But not in the way most people think.
Sun Tzu, who wrote “The Art of War” roughly 2500 years ago, distilled many military strategies that are studied by both military and business leaders everywhere. One of the more well-known psychological warfare tactics he espoused is called 围三阙一 (pronounced: wei san que yi).
In essence, it means when you are besieging a city or a castle, the best way to defeat your enemy while incurring the least amount of cost is to siege three sides but leave one opening. If all four sides are sealed off, your enemy would quickly grow more desperate and determined to fight back. With an opening available, your enemy would waver between fight and escape, making them either easier to defeat by ambush when they try to escape or more susceptible to surrender when there’s hope to stay alive.
This tactic is what AI entrepreneurs in China fear the most when it comes to how US technology sanctions would hurt their dreams and ambitions. The founder and CEO of Zhipu AI, Zhang Peng, articulated this view in an interview last month. Others have shared the same view in private. (Zhipu AI is a Tsinghua University spinout foundation AI model unicorn backed by Alibaba, Tencent, HongShan, Hillhouse, Xiaomi, Meituan, and many others. It is often considered the Chinese startup closest to OpenAI, given its deep academic roots and commitment to developing AGI.)
One of the unintended, second-order consequences of the expanded US AI chips sanctions instituted in early October may be that Zhang Peng and his fellow entrepreneurs in China will have less to fear about this worst case scenario. The sanctions may have become so stringent that it is akin to a complete, four-sided siege with no opening left. Evidence of this is showing up in just how quickly Nvidia’s China business is shrinking.
Nvidia’s China Sacrifice
Nvidia reported its Q3 earnings right before the US Thanksgiving holidays. The main storyline was that its China business, which accounts for 20-25% of its total revenue, may soon disappear. (Fabricated Knowledge has a good summary of the earnings main takeaways.)
In its effort to comply with the new round of US sanctions while still pursuing one of its largest markets, Nvidia is once again developing alternative products just for the China market. However, the key one that’s most relevant to generative AI, the H20 – a lower-performing, more compliant alternative to its H200 series that may perform better than its H100 chips according to SemiAnalysis – is delayed until Q1 of 2024.
It’s hard to predict how well the H20 will sell when it becomes available. What does seem clear is that as these sanctions become more strict, the appetite for Chinese firms to throw money at US-designed chips, even if they are compliant, is shrinking, because the specter of more strict sanctions is now always on the horizon.
That appetite did exist not that long ago. There was a striving “black market” for all kinds of banned or soon-to-be-banned Nvidia chips earlier this year, according to ChinaTalk. Stockpiling A800s and H800s, the once-upon-a-time China-compliant alternatives to the A100 and H100 series, has been an important survival strategy for both big firms, like Tencent and Baidu, as well as startups like 01.ai. In fact, Tencent proactively shared its stockpiling of Nvidia chips on its most recent earnings call, which it claimed would be enough to keep developing its own foundation AI model, Hunyuan, for the next two years. Baidu more or less offered the same assurance during its own earnings call a few days later.
The strategic “opening” that was leading Chinese firms to still drink out of the Nvidia well, which would’ve been a correct application of Sun Tzu’s tactic, is closing. Rather than a total siege to choke off China from all US-made options, intentionally and strategically leaving an opening would’ve served America’s geopolitical strategy better. It would’ve been the worst nightmare in the long-term for every AI builder in China. Maybe the H20 could still be that opening. But it’s looking unlikely. (Update: one day after this post was published, Commerce Secretary Gina Raimondo spoke at the Reagan National Defense Forum and explicitly called out Nvidia’s new product modification as “not productive.” Looks like any semblance of an “opening” has closed.)
After this current phase of stockpiling is done, Nvidia’s China business will become a sacrificial lamb. The beneficiary of all this is, of course, Huawei.
Consolidation Around Huawei
The desperation and determination bred from this full-siege technology sanction is embodied in Huawei. This is not to say that Huawei’s own AI chips, the Ascend series, is close to where Nvidia’s products are right now, but simply that Huawei has become the only game in town. There is no more wavering between Huawei and Nvidia.
This direction is starting to show up per Baidu’s recently reported order of Huawei Ascend chips. Baidu buying Huawei chips is remarkable, because Baidu has had its own semiconductor R&D division for a long time, which would’ve been a competitor to Huawei’s chips if US sanctions weren’t an issue. Abandoning its own home-grown division in favor of Huawei’s product is an implicit recognition that domestic competition on critical, choke point technologies is a waste of national resources. Every firm must now consolidate behind the designated national champion for the good of the country.
Huawei is not only receiving implicit support from the purchase orders of would-be domestic competitors. Per Bloomberg’s long piece on the company, it is receiving plenty of explicit subsidies from multiple layers of government, from the usual land and tax breaks, to investments in many smaller suppliers in a vast ecosystem all geared to make Huawei, and China, more self-sufficient.
Huawei is also doing some smart things to expand its reach and surface area in competition with Nvidia. It recently became a Premier Member of the PyTorch Foundation in order to steward the PyTorch stack, an open source development stack initially out of Meta that competes with Nvidia’s CUDA, to work better with its Ascend chips.
Huawei’s advancement, showcased in its Mate 60 Pro smartphone, was the big surprise development that catalyzed much of the October expanded sanctions. Was this really a surprise though? Not according to Jensen Huang, Nvidia’s co-founder and CEO, who said during his interview at the New York Times DealBook Summit that “I don’t think anyone in the industry was surprised…”:
When asked about the new round of US sanctions that are throttling Nvidia’s China business, Huang was on message, compliant, and paid due respect to the need to balance business growth with national security concerns. But his tone and demeanor was no doubt resigned, ending his answer with “what can you do?” followed by a shrug.
His resignation was understandable. Nvidia is not getting a dime of subsidy from the US government, while it kills one of its largest markets and billions in revenue. And these changes, while painful to China in the short-term, may end up undermining the US geopolitical and strategic positioning in the long-term.
If Nvidia has a somewhat restricted but still viable business in China, the US can predictably know how far behind China’s AI computing power is by simply tracking what products are getting sold there – the intentional “opening”. If Nvidia no longer has a viable business there, with the whole market consolidating around the black box of Huawei, there will only be more, unpleasant surprises in the future.
To truly advance US long-term interests vis-a-vis China, it may serve folks in DC well to brush up on some Sun Tzu.
英伟达的中国市场对美国地缘政治定位至关重要
《孙子兵法》中有一条很有名的心理战术叫：围三阙一。
其本质含义就是，当围攻一个城市或城堡时，以最小的代价击败敌人的最佳方式是围攻三面但留一个出口。如果四面都封死，敌人会迅速变得绝望，破釜沉舟，下决心死战一场。如果有个出口，敌人会在反抗和逃跑之间摇摆，这使得他们要么在尝试逃跑时更容易被伏击捉捕，要么给予生存的欲望更易于投降。
这一策略也正是在中国的AI创业者们最担心的，尤其在评论美国科技制裁如何摧毁他们的梦想和雄心时。智谱AI的创始人兼CEO张鹏在上个月的一次采访中就表达了这种观点。许多业界人士在私下也有同感。（智谱AI是自清华脱出来的一家做基础AI模型的独角兽，得到了阿里、腾讯、红杉、高瓴、小米、美团等许多公司和顶级基金的支持。鉴于其深厚的学术根基和对研发AGI的雄心，它常被认为是最接近OpenAI的中国初创公司。）
10月初美国扩大AI芯片制裁的一个意外的二阶后果可能是，张鹏和其他中国的企业家同行将不必担心被 “围三阙一” 了。制裁可能已严格到了四面围攻的程度，而没有留下任何出口。这一现象正在英伟达的中国业务迅速下滑的速度中显现出来。
英伟达牺牲中国
英伟达在美国感恩节假期前公布了第三季度财报。最重点的takeaway就是其中国业务，目前占其总收入的20-25%，可能很快就会消失。（《Fabricated Knowledge》对财报的要点做了很好的总结。）
为了遵守美国新一轮制裁，但同时还保住其最大市场之一，英伟达再次特意为中国市场开发了替代品。然而，最关键的替代品之一，H20（一个性能较低、更符合要求的H200系列替代产品，据SemiAnalysis 称可能比H100芯片效果更好）的推出被推迟到2024年第一季度。
很难预测H20出炉后的销量如何。更明显的是，随着这些制裁变得更加严格，中国科技公司对购买美国设计的芯片的胃口也在缩减，即使它们目前是合规的，因为未来更严格的新一轮制裁的阴云总会存在。
买美国芯片的胃口，即使是替代品，上半年还一直有。据《ChinaTalk》报道，各种被禁或即将被禁的英伟达芯片在 “黑市”上一直畅销。为A800和H800（曾经是合规可以出口给中国的A100和H100系列替代品）囤积库存，对于腾讯、百度等大公司以及像零一万物这样的创业公司来说，则是重要的生存发展策略。腾讯在其最近的财报中主动分享了对英伟达芯片的囤积，声称囤的货足以支持其在未来两年内继续开发自研的“混元”基础AI模型。百度在几天后的财报中也或多或少提供了同样的信息来保证资本市场对它的看好。
一直把中国公司引向英伟达这个水井中继续汲水的战略“出口”，本是 “围三阙一” 策略的正确运用，但现在出口正在关闭。与其完全围困以切断中国与所有美国芯片产品，有意识和战略性地留下一个出口本来对美国的地缘政治战略和定位更有力。这也是在中国每位从事AI事业人士的噩梦。也许H20还是个出口，但可能性越来越小。
在当前短期的囤积阶段结束后，英伟达的中国业务将成为地缘政治的牺牲品。最终受益者当然就是华为。
围绕华为的整合
全面围剿式的技术制裁所孕育的绝望与决心体现在华为。我并不是说华为自研的AI芯片，已经能与英伟达的产品媲美，而仅是华为已经成为中国各大公司和企业的唯一选择，不会再有在华为和英伟达之间犹豫不决的问题了。
这一趋势已开始显现，如最近有报道称百度开始给华为昇腾系列芯片下订单。百度购买华为芯片很值得关注，因为百度长期拥有自己的半导体研发部门，如果不是美国的制裁，它本可以成为华为芯片的竞争对手。放弃自研项目转而支持华为的产品，这背后的暗示就是：在关键的、“卡脖子”的技术上的国内竞争是对国家资源的浪费。每个公司现在都必须围绕指定的 “国家队” 整合，以利于国家的宏观利益。
华为不仅从国内竞争对手的订单中得到间接性的支持。据彭博社最近的一篇长篇报道，华为已从多个政府层面获得了大量的直接支持和补贴，从常见的土地和税收优惠，到对培养生态和产业链中众多小供应商的投资，都是为了华为（和整个中国）能早一天自立。
华为还在与英伟达的竞争中做出了一些聪明的举措来扩大其影响力和竞争力。它最近成为了PyTorch基金会的高级会员。PyTorch则是最初来自Meta的开源开发堆栈，与英伟达的CUDA对标。华为正在以影响PyTorch技术栈的方式，来使昇腾芯片的开发者生态扩大，与英伟达的CUDA竞争。
华为的发展，在其Mate 60 Pro智能手机中淋漓尽致，也刺激了美国政府因而在10月份扩大制裁的范围。但一发展真是个惊讶吗？据英伟达联合创始人兼CEO黄仁勋在《纽约时报DealBook峰会》上接受采访时的说法，“我不认为业内任何人感到了惊讶……”：
当被问及新一轮美国制裁对英伟达中国业务的限制时，黄仁勋表现出一致的态度，遵守规定，并对平衡业务增长与国家安全考虑表达了应有的尊重。但他的语气和举止无疑是一种无奈，以“又能怎么办？”和一声叹息结束了他的回答。
他的无奈是可以理解的。英伟达没有从美国政府那里拿到一分钱的补贴，同时却损失了其最大的市场之一和数十亿的收入。而这些政策的演变，尽管短期内会让中国很痛苦，但最终可能会对美国长远的地缘政治和战略定位有更大的损失。
如果英伟达在中国仍可以保持一些业务，又能合规，美国政府可以通过跟踪那里销售的产品来更准确地了解中国AI算力落后多远 — 故意留下“开口”的好处。如果英伟达在中国的业务基本消失，整个市场围绕华为这个黑匣子整合，未来只会有更多的“惊喜”。
如果想真正的推进美国的长期利益，华盛顿的官员们可能需要重温一些《孙子兵法》了。