文 | 版面之外，作者｜画画
AI 竞争打到今天，连德国也坐不住了。
最近，德国总理默茨在柏林迎来英伟达创始人黄仁勋。西门子、奔驰、奥迪等德国工业巨头的代表悉数到场，话题指向同一件事：AI 如何进入工厂，改变制造业。
这一幕的背后，藏着欧洲深层的焦虑。
两百年前，欧洲用蒸汽机和内燃机改变了人类命运，但到了互联网和芯片时代，话语权逐渐落入了美国手中。
如今 AI 浪潮席卷全球，美国拿着芯片和模型大吃红利，中国把 AI 带进产业场景，日韩则占据半导体材料与存储的供应链上游。
只有欧洲，仿佛被困在了上个时代的辉煌里。
一、工业老大哥的"失落三十年"
今天欧洲的落寞，早在几十年前就已经埋下了伏笔。
18 世纪，英国人用蒸汽机把手工作坊推进了工厂时代，19 世纪，德国人用西门子的电机、巴斯夫的化学实验室，建立起严密的实验室到车间的工业帝国。这套把科研快速转化为产品的机制，曾是欧洲最硬的底气。
当美国在硅谷用风险投资催生出谷歌、微软和亚马逊，依托统一的本土市场把一项技术迅速铺向全球时，欧洲却陷入了长期的内耗。它有全球最严苛的法律体系、被割裂的 27 个成员国市场，以及一套对试错并不友好的资本体系。
荷兰的 ASML 造出了当今最精密的极紫外光刻机，但前 CEO 温宁克在离任前曾大声疾呼，欧洲正在失去创新的灵魂，因为这里的人习惯了在规章制度里打转，而不是去开创未来。
这种脱节在生成式 AI 爆发时达到了一个高峰。
二、全球牌桌上的四国演义
如果把 AI 产业拆开来看，这场全球博弈从来不是单一技术的比拼，本质上是一场场残酷的供应链割据战。
美国玩的是霸道的生态闭环。英伟达卖 GPU，台积电代工，微软和谷歌搭数据中心，OpenAI、Anthropic 攻坚大模型。
英伟达最可怕的不是把显卡卖得有多贵，更在于围绕 CUDA 建立的软件生态。开发者多年积累的代码、开发工具与底层调优经验越多，迁移到其他平台的成本就越高。芯片由此不再只是硬件产品，也成为了整套计算体系的入口。
日本未必掌握 AI 最受关注的大模型产品，却在若干难以绕开的底层制造环节占据着重要位置。信越化学与胜高的半导体级硅晶圆、东京电子的涂布与刻蚀设备、爱德万测试的芯片测试系统，构成了芯片制造的基础设施。
日本并没有放弃自身 AI 产业，而是将其产业优势沉淀在了供应链中下游那些门槛和替代成本高的核心环节。
韩国则演了一出依靠 AI 翻盘的戏码。
2023 年，全球存储芯片市场遭遇严重周期性下行，SK 海力士承受了巨大的利润压力。但在巨额亏损的阴影下，公司依然维持了对高带宽内存（HBM）的研发与产能投入。
随着 AI 加速器对高吞吐内存带宽的需求爆发，SK 海力士凭借在 HBM3 与 HBM3e 上的先发优势，成为了英伟达等芯片厂商的核心供应商。到了 2026 年，其市值历史性地突破 1 万亿美元。
中国则展现了推理降本与场景工程的双重推动力。2025 年初 DeepSeek R1 模型的横空出世，给硅谷的算力阵营浇了一盆冷水。它用极低的价格和巧妙的架构，打破了只有千亿美金才能堆出领先模型的门槛。
与此同时，模型使用成本的下降，进一步激活了中国庞大的制造业体系，为 AI 提供了丰富的试验场，也让经过验证的方案有机会快速复制到更多生产环节。
反观欧洲，虽然有 ASML 的光刻机、法国 Mistral 的开源模型，却更像是一盘散沙。
三、规矩太多，钱太少
欧洲其实不是没有试过拯救自己。
2024 年起，Mistral 先后引入了微软 1500 万欧元的战略投资，并将其大模型深度分发绑定在微软 Azure 云平台上，后续英伟达等硅谷巨头也相继跟进。
更大的尴尬藏在欧洲引以为傲的工厂里。
按照欧盟官方的规划，政府准备动用数百亿欧元在各国建造"AI 超级工厂"，希望用公共资本帮企业补上算力的缺口。
然而，当技术方案真正落地到工厂层面时，工程人员面临的往往是更基础的数字化断层。大量中小型工厂虽然拥有极高精度的机床，但底层工控协议（如 OPC UA）未能统一，老旧设备缺乏标准化数据采集接口，数据孤岛普遍存在。
欧盟委员会发布的数据直观地呈现了这一差距。
截至 2024 年，在欧盟范围内真正将 AI 技术深度集成到核心业务流程中的企业占比仅有约 8%。经合组织的数据显示，欧洲制造企业里，真正把 AI 应用到核心生产线上的连 10% 都不到。
四、黄仁勋的商业算计
正是在这种被动与焦虑的氛围下，黄仁勋带着英伟达的方案来到了柏林。
英伟达宣布要在欧洲追加近 100 亿美元投资，不仅投向德国的机器人和医疗 AI 初创公司，更直接与西门子、奥迪捆绑，要在德国推广物理 AI。
以西门子与英伟达合作的工业仿真平台为例，企业可以在虚拟空间中构建产线的数字孪生模型，测试生产流程与设备配置，从而在真实部署前大幅降低试错成本。
对于 GDP 增速连年低迷、被高能源成本和出口压力压得喘不过气来的德国而言，借助 AI 提升全要素生产率，是维持其工业地位的重要选项。
但这种合作也带来了新的议题：如果欧洲工业界的软件仿真、机器人控制与算法训练都建立在英伟达的系统之上，英伟达获得的就不只是一笔芯片销售额，更可能演变为工业系统的基础平台。
五、逃离硅谷
欧洲终于清醒地认识到，在通用大模型与纯算力堆砌的赛道上去硬碰硬，并非其优势所在。它开始把更多的筹码，押注在自己最擅长的领域：工业自动化、医疗健康、精密机械以及科学计算。
荷兰的光刻机、德国的机械臂、法国的航空航天，这些沉淀了数百年的工业机密，是硅谷用代码无论如何也无法在短时间内模拟出来的真实数据。
欧洲唯有将这些工业经验转化为可训练的数据与可规模化的商业产品，并打破 27 国之间的合规与市场壁垒，才能真正将老资产变成新引擎。
全球 AI 比赛的下半场，已经从虚拟的数字世界，蔓延到了真实的物理世界。
黄仁勋带着算力敲开了欧洲的大门，而欧洲能否重新站稳脚跟，取决于它能否在硅谷的计算体系中，找到属于自己的不可替代性。
【版面之外】的话：
技术革命从不按资历排座次。
一个时代积累的优势，到了下一个时代，可能是起点，也可能成为惯性。
工业史不会替任何人保留席位。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Outside of the page, the author paints.
AI competes today, and even Germany can't sit.
Recently, German Chancellor Metz welcomed the founder of Young Weida in Berlin. Representatives of German industrial giants such as Siemens, Mercedes and Audi were all present and the subject was the same: AI, how to get into the factory and change the manufacturing industry.
Behind this scene lies deep European anxiety.
Two hundred years ago, Europe changed human destiny with steam and internal combustion engines, but in the age of the Internet and chips, the voice gradually fell into the hands of the United States.
Today, the AI waves are sweeping the world, with the US taking a lot of dividends with chips and models, China taking AI into the industrial scene, and Japan and Korea taking over the supply chain of semiconductor materials and storage.
Only Europe, as if it were trapped in the glory of the last era.
I. "Thirty Years Lost"
The fall of Europe today has been buried decades ago.
In the 18th century, the British advanced the factory age with a steam machine handle, and in the 19th century, the Germans built a tight laboratory to the industrial empire of the workshop with a Siemens electric power machine and a Basff chemical laboratory. This mechanism for the rapid transformation of scientific research into a product was once the strongest in Europe.
Europe was caught in a long-term internal consumption when the United States used risk investments in the Silicon Valley to generate Google, Microsoft and Amazon, and to move a technology around the globe in a unified local market. It has the toughest legal system in the world, a fragmented market of 27 member countries, and a capital system that is not friendly to the wrong.
The Dutch ASML created the most sophisticated ultra-violet machine of the day, but the former CEO Winningk cried out before he left, and Europe is losing its innovative soul, because people here are used to turning around in regulations, not to creating the future.
This disconnect reached a peak at the time of the generation of AI.
II. THE QOURS OF THE GLOBAL PLACE
If the AI industry is to be detached, this global game has never been a single-technology match and is essentially a brutal supply chain battle.
America plays the hegemonic eco-closure. GPU for Young Weeda, a desk electrician, Microsoft and Google to the data centre, Open AI, Anthropic to fight LLM.
The scariest thing about Yvette is not how expensive GPU is, but the software ecology around CUDA. The more codes, tools and bottom-up experience the developers have accumulated over the years, the higher the cost of migrating to other platforms. The chips are thus no longer just a hardware product, but also an entry point for the entire computing system.
Japan does not necessarily own AI's most important LLM product, but it has an important place in a number of bottom manufacturing chains that are difficult to circumvent. The semiconductor silica round, the Tokyo electronic coating and corrosive equipment, and the chip testing system tested in Edwan constitute the infrastructure for chip manufacturing.
Japan has not abandoned its AI industry, but has deposited its industrial advantage at the heart of the thresholds and alternative costs downstream and downstream of the supply chain.
South Korea has played a game of relying on AI overboard.
In 2023, the global storage chip market suffered a severe cyclical downturn, and SK Hercules was under tremendous profit pressure. In the shadow of huge losses, however, companies have maintained R & D and capacity inputs for high bandwidth memory (HBM).
As the demand for high-swallow RAM bandwidth erupted in AI, SK Hercules became a core supplier of chips such as British Weeda by pre-emptive advantage on HBM3 and HBM3e. By 2026, its market value had historically surpassed $1 trillion.
China, for its part, has shown a dual thrust of reasoning and scenario engineering. Early in 2025, DeepSeek R1 was born out of the air, pouring cold water into Silicon Valley's calculator camp. It breaks the threshold of a few hundred billion dollars to build up a leading model with extremely low prices and sophisticated structures.
At the same time, the decline in the cost of using models has further activated China’s vast manufacturing system, providing AI with a rich testing ground, and giving certified programmes the opportunity to quickly replicate more production chains.
Looking back at Europe, although there is an ASML carving machine, an open source model for France's Mistral, it's more like a dispersed sand.
Three, too many rules, too little money.
Europe has not really tried to save itself.
Since 2024, Mistral has introduced a strategic investment of 15 million euros in Microsoft and has tied its LLM deep distribution to the Microsoft Azure cloud platform, followed by Silicon Valley giants like Young Weidar.
The greater embarrassment lies in the proud factories in Europe.
According to official European Union plans, the Government intends to use hundreds of billion euros to build "AI super-workshops" in various countries, hoping that public capital will help businesses to bridge the gap in capacity.
However, when technology programmes actually land at the factory level, engineers often face more basic digital faults. While a large number of small and medium-sized plants have very high-precision machine beds, bottom control agreements (e.g., OPC UA) are not harmonized, old equipment lacks standardized data-gathering interfaces and data isolation is widespread.
The data of the European Commission publish provide a visual picture of this gap.
By 2024, only about 8 per cent of enterprises in the European Union had actually integrated AI technology into core business processes. OECD data show that less than 10 per cent of European manufacturing firms actually apply AI to core production lines.
IV. HONG In-hoon ' s business calculations
It was in this atmosphere of passivity and anxiety that Huang In-hoon came to Berlin with the Young Wida programme.
In Weida announced an additional investment of nearly $10 billion in Europe, not only to German robotics and medical start-ups, but also directly to Siemens and Audi, to promote physical AI in Germany.
In the case of the Siemens industrial simulation platform, which works with Weeda, companies can build digital twinning models of production lines in virtual spaces and test production processes and equipment configurations, thereby significantly reducing trial error costs before real deployment.
For Germany, where GDP growth has been low for years, and is overwhelmed by high energy costs and export pressures, the use of AI to increase full factor productivity is an important option for maintaining its industrial status.
But this cooperation also brings new issues: If European industry software simulations, robotic control and algorithm training are based on the system of Weeda, it is not just a chip sale, but it is more likely to evolve into a basic platform for industrial systems.
Escape from Silicon Valley
Europe has finally come to realize that it is not its advantage to hit hard on the general LLM and pure-calculations tracks. It has begun to place more leverage in its areas of expertise: industrial automation, medical health, precision machinery and scientific computing.
The Dutch carving machines, Germany ' s mechanical arm and French aerospace, which have deposited industrial secrets for hundreds of years, are true data that Silicon Valley codes cannot in any case simulate in a short time.
Only by transforming these industrial experiences into trainingable data and scalable commercial products, and by breaking compliance and market barriers between 27 countries, can Europe truly turn old assets into new engines.
The second half of the global AI competition has spread from the virtual digital world to the real physical world.
Wong In-hoon knocked on Europe’s door with his arithmetic, and Europe’s ability to get back on its feet depends on it finding its own irreplaceable system in Silicon Valley’s computing system.
In the words:
The technological revolution has never been ranked according to seniority.
The advantages accumulated by an era can be a starting point and possibly inertia in the next era.
Industrial history does not reserve seats for anyone.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.