China's Controversial AI Race: A Darker New Phase
By Ali Sadikin Ma · · Updated
Category: Technology
China has nearly closed the AI benchmark gap to just 2.7 percentage points behind the US while spending 23 times less—a feat documented by CFR, CSIS, and Stanford HAI as partly achieved through systematic IP distillation attacks, chip smuggling networks, and open-source soft power dominance. Chinese AI companies behind Xinjiang's surveillance infrastructure are now exporting it to 40+ countries, while Alibaba's Qwen commands over 50% of global open-source AI downloads. The race has fundamentally shifted from model capability to infrastructure dependency, with starkly different implications for developers, enterprise decision-makers, and emerging market governments.
China's AI might be cheaper, faster—and built on your data.
That's not an assumption. That's the conclusion drawn from official reports by CFR, CSIS, and Stanford HAI published between 2025 and 2026.
Three things most people still don't realize about this controversial China AI race:
One: how exactly did China close the AI performance gap from 31 points down to 2.7—while spending 23 times less than the US? Two: where is the surveillance technology from Xinjiang going now? Three: why did Singapore and Malaysia just switch their sovereign AI to Chinese-built models?
None of those three questions have a comfortable answer.
But let's start with the number that's hardest to ignore:
DeepSeek V4-Pro is priced at $3.48 per million tokens. GPT-5.5 runs $30.21 per million tokens. That's a 35x price gap—per Forbes, April 2026.
And if you think this is just about a cheaper price tag:
Keep reading.
The Numbers Look Reasonable—Until You Read the Fine Print
The United States invested $285.9 billion in private AI in 2025, while China invested just $12.4 billion—23 times less. But as of March 2026, the performance gap between each country's top AI models had shrunk from 17–31 percentage points in 2023 to just 2.7 points, according to the Stanford HAI 2026 AI Index Report.
A near-zero gap—on one-twenty-third of the US budget.
How is that even possible?
DeepSeek-R1 was developed in two months for under $6 million using Nvidia A800 chips. It scored 97.4% on the MATH benchmark—surpassing OpenAI o1. When the news broke, Nvidia's stock dropped 18% in a single trading day, per CSIS analysis.
That leaves two possibilities:
First—China discovered engineering efficiencies that no one had thought of before. Second—there's something that didn't make it into their official reports.
The answer is both.
Liang Wenfeng, DeepSeek's CEO, said in an interview cited by CSIS: “Money was never our problem; the advanced chip export bans were.”
Notice what he didn't say. Because this controversial China AI race always has layers that never show up in any press release.
How China Closed the Gap: The Part Nobody Wants to Admit
DeepSeek admitted in September 2025 that their R1 model “inadvertently” distilled from ChatGPT and Anthropic's Claude. CFR 2026 documented something far more systematic: 24,000 fake accounts and 16 million interactions with OpenAI and Anthropic models to extract capabilities those US models had already learned.
Early versions of the DeepSeek chatbot even answered “I am ChatGPT” when asked—per ChinaFile's 2025 report.
Jessica Brandt, CFR Senior Fellow for Technology and National Security, said it plainly: “V4's capabilities reflect, at least in part, access to US intellectual property obtained illegally.”
But that's just one channel.
On the chip side: CSIS identified at least 8 separate smuggling networks transporting Nvidia H100 chips to China—each involving transactions worth over $100 million. Singapore was found to account for 18% of Nvidia's revenue but only 2% of actual shipping destinations—a clear signal that chips were being rerouted before reaching their final destinations.
The CEO of Scale AI confirmed that DeepSeek has access to approximately 50,000 H100 chips—chips that are officially banned from export to China.
And there's still one more layer:
Cybersecurity firm Feroot Security found heavily obfuscated hidden code inside the DeepSeek app—code that links user login data to China Mobile, a company banned from operating in the US due to its ties to the Chinese military, per Forbes 2025.
But that's still not the most alarming part of this controversial China AI race.
The Controversial China AI Race: Surveillance Technology Exported to 40+ Countries
Xinjiang's AI surveillance system isn't a local project—it's an architecture that's been exported. SenseTime, Megvii, and Yitu Technology, the Chinese AI companies that supplied facial recognition systems for Xinjiang, are now exporting their surveillance technology to more than 40 countries, according to CSIS 2024 analysis.
The domestic scale alone already exceeds what most people imagine:
2.6 million CCTV cameras. Approximately 36 million DNA samples collected. 98% biometric coverage across southern Xinjiang—per IHS Markit 2025 data and Human Rights Watch.
IJOP 2.0, the latest system, integrates real-time financial transaction data and uses ML trained on historical detention data. It's a feedback loop that continuously updates itself.
Darren Byler, author of In the Camps: China’s High-Tech Penal Colony, wrote in 2025: “The surveillance system in Xinjiang is no longer the exception—it has become the new normal.”
More than 540,000 Uyghurs are estimated to be in formal detention as of 2024, according to Adrian Zenz's data from the Jamestown Foundation.
But here's what almost never gets written clearly enough:
Experts warn that the facial recognition systems being sold today need only “a few additional lines of code” to become targeting systems—this is a technical assessment from security researchers, not just an analogy.
This is the dimension of the controversial China AI race that almost never gets mentioned in the same breath as benchmark numbers.
And while that's happening, the other side of this race is quietly winning the global market in a very different way.
Open Source Is China's New Soft Power—and It's Already Working
On the other side of the controversial China AI race, there's a strategy far more subtle than distillation attacks and chip smuggling. Alibaba's Qwen reached 1 billion cumulative downloads on Hugging Face as of March 2026—faster than any open-source model family in history—and now accounts for more than 50% of all global open-source AI model downloads, per Forbes 2026. For the first time, China is leading the global AI download market.
Chinese open-weight models surpassed the US share in global downloads in August 2025: 17.1% for China versus 15.86% for the US—per MIT Technology Review. A first in history.
22 of the 50 most-used generative AI apps in the world are Chinese-made. 19 of those 22 are used primarily outside of China—even as the governments of the US, Canada, Australia, Czech Republic, and Taiwan have banned their use on government devices.
But this is what truly shifts the narrative:
Singapore announced it's replacing Meta's Llama with Alibaba's Qwen as its sovereign AI model. Malaysia followed suit—the first time major US-aligned governments have made a decision like this.
Kevin Xu, tech investor and former Obama administration official, put it perfectly: “Open source is the soft power equivalent in the tech world.”
But there's a layer that isn't immediately visible:
Research from USC's Information Sciences Institute found that more than 11% of Chinese AI model responses don't match their visible reasoning process—evidence of “soft censorship” that embeds framing without explicit refusal, per ChinaFile 2025.
Qwen has already generated more than 180,000 derivative models on Hugging Face—more than Google and Meta combined. Whoever controls the base model controls the direction of the ecosystem.
That's why the controversial China AI race isn't just about who has the smartest model. And that brings us to the question that actually matters most.
The Real Race Isn't About Benchmarks—It's About Who Controls the Infrastructure
The controversial China-US AI race has shifted from a benchmark competition to an infrastructure competition. ByteDance alone plans to invest $23 billion in AI infrastructure in 2026, per the Financial Times. Meanwhile, China added 543 GW of new energy capacity in 2025—at an expansion cost roughly three times lower than the US.
The US faces a projected 49 GW power shortage for data centers by 2028, according to Morgan Stanley.
On the chip side: domestic Chinese AI chips account for roughly 41% of China's AI chip market in 2025, with around 50% of that coming from Huawei alone. The Huawei Ascend 950PR is scheduled for 750,000 units of production in 2026—per Reuters estimates cited by Brookings.
One number worth remembering:
The Nvidia Blackwell Ultra GB300 delivers 15 petaflops FP4. The Huawei Ascend 950PR sits at 1.56 petaflops—roughly one-tenth. The compute gap is real and still enormous. CSIS analysts describe Huawei's CANN software ecosystem as “years behind CUDA”—still buggy and immature.
But here's the complete picture:
All three loops opened at the start now have answers. On the distillation attack: V4 was open-sourced precisely because the IP was already baked into its weights. On chip smuggling: those smuggling networks exist because China is building domestic alternatives in parallel. On surveillance exports: what's really being exported isn't just surveillance technology—it's infrastructure dependency.
Kyle Chan of Brookings summed it up in Congressional testimony in April 2026: “The winner of the AI race will not be determined solely by who builds the most powerful model, but by who most effectively translates AI into broad economic and social advantage.”
This race is no longer about who's smartest. It's about who becomes the default—and default advantage compounds.
The controversial China AI race has redefined what winning means in the global technology competition.
Where You Stand Determines What This Means for You
This controversial China AI race doesn't have a single conclusion that applies to everyone. What you take away from it depends on where you're sitting right now. The implications differ for every actor involved. Here are the three most relevant types—and what it concretely means for each.
If you're a developer or tech professional
Chinese AI tools—especially Qwen—offer cost advantages that are hard to ignore. Stanford and UC Berkeley have trained top-performing models on Qwen for between $30 and $50. But before you deploy to production in the context of this controversial China AI race, audit the data handling policies. Feroot Security's findings about hidden code in DeepSeek aren't rumors—that's published research from 2025. Weigh the compliance and reputational risks, not just the benchmark numbers.
If you're in a policy or enterprise decision-making role
The enforcement gap is structural, not a temporary weakness. The US Bureau of Industry and Security has fewer than 600 employees and a $200 million budget to oversee global semiconductor trade worth trillions of dollars, per CSIS 2025. The smuggling won't stop on its own. Decisions about AI vendors today are decisions about infrastructure dependency for the next decade.
If you're in a developing country or emerging market
You're caught between two options that both come with hidden geopolitical strings attached. US AI is expensive and comes with pressure to align on policy. Chinese AI is cheap but comes with a censorship architecture that has no appeals process—as Yaqiu Wang of the University of Chicago noted: “US restrictions are usually security-driven. China's are state-driven, with no appeals process.” True digital sovereignty requires a third option that doesn't exist yet.
Which one of these are you?
FAQ: The Questions Behind the Headlines
The controversial China AI race surfaces three questions that keep coming up in tech and policy discussions.
Did DeepSeek actually steal from OpenAI?
CFR 2026 documented “systematic distillation attacks”—24,000 fake accounts and 16 million interactions with OpenAI and Anthropic models to extract their capabilities. DeepSeek itself admitted in September 2025 that R1 “inadvertently” distilled from ChatGPT and Claude. The debate is over the word “inadvertently”—the underlying facts are not in dispute.
Should I be worried about using Chinese AI apps?
It depends on the context. For non-sensitive personal use, the risk profile is different from enterprise use. Feroot Security found hidden code in DeepSeek linking login data to China Mobile—a company banned in the US due to its ties to the Chinese military. For enterprise or sensitive data, technical due diligence is mandatory before any deployment.
Has China already won the AI race?
Depends on your definition. The frontier model benchmark gap is down to 2.7 points per Stanford HAI 2026—razor-thin. But on global open-source adoption, patent volume, and infrastructure capacity, China is already ahead on several metrics. Ryan Fedasiuk of the American Enterprise Institute put it well: this race is now about markets and product sales, not just model capability.
Share this with your team before your next AI tools decision meeting—benchmark numbers are only half the story.
Or save it for the next time someone tells you Chinese AI is just cheaper. This isn't just about price. It's a different game entirely.