ADeepSeek and the Open-Source AI Revolution: How China Quietly Rewrote the Rules深度搜索与开源人工智能革命：中国如何悄然改写规则tificial-Intelligence

ChinaIn June 2024, Kuaishou quietly launched a video generation model called Kling AI. Most of Silicon Valley barely registered it. Eighteen months later, the same model is reportedly being spun out as an independent company valued at $20 billion, and the video generation race has fundamentally changed.

The numbers behind Kling’s potential IPO are eye-catching: Kuaishou is said to be seeking a 20 billion dollar valuation for Kling AI in a Pre-IPO round that could raise 2 billion dollars in fresh capital. If that valuation holds, it would make Kling the most valuable standalone video generation company on earth — more valuable than Runway, more valuable than Pika, more valuable than any independent competitor currently operating outside the major tech conglomerates.

But the valuation is almost beside the point. What matters is what the valuation represents: the realization that video generation is not a feature. It’s a platform. And China may have gotten there first.

The Architecture of a Platform Shift

Every major platform shift in technology history — from the PC to the internet to mobile — shared a common characteristic: the infrastructure layer eventually separated from the application layer. Companies that tried to do both lost to companies that specialized in one or the other. The infrastructure winners became utilities; the application winners became consumer giants.

Video generation is approaching the same inflection point. Right now, most video AI capabilities are embedded within larger creative suites or consumer applications. But as models improve and costs drop, a standalone infrastructure layer for video generation is emerging — one that any creator, studio, advertiser, or software developer can build on without needing to train their own model.

Kling AI is positioning itself to be that infrastructure layer. Kuaishou didn’t just build a product. It built a model that can generate high-fidelity, long-duration video from text and image prompts, and it is now building the API partnerships, developer tools, and enterprise integrations necessary to become the default video generation engine for developers globally.

This is a strategy that took OpenAI years to articulate with GPT models. Kuaishou appears to be executing the same playbook with unusual speed.

Why Kuaishou Had Structural Advantages Nobody Talked About

Kuaishou entered the video AI race with advantages that most Western observers failed to appreciate. The company operates one of the world’s largest short-video platforms, which means it has access to something enormously valuable in the video AI context: billions of real-world video clips that have already been created, categorized, and consented to by their creators.

Training a video generation model requires footage. Real footage. Diverse, high-quality, properly licensed footage. Kuaishou has been sitting on a library that most Western AI companies would spend hundreds of millions of dollars and years of effort to replicate. The company’s decision to leverage this asset for model training wasn’t just opportunistic — it was arguably the single most important strategic decision in its AI trajectory.

Beyond training data, Kuaishou had the GPU infrastructure. Years of building out compute capacity for its recommendation algorithms and content moderation systems meant that when the video generation opportunity materialized, Kuaishou didn’t need to scramble to secure GPU access. The clusters were already there. The expertise in operating them at scale was already in-house. All that was missing was the model architecture — and given the competitive labor market for AI researchers, that gap closed faster than anyone expected.

The Sora Comparison Is No Longer Favorable to OpenAI

When OpenAI launched Sora in February 2024, it was widely described as a watershed moment — a demonstration that American AI still led the world in foundational breakthroughs. The comparisons to Kling AI at the time were unflattering to the Chinese competitor.

Eighteen months later, the comparison has flipped. Kuaishou has shipped Kling with capabilities that OpenAI has yet to make commercially available at scale: longer duration generation, better character consistency across scenes, native multilingual support, and an API that is already integrated into production applications used by millions of creators.

OpenAI’s Sora remains impressive as a research demonstration. Kling is a product. The difference matters enormously in a market where developers need reliability, pricing predictability, and support — not flashy demos.

The Valuation Reflects Scarcity and Timing

A $20 billion valuation for Kling AI is not without controversy. Skeptics will argue that video generation is still a niche market, that the number of production deployments remains small, and that the revenue base doesn’t yet justify such a high multiple. These are fair points.

But the valuation is not a reflection of current revenue. It’s a reflection of what the market believes the revenue will be in three to five years — and the scarcity premium associated with getting access before that future arrives.

There are very few independent video generation companies with the scale, infrastructure, and commercial traction that Kling has achieved. In a market where investors are desperate for exposure to AI infrastructure plays that aren’t Microsoft, Google, or Meta, Kling represents one of the few genuine alternatives. Scarcity drives valuation when the story is compelling. The Kling story is compelling.

What This Means for the Broader AI Investment Landscape

The Kling IPO — if it proceeds — will be a significant data point for the global AI investment market. A successful debut would validate the thesis that Chinese AI companies can build independent, internationally competitive businesses outside the shadow of their domestic tech giants. It would also signal that the next wave of AI infrastructure investment may be more geographically diverse than the current market assumes.

For venture capitalists, the implications are nuanced. If a Kuaishou spinoff can command a $20 billion valuation, what does that mean for the dozens of smaller video AI companies, the robotics AI firms, and the enterprise AI application builders that are currently raising seed and Series A rounds? The answer is probably upward pressure on valuations across the board — but only for companies that can demonstrate genuine technical differentiation and real commercial traction. The days when an AI company could raise on a slide deck and a founder’s pedigree are probably over. Kling has set a new bar: show us the model, show us the API, show us the revenue.

The Real Story Is the Speed

Step back from the valuation and the competitive comparisons and the investment thesis, and the most remarkable thing about Kling AI is the speed. Kuaishou went from launch to potential $20 billion spinoff in under two years. The model progressed from initial release to global API availability in a timeframe that would be aggressive for a software update at a large tech company, let alone a foundational AI product.

That speed is not accidental. It reflects a combination of technical readiness, data advantage, compute availability, and organizational will that is very difficult to replicate quickly. OpenAI, Google, and Meta are not standing still — but neither is the rest of the world assuming they will always lead.

The Kling story is ultimately a story about how quickly the assumptions that governed AI for the last several years are being dismantled. The geography of AI leadership is not fixed. The competitive moats that seemed permanent are proving surprisingly permeable. And the next chapter of the AI story may be written as much in Beijing as in San Francisco.

Open-Source

Technology

Innovation

rDeepSeek and the Open-Source AI Revolution: How China Quietly Rewrote the Rules

In January 2025, a small Chinese AI lab released a model that made Silicon Valley pause. DeepSeek-R1 wasn’t just another entrant in the AI race — it was a statement. An open-source mod Medium Integration Tokenel matching the performance of GPT-4, released at a fraction of the cost, with weights freely available to anyone. The response from Western tech circles ranged from disbelief to alarm.2025 年 1 月，一家小型的中国人工智能实验室发布了一款模型，让硅谷为之震惊。DeepSeek-R1 不只是人工智能竞赛中的又一参与者——它更是一种宣言。这是一款开源模型，性能可与 GPT-4 相媲美，发布成本却只是其一小部分，而且权重对所有人免费开放。西方科技圈对此的反应从怀疑到警觉不等。In January 2025, a small Chinese AI lab released a model that made Silicon Valley pause. DeepSeek-R1 wasn’ t just another entrant in the AI race — It was a statement. An open-source mod, Medium Integration Tokenel, matching the performance of GPT-4, was released at a fraction of the cost, with weights freely available to anyone. The response from Western tech circles ranged from disbelief to alarm. In January 2025, a small Chinese AI laboratory released a model that shocked Silicon Valley. DeepSeek-R1 was not just another participant in the AI race - — It is also a declaration. This is an open-source model with performance comparable to GPT-4, but with a release cost that is only a fraction of it, and the weights are freely accessible to all. The reaction from the Western tech circle ranges from skepticism to alertness.

But the story of DeepSeek is bigger than one company. It represents a fundamental shift in how AI development works — and who gets to participate in it. China, long seen as a fast-follower in AI, is increasingly setting the agenda. And the world is still catching up to what that means.但 DeepSeek 的故事远不止一家公司。它代表着人工智能开发方式的根本转变——以及谁能够参与其中。长期以来被视为人工智能领域“快跟随者”的中国，正越来越多地引领着议程。而世界仍在努力适应这意味着什么。

The $6 Million Wake-Up Call六百万美元的警钟

When DeepSeek released R1, the figures attached to its development were almost impossibly low. Estimated training cost: roughly $6 million. By contrast, GPT-4 reportedly cost upward of $100 million to train. The implication was staggering — frontier AI development might not require the massive capital expenditures that Western companies had normalized.当 DeepSeek 发布 R1 时，其开发成本数据低得令人难以置信。据估计，训练成本约为 600 万美元。相比之下，据报道 GPT-4 的训练成本超过 1 亿美元。这意味着前沿人工智能开发或许并不需要西方公司所习惯的巨额资本投入。

The numbers sparked immediate debate. Some dismissed them as accounting tricks. Others pointed out that DeepSeek’s claims didn’t account for hardware costs, research salaries, or the enormous preceding investments in data and infrastructure. Regardless of the precise figure, the directional signal was clear: the cost curve for AI capability was bending sharply downward.这些数字立即引发了激烈的争论。有人认为这只是会计上的花招。也有人指出，DeepSeek 的说法没有将硬件成本、研发人员薪酬以及此前在数据和基础设施方面的巨额投入计算在内。不管确切数字如何，一个方向性的信号是明确的：人工智能能力的成本曲线正在急剧下降。

What made this particularly significant was the open-source nature of the release. DeepSeek didn’t just publish a paper or launch a closed API. They released the model weights, the technical details, and the training methodology. Anyone with sufficient hardware could download, run, fine-tune, and build on top of R1. This isn’t just disruption — it’s a structural change in how AI knowledge spreads.这一举措意义非凡的原因在于其开源的性质。DeepSeek 不仅发表了一篇论文或推出一个封闭的 API，还发布了模型权重、技术细节以及训练方法。任何拥有足够硬件的人都可以下载、运行、微调并在此基础上进行开发。这不仅仅是颠覆，更是人工智能知识传播方式的结构性变革。

China’s Long Game in Open-Source AI中国在开源人工智能领域的长远布局

DeepSeek isn’t an anomaly. It’s the culmination of years of deliberate strategy. Chinese AI research has been investing heavily in open-source contributions for half a decade. Models like Qwen (Alibaba), GLM, and Yi have quietly accumulated stars on GitHub and usage across global developer communities.DeepSeek 并非特例。它是多年精心布局的成果。过去五年，中国的人工智能研究一直在大力投入开源贡献。像 Qwen（阿里巴巴）、GLM 和 Yi 这样的模型在 GitHub 上悄然积累了众多星标，并在全球开发者社区中得到广泛应用。

The motivation is partly economic and partly strategic. An open-source ecosystem doesn’t require expensive API subscriptions or licensing deals. It builds the infrastructure for domestic talent to train, experiment, and innovate. And critically, it creates facts on the ground — standards, use cases, developer familiarity — that don’t disappear when a company’s commercial strategy shifts.其动机部分出于经济考量，部分出于战略考虑。开源生态系统无需昂贵的 API 订阅或授权协议。它为国内人才提供了培训、实验和创新的基础设施。而且至关重要的是，它创造了实实在在的事实——标准、用例、开发者熟悉度——这些不会因公司商业策略的转变而消失。

There’s also a geopolitical dimension. Open-source models can’t be export-controlled the way closed APIs can. A model with weights available on Hugging Face is, by definition, available everywhere. Chinese AI companies discovered that open-source was a way to achieve global influence without needing American chip imports or Silicon Valley validation.这当中也存在地缘政治层面的因素。开源模型无法像封闭的 API 那样受到出口管制。在 Hugging Face 上可以获取权重的模型，从定义上讲，是无处不在的。中国的 AI 公司发现，开源是一种无需依赖美国芯片进口或硅谷认可就能实现全球影响力的方式。

The Global Developer Ecosystem Response全球开发者生态系统响应

The response from developers worldwide was immediate and enthusiastic. Within weeks of the R1 release, the model had been downloaded millions of times, fine-tuned for dozens of languages and specialized domains, and integrated into tools ranging from coding assistants to scientific research pipelines.全球开发者们的反应迅速而热烈。在 R1 版本发布后的几周内，该模型已被下载数百万次，针对数十种语言和专业领域进行了微调，并被整合到从编码助手到科研流水线等各种工具中。

This is the power of open — it attracts contributions that closed systems can’t. A global community of researchers, hobbyists, and companies now had skin in the game. Bugs got fixed. Performance improved. Specialized variants proliferated. The model became a platform rather than a product.这就是开放的力量——它能吸引封闭系统无法吸引的贡献。全球的研究人员、业余爱好者和公司都参与其中。漏洞得以修复。性能得以提升。各种专门的变体层出不穷。该模型从一个产品变成了一个平台。

For developers in the Global South especially, this mattered enormously. Running a frontier model through a paid API is expensive. Running it locally on commodity hardware — now possible with optimized versions of R1 — democratizes access in ways that closed models never could. A startup in Lagos or Bangalore could now build on the same foundation as a team in San Francisco.对于全球南方的开发者来说，这一点尤为重要。通过付费 API 运行前沿模型成本高昂。而如今，借助优化版的 R1，在普通硬件上本地运行模型成为可能，这在很大程度上打破了访问壁垒，这是封闭模型永远无法做到的。拉各斯或班加罗尔的初创企业现在可以与旧金山的团队站在同一起跑线上进行开发。

What Silicon Valley Gets Wrong About China AI硅谷对中国人工智能的误解

Western analysis of Chinese AI tends to oscillate between two errors. The first is dismissal — Chinese AI companies are forever catching up, their progress derivative, their best models years behind the frontier. The second is alarm — China is about to overtake the US through sheer scale, data advantage, or government mandate.西方对中国人工智能的分析往往在两个错误之间摇摆不定。第一个错误是轻视——中国的 AI 公司永远在追赶，其进展是衍生的，其最好的模型也落后前沿数年。第二个错误是惊慌——中国即将凭借规模、数据优势或政府指令超越美国。

Both miss the actual mechanism of Chinese AI progress. It isn’t just about building bigger models or collecting more data. It’s about building ecosystems. The Chinese AI approach, at its best, combines rapid iteration, aggressive open-source publication, massive talent pools, and a manufacturing mentality that turns AI research into deployable infrastructure at unprecedented speed.这两种观点都未能准确把握中国人工智能发展的实际机制。这并非仅仅在于构建更大的模型或收集更多的数据，而是关乎构建生态系统。中国人工智能发展的最佳实践在于将快速迭代、积极的开源发布、庞大的人才库以及将人工智能研究转化为可部署基础设施的制造业思维相结合，从而以前所未有的速度推进发展。

This doesn’t mean Chinese AI is superior across the board. Closed frontier models from American companies still lead on certain benchmarks, and the research culture of openness has trade-offs. But the gap is narrower than it was, and it’s closing faster than most Western observers expected.这并不意味着中国的人工智能在所有方面都更胜一筹。美国公司的封闭式前沿模型在某些基准测试中仍处于领先地位，而且开放的研究文化也有其利弊。但差距已经比过去小了，而且缩小的速度比大多数西方观察家预期的要快。

The Implications for the AI Race人工智能竞赛的影响

If the cost of frontier AI capability continues to fall, the implications for the ‘AI race’ narrative are profound. Races assume that winning requires crossing a finish line first. But if the finish line keeps moving backward — if the frontier expands faster than anyone can run — then the entire framing shifts.如果前沿人工智能能力的成本持续下降，那么对于“人工智能竞赛”的说法影响将是深远的。竞赛意味着获胜需要率先冲过终点线。但如果终点线不断后移——如果前沿的发展速度超过了任何人的追赶速度——那么整个框架就会发生转变。

What matters increasingly isn’t who trains the biggest closed model, but who builds the most useful applications on top of open infrastructure. Who creates the tooling, the fine-tuned variants, the specialized solutions? In that race, the advantages of incumbency are smaller, and the advantages of ecosystem size are larger.如今愈发重要的是，谁能在开放基础设施之上构建出最有用的应用程序，而非谁训练出了最大的封闭模型。谁在打造工具、微调变体、提供专业解决方案？在这场竞赛中，先发优势较小，而生态系统规模的优势则更大。

China, with its enormous developer community, its manufacturing and enterprise ecosystem, and its growing open-source culture, is unusually well-positioned for this version of the competition. The DeepSeek moment wasn’t just a milestone — it was a preview.中国拥有庞大的开发者群体、制造业和企业生态系统，以及日益浓厚的开源文化，在这场竞争中处于得天独厚的位置。DeepSeek 时刻不仅是一个里程碑，更是一次预演。

The Road Ahead 前面的路

None of this means the US AI ecosystem is doomed. The US still leads in fundamental research, in chip design, in the venture and talent infrastructure that turns ideas into companies. The question isn’t which country wins, but how the dynamics of AI development shift when open-source becomes a first-tier player alongside closed development.这并不意味着美国的人工智能生态系统注定要失败。美国在基础研究、芯片设计以及将创意转化为企业的风险投资和人才基础设施方面仍处于领先地位。问题不在于哪个国家会胜出，而在于当开源成为与封闭式开发并驾齐驱的一流参与者时，人工智能发展的动态会如何变化。

The more likely outcome is bifurcation: a world where closed frontier models handle the most sensitive, highest-stakes applications, while open models handle everything else — which turns out to be most of the economy. In that world, China’s open-source push is a bet on being indispensable in the ‘everything else’ category.更有可能出现的情况是分化：一个世界里，封闭的前沿模型处理最敏感、风险最高的应用，而开放模型处理其他所有应用——结果发现这涵盖了大部分经济领域。在这样的世界里，中国推动开源是押注于在“其他所有”这一类别中成为不可或缺的存在。

Whether that’s a winning strategy depends on what you think AI is ultimately for. If you believe AI’s greatest impact will come through commoditized infrastructure, global accessibility, and application-layer innovation, China’s open-source orientation looks prescient.

What’s certain is that the question is no longer whether China matters in AI. It does. The question is what kind of influence it will have — and whether the rest of the world is paying attention.

The DeepSeek moment was a wake-up call. The alarm has been sounded. Now comes the harder part: figuring out what to do next.

AI

China

Open-Source

Technology

Innovation