China’s open-weights AI strategy is winning
China’s open-weights AI strategy is winning
中国的开源权重 AI 战略正在胜出
AI models, as a product in themselves, have very little moat beyond what amounts to brand loyalty and superficial switching costs. Instead, the moat is in the enterprise services that sit around them: the deals and contracts, connectivity with enterprise systems, and quality of life features in an enterprise context.
作为产品本身，AI 模型除了品牌忠诚度和表面的转换成本外，几乎没有什么护城河。真正的护城河在于围绕它们的企业服务：交易与合同、与企业系统的连接性，以及企业环境下的用户体验功能。
If we consider the models themselves, it’s easy to switch between them: someone could be using ChatGPT today and Claude tomorrow, with very little impact on their workflows. This is particularly true in the engineering world, where models are accessed via API: you can swap out the API and use the same prompt.
如果我们只看模型本身，切换起来非常容易：一个人今天可以用 ChatGPT，明天就可以换成 Claude，对工作流程几乎没有影响。在工程领域尤其如此，因为模型是通过 API 访问的：你可以直接替换 API 并使用相同的提示词（prompt）。
Those companies can make deals to lock their customers in, but in practice there’s very little long-term technical incentive to use one vendor over another. You pick the best model for your needs and change models and vendors if another one becomes better.
这些公司可以通过交易来锁定客户，但在实践中，长期来看，使用某一家供应商而非另一家的技术动机微乎其微。你只需根据需求选择最好的模型，如果另一个模型表现更好，你随时可以更换模型和供应商。
The US government has placed export controls on GPUs. There are also strong regulations that (reasonably) prevent sharing certain kinds of data with Chinese servers. The result is that while Chinese companies have enough compute to train models, they can’t really provide the kinds of global-scale centralized services that we see from OpenAI and Anthropic — at least, not in the same way.
美国政府对 GPU 实施了出口管制。此外，还有严格的法规（合理地）禁止与中国服务器共享某些类型的数据。结果是，尽管中国公司拥有足够的算力来训练模型，但他们无法真正提供像 OpenAI 和 Anthropic 那种全球规模的中心化服务——至少无法以同样的方式提供。
And open almost always wins when it comes to infrastructure adoption. Open technologies can be used permissionlessly and therefore can be at the center of more innovation. You can host them where you want, experiment with them, alter them, and tweak to fit your use case. Open weights models are not open source, but they are portable and permissionless.
在基础设施采用方面，开放几乎总是赢家。开放技术可以无需许可地使用，因此能够成为更多创新的中心。你可以将它们托管在任何地方，进行实验、修改并根据你的用例进行调整。开源权重模型虽然不是严格意义上的“开源”，但它们具有可移植性和无需许可的特性。
With all this in mind, it makes sense for China to release its AI models openly. It turns a US-created compute disadvantage into a distribution advantage; it commoditizes the layer where American companies make money; and it creates a far more effective global ecosystem than could be established through locked-in, centralized services. It’s obvious to me that there are ecosystem benefits throughout China, from manufacturing to scientific research; every sector can just plug in these models.
考虑到这一点，中国选择公开其 AI 模型是合乎逻辑的。这将其在算力上的劣势（由美国造成）转化为分发优势；它将美国公司赖以盈利的层级商品化；并创造了一个比通过锁定、中心化服务所能建立的更有效的全球生态系统。对我来说显而易见的是，中国从制造业到科学研究的各个领域都从中获得了生态红利；每个行业都可以直接接入这些模型。
The saving grace for American companies has been that US frontier models have outperformed open ones. That gap is now closing: “Moonshot and Alibaba unveiled models they claim can go toe-to-toe with the best from OpenAI and Anthropic at a fraction of the cost. The rapid-fire releases suggest America’s lead at the AI frontier is increasingly tight, just as the technology is becoming central to national security, economic power, and geopolitical influence.”
美国公司的救命稻草一直是其前沿模型优于开源模型。但这一差距正在缩小：“月之暗面（Moonshot）和阿里巴巴发布了声称能与 OpenAI 和 Anthropic 最强模型相媲美的模型，且成本仅为后者的一小部分。这种密集发布表明，美国在 AI 前沿领域的领先优势正变得越来越紧迫，而此时该技术正成为国家安全、经济实力和地缘政治影响力的核心。”
Even without these new capabilities, the strategy has already been working. a16z partner Martin Casado noted in the Economist that there’s an 80% chance that any given startup is using Chinese models, and Chinese models are poised to take the lead.
即使没有这些新能力，该战略也已经奏效。a16z 合伙人 Martin Casado 在《经济学人》中指出，任何一家初创公司使用中国模型的概率高达 80%，且中国模型正准备占据主导地位。
It’s worth taking a step back and considering the surprising underlying dynamics. We think of China as being a locked-down society — and it is in many ways. I have serious concerns about how these models might reflect Chinese government perspectives (try asking them about Tiananmen Square). But it’s American companies that are keeping tight control of their technology rather than releasing it as openly as possible. This is in stark contrast to the strategy behind US government support for the open internet, for example.
值得退后一步，思考一下其中令人惊讶的潜在动态。我们认为中国是一个封闭的社会——在许多方面确实如此。我非常担心这些模型如何反映中国政府的观点（试着问问它们关于天安门广场的问题）。但实际上，是美国公司在严格控制其技术，而不是尽可能开放地发布。这与美国政府支持开放互联网背后的战略形成了鲜明对比。
Locked-down business practices for a technology with no real moat but significant potential ecosystem benefits is an obviously losing strategy; permissively releasing it with an open, collaborative approach is obviously a winning one. But the incentives in the US aren’t there: instead, these companies are forced to chase first-order profits rather than ecosystem benefits, and the government tries to put its finger on the scale through forcible measures like tight export controls. We should consider what would need to change to make those incentives more aligned. That’s particularly important given how much of the US economy is currently driven by AI spending. If the bottom falls out of that spending — and I think it clearly will, given the dynamics — the outcome could be severe.
对于一种没有真正护城河但具有巨大潜在生态效益的技术，采取封闭的商业做法显然是失败的策略；而以开放、协作的方式宽容地发布它显然是赢家策略。但美国的激励机制并不支持这一点：相反，这些公司被迫追求短期利润而非生态效益，政府则试图通过严格的出口管制等强制手段来干预市场。我们应该考虑需要做出什么改变，才能使这些激励机制更加一致。考虑到目前美国经济在多大程度上由 AI 支出驱动，这一点尤为重要。如果这些支出崩盘——考虑到目前的动态，我认为这显然会发生——后果可能会很严重。
I care about having open technology that can be run in the public interest, aligned with the public’s values. Threads like public AI, federated services, and open research have traction but need backing. Getting there in the US needs more nuanced strategy and support than we’re seeing today.
我关心的是拥有能够为公共利益服务、符合公众价值观的开放技术。诸如公共 AI、联邦服务和开放研究等方向虽然有进展，但需要支持。在美国实现这一目标，需要比我们今天所看到的更细致的战略和支持。