个人资料
正文

芯片战争

(2023-09-24 16:33:09) 下一个

"这是一种战争行为":解码美国对华芯片封锁行动

ALEX W. PALMER  

英伟达H100张量核心GPU被用于大型人工智能、高性能计算和数据分析工作。
 
去年10月,美国工业与安全局(Bureau of Industry and Security,简称BIS)发布了一相当于向中国发起经济战的宣战文书的文件,这份139页的文件里充斥着繁琐官僚术语和详尽的技术细节。这样重量级的行动因其来源相对不为人知而变得更加引人注目。BIS规模很小,是美国商务部的13个局之一,也是资金规模最小的联邦部门:2022年的预算略高于1.4亿美元,约为一个爱国者防空导弹连成本的八分之一。该局雇用了大约350名特工和官员,他们共同监控世界各地在进行的价值数万亿美元的交易。
在冷战最激烈的时期,对苏联阵营的出口管制最为严格,BIS作为西方国防的重要枢纽,每年处理多达10万个出口许可证。在1990年代相对和平稳定的时期,BIS失去了一些存在的理由,也失去了一些人员和资金,许可证数量减少到每年大约1万个。如今,这一数字已达到4万,并且还在不断增加。BIS比以往任何时候都更加忙碌,它有一份庞大的贸易黑名单——即实体名单(目前有662页,并且还在不断增加)、众多先前存在的多边出口管制协议以及针对俄罗斯和中国的持续行动。“我们把100%的时间花在对俄罗斯的制裁上,另外100%的时间花在中国上,还有100%的时间花在其他事情上,”负责出口管理的商务部副助理部长马特·博尔曼表示。
最近几年,半导体芯片已成为BIS工作的核心。芯片是现代经济的命脉,也是从iPhone到吐司机、从数据中心到信用卡等所有电子设备和系统的大脑。一辆新车可能有一千多个芯片,每个芯片管理车辆操作的不同方面。半导体也是量子计算和人工智能等有望在下个世纪彻底改变生活的创新技术背后的驱动力。例如,据报道,OpenAI对ChatGPT的训练是在1万个目前最先进的芯片上进行的
美国政府在10月7日通过出口管制宣布其意图削弱中国生产甚至购买最高端芯片的能力。该措施的逻辑很简单:先进芯片及其驱动的超级计算机和人工智能系统可被用于生产新的武器和监视设备。然而其影响范围极广,意义极大,目标是远不仅限于中国的诸多安全国家。“这里的关键是要明白,美国想要影响中国的人工智能产业,”华盛顿战略与国际研究中心瓦德瓦尼人工智能和先进技术中心主任格雷戈里·C·艾伦说。“半导体这方面是实现这一目标的手段。”
尽管10月7日的管制措施是以更新出口规则的低调形式实施,但本质上是为了根除中国整个先进技术生态系统。“10月7日体现的新政策是:我们不仅不会允许中国在技术上取得任何进展,我们还将积极扭转他们目前的技术水平,”艾伦说。EvercoreISI的高级半导体分析师C·J·缪斯这样说道:“如果你五年前告诉我这些规则,我会告诉你这是一种战争行为——我们肯定是在战争状态”。
如果这些控制措施成功,可能会影响中国一代人的进步;如果失败,可能会产生惊人的适得其反的结果,美国在极力避免的那个未来会更快发生。这一结果可能会影响未来几十年的中美竞争以及全球秩序的未来。“从2022年开始,有两个日期将被历史铭记,”艾伦说。“第一个是2月24日,俄罗斯入侵乌克兰;第二个是10月7日。”
尽管半导体的设计非常复杂,但从某种意义上说也非常简单:在微小硅片上雕刻大量的电路。被称为晶体管的开关控制着电路的通断。当电路接通时是1;当电路断开时是0。第一批芯片于1950年代末发明,只包含少量晶体管。如今,新型智能手机中的主要半导体拥有100至200亿个晶体管,每个晶体管只有病毒大小,像层叠蛋糕一样雕刻在硅结构中。
著名的摩尔定律描述了过去60年的进步速度,该定律指出,芯片上可安装的晶体管数量大约每两年增加一倍。《芯片战争》(Chip War)一书的作者、塔夫茨大学弗莱彻学院国际历史副教授克里斯·米勒常常指出,如果飞机的改进速度与芯片相同,那么它们现在的飞行速度会是光速的好几倍。人类文明史上没有任何技术能够与计算能力的惊人提升相媲美。
半导体制造工厂——晶圆厂——是世界上最昂贵的工厂,进行着有史以来最复杂的制造,其生产规模是任何其他设备都无法达到的。与此同时,更广泛的芯片行业是一个相互依存的网络,高度专业化的地区和公司遍布全球,极长和复杂的供应链实现了这种壮举的可能性——换句话说,这是全球化的产物。米勒说:“很难想象,如果没有世界上最聪明的人共同努力,他们如何能够达到这样的能力。”然而,正是这种相互关联性使得该行业非常容易受到拜登政府正在推行的这类法规的影响。
只有少数公司能够在前沿竞争,而突破需要花费数十亿美元和数十年的研究。这就造成了一个由一系列瓶颈构成的行业。最著名的例子是荷兰制造集团阿斯麦(ASML)制造的极紫外光(EUV)刻机,用于打印芯片的各层。1997年,阿斯麦聘请了拥有物理学博士学位的年轻工程师乔斯·本肖普,率先创制出一个新系统,将帮助阿斯麦半导体行业的客户打印比以往更小、更快和更密集的芯片。单单是为了论证分配一个小团队从事该项目的必要性,就进行了四年的概念验证,然后该团队又花了五年时间来建造一台原型机。2010年12月,在一个韩国研究设施内,原型机的更新版TWINSCANNXE:3100终于首次成功试运行。距离第一批支持极紫外的产品上市还需要近十年的时间。
最新版本的机器可以制作小至10纳米的结构;相比之下,人类红细胞的直径约为7000纳米。它使用激光产生比太阳表面温度高40倍的等离子体,发出人眼看不见的极紫外光,通过一系列反射镜折射到硅芯片上。该激光器来自德国一家公司,共有457329个部件;一个完整的EUV光刻机有超过10万个复杂程度类似的组件。
EUV只是整个工艺的一部分:一个尖端的晶圆厂可能包括500多台机器和1000个步骤。然而,单是EUV就几乎是一项奇迹般的人类成就,它能以难以想象的规模和精度工作。“我真的相信,我们的机器是人类制造过的最复杂的东西,”现任阿斯麦公司技术副总裁的本肖普说。今天,在TWINSCAN首次试运行十多年后,没有其他公司能够重现阿斯麦的成就。
通过挤压该行业的天然瓶颈,拜登政府的目标是阻止中国进入芯片技术的未来。其影响将远远超出削弱中国的军事进步,还将威胁到中国的经济增长和科学领导地位。“我们说过,中国不应该在一些关键技术领域取得进展,”新美国安全中心高级研究员、前美国贸易官员艾米丽·基尔克雷斯表示。“而这些领域恰好是未来经济增长和发展的动力所在。”今天,科学进步通常是通过模拟和分析大量数据来实现,而不是通过反复试验。模拟被用于发现新的救命药物,模拟气候变化的未来,探索星系碰撞的行为,以及高超音速导弹和核爆炸的物理学。
田纳西大学创新计算实验室的创始主任杰克·唐加拉告诉我:“拥有最好超级计算机的人可以从事最好的科学研究。”唐加拉运行着一个名为TOP500的项目,每两年对世界上最快的超级计算机进行排名。截至今年6月,中国占据134个席位,而美国占据150个。但这不反映全貌:在2020年左右,中国提交的计算机数量大幅下滑,这在唐加拉看来是希望避免引起不必要的关注。关于新型超级计算机的传言在科学论文和研究公告中泄露出来,引发观察人士猜测竞争的真实状态——以及假定的中国领先优势的规模。“这很惊人,因为2001年中国的计算机还没有上榜,”唐加拉说。“现在他们已经成长到占据主导地位的地步。”
然而,在中国强大的背后是一个关键的弱点:为中国最先进的项目和机构提供动力的几乎所有芯片都与美国技术密不可分。“整个行业只有在有美国参与的情况下才能运转,”米勒说。“在每个接近尖端技术很近的设施中,整个过程都有美国的工具、美国的设计软件和美国的知识产权。”尽管中国政府几十年来一直在努力,并在“自主创新”方面投入了数以百亿计的美元,但这个问题仍然很严重。2020年,中国国内芯片生产商只供应了全国总需求的15.9%。就在今年4月,中国进口半导体的花费还超过了进口石油。
美国对全球半导体市场进行完全掌控是在2019年,当时特朗普政府将中国主要电信制造商华为列入实体名单。虽然表面上看,将华为列入实体名单是对其违规行为的惩罚——华为曾被发现向伊朗出售受制裁的材料——但战略利益立即变得显而易见。由于无法获得美国的半导体、软件和其他必需品,华为这家全球最大的电信设备生产商只能艰难生存。“对华为的制裁立即拉开了帷幕,”卡内基国际和平基金会研究中国科技生态系统的研究员马特·希恩说。“中国科技巨头使用的芯片都是美国制造的,或者含有大量美国部件。”
长期以来,出口管制法律一直被视为尘封已久的神秘领域,不太可能被施用于美国国力投射的实践中。但在华为之后,美国发现自己在半导体供应链中的主导地位是一个尚未利用起来的强大杠杆。三家位于美国的公司主导着芯片设计软件市场,该软件用于为新芯片配置数十亿个晶体管。先进芯片制造工具的市场也同样集中,只有少数几家公司能够有效地垄断重要的机器或工艺,而这些公司几乎都是美国公司或依赖美国零部件的公司。供应链的每一步都贯穿美国、美国的条约盟国或台湾,所有这些国家都在美国主导的生态系统中运作。“我们是偶然发现的,”希恩说。“在我们真正知道如何使用这些武器之前,我们就开始使用它们了。”
2020年5月,特朗普政府进一步收紧政策,这次是让华为受制于“外国直接产品规则”,这是一项出口管制法律中曾被认为晦涩难懂的条款。在该条款下,使用美国技术或软件生产的外国产品都要受到美国的管制。这是对治外法权的全面实践:即便一件商品在美国境外制造和运输,从未进入过美国国界,其最终成品并不包含美国原产零部件或技术,但它依然可以被视为美国产品。
对华为而言,该条款的实施意味着公司的半导体来源基本被切断。“这一规则让全球所有半导体都受制于美国法律,因为全球所有芯片代工厂都至少在一定程度上使用了美国的设备,”曾在BIS负责出口管理事务的前助理商务部长凯文·沃尔夫表示。“哪怕你的代工厂里只有一种美国设备,而其他非美国设备有一百种,但整条生产线上的晶圆就跟美国沾了边。”
根据市场分析公司Canalys的数据,华为在2020年是全球最大的智能手机销售商,占据市场份额达18%,甚至超过苹果和三星。2021年,华为的营收下降近三分之一,靠出售旗下一个智能手机品牌才得以维持。到2022年,华为全球市场份额已降至2%。
10月7日法规代表了美国政策制定者对半导体、供应链以及美国国力问题的全面认知。这些措施被宣布为“暂行最终规则”,意味着它们将立即生效——这是在对控制华为的问题中觉察到的一个弱点做出直接应对。“在华为规定生效前已经发出了很多通知,让华为有了提前储备的时间,”参与制定10月7日法规的前国家安全委员会国际经济高级主任彼得·哈勒尔表示。“这是战术上的教训——出其不意是必要的。”更重要的是,美国发现拖垮一家公司——不管规模有多大——只会为新的竞争对手创造捷足先登的空间。美国需要采取更为全面的手段。“特朗普政府针对的是企业,”CSIS专家艾伦表示。“拜登政府打击的是行业。”
这些规定对于半导体供应链的深刻影响堪称前所未有。中国不仅无法进口最先进的芯片,还无法获得自主研发先进半导体和超级计算机所需的投入,甚至不能获取可用于生产半导体制造设备的美国原产零部件、技术和软件,这些设备本可让中国最终建成自己的晶圆厂,造出自己的芯片。“这是‘一锅端’的策略,”前BIS官员沃尔夫说。部分规定是全新创举,比如任何“美国人”——包括企业、个人以及绿卡持有者和永久居民——的行为都将受到限制。10月7日之后,美国人不再被允许从事任何支持在中国生产先进半导体的活动,无论是维修中国晶圆厂设备、提供建议、甚至连授权向中国半导体制造交付产品也不行。
采取单方面行动的决定是一场外交赌博。尽管美国掌控了半导体全球供应链上的诸多关键瓶颈,但其他国家——特别是台湾、日本与荷兰——在制造过程中同样关键的部门也占据了主导地位。如果这些国家和以前一样继续向中国出售产品,那10月7日的管制措施基本等于形同虚设。但在1月底,拜登政府与日本及荷兰达成协议,将对半导体或半导体制造设备实施类似的管控
台湾在此前几个月管制措施刚宣布的时候就签署了协议。该岛屿是芯片制造巨头:每年半导体产量占到全球的三分之二,其中90%都是最先进的半导体。台湾的大部分半导体产出都来自台积电这一家公司,它是全亚洲市值最高的上市企业,也是全球最先进的半导体制造商。光是台积电一家企业就已占据全球芯片代工制造市场总量的三分之一。(相较之下,石油输出国组织控制的全球石油市场份额约为40%。)
对美国来说,在全球芯片制造中发挥核心作用的台湾是不可或缺的。如果这座岛上的晶圆厂被中国占据,或是在被入侵期间关停,将会给全球经济带来灾难性后果。台湾在芯片产业中掌握的命门有时被称为“硅盾”,这是该岛屿对抗中国攻击最为有力的震慑,也是它在被中国入侵时得到美国帮助的终极保障。
但美台之间的伙伴关系并不对等。尽管台湾的芯片制造水平无可比拟,但按营收计算,其全球市场份额还不到10%。大部分销售额(在2022年达到40%)都流向将芯片制造出口到台湾的美国企业,这与美国服装设计师通过实际上在海外缝制的产品销售而获利的方式非常相似。在战略上,美国政策制定者将美国对台湾的依赖视为一种不可接受的风险。他们一直在推动台积电在美国建造更多晶圆厂,这是让更多半导体制造靠近美国海岸的更广泛战略的一部分。
由于担心惹恼这个最强大盟友兼最大武器供应国,台湾别无选择,只能顺从;但随着美国不断采取削弱该岛屿优势地位的举措,台湾正在让自己的处境愈发岌岌可危。在最坏的情况下,台湾在芯片上的主导优势可能只会招致更多破坏:一些美国评论人士和战争叫嚣者曾提出设想,如果中国入侵台湾,美国应摧毁台积电的晶圆厂,以防其落入中国之手。
试图控制半导体全球流通的一个难点在于,半导体非常小、重量轻、价值又高。“走私者就喜欢这样的货品,”艾伦说。但中国需要大量芯片来驱动大型数据中心及配备尖端计算机的设施,这让采购面临极大困难。“那都是不能移动的大型建筑,”米勒说。“非常利于美国情报机构探查。”半导体市场的构成也给试图规避管制的所有人制造了阻碍:有能力生产尖端芯片的企业数量极为有限,而在这些企业有芯片购买记录的买方也不多。
但执行管制的系统也存在漏洞,已经引来中国企业的试探。今年3月,活跃于云计算和服务器制造领域的中国企业浪潮集团被列入实体名单。但据《华尔街日报》报道,该公司至少有一家子公司未被列入名单,美国企业向其出售产品依然不受阻碍。
芯片在中国流通的路线也更加迂回。上个月,路透社报道了深圳高端芯片黑市繁荣的情况,许多零售商都宣称可以供应A100芯片,这是美企英伟达制造的一款功能强大的芯片。美国政府探查并阻止这种实体交易的能力是有限的:BIS在中国仅有三名执法人员。但黑市的存在其实正是管控奏效的早期结果。路透社采访的零售商声称,这些芯片只能小批量供应,可能是禁令生效前运往中国的库存。“这凸显了管制措施正在发挥作用,”一位要求匿名以便坦率评估美国政策的行业高管告诉我。“如果芯片可以自由流通,他们就不会这么做了。”
可以将芯片控制权之争看作一场对规范的考验。在西方,合规的责任将主要落在私营企业身上。“产业是我们的主要防线,”BIS出口管理助理部长西娅·罗兹曼·肯德勒表示。“我们政府机构会尽可能出台清晰、简明且有效的规定,但遵守和践行这些规定的责任在于行业。”为保证管制措施取得成效,美国的产业至少得在短期内采取自损八百的举措,与一部分利润丰厚的中国市场进行切割。美国企业将有充分的理由在合法边缘试探,而中国企业也有充分的动机来钻制度空子,为美国企业提供批准出售所需的信息。
对中国来说,科技自主争夺战可能为其带来前所未有的挑战。中国能够取得成功的特质——坚固的政治意志、源源不断的资金和围绕关键目标的全社会动员——可能会成为其致命的弱点。过去几年,就在发展国内半导体产业的努力变得更加紧迫之时,至少有六个耗资数十亿美元的芯片项目宣告失败,大批高管因涉嫌腐败接受调查。与此同时,无数企业涌入半导体行业,其中一些几乎完全是芯片外行,仅仅是为了拿到唾手可得的政府拨款。
“政治领导人或企业高管很容易认为,只要投入足够的资金和技术人员,就能解决这个问题,”前白宫科学和技术政策办公室副主任杰森·马西尼表示。但极度复杂的科学问题和遍布全球的供应链是很难移植的。“某种程度上,那是要复制整个人类文明,”马西尼说。
但若要说有哪个国家能够克服这样的挑战,那很可能非中国莫属。10月7日的出口管制虽然在可预见的未来对中国的先进芯片制造能力造成重大打击,但最后却可能刺激其实现长期的增长。当中国企业能够接触到先进西方芯片和供应商,国内制造商是找不到什么业务的。现在,中国企业不团结起来实现创新,下场就是一起覆灭。“我们消除了选项,”基尔克雷斯说。“以前他们可以在国家韧性和商业发展之间做选择,现在这个选择已经不复存在。”如果中国每年用于芯片进口的4000亿美元有很大一部分转而用于国内,其国内芯片企业可能最终将得到迎头赶上的手段和动力。
华为的案例可能会再一次带来启发。在美国制裁和中国严格疫情防控的双重打击下,华为在2022年的利润比上一年锐减70%。但仍能看到绝处逢生的迹象:尽管利润大幅下降,但营收却略有增长,而且华为在被禁止使用安卓系统后开发的鸿蒙操作系统已有超过3.3亿装机量,其中大部分在中国。华为仍是全球研发投入最大的企业之一,去年的研发预算为240亿美元,研发团队规模超过10万人。
对创新的重视势在必行。没有了美国的芯片和技术,华为被迫重新设计和制造了所有旗舰产品,以确保这些产品不包含美国零部件。该公司正在以一己之力带动整个国内供应链的发展,派遣自家工程师帮助培训和提升中国供应商的水平,华为曾不愿与这些供应商合作,转而选择外国供应商。最近,华为宣称在用于生产先进半导体的电子设计软件上取得了重大突破,尽管这种半导体的体积与美国比仍有几代的差距,但仍让华为进一步拉开了领先其他中国企业的身位。如果华为能够成功,它可能会突破美国制裁,比以往任何时候都更加强大和坚韧。
这些管制措施不能一劳永逸地遏制中国。即便在最理想的情况下,它们也只是一种拖延战术,旨在为美国及其盟友提供扩大关键技术领先地位的空间。问题在于,BIS能为西方争取到的时间到底有多少。“在这个行业,一击即中不等于成功,”负责出口执法的助理部长马特·阿克塞尔罗德表示。“我们的目标是能阻止多少算多少。”
我在商务部大楼的办公室与阿克塞尔罗德和负责出口管理的罗兹曼·肯德勒见面,这里能够俯瞰华盛顿特区中心的椭圆形草坪。只需几分钟时间就能走完几乎整个BIS总部。哪怕清楚这些规定不要求滴水不漏的执行,但我依然怀疑,让工业与安全局对抗整个中国政府的力量恐怕是不公平的。BIS哪有什么胜算呢?它的行动速度如何赶得上对方呢?BIS怎么可能像中国那样关注芯片,并为此投入那么多资金呢?对中国来说,芯片的未来是关乎国运的大事。
短暂沉默之后,罗兹曼·肯德勒轻声做出了答复。“这可能也关乎我们的存亡,”她说。

Alex W. Palmer是《纽约时报杂志》的特约撰稿人

'An Act of War' : Inside America's Silicon Blockade Against China

The Biden administration thinks it can preserve America’s technological primacy by cutting China off from advanced computer chips. Could the plan backfire?

By Alex W. Palmer

Last October, the United States Bureau of Industry and Security issued a document that — underneath its 139 pages of dense bureaucratic jargon and minute technical detail — amounted to a declaration of economic war on China. The magnitude of the act was made all the more remarkable by the relative obscurity of its source. One of 13 bureaus within the Department of Commerce, the smallest federal department by funding, B.I.S. is tiny: Its budget for 2022 was just over $140 million, about one-eighth the cost of a single Patriot air-defense missile battery. The bureau employs approximately 350 agents and officers, who collectively monitor trillions of dollars’ worth of transactions taking place all around the world.

During the height of the Cold War, when export controls to the Soviet bloc were at their strictest, B.I.S. was a critical hub in the Western defenses, processing up to 100,000 export licenses annually. During the relative peace and stability of the 1990s, the bureau lost some of its raison d’être — as well as staff and funding — and licenses shriveled to roughly 10,000 per year. Today, the number is 40,000 and climbing. With a sprawling trade blacklist known as the entity list (currently 662 pages and counting), numerous pre-existing multilateral export-control agreements and ongoing actions against Russia and China, B.I.S. is busier than ever. “We spend 100 percent of our time on Russia sanctions, another 100 percent on China and the other 100 percent on everything else,” says Matt Borman, the deputy assistant secretary of commerce for export administration.

In recent years, semiconductor chips have become central to the bureau’s work. Chips are the lifeblood of the modern economy, and the brains of every electronic device and system, from iPhones to toasters, data centers to credit cards. A new car might have more than a thousand chips, each one managing a different facet of the vehicle’s operation. Semiconductors are also the driving force behind the innovations poised to revolutionize life over the next century, like quantum computing and artificial intelligence. OpenAI’s ChatGPT, for example, was reportedly trained on 10,000 of the most advanced chips currently available.

 

With the Oct. 7 export controls, the United States government announced its intent to cripple China’s ability to produce, or even purchase, the highest-end chips. The logic of the measure was straightforward: Advanced chips, and the supercomputers and A.I. systems they power, enable the production of new weapons and surveillance apparatuses. In their reach and meaning, however, the measures could hardly have been more sweeping, taking aim at a target far broader than the Chinese security state. “The key here is to understand that the U.S. wanted to impact China’s A.I. industry,” says Gregory C. Allen, director of the Wadhwani Center for A.I. and Advanced Technologies at the Center for Strategic and International Studies in Washington. “The semiconductor stuff is the means to that end.”

Though delivered in the unassuming form of updated export rules, the Oct. 7 controls essentially seek to eradicate, root and branch, China’s entire ecosystem of advanced technology. “The new policy embodied in Oct. 7 is: Not only are we not going to allow China to progress any further technologically, we are going to actively reverse their current state of the art,” Allen says. C.J. Muse, a senior semiconductor analyst at Evercore ISI, put it this way: “If you’d told me about these rules five years ago, I would’ve told you that’s an act of war — we’d have to be at war.”

If the controls are successful, they could handicap China for a generation; if they fail, they may backfire spectacularly, hastening the very future the United States is trying desperately to avoid. The outcome will likely shape U.S.-China competition, and the future of the global order, for decades to come. “There are two dates that will echo in history from 2022,” Allen says. “The first is Feb. 24, when Russia invaded Ukraine; and the second is Oct. 7.”

Despite the immense intricacy of their design, semiconductors are, in a sense, quite simple: tiny pieces of silicon carved with arrays of circuits. The circuits flip on and off based on the activity of switches called transistors. When a circuit is on, it produces a one; off, a zero. The first chips, invented in the late 1950s, held only a handful of transistors. Today the primary semiconductor in a new smartphone has between 10 and 20 billion transistors, each about the size of a virus, carved like a layer cake into the structure of the silicon.

The rate of progress over the last six decades has been famously described by Moore’s Law, which observed that the number of transistors that can be fit on a chip has roughly doubled every two years. Chris Miller, author of the book “Chip War” and an associate professor of international history at the Fletcher School at Tufts University, likes to note that if airplanes had improved at the same rate as chips, they’d now be flying at several times the speed of light. No technology in the history of human civilization has ever matched the breathtaking ascent of computing power.

 

Semiconductor-manufacturing plants, known as fabs, are the most expensive factories in the world, conducting the most complex manufacturing ever accomplished, at a scale of production never before achieved with any other device. The wider chip industry, meanwhile, is a web of mutual interdependence, spread all over the planet in highly specialized regions and companies, its feats made possible by supply chains of exceptional length and complexity — a poster child, in other words, for globalization. “It’s hard to imagine how the capabilities they’ve reached would be possible without access to the smartest minds in the world all working together,” Miller says. And yet it is this same interconnectedness that makes the industry vulnerable to regulations like those the Biden administration is pursuing.

Only a small handful of companies can compete at the cutting edge, where breakthroughs cost billions of dollars and decades of research. The result is an industry structured as a series of choke points. The best-known example is the extreme ultraviolet (EUV) lithography machine made by ASML, a Dutch manufacturing conglomerate, which is used to print out the layers of a chip. In 1997, ASML hired Jos Benschop, a young engineer with a Ph.D. in physics, to spearhead the creation of a new system, one that would help ASML’s customers in the semiconductor industry print smaller, faster and denser chips than ever before. It took four years to achieve the proof of concept necessary to even justify assigning a small team to the task, and then another five years for the team to build a prototype machine. In December 2010, at a research facility in South Korea, an updated prototype, a TWINSCAN NXE:3100, finally had its first successful test run. It would be nearly another decade before the first EUV-enabled products would go to market.

‘I truly believe our machine is the most complex thing mankind has ever produced.’

The newest version of the machine can craft structures as small as 10 nanometers; a human red blood cell, by comparison, is about 7,000 nanometers across. It uses a laser to create plasma 40 times hotter than the surface of the sun, which emits extreme ultraviolet light — invisible to the human eye — that is reflected onto a silicon chip by a series of mirrors. The laser is sourced from a German company and has 457,329 pieces; an entire EUV has more than 100,000 components of similar intricacy.

An EUV is just one part of the process: A cutting-edge fab can include more than 500 machines and 1,000 steps. And yet an EUV alone is a nearly miraculous human achievement, capable of working at scales and precisions that are difficult to fathom. “I truly believe our machine is the most complex thing mankind has ever produced,” says Benschop, now ASML’s corporate vice president of technology. Today, more than a decade since the TWINSCAN’s first test run, no other company has been able to recreate ASML’s achievement.

 

By squeezing on the industry’s natural choke points, the Biden administration aims to block China from the future of chip technology. The effects will go far beyond cutting into Chinese military advancements, threatening the country’s economic growth and scientific leadership too. “We said there are key tech areas that China should not advance in,” says Emily Kilcrease, a senior fellow at the Center for a New American Security and a former U.S. trade official. “And those happen to be the areas that will power future economic growth and development.” Today, scientific advances are often made by running simulations and analyzing huge amounts of data, rather than through trial-and-error experiments. Simulations are used to discover new lifesaving drugs, to model the future of climate change and to explore the behavior of colliding galaxies — as well as the physics of hypersonic missiles and nuclear explosions.

“The person with the best supercomputer can do the best science,” Jack Dongarra, founding director of the Innovative Computing Laboratory at the University of Tennessee, told me. Dongarra runs a program called the TOP500, which offers a biannual ranking of the fastest supercomputers in the world. As of June, China claims 134 spots, compared with 150 for the U.S. But the picture is incomplete: Around 2020, China’s submissions plummeted in a way that suggested to Dongarra a desire to avoid attracting unwanted attention. Rumors of new supercomputers leak out in scientific papers and research announcements, leaving observers to guess at the true state of the competition — and the size of China’s presumed lead. “It’s striking because in 2001 China had no computers on the list,” Dongarra says. “Now they’ve grown to the point that they dominate it.”

Yet beneath China’s strength is a crucial vulnerability: Nearly all the chips that power the country’s most advanced projects and institutions are inexorably tied to U.S. technology. “The entire industry can only function with U.S. inputs,” Miller says. “In every facility that’s remotely close to the cutting edge, there’s U.S. tools, U.S. design software and U.S. intellectual property throughout the process.” Despite decades of effort by the Chinese government, and tens of billions of dollars spent on “indigenous innovation,” the problem remains acute. In 2020, China’s domestic chip producers supplied just 15.9 percent of the country’s overall demand. As recently as April, China spent more money importing semiconductors than it did oil.

America fully grasped its power over the global semiconductor market in 2019, when the Trump administration added Huawei, a major Chinese telecommunications maker, to the entity list. Though the listing was ostensibly punishment for a criminal violation — Huawei had been caught selling sanctioned materials to Iran — the strategic benefits became immediately obvious. Without access to U.S. semiconductors, software and other essential supplies, Huawei, the largest telecommunications-equipment producer in the world, was left struggling to survive. “The Huawei sanctions immediately pulled back the curtain,” says Matt Sheehan, a fellow at the Carnegie Endowment for International Peace who studies China’s tech ecosystem. “Chinese tech giants are running on chips that are made in America or have deep American components.”

Export-control law had long been seen as a dusty, arcane backwater, far removed from the actual exercise of American power. But after Huawei, the United States discovered that its primacy in the semiconductor supply chain was a rich source of untapped leverage. Three firms, all located in the U.S., dominate the market for chip-design software, which is used to arrange the billions of transistors that fit on a new chip. The market for advanced chip-manufacturing tools is similarly concentrated, with a handful of companies able to claim effective monopolies over essential machines or processes — and nearly all of these companies are American or dependent on American components. At every step, the supply chain runs through the U.S., U.S. treaty allies or Taiwan, all of them operating in a U.S.-dominated ecosystem. “We stumbled into it,” Sheehan says. “We started using these weapons before we really knew how to use them.”

 

In May 2020, the Trump administration tightened the screws further, this time by making Huawei subject to a formerly obscure provision of export-control law called the foreign direct product rule. Under the F.D.P.R., foreign-made items are subject to American controls if they were produced using American technology or software. It is a sweeping assertion of extraterritorial power: Even if an item is made and shipped outside the United States, never once crossing the country’s borders, and contains no U.S.-origin components or technology in the final product, it can still be considered an American good.

For Huawei, the application of the F.D.P.R. meant the company was virtually cut off from semiconductors. “That rule subjected all semiconductors on the planet to American law, because every foundry on the planet uses U.S. tools at least in part,” Kevin Wolf, a former assistant secretary of commerce for export administration at the B.I.S., says. “If you have one U.S. tool and 100 non-American tools in your fab, that taints any wafer moving across the line.”

In 2020, according to the market-analysis firm Canalys, Huawei was the largest smartphone seller in the world, with an 18 percent market share, besting even Apple and Samsung. Huawei’s revenues plunged by nearly a third in 2021, and the company sold off one of its smartphone brands in a bid to stay afloat. By 2022, its share had fallen to 2 percent.

The Oct. 7 rules represented the sum of everything U.S. policymakers had learned about semiconductors, supply chains and American power. The measures were announced as an “interim final rule,” meaning they took effect immediately — a direct reaction to a perceived weakness in the Huawei controls. “There was a lot of notice before the Huawei rule came into effect, and they spent the time beforehand stockpiling,” says Peter Harrell, a former senior director for international economics at the National Security Council who was involved in crafting the Oct. 7 rules. “That was a tactical lesson — that you need the element of surprise.” More important, the United States had learned that hobbling one company, however large, simply created room for new competitors to step in. A more comprehensive approach would be needed. “The Trump administration went after companies,” says Allen, the CSIS expert. “The Biden administration is going after industries.”

The rules went deeper into the semiconductor supply chain than any previous measure. China was cut off not just from importing the most advanced chips, but also from acquiring the inputs to develop its own advanced semiconductors and supercomputers, and even from the U.S.-origin components, technology and software that could be used to produce semiconductor-manufacturing equipment to eventually build their own fabs to make their own chips. “It was an ‘all of the above’ strategy,” Wolf, the former B.I.S. official, says. Some elements were entirely novel, like a restriction on the activity of any “U.S. persons” — companies and citizens, as well as green-card holders and permanent residents. After Oct. 7, U.S. persons are no longer allowed to engage in any activity that supports the production of advanced semiconductors in China, whether by maintaining or repairing equipment in a Chinese fab, offering advice or even authorizing deliveries to a Chinese semiconductor manufacturer.

 

The decision to act unilaterally was a diplomatic gamble. Though the United States controls a number of key choke points in the global supply chain, other countries — particularly Taiwan, Japan and the Netherlands — hold dominance over similarly crucial sectors of the manufacturing process. Had those countries continued to sell to China as before, it would have rendered the Oct. 7 controls nearly useless. But in late January, the Biden administration reached an agreement with Japan and the Netherlands, under which they would implement similar controls on semiconductors or semiconductor-manufacturing equipment.

Taiwan had already signed on months earlier, as soon as the controls were announced. The island is a chip-manufacturing juggernaut: It produces almost two-thirds of the world’s semiconductors annually, and over 90 percent of the most advanced ones. Much of that output is thanks to a single firm, TSMC, the most valuable public company in all of Asia and the most advanced semiconductor manufacturer in the world. By itself, TSMC accounts for about a third of the total global market for contract chip fabrication. (OPEC, by comparison, controls about 40 percent of the global oil market.)

‘At some point, you’re replicating all of human civilization.’

Taiwan’s central role in global chip production makes it indispensable to the United States. If the island’s fabs were to be captured by China, or knocked offline during an invasion, the costs to the global economy would be catastrophic. Taiwan’s chips stranglehold is sometimes called its “silicon shield” — the island’s most formidable deterrent against a Chinese attack, and its best assurance of American help in the event of a Chinese invasion.

But the partnership between the U.S. and Taiwan is an unequal one. Though Taiwan is unmatched in chip manufacturing, it captures less than 10 percent of the global market by revenue. The bulk of sales — 40 percent in 2022 — go to the American firms that export their chip manufacturing to Taiwan, in much the same way that American clothes designers profit from the sale of items that are actually sewn overseas. Strategically, American policymakers see the U.S.’s dependence on Taiwan as an unacceptable risk. They have pushed for TSMC to build more fabs in the U.S., as part of a broader strategy to locate more semiconductor manufacturing closer to American shores.

 

Taiwan has no choice but to comply, for fear of upsetting its most powerful ally and largest arms supplier; but with every move to erode the island’s pre-eminence, it makes itself more vulnerable. In the worst case, Taiwan’s chip chokehold may only invite more destruction: Some American commentators and war-gamers have suggested that, if China does invade, the U.S. should destroy TSMC’s fabs to stop them from falling under China’s control.

One problem with trying to control the global flow of semiconductors is that they’re very small, lightweight and valuable. “Smugglers love stuff like that,” Allen says. But China needs chips in large quantities to power massive data centers and facilities housing cutting-edge computers — and that makes their procurement uniquely challenging. “Those are large buildings, and they don’t move,” Miller says. “It’s uniquely suited to be understood by U.S. intelligence.” The structure of the market will also present a hurdle to anyone trying to circumvent the regulations: The number of companies capable of producing cutting-edge chips is extremely limited, and the number of buyers with a history of purchasing from them is also small.

But there are also loopholes in the enforcement system, which Chinese companies are already probing. In March, Inspur Group, a Chinese conglomerate active in cloud computing and server manufacturing, was added to the entity list. But according to The Wall Street Journal, at least one of the company’s affiliates was not included in the listing, allowing American businesses to sell to the subsidiary unimpeded.

Chips are moving through China by more circuitous routes as well. Last month, Reuters reported on a booming underground trade in high-end chips in Shenzhen, with multiple retailers touting their ability to supply the A100, a powerful chip made by the American company Nvidia. The U.S. government’s ability to detect and prevent these types of hand-to-hand sales is limited: B.I.S. has only three enforcement agents stationed in China. But the existence of the underground market was, in fact, an early signal of the controls’ efficacy. According to retailers interviewed by Reuters, the chips were available only in small batches, perhaps from stocks shipped to China before the ban took effect. “It highlights that the controls are working,” an industry executive, who requested anonymity in order to candidly assess American policy, told me. “They wouldn’t be doing that if chips flowed freely.”

The battle over the controls may serve as a kind of civilizational test. In the West, the onus of compliance will fall largely on private companies. “Industry is our primary line of defense,” says Thea Rozman Kendler, the assistant secretary of export administration at B.I.S. “We can do whatever we can in government to promulgate clear and concise and effective rules, but it’s industry that’s responsible for compliance and putting those rules into effect.” For the controls to succeed, American industry will need to engage in actions that are, at least in the short-term, self-sabotaging, shutting off a piece of the lucrative Chinese market. Companies will have ample reason to operate as close to the edge of legality as possible, and their Chinese counterparts will have every incentive to game the system and feed them the information needed to approve a sale.

 

For China, the race for technological self-sufficiency presents perhaps a greater challenge than any the country has faced. The very traits that make China’s success possible — iron political will, endless money and a whole-of-society mobilization around key goals — are just as likely to prove its Achilles’ heel. In the last several years, as the push to develop a domestic semiconductor industry has taken on new urgency, at least six multibillion-dollar chip projects have failed and a number of executives have been put under investigation for corruption. Tens of thousands of companies, meanwhile, have flooded into the semiconductor industry, some of them with little or no expertise in chips, solely in search of easy government money.

“It’s easy for political leaders or executives to think if we throw enough money and engineers at this problem, we’ll solve it,” Jason Matheny, former deputy director of the White House Office of Science and Technology Policy, says. But the immense complexity of the science and the globe-spanning supply chains are difficult to imitate. “At some point,” says Matheny, “you’re replicating all of human civilization.”

Yet if any country can overcome such a challenge, it is likely to be China. The Oct. 7 export controls, while crippling China’s advanced chip-making ability for the foreseeable future, may end up spurring long-term growth. When Chinese companies had access to superior Western chips and suppliers, domestic manufacturers struggled to find business. Now Chinese companies must innovate together or die. “We’ve removed choice,” Kilcrease says. “Before they could choose between national resiliency and commercial motivations, and now they don’t have that choice.” Should a large share of China’s $400 billion in annual chip imports be turned inward, domestic chip companies may finally have the means and motivation to catch up.

Huawei may prove instructive once again. Battered by American sanctions and China’s strict pandemic controls, the company’s 2022 profits fell by a staggering 70 percent compared with the previous year. But there are signs of life: Despite the plunge in profits, revenues rose slightly, and the company’s operating system, HarmonyOS — which it developed after being cut off from using Android — has been installed on more than 330 million devices, mostly in China. Huawei remains one of the world’s biggest spenders on research and development, with a budget of about $24 billion last year and a research team of over 100,000 employees.

The emphasis on innovation is by necessity. Bereft of American chips and technology, Huawei has been forced to redesign and remanufacture all of its legacy products to ensure they contain no American components. The company is dragging along an entire domestic supply chain in its wake, sending its own engineers to help train and upscale Chinese suppliers it once shunned in favor of foreign alternatives. Recently, Huawei claimed that it had made significant breakthroughs in the electronic design software used to produce advanced semiconductors at a size that, though still a few generations behind the U.S., would put it further along than any other Chinese company. If Huawei manages to succeed, it could emerge from American sanctions stronger and more resilient than ever.

 

The controls will not stop China permanently. Even in the best case, they’re a delay tactic, meant to offer the U.S. and its allies space to expand their lead in key technologies. The question is how much time B.I.S. can buy for the West. “This isn’t the type of business where success is batting one thousand,” said Matt Axelrod, the assistant secretary for export enforcement. “Our goal is to stop as much as possible.”

I was meeting with Axelrod and Rozman Kendler, the export administration chief, at the Commerce Department building, in an office overlooking the Ellipse in downtown Washington, D.C. It had taken just a few minutes to walk nearly the entire length of B.I.S.’s headquarters. Even allowing that enforcement need not be perfect, I wondered whether this was a fair fight — the Bureau of Industry and Security versus the full weight of the Chinese government. How could B.I.S. win? How could it hope to move as quickly? How could B.I.S. possibly put as much money behind the effort, and care as much about chips as China does? The future of chips was life or death for China.

There were a few seconds of silence before Rozman Kendler answered, in a quiet voice. “It’s probably life or death for us too,” she said.


Alex W. Palmer is a contributing writer for the magazine. He last wrote about the rise of TikTok. Grant Cornett is an artist who resides in the Catskill Mountains. His work focuses on objects and their relation to light and time in natural settings and more composed commercial projects.

A correction was made on 
July 25, 2023

An earlier version of this article misidentified the way in which extreme ultraviolet light is manipulated during the manufacturing of semiconductor chips. EUV is reflected, not refracted, during the process.

[ 打印 ]
阅读 ()评论 (0)
评论
目前还没有任何评论
登录后才可评论.