One Year After DeepSeek Shook the World: The Real Challenges Facing China's AI Industry

The “Sputnik Moment” That Changed Everything

On January 20, 2025, as a new US president took office and dominated global headlines, a little-known AI startup in Hangzhou quietly released a model that shook Silicon Valley to its core.

DeepSeek R1 — a reasoning model released under the permissive MIT license, trained for roughly $6 million, with performance comparable to OpenAI’s flagship products. Wall Street dubbed it the “DeepSeek Shock.” Marc Andreessen called it AI’s “Sputnik moment.”

One year later, as the initial euphoria has settled, a clearer picture emerges. DeepSeek’s impact extends far beyond a single model release. It reshaped China’s AI development trajectory, transformed the global open-source landscape, and laid bare the structural challenges that China’s AI industry must overcome.

The State of Play: From “Can We?” to “How Well?”

An Open-Source Ecosystem in Overdrive

DeepSeek R1’s most enduring legacy is that it simultaneously lowered three critical barriers:

Technical: By open-sourcing its reasoning pathways and post-training methods, R1 turned advanced reasoning capabilities — previously locked behind closed-source APIs — into downloadable, distillable engineering assets.

Application: The MIT license made free use, modification, and redistribution straightforward. Companies that depended on proprietary models began integrating R1 directly into production environments.

Psychological: The question shifted from “Can we build competitive AI?” to “How do we build it better?” — a fundamentally different starting point for strategic decision-making.

The result was explosive growth. By August 2025, Chinese-developed models occupied 9 out of the top 10 spots on Hugging Face’s open-source model leaderboard.

International Recognition

• October 2025: DeepSeek’s open-source LLM was named one of the “2025 Top 10 Global Engineering Achievements” by Engineering, the journal of the Chinese Academy of Engineering
• December 2025: Founder Liang Wenfeng was named to Nature’s annual list of Top 10 science newsmakers

Big Tech Responds: Alibaba, Baidu, Tencent Go All In

DeepSeek’s breakthrough served as a starting gun for China’s tech giants:

• Alibaba released the Qwen series, excelling across multiple benchmarks
• Baidu continued iterating on ERNIE Bot, maintaining competitiveness in Chinese-language scenarios
• Tencent’s Hunyuan model accelerated deployment in social and gaming contexts

But beneath the surface-level optimism, a more nuanced reality demands attention.

Behind the Numbers: What the Data Actually Tells Us

Can the Cost Advantage Last?

DeepSeek R1’s $6 million training cost was genuinely shocking. However, two factors are often overlooked:

1. Cumulative expertise: This cost reflected years of accumulated technical know-how, not a cold start
2. The compute ceiling: Chinese companies currently operate primarily with export-restricted H800/H20 chips. While adequate for inference and moderate-scale training, these chips will become increasingly constraining as model scales grow further

The Compute Chasm: An Underestimated Long-Term Challenge

US investment in AI compute is staggering. The Trump-era Stargate Project envisions hundreds of billions of dollars in AI infrastructure. China, by contrast, remains constrained by semiconductor export controls — particularly the most advanced GPUs.

While reports in late 2025 suggested China had developed an EUV lithography prototype entering testing, the path from prototype to mass production remains long and uncertain.

The Monetization Problem: Who’s Actually Making Money?

This is perhaps the most underestimated challenge in Chinese AI. Model capability is the input; commercial viability is the output — but the causal chain doesn’t work automatically.

Current business models include:

• API access fees: Already caught in a race to the bottom on pricing
• Enterprise customization: Labor-intensive and difficult to scale
• Open-source ecosystem monetization: Paths to revenue remain unclear

Compared to OpenAI’s estimated multi-billion dollar annualized revenue, Chinese AI companies have significant ground to cover.

The Real Challenges: It’s Not About Models — It’s About Ecosystems

Challenge 1: The “Inverted Pyramid” of Research Talent

China has an abundance of AI application talent, but top-tier researchers in fundamental areas — novel architectures, training methodology breakthroughs, theoretical frontiers — remain scarce. DeepSeek’s success depended heavily on “tech obsessives” like Liang Wenfeng, but founders of that caliber are rare and not easily replicated.

Challenge 2: Invisible Barriers in Global Ecosystems

Open-source lowered the technology barrier, but not the ecosystem barrier. Foundational platforms and toolchains — Hugging Face, GitHub, NVIDIA’s CUDA — remain Western-dominated. Geopolitical factors are actively pushing Western AI communities to seek alternatives to Chinese models, creating non-technical barriers that may grow over time.

As one analysis from the Beijing Academy of Artificial Intelligence noted: “Geopolitical factors have largely driven open-source model adoption; while Chinese models continue to dominate benchmarks throughout 2025, Western AI communities are actively seeking commercially deployable alternatives.”

Challenge 3: From “AI+” to “+AI” — The Mindset Shift

China’s 2017 “AI+” strategy centered on using AI to empower traditional industries. In practice, many companies remain stuck in “AI for AI’s sake.” True industrial AI requires deep domain understanding combined with clever tool deployment — not the most powerful general model, but the right tool for the specific job.

Looking Ahead: Cautious Optimism

DeepSeek proved that resource-constrained environments can still produce breakthroughs through open-source development and rapid iteration. This is China’s greatest AI asset — not compute power, not data volume, but engineering efficiency and iteration speed.

Key trends to watch over the next 1–2 years:

1. Blurring lines between open and closed source: The DeepSeek model of open-source + commercial support may become the industry norm
2. Edge AI explosion: Compute constraints are paradoxically driving innovation in on-device deployment
3. From general to specialized: Vertical-domain models will become the primary battleground for commercialization
4. The Global South opportunity: Low-cost AI solutions like DeepSeek’s may find widespread adoption in developing nations

Conclusion

One year ago, DeepSeek R1 was more than a model release — it was a declaration of confidence from China’s AI industry. But the real test begins now.

Technical breakthroughs are the starting point. Ecosystem building is the middle. Commercial viability is the finish line. Chinese AI has navigated the first step brilliantly. The road ahead demands more patience, more sustained investment, and less hype.

The AI race, after all, is not a sprint — it’s a marathon with no finish line.

---

This article was originally published by 鲲鹏AI探索局. Data sources include Hugging Face, the Chinese Academy of Sciences, Nikkei Asia, Bloomberg, and other public reports.