according to Lin Junyang of Alibaba’s Qwen team, who called that estimate “highly optimistic.” The 3 core obstacles are compute capacity 10–100× smaller than U.S. firms, no domestic EUV chip manufacturing tools, and limited foundational research investment.

Why can’t China just build its own AI chips?

Advanced chip manufacturing requires extreme-ultraviolet (EUV) lithography machines. ASML, the Dutch company that builds them, is the only commercial supplier. Export controls from the U.S. and its allies prevent ASML from selling EUV machines to China. SMIC, China’s leading foundry, is stuck at 7nm process technology as a result, while TSMC and Samsung manufacture chips at 3nm and below.

How does Huawei’s AI chip compare to Nvidia’s?

Nvidia’s best chips are currently about 5 times more powerful than Huawei’s best AI chips. By 2027, that gap will reach 17 times. Huawei’s own public roadmap shows its 2026 chip will actually perform worse than its current best an unusual regression that reflects manufacturing constraints at SMIC.

Is China making any real progress in AI?

Yes. China leads in open-source LLM development, with models like Qwen and DeepSeek closing the public benchmark gap to 6–12 months in 2025, down from 3 years in 2023. Chinese researchers also excel at compute-efficient training. The gap that matters is in closed-frontier models and raw compute infrastructure not in published benchmark scores.

What did Tencent’s AI chief say about China’s chances?

Yao Shunyu, Tencent’s chief AI scientist and a former OpenAI researcher, said China could lead global AI in 3 to 5 years only if it addresses 4 structural problems: lack of domestic EUV lithography, slow enterprise AI adoption, limited foundational research, and a shortage of the risk-taking culture needed to define new AI paradigms not just optimize existing ones.

What is the closed-frontier gap in AI?

The closed-frontier gap is the difference between AI models that U.S. labs publish publicly and the more advanced models they test and refine internally. Public benchmarks only compare published models. Because OpenAI, Anthropic, and Google DeepMind keep their frontier models private, Chinese teams improving against public benchmarks are chasing a target that has already moved. Tang Jie of Zhipu AI identified this as the reason the real AI gap may be growing even as benchmark scores look closer.

How much does the U.S. spend on AI compared to China?

5 U.S. companies alone Meta, Alphabet, Microsoft, Amazon, and Oracle are expected to spend over $450 billion on AI-specific capital expenditures in 2026. That exceeds the entire Apollo program budget (inflation-adjusted) by more than $100 billion. Chinese AI investment is growing but remains a fraction of this scale.

Bottom Line

China’s AI industry is not stagnating. Open-source models, compute efficiency, and enterprise deployment are all advancing fast. But the 3 researchers who know the field best Lin Junyang, Tang Jie, and Yao Shunyu agree that the structural obstacles between China and U.S. AI leadership are real, large, and growing, not shrinking.

The compute gap is not a funding problem that IPOs solve. The chip gap is not an engineering problem that more engineers solve. The closed-frontier gap is not a benchmark problem at all. All 3 require systemic changes that will take longer than 5 years. What China can do and is doing is build the world’s best efficiency-first AI ecosystem. That has real value. It just will not produce the next foundational AI breakthrough. That remains a U.S. advantage for the foreseeable future.